release v0.3.6.6

Lkhphuc lr sched

.gitignore

@@ -8,6 +8,8 @@ datasets/
 model_weights/
 app/models/
 pip-wheel-metadata/
+test_tube_exp/
+tests/tests_tt_dir/
 
 # Byte-compiled / optimized / DLL files
 __pycache__/
@@ -118,4 +120,4 @@ ENV/
 .mypy_cache/
 
 # data
-mnist/
+mnist/

.readthedocs.yml

@@ -0,0 +1,19 @@
+# .readthedocs.yml
+# Read the Docs configuration file
+# See https://docs.readthedocs.io/en/stable/config-file/v2.html for details
+
+# Required
+version: 2
+
+# Build documentation with MkDocs
+mkdocs:
+  configuration: mkdocs.yml
+
+# Optionally build your docs in additional formats such as PDF and ePub
+formats: all
+
+# Optionally set the version of Python and requirements required to build your docs
+python:
+  version: 3.7
+  install:
+    - requirements: docs/doc_requirements.txt

.travis.yml

@@ -0,0 +1,16 @@
+language: python
+python:
+  - "3.7"
+# command to install dependencies
+cache: pip
+install:
+  - pip install -e .
+  - pip install -r requirements.txt
+  - pip install -U numpy
+
+# keep build from timing out
+dist: xenial
+
+# command to run tests
+script:
+  - py.test # or py.test for Python versions 3.5 and below

README.md

@@ -9,36 +9,136 @@
 <p align="center">
   The Keras for ML researchers using PyTorch. More control. Less boilerplate.    
 </p>
+
 <p align="center">
   <a href="https://badge.fury.io/py/pytorch-lightning"><img src="https://badge.fury.io/py/pytorch-lightning.svg" alt="PyPI version" height="18"></a>
-<!--   <a href="https://travis-ci.org/williamFalcon/test-tube"><img src="https://travis-ci.org/williamFalcon/pytorch-lightning.svg?branch=master"></a> -->
+  <a href="https://pepy.tech/project/pytorch-lightning"><img src="https://pepy.tech/badge/pytorch-lightning" alt="PyPI version" height="18"></a>
+  <a href="https://github.com/williamFalcon/pytorch-lightning/tree/master/tests"><img src="https://github.com/williamFalcon/pytorch-lightning/blob/master/coverage.svg"></a>
+  <a href="https://travis-ci.org/williamFalcon/pytorch-lightning"><img src="https://travis-ci.org/williamFalcon/pytorch-lightning.svg?branch=master"></a>
+  <a href="https://williamfalcon.github.io/pytorch-lightning/"><img src="https://readthedocs.org/projects/pytorch-lightning/badge/?version=latest"></a>
   <a href="https://github.com/williamFalcon/pytorch-lightning/blob/master/COPYING"><img src="https://img.shields.io/badge/License-MIT-yellow.svg"></a>
 </p>   
 
 ```bash
-pip install pytorch-lightning    
+pip install pytorch-lightning  
 ```
 
 ## Docs   
 **[View the docs here](https://williamfalcon.github.io/pytorch-lightning/)**
 
 ## What is it?  
-Keras and fast.ai are too abstract for researchers. Lightning abstracts the full training loop but gives you control in the critical points.   
+Lightning defers training and validation loop logic to you. It guarantees correct, modern best practices for the core training logic.
 
 
 ## Why do I want to use lightning?
-Because you don't want to define a training loop, validation loop, gradient clipping, checkpointing, loading, 
-gpu training, etc... every time you start a project. Let lightning handle all of that for you! Just define your 
-data and what happens in the training, testing and validation loop and lightning will do the rest.
+When starting a new project the last thing you want to do is recode a training loop, model loading/saving, distributed training, when to validate, etc... You're likely to spend a long time ironing out all the bugs without even getting to the core of your research.
+
+With lightning, you guarantee those parts of your code work so you can focus on what the meat of the research: Data and training, validation loop logic. Don't worry about multiple gpus or speeding up your code, lightning will do that for you!
+
+## How do I do use it?   
 
 To use lightning do 2 things:  
-1. [Define a Trainer](https://github.com/williamFalcon/pytorch-lightning/blob/master/examples/new_project_templates/trainer_cpu_template.py).   
-2. [Define a LightningModel](https://github.com/williamFalcon/pytorch-lightning/blob/master/examples/new_project_templates/lightning_module_template.py).     
+1. [Define a LightningModel](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/)         
+```python
+import os
+import torch
+from torch.nn import functional as F
+from torch.utils.data import DataLoader
+from torchvision.datasets import MNIST
+import torchvision.transforms as transforms
+
+import pytorch_lightning as ptl
+
+class CoolModel(ptl.LightningModule):
+
+    def __init__(self):
+        super(CoolModel, self).__init__()
+        # not the best model...
+        self.l1 = torch.nn.Linear(28 * 28, 10)
+
+    def forward(self, x):
+        return torch.relu(self.l1(x.view(x.size(0), -1)))
+
+    def my_loss(self, y_hat, y):
+        return F.cross_entropy(y_hat, y)
+
+    def training_step(self, batch, batch_nb):
+        x, y = batch
+        y_hat = self.forward(x)
+        return {'loss': self.my_loss(y_hat, y)}
+
+    def validation_step(self, batch, batch_nb):
+        x, y = batch
+        y_hat = self.forward(x)
+        return {'val_loss': self.my_loss(y_hat, y)}
+
+    def validation_end(self, outputs):
+        avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
+        return {'avg_val_loss': avg_loss}
+
+    def configure_optimizers(self):
+        return [torch.optim.Adam(self.parameters(), lr=0.02)]
+
+    @ptl.data_loader
+    def tng_dataloader(self):
+        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)
+
+    @ptl.data_loader
+    def val_dataloader(self):
+        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)
+
+    @ptl.data_loader
+    def test_dataloader(self):
+        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)
+```
+
+2. Fit with a [trainer](https://williamfalcon.github.io/pytorch-lightning/Trainer/)    
+```python
+from pytorch_lightning import Trainer
+from test_tube import Experiment
+
+model = CoolModel()
+exp = Experiment(save_dir=os.getcwd())
+
+# train on cpu using only 10% of the data (for demo purposes)
+trainer = Trainer(experiment=exp, max_nb_epochs=1, train_percent_check=0.1)
+
+# train on 4 gpus
+# trainer = Trainer(experiment=exp, max_nb_epochs=1, gpus=[0, 1, 2, 3])
+
+# train on 32 gpus across 4 nodes (make sure to submit appropriate SLURM job)
+# trainer = Trainer(experiment=exp, max_nb_epochs=1, gpus=[0, 1, 2, 3, 4, 5, 6, 7], nb_gpu_nodes=4)
+
+# train (1 epoch only here for demo)
+trainer.fit(model)
+
+# view tensorflow logs 
+print(f'View tensorboard logs by running\ntensorboard --logdir {os.getcwd()}')
+print('and going to http://localhost:6006 on your browser')
+```
+
 
 ## What does lightning control for me?
-Everything! Except the following three things:
+Everything!    
+Except for these 6 core functions which you define:    
 
-**What happens in the training loop**    
+```{.python}
+# what to do in the training loop
+def training_step(self, data_batch, batch_nb):
+
+# what to do in the validation loop
+def validation_step(self, data_batch, batch_nb):
+
+# how to aggregate validation_step outputs
+def validation_end(self, outputs):
+
+# and your dataloaders
+def tng_dataloader():
+def val_dataloader():
+def test_dataloader():
+```
+
+**Could be as complex as seq-2-seq + attention**    
 
 ```python
 # define what happens for training here
@@ -46,25 +146,39 @@ def training_step(self, data_batch, batch_nb):
     x, y = data_batch
     
     # define your own forward and loss calculation
-    out = self.forward(x)
-    loss = my_loss(out, y)
+    hidden_states = self.encoder(x)
+     
+    # even as complex as a seq-2seq + attn model
+    # (this is just a toy, non-working example to illustrate)
+    start_token = '<SOS>'
+    last_hidden = torch.zeros(...)
+    loss = 0
+    for step in range(max_seq_len):
+        attn_context = self.attention_nn(hidden_states, start_token)
+        pred = self.decoder(start_token, attn_context, last_hidden) 
+        last_hidden = pred
+        pred = self.predict_nn(pred)
+        loss += self.loss(last_hidden, y[step])
+        
+    #toy example as well
+    loss = loss / max_seq_len
     return {'loss': loss} 
 ```
 
-**What happens in the validation loop**      
+**Or as basic as CNN image classification**      
 
 ```python
 # define what happens for validation here
 def validation_step(self, data_batch, batch_nb):    
     x, y = data_batch
     
-    # define your own forward and loss calculation
+    # or as basic as a CNN classification
     out = self.forward(x)
     loss = my_loss(out, y)
     return {'loss': loss} 
 ```
 
-**And what to do with the output of all validation batches**    
+**And you also decide how to collate the output of all validation steps**    
 
 ```python
 def validation_end(self, outputs):
@@ -84,22 +198,53 @@ def validation_end(self, outputs):
     tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()}
     return tqdm_dic
 ```
+   
+## Tensorboard    
+Lightning is fully integrated with tensorboard.   
 
-## Lightning gives you options to control the following:
+<p align="center">
+  <a href="https://williamfalcon.github.io/pytorch-lightning/">
+    <img alt="" src="https://github.com/williamFalcon/pytorch-lightning/blob/master/docs/source/_static/tf_loss.png" width="900px">
+  </a>
+</p>
 
-**Checkpointing**    
+Lightning also adds a text column with all the hyperparameters for this experiment.      
 
-- Model saving
-- Model loading 
+<p align="center">
+  <a href="https://williamfalcon.github.io/pytorch-lightning/">
+        <img alt="" src="https://github.com/williamFalcon/pytorch-lightning/blob/master/docs/source/_static/tf_tags.png" width="900px">
+  </a>
+</p>
 
-**Computing cluster (SLURM)**    
+Simply note the path you set for the Experiment    
+``` {.python}   
+from test_tube import Experiment
+from pytorch-lightning import  Trainer
 
-- Automatic checkpointing   
-- Automatic saving, loading  
-- Running grid search on a cluster 
-- Walltime auto-resubmit   
+exp = Experiment(save_dir='/some/path')
+trainer = Trainer(experiment=exp)
+...
+```   
 
-**Debugging**  
+And run tensorboard from that dir   
+```bash
+tensorboard --logdir /some/path     
+```    
+
+## Lightning automates all of the following ([each is also configurable](https://williamfalcon.github.io/pytorch-lightning/Trainer/)):
+
+
+###### Checkpointing    
+
+- [Model saving](https://williamfalcon.github.io/pytorch-lightning/Trainer/Checkpointing/#model-saving)
+- [Model loading](https://williamfalcon.github.io/pytorch-lightning/LightningModule/methods/#load-from-metrics) 
+
+###### Computing cluster (SLURM)    
+
+- [Running grid search on a cluster](https://williamfalcon.github.io/pytorch-lightning/Trainer/SLURM%20Managed%20Cluster#running-grid-search-on-a-cluster) 
+- [Walltime auto-resubmit](https://williamfalcon.github.io/pytorch-lightning/Trainer/SLURM%20Managed%20Cluster#walltime-auto-resubmit)   
+
+###### Debugging  
 
 - [Fast dev run](https://williamfalcon.github.io/pytorch-lightning/Trainer/debugging/#fast-dev-run)
 - [Inspect gradient norms](https://williamfalcon.github.io/pytorch-lightning/Trainer/debugging/#inspect-gradient-norms)
@@ -109,7 +254,7 @@ def validation_end(self, outputs):
 - [Pring which gradients are nan](https://williamfalcon.github.io/pytorch-lightning/Trainer/debugging/#print-which-gradients-are-nan)
 
 
-**Distributed training**    
+###### Distributed training    
 
 - [16-bit mixed precision](https://williamfalcon.github.io/pytorch-lightning/Trainer/Distributed%20training/#16-bit-mixed-precision)
 - [Multi-GPU](https://williamfalcon.github.io/pytorch-lightning/Trainer/Distributed%20training/#Multi-GPU)
@@ -118,34 +263,38 @@ def validation_end(self, outputs):
 - [Self-balancing architecture](https://williamfalcon.github.io/pytorch-lightning/Trainer/Distributed%20training/#self-balancing-architecture)
 
 
-**Experiment Logging**   
+###### Experiment Logging   
 
 - [Display metrics in progress bar](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#display-metrics-in-progress-bar)
-- Log arbitrary metrics
 - [Log metric row every k batches](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#log-metric-row-every-k-batches)
-- [Process position](Logging/#process-position)
+- [Process position](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#process-position)
+- [Tensorboard support](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#tensorboard-support)
 - [Save a snapshot of all hyperparameters](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#save-a-snapshot-of-all-hyperparameters) 
 - [Snapshot code for a training run](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#snapshot-code-for-a-training-run) 
 - [Write logs file to csv every k batches](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#write-logs-file-to-csv-every-k-batches)
 
-**Training loop**    
+###### Training loop    
 
 - [Accumulate gradients](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#accumulated-gradients)
-- [Anneal Learning rate](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#anneal-learning-rate)
 - [Force training for min or max epochs](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#force-training-for-min-or-max-epochs)
 - [Force disable early stop](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#force-disable-early-stop)
-- [Use multiple optimizers (like GANs)](https://williamfalcon.github.io/pytorch-lightning/Pytorch-Lightning/LightningModule/#configure_optimizers)
+- [Gradient Clipping](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#gradient-clipping)
+- [Hooks](https://williamfalcon.github.io/pytorch-lightning/Trainer/hooks/)
+- [Learning rate scheduling](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/#configure_optimizers)    
+- [Use multiple optimizers (like GANs)](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/#configure_optimizers)
 - [Set how much of the training set to check (1-100%)](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#set-how-much-of-the-training-set-to-check)
 
-**Validation loop**    
+###### Validation loop    
 
 - [Check validation every n epochs](https://williamfalcon.github.io/pytorch-lightning/Trainer/Validation%20loop/#check-validation-every-n-epochs)
+- [Hooks](https://williamfalcon.github.io/pytorch-lightning/Trainer/hooks/)
 - [Set how much of the validation set to check](https://williamfalcon.github.io/pytorch-lightning/Trainer/Validation%20loop/#set-how-much-of-the-validation-set-to-check)
 - [Set how much of the test set to check](https://williamfalcon.github.io/pytorch-lightning/Trainer/Validation%20loop/#set-how-much-of-the-test-set-to-check)
 - [Set validation check frequency within 1 training epoch](https://williamfalcon.github.io/pytorch-lightning/Trainer/Validation%20loop/#set-validation-check-frequency-within-1-training-epoch)
 - [Set the number of validation sanity steps](https://williamfalcon.github.io/pytorch-lightning/Trainer/Validation%20loop/#set-the-number-of-validation-sanity-steps)
 
 
+
 ## Demo
 ```bash
 # install lightning
@@ -153,19 +302,22 @@ pip install pytorch-lightning
 
 # clone lightning for the demo
 git clone https://github.com/williamFalcon/pytorch-lightning.git
-cd examples/new_project_templates/
+cd pytorch_lightning/examples/new_project_templates/
 
-# run demo (on cpu)
-python trainer_gpu_cluster_template.py
+# all of the following demos use the SAME model to show no modification needs to be made to your code
+
+# train on cpu 
+python single_cpu_template.py
+
+# train on multiple-gpus 
+python single_gpu_node_template.py --gpus "0,1"
+
+# train on 32 gpus on a cluster (run on a SLURM managed cluster)
+python multi_node_cluster_template.py --nb_gpu_nodes 4 --gpus '0,1,2,3,4,5,6,7'
 ```
 
-Without changing the model AT ALL, you can run the model on a single gpu, over multiple gpus, or over multiple nodes.
+## Bleeding edge
+If you can't wait for the next release, install the most up to date code with:  
 ```bash
-# run a grid search on two gpus
-python fully_featured_trainer.py --gpus "0;1"
-
-# run single model on multiple gpus
-python fully_featured_trainer.py --gpus "0;1" --interactive
+pip install git+https://github.com/williamFalcon/pytorch-lightning.git@master --upgrade
 ```
-
-

coverage.svg

@@ -0,0 +1,21 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<svg xmlns="http://www.w3.org/2000/svg" width="99" height="20">
+    <linearGradient id="b" x2="0" y2="100%">
+        <stop offset="0" stop-color="#bbb" stop-opacity=".1"/>
+        <stop offset="1" stop-opacity=".1"/>
+    </linearGradient>
+    <mask id="a">
+        <rect width="99" height="20" rx="3" fill="#fff"/>
+    </mask>
+    <g mask="url(#a)">
+        <path fill="#555" d="M0 0h63v20H0z"/>
+        <path fill="#4c1" d="M63 0h36v20H63z"/>
+        <path fill="url(#b)" d="M0 0h99v20H0z"/>
+    </g>
+    <g fill="#fff" text-anchor="middle" font-family="DejaVu Sans,Verdana,Geneva,sans-serif" font-size="11">
+        <text x="31.5" y="15" fill="#010101" fill-opacity=".3">coverage</text>
+        <text x="31.5" y="14">coverage</text>
+        <text x="80" y="15" fill="#010101" fill-opacity=".3">99%</text>
+        <text x="80" y="14">99%</text>
+    </g>
+</svg>

docs/LightningModule/RequiredTrainerInterface.md

@@ -1,30 +1,85 @@
-# Lightning module  
+# Lightning Module interface
 [[Github Code](https://github.com/williamFalcon/pytorch-lightning/blob/master/pytorch_lightning/root_module/root_module.py)]
 
 A lightning module is a strict superclass of nn.Module, it provides a standard interface for the trainer to interact with the model.
 
-The easiest thing to do is copy [this template](../../examples/new_project_templates/lightning_module_template.py) and modify accordingly. 
+The easiest thing to do is copy the [minimal example](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/#minimal-example) below and modify accordingly. 
 
 Otherwise, to Define a Lightning Module, implement the following methods:
 
 **Required**:  
 
-- [training_step](LightningModule.md#training_step)   
-- [validation_step](LightningModule.md#validation_step)
-- [validation_end](LightningModule.md#validation_end)
+- [training_step](RequiredTrainerInterface.md#training_step)   
+- [validation_step](RequiredTrainerInterface.md#validation_step)
+- [validation_end](RequiredTrainerInterface.md#validation_end)
     
-- [configure_optimizers](LightningModule.md#configure_optimizers)
-- [get_save_dict](LightningModule.md#get_save_dict)
-- [load_model_specific](LightningModule.md#load_model_specific)
+- [configure_optimizers](RequiredTrainerInterface.md#configure_optimizers)
 
-- [tng_dataloader](LightningModule.md#tng_dataloader)
-- [tng_dataloader](LightningModule.md#tng_dataloader)
-- [test_dataloader](LightningModule.md#test_dataloader)
+- [tng_dataloader](RequiredTrainerInterface.md#tng_dataloader)
+- [tng_dataloader](RequiredTrainerInterface.md#tng_dataloader)
+- [test_dataloader](RequiredTrainerInterface.md#test_dataloader)
 
 **Optional**:   
 
-- [update_tng_log_metrics](LightningModule.md#update_tng_log_metrics)
-- [add_model_specific_args](LightningModule.md#add_model_specific_args)
+- [on_save_checkpoint](RequiredTrainerInterface.md#on_save_checkpoint)
+- [on_load_checkpoint](RequiredTrainerInterface.md#on_load_checkpoint)
+- [update_tng_log_metrics](RequiredTrainerInterface.md#update_tng_log_metrics)
+- [add_model_specific_args](RequiredTrainerInterface.md#add_model_specific_args)
+
+---
+### Minimal example   
+```python
+import os
+import torch
+from torch.nn import functional as F
+from torch.utils.data import DataLoader
+from torchvision.datasets import MNIST
+import torchvision.transforms as transforms
+
+import pytorch_lightning as ptl
+
+class CoolModel(ptl.LightningModule):
+
+    def __init__(self):
+        super(CoolModel, self).__init__()
+        # not the best model...
+        self.l1 = torch.nn.Linear(28 * 28, 10)
+
+    def forward(self, x):
+        return torch.relu(self.l1(x.view(x.size(0), -1)))
+
+    def my_loss(self, y_hat, y):
+        return F.cross_entropy(y_hat, y)
+
+    def training_step(self, batch, batch_nb):
+        x, y = batch
+        y_hat = self.forward(x)
+        return {'loss': self.my_loss(y_hat, y)}
+
+    def validation_step(self, batch, batch_nb):
+        x, y = batch
+        y_hat = self.forward(x)
+        return {'val_loss': self.my_loss(y_hat, y)}
+
+    def validation_end(self, outputs):
+        avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
+        return {'avg_val_loss': avg_loss}
+
+    def configure_optimizers(self):
+        return [torch.optim.Adam(self.parameters(), lr=0.02)]
+
+    @ptl.data_loader
+    def tng_dataloader(self):
+        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)
+
+    @ptl.data_loader
+    def val_dataloader(self):
+        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)
+
+    @ptl.data_loader
+    def test_dataloader(self):
+        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)
+```
 
 ---
 
@@ -170,57 +225,59 @@ def validation_end(self, outputs):
 def configure_optimizers(self)
 ```
 
-Set up as many optimizers as you need. Normally you'd need one. But in the case of GANs or something more esoteric you might have multiple. 
-Lightning will call .backward() and .step() on each one.  If you use 16 bit precision it will also handle that.
+Set up as many optimizers and (optionally) learning rate schedulers as you need. Normally you'd need one. But in the case of GANs or something more esoteric you might have multiple. 
+Lightning will call .backward() and .step() on each one in every epoch.  If you use 16 bit precision it will also handle that.
 
 
 ##### Return
-List - List of optimizers
+List or Tuple - List of optimizers with an optional second list of learning-rate schedulers
 
 **Example**
 
 ``` {.python}
 # most cases
 def configure_optimizers(self):
-    opt = Adam(lr=0.01)
+    opt = Adam(self.parameters(), lr=0.01)
     return [opt]
     
-# gan example
+# gan example, with scheduler for discriminator
 def configure_optimizers(self):
-    generator_opt = Adam(lr=0.01)
-    disriminator_opt = Adam(lr=0.02)
-    return [generator_opt, disriminator_opt] 
+    generator_opt = Adam(self.model_gen.parameters(), lr=0.01)
+    disriminator_opt = Adam(self.model_disc.parameters(), lr=0.02)
+    discriminator_sched = CosineAnnealing(discriminator_opt, T_max=10)
+    return [generator_opt, disriminator_opt], [discriminator_sched]
 ```
 
 --- 
-### get_save_dict 
+### on_save_checkpoint 
 
 ``` {.python}
-def get_save_dict(self)
+def on_save_checkpoint(self, checkpoint)
 ```
-Called by lightning to checkpoint your model. Lightning saves current epoch, current batch nb, etc...
-All you have to return is what specifically about your lightning model you want to checkpoint.
+Called by lightning to checkpoint your model. Lightning saves the training state (current epoch, global_step, etc)
+and also saves the model state_dict. If you want to save anything else, use this method to add your own
+key-value pair.
 
 ##### Return
-Dictionary - No required keys. Most of the time as described in this example.   
+Nothing
 
 **Example**
 
 ``` {.python}
-def get_save_dict(self):
-    # 99% of use cases this is all you need to return
-    checkpoint = {'state_dict': self.state_dict()}
-    return checkpoint
+def on_save_checkpoint(self, checkpoint):
+    # 99% of use cases you don't need to implement this method 
+    checkpoint['something_cool_i_want_to_save'] = my_cool_pickable_object
 ```
 
 --- 
-### load_model_specific 
+### on_load_checkpoint 
 
 ``` {.python}
-def load_model_specific(self, checkpoint)
+def on_load_checkpoint(self, checkpoint)
 ```
-Called by lightning to restore your model. This is your chance to restore your model using the keys you added in get_save_dict.
-Lightning will automatically restore current epoch, batch nb, etc. 
+Called by lightning to restore your model. Lighting auto-restores global step, epoch, etc...
+It also restores the model state_dict.
+If you saved something with **on_save_checkpoint** this is your chance to restore this.
 
 ##### Return
 Nothing  
@@ -228,19 +285,19 @@ Nothing
 **Example**
 
 ``` {.python}
-def load_model_specific(self, checkpoint):
-    # you defined 'state_dict' in get_save_dict()
-    self.load_state_dict(checkpoint['state_dict'])
+def on_load_checkpoint(self, checkpoint):
+    # 99% of the time you don't need to implement this method
+    self.something_cool_i_want_to_save = checkpoint['something_cool_i_want_to_save']
 ```
 
 --- 
 ### tng_dataloader 
 
 ``` {.python}
-@property
+@ptl.data_loader
 def tng_dataloader(self)
 ```
-Called by lightning during training loop. Define it as a property.
+Called by lightning during training loop. Make sure to use the @ptl.data_loader decorator, this ensures not calling this function until the data are needed.
 
 ##### Return
 Pytorch DataLoader
@@ -248,32 +305,26 @@ Pytorch DataLoader
 **Example**
 
 ``` {.python}
-@property
+@ptl.data_loader
 def tng_dataloader(self):
-    if self._tng_dataloader is None:
-        try:
-            transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
-            dataset = MNIST(root='/path/to/mnist/', train=True, transform=transform, download=True)
-            loader = torch.utils.data.DataLoader(
-                dataset=dataset,
-                batch_size=self.hparams.batch_size,
-                shuffle=True
-            )
-            self._tng_dataloader = loader
-        except Exception as e:
-            raise e
-            
-    return self._tng_dataloader
+    transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
+    dataset = MNIST(root='/path/to/mnist/', train=True, transform=transform, download=True)
+    loader = torch.utils.data.DataLoader(
+        dataset=dataset,
+        batch_size=self.hparams.batch_size,
+        shuffle=True
+    )
+    return loader
 ```
 
 --- 
 ### val_dataloader 
 
 ``` {.python}
-@property
+@ptl.data_loader
 def tng_dataloader(self)
 ```
-Called by lightning during validation loop. Define it as a property.
+Called by lightning during validation loop. Make sure to use the @ptl.data_loader decorator, this ensures not calling this function until the data are needed.
 
 ##### Return
 Pytorch DataLoader
@@ -281,32 +332,27 @@ Pytorch DataLoader
 **Example**
 
 ``` {.python}
-@property
+@ptl.data_loader
 def val_dataloader(self):
-    if self._val_dataloader is None:
-        try:
-            transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
-            dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True)
-            loader = torch.utils.data.DataLoader(
-                dataset=dataset,
-                batch_size=self.hparams.batch_size,
-                shuffle=True
-            )
-            self._val_dataloader = loader
-        except Exception as e:
-            raise e
-            
-    return self._val_dataloader
+    transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
+    dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True)
+    loader = torch.utils.data.DataLoader(
+        dataset=dataset,
+        batch_size=self.hparams.batch_size,
+        shuffle=True
+    )
+    
+    return loader
 ```
 
 --- 
 ### test_dataloader 
 
 ``` {.python}
-@property
+@ptl.data_loader
 def test_dataloader(self)
 ```
-Called by lightning during test loop. Define it as a property.
+Called by lightning during test loop. Make sure to use the @ptl.data_loader decorator, this ensures not calling this function until the data are needed.
 
 ##### Return
 Pytorch DataLoader
@@ -314,22 +360,17 @@ Pytorch DataLoader
 **Example**
 
 ``` {.python}
-@property
+@ptl.data_loader
 def test_dataloader(self):
-    if self._test_dataloader is None:
-        try:
-            transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
-            dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True)
-            loader = torch.utils.data.DataLoader(
-                dataset=dataset,
-                batch_size=self.hparams.batch_size,
-                shuffle=True
-            )
-            self._test_dataloader = loader
-        except Exception as e:
-            raise e
-            
-    return self._test_dataloader
+    transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
+    dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True)
+    loader = torch.utils.data.DataLoader(
+        dataset=dataset,
+        batch_size=self.hparams.batch_size,
+        shuffle=True
+    )
+    
+    return loader
 ```
 
 --- 
@@ -361,7 +402,7 @@ def add_model_specific_args(parent_parser, root_dir)
 ```
 Lightning has a list of default argparse commands.
 This method is your chance to add or modify commands specific to your model.
-The argument parser is available anywhere in your model by calling self.hparams
+The [hyperparameter argument parser](https://williamfalcon.github.io/test-tube/hyperparameter_optimization/HyperOptArgumentParser/) is available anywhere in your model by calling self.hparams.
 
 ##### Return
 An argument parser

docs/LightningModule/methods.md

@@ -0,0 +1,50 @@
+Lightning modules are strict superclasses of torch.nn.Module. A LightningModule offers the following in addition to that API.
+
+---    
+### freeze 
+Freeze all params for inference
+```{.python}
+model = MyLightningModule(...)
+model.freeze()
+```
+
+---    
+### load_from_metrics
+This is the easiest/fastest way which uses the meta_tags.csv file from test-tube to rebuild the model.
+The meta_tags.csv file can be found in the test-tube experiment save_dir.       
+
+```{.python}
+pretrained_model = MyLightningModule.load_from_metrics(
+    weights_path='/path/to/pytorch_checkpoint.ckpt',
+    tags_csv='/path/to/test_tube/experiment/version/meta_tags.csv',
+    on_gpu=True,
+    map_location=None
+)
+    
+# predict
+pretrained_model.eval()
+pretrained_model.freeze()
+y_hat = pretrained_model(x)
+```
+
+**Params**    
+
+| Param  | description  |
+|---|---|
+|  weights_path | Path to a pytorch checkpoint  |
+|  tags_csv | Path to meta_tags.csv file generated by the test-tube Experiment  |
+|  on_gpu | if True, puts model on GPU. Make sure to use transforms option if model devices have changed  |
+|  map_location | A dictionary mapping saved weight GPU devices to new GPU devices |
+
+**Returns**    
+
+LightningModule - The pretrained LightningModule
+
+---    
+### unfreeze 
+Unfreeze all params for inference
+```{.python}
+model = MyLightningModule(...)
+model.unfreeze()
+```
+

docs/LightningModule/properties.md

@@ -0,0 +1,40 @@
+A LightningModule has the following properties which you can access at any time
+
+--- 
+#### current_epoch
+The current epoch   
+
+---
+#### dtype    
+Current dtype    
+
+---
+#### experiment    
+An instance of test-tube Experiment which you can use to log anything for tensorboarX.   
+```{.python}
+self.experiment.add_embedding(...)   
+self.experiment.log({'val_loss': 0.9})   
+self.experiment.add_scalars(...)   
+```
+
+--- 
+#### global_step   
+Total training batches seen across all epochs   
+
+--- 
+#### gradient_clip   
+The current gradient clip value    
+
+---
+#### on_gpu    
+True if your model is currently running on GPUs. Useful to set flags around the LightningModule for different CPU vs GPU behavior.    
+
+---
+#### trainer    
+Last resort access to any state the trainer has. Changing certain properties here could affect your training run.
+```{.python}   
+self.trainer.optimizers
+self.trainer.current_epoch
+...   
+```   
+

docs/Trainer/Checkpointing.md

@@ -0,0 +1,22 @@
+Lightning can automate saving and loading checkpoints.
+
+---
+### Model saving
+To enable checkpointing, define the checkpoint callback and give it to the trainer.
+
+``` {.python}
+from pytorch_lightning.utils.pt_callbacks import ModelCheckpoint
+
+checkpoint_callback = ModelCheckpoint(
+    filepath='/path/to/store/weights.ckpt',
+    save_best_only=True,
+    verbose=True,
+    monitor='val_loss',
+    mode='min'
+)
+
+trainer = Trainer(checkpoint_callback=checkpoint_callback)
+```
+
+
+

docs/Trainer/Distributed training.md

@@ -3,6 +3,36 @@ Lightning makes multi-gpu training and 16 bit training trivial.
 *Note:*   
 None of the flags below require changing anything about your lightningModel definition. 
 
+---
+#### Choosing a backend  
+Lightning supports two backends. DataParallel and DistributedDataParallel. Both can be used for single-node multi-GPU training.
+For multi-node training you must use DistributedDataParallel.   
+
+You can toggle between each mode by setting this flag.
+``` {.python}
+# DEFAULT uses DataParallel
+trainer = Trainer(distributed_backend='dp')
+
+# change to distributed data parallel
+trainer = Trainer(distributed_backend='ddp')
+```
+
+If you request multiple nodes, the back-end will auto-switch to ddp.
+We recommend you use DistributedDataparallel even for single-node multi-GPU training. It is MUCH faster than DP but *may*
+have configuration issues depending on your cluster.
+
+For a deeper understanding of what lightning is doing, feel free to read [this guide](https://medium.com/@_willfalcon/9-tips-for-training-lightning-fast-neural-networks-in-pytorch-8e63a502f565).   
+
+---
+#### CUDA flags   
+CUDA flags make certain GPUs visible to your script. 
+Lightning sets these for you automatically, there's NO NEED to do this yourself.
+```python
+# lightning will set according to what you give the trainer
+# os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+# os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+```
+
 ---
 #### 16-bit mixed precision
 16 bit precision can cut your memory footprint by half. If using volta architecture GPUs it can give a dramatic training speed-up as well.    
@@ -23,10 +53,6 @@ trainer = Trainer(amp_level='O2', use_amp=False)
 #### Single-gpu
 Make sure you're on a GPU machine. 
 ```python
-# set these flags
-os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
-os.environ["CUDA_VISIBLE_DEVICES"] = "0"
-
 # DEFAULT
 trainer = Trainer(gpus=[0])
 ```
@@ -36,17 +62,63 @@ trainer = Trainer(gpus=[0])
 Make sure you're on a GPU machine. You can set as many GPUs as you want.
 In this setting, the model will run on all 8 GPUs at once using DataParallel under the hood.
 ```python
-# set these flags
-os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
-os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3,4,5,6,7"
+# to use DataParallel (default)
+trainer = Trainer(gpus=[0,1,2,3,4,5,6,7], distributed_backend='dp')
 
-# DEFAULT
-trainer = Trainer(gpus=[0,1,2,3,4,5,6,7])
+# RECOMMENDED use DistributedDataParallel
+trainer = Trainer(gpus=[0,1,2,3,4,5,6,7], distributed_backend='ddp')
 ```
 
 ---
 #### Multi-node
-COMING SOON.
+Multi-node training is easily done by specifying these flags.
+```python
+# train on 12*8 GPUs
+trainer = Trainer(gpus=[0,1,2,3,4,5,6,7], nb_gpu_nodes=12)
+```
+
+In addition, make sure to set up your SLURM job correctly via the [SlurmClusterObject](https://williamfalcon.github.io/test-tube/hpc/SlurmCluster/). In particular, specify the number of tasks per node correctly.
+
+```python
+cluster = SlurmCluster(
+    hyperparam_optimizer=test_tube.HyperOptArgumentParser(),
+    log_path='/some/path/to/save',
+)
+
+# OPTIONAL FLAGS WHICH MAY BE CLUSTER DEPENDENT
+# which interface your nodes use for communication
+cluster.add_command('export NCCL_SOCKET_IFNAME=^docker0,lo')
+
+# see output of the NCCL connection process
+# NCCL is how the nodes talk to each other
+cluster.add_command('export NCCL_DEBUG=INFO')
+
+# setting a master port here is a good idea.
+cluster.add_command(f'export MASTER_PORT={PORT}')
+
+# good to load the latest NCCL version
+cluster.load_modules(['NCCL/2.4.7-1-cuda.10.0'])
+
+# configure cluster
+cluster.per_experiment_nb_nodes = 12 
+cluster.per_experiment_nb_gpus = 8
+
+cluster.add_slurm_cmd(cmd='ntasks-per-node', value=8, comment='1 task per gpu')
+```
+
+Finally, make sure to add a distributed sampler to your dataset. The distributed sampler copies a 
+portion of your dataset onto each GPU. (World_size = gpus_per_node * nb_nodes).   
+
+```python
+# ie: this:
+dataset = myDataset()
+dataloader = Dataloader(dataset)
+
+# becomes:
+dataset = myDataset()
+dist_sampler = torch.utils.data.distributed.DistributedSampler(dataset)
+dataloader = Dataloader(dataset, sampler=dist_sampler)
+```
 
 ---
 #### Self-balancing architecture

docs/Trainer/Logging.md

@@ -50,6 +50,33 @@ exp = Experiment(create_git_tag=True)
 Trainer(experiment=exp)
 ```
 
+---
+### Tensorboard support   
+The experiment object is a strict subclass of Pytorch SummaryWriter. However, this class
+also snapshots every detail about the experiment (data folder paths, code, hyperparams),
+and allows you to visualize it using tensorboard.
+``` {.python}
+from test_tube import Experiment, HyperOptArgumentParser
+
+# exp hyperparams
+args = HyperOptArgumentParser()
+hparams = args.parse_args()
+
+# this is a summaryWriter with nicer logging structure
+exp = Experiment(save_dir='/some/path', create_git_tag=True)
+
+# track experiment details (must be ArgumentParser or HyperOptArgumentParser).
+# each option in the parser is tracked
+exp.argparse(hparams)
+exp.tag({'description': 'running demo'})
+
+# trainer uses the exp object to log exp data
+trainer = Trainer(experiment=exp)
+trainer.fit(model)
+
+# view logs at:
+# tensorboard --logdir /some/path   
+```
 
 ---
 #### Write logs file to csv every k batches 

docs/Trainer/SLURM Managed Cluster.md

@@ -24,7 +24,7 @@ hparams = parser.parse_args()
 ```    
     
      
-(2). Define the cluster options (over 5 nodes and 8 gpus)    
+(2). Define the cluster options in the [SlurmCluster object](https://williamfalcon.github.io/test-tube/hpc/SlurmCluster/) (over 5 nodes and 8 gpus)    
 
 ```{.python}
 from test_tube.hpc import SlurmCluster

docs/Trainer/Training Loop.md

@@ -1,4 +1,4 @@
-The lightning training loop handles everything except the actual computations of your model. To decide what will happen in your training loop, define the [training_step function](../../Pytorch-lightning/LightningModule/#training_step).
+The lightning training loop handles everything except the actual computations of your model. To decide what will happen in your training loop, define the [training_step function](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/#training_step).
 
 Below are all the things lightning automates for you in the training loop.
 
@@ -11,17 +11,6 @@ Accumulated gradients runs K small batches of size N before doing a backwards pa
 trainer = Trainer(accumulate_grad_batches=1)
 ```
 
----
-#### Anneal Learning rate
-Cut the learning rate by 10 at every epoch listed in this list.
-``` {.python}
-# DEFAULT (don't anneal)
-trainer = Trainer(lr_scheduler_milestones=None)
-
-# cut LR by 10 at 100, 200, and 300 epochs 
-trainer = Trainer(lr_scheduler_milestones=[100, 200, 300])
-```
-
 ---
 #### Force training for min or max epochs
 It can be useful to force training for a minimum number of epochs or limit to a max number
@@ -38,6 +27,16 @@ Use this to turn off early stopping and run training to the [max_epoch](#force-t
 trainer = Trainer(enable_early_stop=True)
 ```
 
+---
+#### Gradient Clipping 
+Use this to turn off early stopping and run training to the [max_epoch](#force-training-for-min-or-max-epochs)
+``` {.python}
+# DEFAULT (ie: don't clip)
+trainer = Trainer(gradient_clip=0)
+```
+
+
+
 ---
 #### Inspect gradient norms
 Looking at grad norms can help you figure out where training might be going wrong.

docs/Trainer/Validation loop.md

@@ -1,4 +1,4 @@
-The lightning validation loop handles everything except the actual computations of your model. To decide what will happen in your validation loop, define the [validation_step function](../../Pytorch-lightning/LightningModule/#validation_step).
+The lightning validation loop handles everything except the actual computations of your model. To decide what will happen in your validation loop, define the [validation_step function](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/#validation_step).
 Below are all the things lightning automates for you in the validation loop.
 
 **Note**   

docs/Trainer/hooks.md

@@ -0,0 +1,86 @@
+# Hooks
+[[Github Code](https://github.com/williamFalcon/pytorch-lightning/blob/master/pytorch_lightning/root_module/hooks.py)]   
+
+There are cases when you might want to do something different at different parts of the training/validation loop.
+To enable a hook, simply override the method in your LightningModule and the trainer will call it at the correct time.
+
+**Contributing** If there's a hook you'd like to add, simply:    
+1. Fork PytorchLightning.    
+2. Add the hook [here](https://github.com/williamFalcon/pytorch-lightning/blob/master/pytorch_lightning/root_module/hooks.py).       
+3. Add the correct place in the [Trainer](https://github.com/williamFalcon/pytorch-lightning/blob/master/pytorch_lightning/models/trainer.py) where it should be called.    
+
+---
+#### on_epoch_start
+Called in the training loop at the very beginning of the epoch.   
+```python
+def on_epoch_start(self):
+    # do something when the epoch starts
+```
+
+---
+#### on_batch_end
+Called in the training loop at the very end of the epoch.   
+```python
+def on_epoch_end(self):
+    # do something when the epoch ends 
+```
+
+---
+#### on_batch_start
+Called in the training loop before anything happens for that batch.   
+```python
+def on_batch_start(self):
+    # do something when the batch starts
+```
+
+---
+#### on_pre_performance_check
+Called at the very beginning of the validation loop.   
+```python
+def on_pre_performance_check(self):
+    # do something before validation starts 
+```
+
+---
+#### on_post_performance_check
+Called at the very end of the validation loop.   
+```python
+def on_post_performance_check(self):
+    # do something before validation end
+```
+
+---
+#### on_tng_metrics
+Called in the training loop, right before metrics are logged.
+Although you can log at any time by using self.experiment, you can use
+this callback to modify what will be logged.
+```python
+def on_tng_metrics(self, metrics):
+    # do something before validation end
+```
+
+---
+#### on_before_zero_grad
+Called in the training loop after taking an optimizer step and before zeroing grads.
+Good place to inspect weight information with weights updated.
+
+Called once per optimizer
+```python
+def on_before_zero_grad(self, optimizer):
+    # do something with the optimizer or inspect it. 
+```
+
+---
+#### on_after_backward
+Called in the training loop after model.backward()
+This is the ideal place to inspect or log gradient information 
+```python
+def on_after_backward(self):
+    # example to inspect gradient information in tensorboard
+    if self.trainer.global_step % 25 == 0:  # don't make the tf file huge
+        params = self.state_dict()
+        for k, v in params.items():
+            grads = v
+            name = k
+            self.experiment.add_histogram(tag=name, values=grads, global_step=self.trainer.global_step)
+```

docs/Trainer/index.md

@@ -24,10 +24,8 @@ But of course the fun is in all the advanced things it can do:
 
 **Computing cluster (SLURM)**    
 
-- Automatic checkpointing   
-- Automatic saving, loading  
-- Running grid search on a cluster
-- Walltime auto-resubmit   
+- [Running grid search on a cluster](SLURM%20Managed%20Cluster/#running-grid-search-on-a-cluster)
+- [Walltime auto-resubmit](SLURM%20Managed%20Cluster/#walltime-auto-resubmit)   
 
 **Debugging**  
 
@@ -51,25 +49,28 @@ But of course the fun is in all the advanced things it can do:
 **Experiment Logging**   
 
 - [Display metrics in progress bar](Logging/#display-metrics-in-progress-bar)
-- Log arbitrary metrics
 - [Log metric row every k batches](Logging/#log-metric-row-every-k-batches)
 - [Process position](Logging/#process-position)
+- [Tensorboard support](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#tensorboard-support)
 - [Save a snapshot of all hyperparameters](Logging/#save-a-snapshot-of-all-hyperparameters) 
 - [Snapshot code for a training run](Logging/#snapshot-code-for-a-training-run) 
 - [Write logs file to csv every k batches](Logging/#write-logs-file-to-csv-every-k-batches)
 
 **Training loop**    
 
-- [Accumulate gradients](Training%20Loop/#accumulated-gradients)
-- [Anneal Learning rate](Training%20Loop/#anneal-learning-rate)
-- [Force training for min or max epochs](Training%20Loop/#force-training-for-min-or-max-epochs)
-- [Force disable early stop](Training%20Loop/#force-disable-early-stop)
-- [Use multiple optimizers (like GANs)](../Pytorch-lightning/LightningModule/#configure_optimizers)
-- [Set how much of the training set to check (1-100%)](Training%20Loop/#set-how-much-of-the-training-set-to-check)
+- [Accumulate gradients](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#accumulated-gradients)
+- [Force training for min or max epochs](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#force-training-for-min-or-max-epochs)
+- [Force disable early stop](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#force-disable-early-stop)
+- [Gradient Clipping](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#gradient-clipping)
+- [Hooks](hooks)
+- [Learning rate scheduling](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/#configure_optimizers)
+- [Use multiple optimizers (like GANs)](https://williamfalcon.github.io/pytorch-lightning/Pytorch-Lightning/LightningModule/#configure_optimizers)
+- [Set how much of the training set to check (1-100%)](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#set-how-much-of-the-training-set-to-check)
 
 **Validation loop**    
 
 - [Check validation every n epochs](Validation%20Loop/#check-validation-every-n-epochs)
+- [Hooks](hooks)
 - [Set how much of the validation set to check](Validation%20Loop/#set-how-much-of-the-validation-set-to-check)
 - [Set how much of the test set to check](Validation%20Loop/#set-how-much-of-the-test-set-to-check)
 - [Set validation check frequency within 1 training epoch](Validation%20Loop/#set-validation-check-frequency-within-1-training-epoch)

docs/doc_requirements.txt

@@ -0,0 +1 @@
+mkdocs-material==4.4.0

docs/examples/Examples.md

@@ -0,0 +1,171 @@
+### Template model definition
+In 99% of cases you want to just copy [this template](https://github.com/williamFalcon/pytorch-lightning/blob/master/examples/new_project_templates/lightning_module_template.py) to start a new lightningModule and change the core of what your model is actually trying to do.
+
+```bash
+# get a copy of the module template
+wget https://github.com/williamFalcon/pytorch-lightning/blob/master/examples/new_project_templates/lightning_module_template.py
+```
+
+---    
+### Trainer Example 
+
+** \_\_main__ function**    
+
+Normally, we want to let the \_\_main__ function start the training.
+Inside the main we parse training arguments with whatever hyperparameters we want. Your LightningModule will have a 
+chance to add hyperparameters.   
+
+```{.python}
+from test_tube import HyperOptArgumentParser
+
+if __name__ == '__main__':
+
+    # use default args given by lightning
+    root_dir = os.path.split(os.path.dirname(sys.modules['__main__'].__file__))[0]
+    parent_parser = HyperOptArgumentParser(strategy='random_search', add_help=False)
+    add_default_args(parent_parser, root_dir)
+
+    # allow model to overwrite or extend args
+    parser = ExampleModel.add_model_specific_args(parent_parser)
+    hyperparams = parser.parse_args()
+
+    # train model
+    main(hyperparams)
+```
+**Main Function**      
+
+The main function is your entry into the program. This is where you init your model, checkpoint directory, and launch the training.
+The main function should have 3 arguments:   
+- hparams: a configuration of hyperparameters.    
+- slurm_manager: Slurm cluster manager object (can be None)
+- dict: for you to return any values you want (useful in meta-learning, otherwise set to _)    
+
+```{}
+def main(hparams, cluster, results_dict):
+    """
+    Main training routine specific for this project
+    :param hparams:
+    :return:
+    """
+    # init experiment
+    log_dir = os.path.dirname(os.path.realpath(__file__))
+    exp = Experiment(
+        name='test_tube_exp',
+        debug=True,
+        save_dir=log_dir,
+        version=0,
+        autosave=False,
+        description='test demo'
+    )
+    
+    # set the hparams for the experiment
+    exp.argparse(hparams)
+    exp.save()
+
+    # build model
+    model = MyLightningModule(hparams)
+
+    # callbacks
+    early_stop = EarlyStopping(
+        monitor=hparams.early_stop_metric,
+        patience=hparams.early_stop_patience,
+        verbose=True,
+        mode=hparams.early_stop_mode
+    )
+
+    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
+    checkpoint = ModelCheckpoint(
+        filepath=model_save_path,
+        save_function=None,
+        save_best_only=True,
+        verbose=True,
+        monitor=hparams.model_save_monitor_value,
+        mode=hparams.model_save_monitor_mode
+    )
+
+    # configure trainer
+    trainer = Trainer(
+        experiment=exp,
+        cluster=cluster,
+        checkpoint_callback=checkpoint,
+        early_stop_callback=early_stop,
+    )
+
+    # train model
+    trainer.fit(model)
+```
+
+
+
+
+The __main__ function will start training on your **main** function. If you use the HyperParameterOptimizer
+in hyper parameter optimization mode, this main function will get one set of hyperparameters. If you use it as a simple
+argument parser you get the default arguments in the argument parser.
+
+So, calling main(hyperparams) runs the model with the default argparse arguments.       
+```{.python}
+main(hyperparams)
+```
+
+---
+#### CPU hyperparameter search      
+
+```{.python}
+# run a grid search over 20 hyperparameter combinations.
+hyperparams.optimize_parallel_cpu(
+    main_local,
+    nb_trials=20,
+    nb_workers=1
+)
+```
+
+---
+#### Hyperparameter search on a single or multiple GPUs       
+```{.python}
+# run a grid search over 20 hyperparameter combinations.
+hyperparams.optimize_parallel_gpu(
+    main_local,
+    nb_trials=20,
+    nb_workers=1,
+    gpus=[0,1,2,3]
+)
+```
+
+---
+#### Hyperparameter search on a SLURM HPC cluster   
+```{.python}    
+def optimize_on_cluster(hyperparams):
+    # enable cluster training
+    cluster = SlurmCluster(
+        hyperparam_optimizer=hyperparams,
+        log_path=hyperparams.tt_save_path,
+        test_tube_exp_name=hyperparams.tt_name
+    )
+
+    # email for cluster coms
+    cluster.notify_job_status(email='add_email_here', on_done=True, on_fail=True)
+
+    # configure cluster
+    cluster.per_experiment_nb_gpus = hyperparams.per_experiment_nb_gpus
+    cluster.job_time = '48:00:00'
+    cluster.gpu_type = '1080ti'
+    cluster.memory_mb_per_node = 48000
+
+    # any modules for code to run in env
+    cluster.add_command('source activate pytorch_lightning')
+
+    # name of exp
+    job_display_name = hyperparams.tt_name.split('_')[0]
+    job_display_name = job_display_name[0:3]
+
+    # run hopt
+    print('submitting jobs...')
+    cluster.optimize_parallel_cluster_gpu(
+        main,
+        nb_trials=hyperparams.nb_hopt_trials,
+        job_name=job_display_name
+    )
+
+# run cluster hyperparameter search    
+optimize_on_cluster(hyperparams)
+```

docs/index.md

@@ -1,35 +1,40 @@
-# PYTORCH-LIGHTNING DOCUMENTATION
+###### New project Quick Start    
+To start a new project you define two files, a LightningModule and a Trainer file.    
 
-###### Main Docs
-- [LightningModule](Pytorch-Lightning/LightningModule)  
+A separate trainer file allows to run many LightningModules. Each LightningModule has the core
+logic to a particular research project.    
+
+For example, one lightningModule could be an image classifier, the other
+one could be a seq-2-seq model, both (optionally) ran by the same trainer file.
+
+1. [MNIST LightningModule](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/#minimal-example) 
+2. [Trainer](https://williamfalcon.github.io/pytorch-lightning/Trainer/)
+    - [Basic CPU Trainer Template](https://github.com/williamFalcon/pytorch-lightning/blob/master/pytorch_lightning/examples/new_project_templates/single_cpu_template.py) 
+    - [Multi-GPU Trainer Template](https://github.com/williamFalcon/pytorch-lightning/blob/master/pytorch_lightning/examples/new_project_templates/single_gpu_node_template.py)
+    - [GPU cluster Trainer Template](https://github.com/williamFalcon/pytorch-lightning/blob/master/pytorch_lightning/examples/new_project_templates/multi_node_cluster_template.py)
+
+###### Docs shortcuts
+- [LightningModule](LightningModule/RequiredTrainerInterface/)  
 - [Trainer](Trainer/)  
 
-###### New project Quick Start
-1. [Define a LightningModule](https://github.com/williamFalcon/pytorch-lightning/blob/master/examples/new_project_templates/lightning_module_template.py)  
-2. Pick a trainer      
-    - [Basic CPU Trainer](https://github.com/williamFalcon/pytorch-lightning/blob/master/examples/new_project_templates/trainer_cpu_template.py) 
-    - [GPU cluster Trainer](https://github.com/williamFalcon/pytorch-lightning/blob/master/examples/new_project_templates/trainer_gpu_cluster_template.py)
-
 ###### Quick start examples 
-- CPU example   
-- Single GPU example   
-- Multi-gpu example 
-- SLURM cluster grid search example      
+- [CPU example](examples/Examples/#cpu-hyperparameter-search)   
+- [Hyperparameter search on single GPU](examples/Examples/#hyperparameter-search-on-a-single-or-multiple-gpus)    
+- [Hyperparameter search on multiple GPUs on same node](examples/Examples/#hyperparameter-search-on-a-single-or-multiple-gpus)  
+- [Hyperparameter search on a SLURM HPC cluster](examples/Examples/#Hyperparameter search on a SLURM HPC cluster)      
 
 
 ###### Checkpointing    
 
-- Model saving
-- Model loading 
+- [Model saving](https://williamfalcon.github.io/pytorch-lightning/Trainer/Checkpointing/#model-saving)
+- [Model loading](https://williamfalcon.github.io/pytorch-lightning/LightningModule/methods/#load-from-metrics) 
 
-######Computing cluster (SLURM)    
+###### Computing cluster (SLURM)    
 
-- Automatic checkpointing   
-- Automatic saving, loading  
-- Running grid search on a cluster 
-- Walltime auto-resubmit   
+- [Running grid search on a cluster](https://williamfalcon.github.io/pytorch-lightning/Trainer/SLURM%20Managed%20Cluster#running-grid-search-on-a-cluster) 
+- [Walltime auto-resubmit](https://williamfalcon.github.io/pytorch-lightning/Trainer/SLURM%20Managed%20Cluster#walltime-auto-resubmit)   
 
-######Debugging  
+###### Debugging  
 
 - [Fast dev run](https://williamfalcon.github.io/pytorch-lightning/Trainer/debugging/#fast-dev-run)
 - [Inspect gradient norms](https://williamfalcon.github.io/pytorch-lightning/Trainer/debugging/#inspect-gradient-norms)
@@ -39,7 +44,7 @@
 - [Pring which gradients are nan](https://williamfalcon.github.io/pytorch-lightning/Trainer/debugging/#print-which-gradients-are-nan)
 
 
-######Distributed training    
+###### Distributed training    
 
 - [16-bit mixed precision](https://williamfalcon.github.io/pytorch-lightning/Trainer/Distributed%20training/#16-bit-mixed-precision)
 - [Multi-GPU](https://williamfalcon.github.io/pytorch-lightning/Trainer/Distributed%20training/#Multi-GPU)
@@ -48,28 +53,31 @@
 - [Self-balancing architecture](https://williamfalcon.github.io/pytorch-lightning/Trainer/Distributed%20training/#self-balancing-architecture)
 
 
-######Experiment Logging   
+###### Experiment Logging   
 
 - [Display metrics in progress bar](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#display-metrics-in-progress-bar)
-- Log arbitrary metrics
 - [Log metric row every k batches](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#log-metric-row-every-k-batches)
-- [Process position](Logging/#process-position)
+- [Process position](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#process-position)
+- [Tensorboard support](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#tensorboard-support)
 - [Save a snapshot of all hyperparameters](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#save-a-snapshot-of-all-hyperparameters) 
 - [Snapshot code for a training run](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#snapshot-code-for-a-training-run) 
 - [Write logs file to csv every k batches](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#write-logs-file-to-csv-every-k-batches)
 
-######Training loop    
+###### Training loop    
 
 - [Accumulate gradients](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#accumulated-gradients)
-- [Anneal Learning rate](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#anneal-learning-rate)
 - [Force training for min or max epochs](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#force-training-for-min-or-max-epochs)
 - [Force disable early stop](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#force-disable-early-stop)
+- [Gradient Clipping](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#gradient-clipping)
+- [Hooks](https://williamfalcon.github.io/pytorch-lightning/Trainer/hooks/)
+- [Learning rate scheduling](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/#configure_optimizers)
 - [Use multiple optimizers (like GANs)](https://williamfalcon.github.io/pytorch-lightning/Pytorch-Lightning/LightningModule/#configure_optimizers)
 - [Set how much of the training set to check (1-100%)](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#set-how-much-of-the-training-set-to-check)
 
-######Validation loop    
+###### Validation loop    
 
 - [Check validation every n epochs](https://williamfalcon.github.io/pytorch-lightning/Trainer/Validation%20loop/#check-validation-every-n-epochs)
+- [Hooks](https://williamfalcon.github.io/pytorch-lightning/Trainer/hooks/)
 - [Set how much of the validation set to check](https://williamfalcon.github.io/pytorch-lightning/Trainer/Validation%20loop/#set-how-much-of-the-validation-set-to-check)
 - [Set how much of the test set to check](https://williamfalcon.github.io/pytorch-lightning/Trainer/Validation%20loop/#set-how-much-of-the-test-set-to-check)
 - [Set validation check frequency within 1 training epoch](https://williamfalcon.github.io/pytorch-lightning/Trainer/Validation%20loop/#set-validation-check-frequency-within-1-training-epoch)

docs/source/examples/__init__.py

@@ -1 +0,0 @@
-from .example_model import ExampleModel

docs/source/examples/basic_trainer.py

@@ -1,74 +0,0 @@
-import os
-import sys
-
-from test_tube import HyperOptArgumentParser, Experiment
-from pytorch_lightning.models.trainer import Trainer
-from pytorch_lightning.utils.arg_parse import add_default_args
-from pytorch_lightning.callbacks.pt_callbacks import EarlyStopping, ModelCheckpoint
-from docs.source.examples.example_model import ExampleModel
-
-
-def main(hparams):
-    """
-    Main training routine specific for this project
-    :param hparams:
-    :return:
-    """
-    # init experiment
-    exp = Experiment(
-        name=hparams.tt_name,
-        debug=hparams.debug,
-        save_dir=hparams.tt_save_path,
-        version=hparams.hpc_exp_number,
-        autosave=False,
-        description=hparams.tt_description
-    )
-
-    exp.argparse(hparams)
-    exp.save()
-
-    # build model
-    model = ExampleModel(hparams)
-
-    # callbacks
-    early_stop = EarlyStopping(
-        monitor='val_acc',
-        patience=3,
-        mode='min',
-        verbose=True,
-    )
-
-    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
-    checkpoint = ModelCheckpoint(
-        filepath=model_save_path,
-        save_function=None,
-        save_best_only=True,
-        verbose=True,
-        monitor='val_acc',
-        mode='min'
-    )
-
-    # configure trainer
-    trainer = Trainer(
-        experiment=exp,
-        checkpoint_callback=checkpoint,
-        early_stop_callback=early_stop,
-    )
-
-    # train model
-    trainer.fit(model)
-
-
-if __name__ == '__main__':
-
-    # use default args given by lightning
-    root_dir = os.path.split(os.path.dirname(sys.modules['__main__'].__file__))[0]
-    parent_parser = HyperOptArgumentParser(strategy='random_search', add_help=False)
-    add_default_args(parent_parser, root_dir)
-
-    # allow model to overwrite or extend args
-    parser = ExampleModel.add_model_specific_args(parent_parser)
-    hyperparams = parser.parse_args()
-
-    # train model
-    main(hyperparams)

docs/source/examples/example_model.py

@@ -1,211 +0,0 @@
-import torch.nn as nn
-import numpy as np
-from pytorch_lightning.root_module.root_module import LightningModule
-from test_tube import HyperOptArgumentParser
-from torchvision.datasets import MNIST
-import torchvision.transforms as transforms
-import torch
-import torch.nn.functional as F
-import os, pdb
-from collections import OrderedDict
-
-
-class ExampleModel(LightningModule):
-    """
-    Sample model to show how to define a template
-    """
-
-    def __init__(self, hparams):
-        # init superclass
-        super(ExampleModel, self).__init__(hparams)
-
-        self.batch_size = hparams.batch_size
-
-        # build model
-        self.__build_model()
-
-    # ---------------------
-    # MODEL SETUP
-    # ---------------------
-    def __build_model(self):
-        """
-        Layout model
-        :return:
-        """
-        self.c_d1 = nn.Linear(in_features=self.hparams.in_features, out_features=self.hparams.hidden_dim)
-        self.c_d1_bn = nn.BatchNorm1d(self.hparams.hidden_dim)
-        self.c_d1_drop = nn.Dropout(self.hparams.drop_prob)
-
-        self.c_d2 = nn.Linear(in_features=self.hparams.hidden_dim, out_features=self.hparams.out_features)
-
-    # ---------------------
-    # TRAINING
-    # ---------------------
-    def forward(self, x):
-
-        x = self.c_d1(x)
-        x = torch.tanh(x)
-        x = self.c_d1_bn(x)
-        x = self.c_d1_drop(x)
-
-        x = self.c_d2(x)
-        logits = F.log_softmax(x, dim=1)
-
-        return logits
-
-    def loss(self, labels, logits):
-        nll = F.nll_loss(logits, labels)
-        return nll
-
-    def training_step(self, data_batch, batch_i):
-        """
-        Called inside the training loop
-        :param data_batch:
-        :return:
-        """
-        # forward pass
-        x, y = data_batch
-        x = x.view(x.size(0), -1)
-        y_hat = self.forward(x)
-
-        # calculate loss
-        loss_val = self.loss(y, y_hat)
-
-        output = OrderedDict({
-            'loss': loss_val,
-            'tqdm_metrics': {}
-        })
-        return output
-
-    def validation_step(self, data_batch, batch_i):
-        """
-        Called inside the validation loop
-        :param data_batch:
-        :return:
-        """
-        x, y = data_batch
-        x = x.view(x.size(0), -1)
-        y_hat = self.forward(x)
-
-        loss_val = self.loss(y, y_hat)
-
-        # acc
-        labels_hat = torch.argmax(y_hat, dim=1)
-        val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)
-
-        output = OrderedDict({
-            'val_loss': loss_val,
-            'val_acc': torch.tensor(val_acc),
-        })
-        return output
-
-
-    def validation_end(self, outputs):
-        """
-        Called at the end of validation to aggregate outputs
-        :param outputs: list of individual outputs of each validation step
-        :return:
-        """
-        val_loss_mean = 0
-        val_acc_mean = 0
-        for output in outputs:
-            val_loss_mean += output['val_loss']
-            val_acc_mean += output['val_acc']
-
-        val_loss_mean /= len(outputs)
-        val_acc_mean /= len(outputs)
-        tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()}
-        return tqdm_dic
-
-    def update_tng_log_metrics(self, logs):
-        return logs
-
-    # ---------------------
-    # MODEL SAVING
-    # ---------------------
-    def get_save_dict(self):
-        checkpoint = {'state_dict': self.state_dict()}
-        return checkpoint
-
-    def load_model_specific(self, checkpoint):
-        self.load_state_dict(checkpoint['state_dict'])
-        pass
-
-    # ---------------------
-    # TRAINING SETUP
-    # ---------------------
-    def configure_optimizers(self):
-        """
-        return whatever optimizers we want here
-        :return: list of optimizers
-        """
-        optimizer = self.choose_optimizer(self.hparams.optimizer_name, self.parameters(), {'lr': self.hparams.learning_rate}, 'optimizer')
-        self.optimizers = [optimizer]
-        return self.optimizers
-
-    def __dataloader(self, train):
-        # init data generators
-        transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
-
-        dataset = MNIST(root=self.hparams.data_root, train=train, transform=transform, download=True)
-
-        loader = torch.utils.data.DataLoader(
-            dataset=dataset,
-            batch_size=self.hparams.batch_size,
-            shuffle=True
-        )
-
-        return loader
-
-    @property
-    def tng_dataloader(self):
-        if self._tng_dataloader is None:
-            try:
-                self._tng_dataloader = self.__dataloader(train=True)
-            except Exception as e:
-                print(e)
-                raise e
-        return self._tng_dataloader
-
-    @property
-    def val_dataloader(self):
-        if self._val_dataloader is None:
-            try:
-                self._val_dataloader = self.__dataloader(train=False)
-            except Exception as e:
-                print(e)
-                raise e
-        return self._val_dataloader
-
-    @property
-    def test_dataloader(self):
-        if self._test_dataloader is None:
-            try:
-                self._test_dataloader = self.__dataloader(train=False)
-            except Exception as e:
-                print(e)
-                raise e
-        return self._test_dataloader
-
-    @staticmethod
-    def add_model_specific_args(parent_parser, root_dir):
-        parser = HyperOptArgumentParser(strategy=parent_parser.strategy, parents=[parent_parser])
-
-        # param overwrites
-        # parser.set_defaults(gradient_clip=5.0)
-
-        # network params
-        parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=False)
-        parser.add_argument('--in_features', default=28*28)
-        parser.add_argument('--out_features', default=10)
-        parser.add_argument('--hidden_dim', default=50000) # use 500 for CPU, 50000 for GPU to see speed difference
-
-        # data
-        parser.add_argument('--data_root', default=os.path.join(root_dir, 'mnist'), type=str)
-
-        # training params (opt)
-        parser.opt_list('--learning_rate', default=0.001, type=float, options=[0.0001, 0.0005, 0.001, 0.005],
-                        tunable=False)
-        parser.opt_list('--batch_size', default=256, type=int, options=[32, 64, 128, 256], tunable=False)
-        parser.opt_list('--optimizer_name', default='adam', type=str, options=['adam'], tunable=False)
-        return parser

docs/source/examples/fully_featured_trainer.py

@@ -1,210 +0,0 @@
-import os
-import sys
-import numpy as np
-from time import sleep
-import torch
-
-from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
-from pytorch_lightning.models.trainer import Trainer
-from pytorch_lightning.utils.arg_parse import add_default_args
-
-from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
-
-SEED = 2334
-torch.manual_seed(SEED)
-np.random.seed(SEED)
-
-# ---------------------
-# DEFINE MODEL HERE
-# ---------------------
-from docs.source.examples.example_model import ExampleModel
-# ---------------------
-
-AVAILABLE_MODELS = {
-    'model_template': ExampleModel
-}
-
-
-"""
-Allows training by using command line arguments
-Run by:
-# TYPE YOUR RUN COMMAND HERE
-"""
-
-
-def main_local(hparams):
-    main(hparams, None, None)
-
-
-def main(hparams, cluster, results_dict):
-    """
-    Main training routine specific for this project
-    :param hparams:
-    :return:
-    """
-    on_gpu = hparams.gpus is not None and torch.cuda.is_available()
-
-    device = 'cuda' if on_gpu else 'cpu'
-    hparams.__setattr__('device', device)
-    hparams.__setattr__('on_gpu', on_gpu)
-    hparams.__setattr__('nb_gpus', torch.cuda.device_count())
-    hparams.__setattr__('inference_mode', hparams.model_load_weights_path is not None)
-
-    # delay each training start to not overwrite logs
-    process_position, current_gpu = TRAINING_MODEL.get_process_position(hparams.gpus)
-    sleep(process_position + 1)
-
-    # init experiment
-    log_dir = os.path.dirname(os.path.realpath(__file__))
-    exp = Experiment(
-        name='test_tube_exp',
-        debug=True,
-        save_dir=log_dir,
-        version=0,
-        autosave=False,
-        description='test demo'
-    )
-
-    exp.argparse(hparams)
-    exp.save()
-
-    # build model
-    print('loading model...')
-    model = TRAINING_MODEL(hparams)
-    print('model built')
-
-    # callbacks
-    early_stop = EarlyStopping(
-        monitor=hparams.early_stop_metric,
-        patience=hparams.early_stop_patience,
-        verbose=True,
-        mode=hparams.early_stop_mode
-    )
-
-    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
-    checkpoint = ModelCheckpoint(
-        filepath=model_save_path,
-        save_function=None,
-        save_best_only=True,
-        verbose=True,
-        monitor=hparams.model_save_monitor_value,
-        mode=hparams.model_save_monitor_mode
-    )
-
-    # gpus are ; separated for inside a node and , within nodes
-    gpu_list = None
-    if hparams.gpus is not None:
-        gpu_list = [int(x) for x in hparams.gpus.split(';')]
-
-    # configure trainer
-    trainer = Trainer(
-        experiment=exp,
-        cluster=cluster,
-        checkpoint_callback=checkpoint,
-        early_stop_callback=early_stop,
-        gpus=gpu_list
-    )
-
-    # train model
-    trainer.fit(model)
-
-
-def get_default_parser(strategy, root_dir):
-
-    possible_model_names = list(AVAILABLE_MODELS.keys())
-    parser = HyperOptArgumentParser(strategy=strategy, add_help=False)
-    add_default_args(parser, root_dir, possible_model_names=possible_model_names, rand_seed=SEED)
-    return parser
-
-
-def get_model_name(args):
-    for i, arg in enumerate(args):
-        if 'model_name' in arg:
-            return args[i+1]
-
-
-def optimize_on_cluster(hyperparams):
-    # enable cluster training
-    cluster = SlurmCluster(
-        hyperparam_optimizer=hyperparams,
-        log_path=hyperparams.tt_save_path,
-        test_tube_exp_name=hyperparams.tt_name
-    )
-
-    # email for cluster coms
-    cluster.notify_job_status(email='add_email_here', on_done=True, on_fail=True)
-
-    # configure cluster
-    cluster.per_experiment_nb_gpus = hyperparams.per_experiment_nb_gpus
-    cluster.job_time = '48:00:00'
-    cluster.gpu_type = '1080ti'
-    cluster.memory_mb_per_node = 48000
-
-    # any modules for code to run in env
-    cluster.add_command('source activate pytorch_lightning')
-
-    # name of exp
-    job_display_name = hyperparams.tt_name.split('_')[0]
-    job_display_name = job_display_name[0:3]
-
-    # run hopt
-    print('submitting jobs...')
-    cluster.optimize_parallel_cluster_gpu(
-        main,
-        nb_trials=hyperparams.nb_hopt_trials,
-        job_name=job_display_name
-    )
-
-
-if __name__ == '__main__':
-
-    model_name = get_model_name(sys.argv)
-    if model_name is None:
-        model_name = 'model_template'
-
-    # use default args
-    root_dir = os.path.dirname(os.path.realpath(__file__))
-    parent_parser = get_default_parser(strategy='random_search', root_dir=root_dir)
-
-    # allow model to overwrite or extend args
-    TRAINING_MODEL = AVAILABLE_MODELS[model_name]
-    parser = TRAINING_MODEL.add_model_specific_args(parent_parser, root_dir)
-    hyperparams = parser.parse_args()
-
-    # format GPU layout
-    os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
-
-    # ---------------------
-    # RUN TRAINING
-    # ---------------------
-
-    # cluster and CPU
-    if hyperparams.on_cluster:
-        # run on HPC cluster
-        print('RUNNING ON SLURM CLUSTER')
-        gpu_ids = hyperparams.gpus.split(';')
-        os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(gpu_ids)
-        optimize_on_cluster(hyperparams)
-
-    elif hyperparams.gpus is None:
-        # run on cpu
-        print('RUNNING ON CPU')
-        main(hyperparams, None, None)
-
-    # single or multiple GPUs on same machine
-    gpu_ids = hyperparams.gpus.split(';')
-    if hyperparams.interactive:
-        # run on 1 gpu
-        print(f'RUNNING INTERACTIVE MODE ON GPUS. gpu ids: {gpu_ids}')
-        os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(gpu_ids)
-        main(hyperparams, None, None)
-
-    else:
-        # multiple GPUs on same machine
-        print(f'RUNNING MULTI GPU. GPU ids: {gpu_ids}')
-        hyperparams.optimize_parallel_gpu(
-            main_local,
-            gpu_ids=gpu_ids,
-            nb_trials=hyperparams.nb_hopt_trials,
-            nb_workers=len(gpu_ids)
-        )

examples/new_project_templates/trainer_gpu_cluster_template.py

@@ -1,210 +0,0 @@
-import os
-import sys
-import numpy as np
-from time import sleep
-import torch
-
-from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
-from pytorch_lightning.models.trainer import Trainer
-from pytorch_lightning.utils.arg_parse import add_default_args
-
-from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
-
-SEED = 2334
-torch.manual_seed(SEED)
-np.random.seed(SEED)
-
-# ---------------------
-# DEFINE MODEL HERE
-# ---------------------
-from examples.new_project_templates.lightning_module_template import LightningTemplateModel
-# ---------------------
-
-AVAILABLE_MODELS = {
-    'model_template': LightningTemplateModel
-}
-
-
-"""
-Allows training by using command line arguments
-Run by:
-# TYPE YOUR RUN COMMAND HERE
-"""
-
-
-def main_local(hparams):
-    main(hparams, None, None)
-
-
-def main(hparams, cluster, results_dict):
-    """
-    Main training routine specific for this project
-    :param hparams:
-    :return:
-    """
-    on_gpu = hparams.gpus is not None and torch.cuda.is_available()
-
-    device = 'cuda' if on_gpu else 'cpu'
-    hparams.__setattr__('device', device)
-    hparams.__setattr__('on_gpu', on_gpu)
-    hparams.__setattr__('nb_gpus', torch.cuda.device_count())
-    hparams.__setattr__('inference_mode', hparams.model_load_weights_path is not None)
-
-    # delay each training start to not overwrite logs
-    process_position, current_gpu = TRAINING_MODEL.get_process_position(hparams.gpus)
-    sleep(process_position + 1)
-
-    # init experiment
-    log_dir = os.path.dirname(os.path.realpath(__file__))
-    exp = Experiment(
-        name='test_tube_exp',
-        debug=True,
-        save_dir=log_dir,
-        version=0,
-        autosave=False,
-        description='test demo'
-    )
-
-    exp.argparse(hparams)
-    exp.save()
-
-    # build model
-    print('loading model...')
-    model = TRAINING_MODEL(hparams)
-    print('model built')
-
-    # callbacks
-    early_stop = EarlyStopping(
-        monitor=hparams.early_stop_metric,
-        patience=hparams.early_stop_patience,
-        verbose=True,
-        mode=hparams.early_stop_mode
-    )
-
-    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
-    checkpoint = ModelCheckpoint(
-        filepath=model_save_path,
-        save_function=None,
-        save_best_only=True,
-        verbose=True,
-        monitor=hparams.model_save_monitor_value,
-        mode=hparams.model_save_monitor_mode
-    )
-
-    # gpus are ; separated for inside a node and , within nodes
-    gpu_list = None
-    if hparams.gpus is not None:
-        gpu_list = [int(x) for x in hparams.gpus.split(';')]
-
-    # configure trainer
-    trainer = Trainer(
-        experiment=exp,
-        cluster=cluster,
-        checkpoint_callback=checkpoint,
-        early_stop_callback=early_stop,
-        gpus=gpu_list
-    )
-
-    # train model
-    trainer.fit(model)
-
-
-def get_default_parser(strategy, root_dir):
-
-    possible_model_names = list(AVAILABLE_MODELS.keys())
-    parser = HyperOptArgumentParser(strategy=strategy, add_help=False)
-    add_default_args(parser, root_dir, possible_model_names=possible_model_names, rand_seed=SEED)
-    return parser
-
-
-def get_model_name(args):
-    for i, arg in enumerate(args):
-        if 'model_name' in arg:
-            return args[i+1]
-
-
-def optimize_on_cluster(hyperparams):
-    # enable cluster training
-    cluster = SlurmCluster(
-        hyperparam_optimizer=hyperparams,
-        log_path=hyperparams.tt_save_path,
-        test_tube_exp_name=hyperparams.tt_name
-    )
-
-    # email for cluster coms
-    cluster.notify_job_status(email='add_email_here', on_done=True, on_fail=True)
-
-    # configure cluster
-    cluster.per_experiment_nb_gpus = hyperparams.per_experiment_nb_gpus
-    cluster.job_time = '48:00:00'
-    cluster.gpu_type = '1080ti'
-    cluster.memory_mb_per_node = 48000
-
-    # any modules for code to run in env
-    cluster.add_command('source activate pytorch_lightning')
-
-    # name of exp
-    job_display_name = hyperparams.tt_name.split('_')[0]
-    job_display_name = job_display_name[0:3]
-
-    # run hopt
-    print('submitting jobs...')
-    cluster.optimize_parallel_cluster_gpu(
-        main,
-        nb_trials=hyperparams.nb_hopt_trials,
-        job_name=job_display_name
-    )
-
-
-if __name__ == '__main__':
-
-    model_name = get_model_name(sys.argv)
-    if model_name is None:
-        model_name = 'model_template'
-
-    # use default args
-    root_dir = os.path.dirname(os.path.realpath(__file__))
-    parent_parser = get_default_parser(strategy='random_search', root_dir=root_dir)
-
-    # allow model to overwrite or extend args
-    TRAINING_MODEL = AVAILABLE_MODELS[model_name]
-    parser = TRAINING_MODEL.add_model_specific_args(parent_parser, root_dir)
-    hyperparams = parser.parse_args()
-
-    # format GPU layout
-    os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
-
-    # ---------------------
-    # RUN TRAINING
-    # ---------------------
-
-    # cluster and CPU
-    if hyperparams.on_cluster:
-        # run on HPC cluster
-        print('RUNNING ON SLURM CLUSTER')
-        gpu_ids = hyperparams.gpus.split(';')
-        os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(gpu_ids)
-        optimize_on_cluster(hyperparams)
-
-    elif hyperparams.gpus is None:
-        # run on cpu
-        print('RUNNING ON CPU')
-        main(hyperparams, None, None)
-
-    # single or multiple GPUs on same machine
-    gpu_ids = hyperparams.gpus.split(';')
-    if hyperparams.interactive:
-        # run on 1 gpu
-        print(f'RUNNING INTERACTIVE MODE ON GPUS. gpu ids: {gpu_ids}')
-        os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(gpu_ids)
-        main(hyperparams, None, None)
-
-    else:
-        # multiple GPUs on same machine
-        print(f'RUNNING MULTI GPU. GPU ids: {gpu_ids}')
-        hyperparams.optimize_parallel_gpu(
-            main_local,
-            gpu_ids=gpu_ids,
-            nb_trials=hyperparams.nb_hopt_trials,
-            nb_workers=len(gpu_ids)
-        )

mkdocs.yml

@@ -1,9 +1,10 @@
 site_name: Pytorch lightning Documentation
-theme: readthedocs
+theme:
+  name: 'material'
 docs_dir: docs
 repo_url: https://github.com/williamFalcon/pytorch-lightning
 site_dir: 'site'
-site_description: 'Documentation for Pytorch Pytorch-Lightning, the researcher version of keras.'
+site_description: 'Documentation for Pytorch LightningModule, the researcher version of keras.'
 
 dev_addr: '0.0.0.0:8000'
 #google_analytics: ['UA-aasd', 'sitename']

pytorch_lightning/__init__.py

@@ -1 +1,3 @@
-from .models import Trainer
+from .models import Trainer
+from .root_module.root_module import LightningModule
+from .root_module.decorators import data_loader

pytorch_lightning/callbacks/pt_callbacks.py

@@ -1,6 +1,6 @@
 import numpy as np
 import os, shutil
-from pytorch_lightning.pt_overrides.override_data_parallel import LightningDataParallel
+from pytorch_lightning.pt_overrides.override_data_parallel import LightningDistributedDataParallel
 
 
 class Callback(object):
@@ -34,7 +34,7 @@ class Callback(object):
         self.params = params
 
     def set_model(self, model):
-        if type(model) is LightningDataParallel:
+        if type(model) is LightningDistributedDataParallel:
             model = model.module
         self.model = model
 
@@ -170,12 +170,11 @@ class ModelCheckpoint(Callback):
         period: Interval (number of epochs) between checkpoints.
     """
 
-    def __init__(self, filepath, save_function, monitor='val_loss', verbose=0,
+    def __init__(self, filepath, monitor='val_loss', verbose=0,
                  save_best_only=False, save_weights_only=False,
                  mode='auto', period=1, prefix=''):
         super(ModelCheckpoint, self).__init__()
         self.monitor = monitor
-        self.save_function = save_function
         self.verbose = verbose
         self.filepath = filepath
         self.save_best_only = save_best_only

pytorch_lightning/examples/__init__.py

@@ -0,0 +1 @@
+from .new_project_templates.lightning_module_template import LightningTemplateModel

pytorch_lightning/examples/new_project_templates/lightning_module_template.py

@@ -7,7 +7,10 @@ import torch
 import torch.nn.functional as F
 from test_tube import HyperOptArgumentParser
 from torch import optim
+from torch.utils.data import DataLoader
+from torch.utils.data.distributed import DistributedSampler
 
+import pytorch_lightning as ptl
 from pytorch_lightning.root_module.root_module import LightningModule
 
 
@@ -22,10 +25,14 @@ class LightningTemplateModel(LightningModule):
         :param hparams:
         """
         # init superclass
-        super(LightningTemplateModel, self).__init__(hparams)
+        super(LightningTemplateModel, self).__init__()
+        self.hparams = hparams
 
         self.batch_size = hparams.batch_size
 
+        # if you specify an example input, the summary will show input/output for each layer
+        self.example_input_array = torch.rand(5, 28 * 28)
+
         # build model
         self.__build_model()
 
@@ -76,15 +83,21 @@ class LightningTemplateModel(LightningModule):
         # forward pass
         x, y = data_batch
         x = x.view(x.size(0), -1)
+
         y_hat = self.forward(x)
 
         # calculate loss
         loss_val = self.loss(y, y_hat)
 
+        # in DP mode (default) make sure if result is scalar, there's another dim in the beginning
+        if self.trainer.use_dp:
+            loss_val = loss_val.unsqueeze(0)
+
         output = OrderedDict({
-            'loss': loss_val,
-            'tqdm_metrics': {}
+            'loss': loss_val
         })
+
+        # can also return just a scalar instead of a dict (return loss_val)
         return output
 
     def validation_step(self, data_batch, batch_i):
@@ -102,11 +115,22 @@ class LightningTemplateModel(LightningModule):
         # acc
         labels_hat = torch.argmax(y_hat, dim=1)
         val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)
+        val_acc = torch.tensor(val_acc)
+
+        if self.on_gpu:
+            val_acc = val_acc.cuda(loss_val.device.index)
+
+        # in DP mode (default) make sure if result is scalar, there's another dim in the beginning
+        if self.trainer.use_dp:
+            loss_val = loss_val.unsqueeze(0)
+            val_acc = val_acc.unsqueeze(0)
 
         output = OrderedDict({
             'val_loss': loss_val,
-            'val_acc': torch.tensor(val_acc),
+            'val_acc': val_acc,
         })
+
+        # can also return just a scalar instead of a dict (return loss_val)
         return output
 
     def validation_end(self, outputs):
@@ -115,6 +139,10 @@ class LightningTemplateModel(LightningModule):
         :param outputs: list of individual outputs of each validation step
         :return:
         """
+        # if returned a scalar from validation_step, outputs is a list of tensor scalars
+        # we return just the average in this case (if we want)
+        # return torch.stack(outputs).mean()
+
         val_loss_mean = 0
         val_acc_mean = 0
         for output in outputs:
@@ -126,20 +154,6 @@ class LightningTemplateModel(LightningModule):
         tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()}
         return tqdm_dic
 
-    def update_tng_log_metrics(self, logs):
-        return logs
-
-    # ---------------------
-    # MODEL SAVING
-    # ---------------------
-    def get_save_dict(self):
-        checkpoint = {'state_dict': self.state_dict()}
-        return checkpoint
-
-    def load_model_specific(self, checkpoint):
-        self.load_state_dict(checkpoint['state_dict'])
-        pass
-
     # ---------------------
     # TRAINING SETUP
     # ---------------------
@@ -149,54 +163,52 @@ class LightningTemplateModel(LightningModule):
         :return: list of optimizers
         """
         optimizer = optim.Adam(self.parameters(), lr=self.hparams.learning_rate)
-        return [optimizer]
+        scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=10)
+        return [optimizer], [scheduler]
 
     def __dataloader(self, train):
         # init data generators
         transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
-
         dataset = MNIST(root=self.hparams.data_root, train=train, transform=transform, download=True)
 
-        loader = torch.utils.data.DataLoader(
+        # when using multi-node we need to add the datasampler
+        train_sampler = None
+        batch_size = self.hparams.batch_size
+
+        try:
+            if self.on_gpu:
+                train_sampler = DistributedSampler(dataset, rank=self.trainer.proc_rank)
+                batch_size = batch_size // self.trainer.world_size  # scale batch size
+        except Exception as e:
+            pass
+
+        should_shuffle = train_sampler is None
+        loader = DataLoader(
             dataset=dataset,
-            batch_size=self.hparams.batch_size,
-            shuffle=True
+            batch_size=batch_size,
+            shuffle=should_shuffle,
+            sampler=train_sampler
         )
 
         return loader
 
-    @property
+    @ptl.data_loader
     def tng_dataloader(self):
-        if self._tng_dataloader is None:
-            try:
-                self._tng_dataloader = self.__dataloader(train=True)
-            except Exception as e:
-                print(e)
-                raise e
-        return self._tng_dataloader
+        print('tng data loader called')
+        return self.__dataloader(train=True)
 
-    @property
+    @ptl.data_loader
     def val_dataloader(self):
-        if self._val_dataloader is None:
-            try:
-                self._val_dataloader = self.__dataloader(train=False)
-            except Exception as e:
-                print(e)
-                raise e
-        return self._val_dataloader
+        print('val data loader called')
+        return self.__dataloader(train=False)
 
-    @property
+    @ptl.data_loader
     def test_dataloader(self):
-        if self._test_dataloader is None:
-            try:
-                self._test_dataloader = self.__dataloader(train=False)
-            except Exception as e:
-                print(e)
-                raise e
-        return self._test_dataloader
+        print('test data loader called')
+        return self.__dataloader(train=False)
 
     @staticmethod
-    def add_model_specific_args(parent_parser, root_dir):
+    def add_model_specific_args(parent_parser, root_dir): # pragma: no cover
         """
         Parameters you define here will be available to your model through self.hparams
         :param parent_parser:
@@ -209,17 +221,19 @@ class LightningTemplateModel(LightningModule):
         # parser.set_defaults(gradient_clip=5.0)
 
         # network params
-        parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=False)
-        parser.add_argument('--in_features', default=28*28)
-        parser.add_argument('--out_features', default=10)
-        parser.add_argument('--hidden_dim', default=50000) # use 500 for CPU, 50000 for GPU to see speed difference
+        parser.add_argument('--in_features', default=28*28, type=int)
+        parser.add_argument('--out_features', default=10, type=int)
+        parser.add_argument('--hidden_dim', default=50000, type=int) # use 500 for CPU, 50000 for GPU to see speed difference
 
         # data
         parser.add_argument('--data_root', default=os.path.join(root_dir, 'mnist'), type=str)
 
         # training params (opt)
-        parser.opt_list('--learning_rate', default=0.001, type=float, options=[0.0001, 0.0005, 0.001, 0.005],
+        parser.opt_list('--learning_rate', default=0.001*8, type=float, options=[0.0001, 0.0005, 0.001, 0.005],
                         tunable=False)
-        parser.opt_list('--batch_size', default=256, type=int, options=[32, 64, 128, 256], tunable=False)
         parser.opt_list('--optimizer_name', default='adam', type=str, options=['adam'], tunable=False)
+
+        # if using 2 nodes with 4 gpus each the batch size here (256) will be 256 / (2*8) = 16 per gpu
+        parser.opt_list('--batch_size', default=256*8, type=int, options=[32, 64, 128, 256], tunable=False,
+                        help='batch size will be divided over all the gpus being used across all nodes')
         return parser

pytorch_lightning/examples/new_project_templates/multi_node_cluster_template.py

@@ -0,0 +1,172 @@
+import os
+import sys
+import numpy as np
+from time import sleep
+import torch
+
+from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
+from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning.utils.arg_parse import add_default_args
+
+from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
+
+SEED = 2334
+torch.manual_seed(SEED)
+np.random.seed(SEED)
+
+# ---------------------
+# DEFINE MODEL HERE
+# ---------------------
+from lightning_module_template import LightningTemplateModel
+# ---------------------
+
+"""
+Allows training by using command line arguments
+Run by:
+# TYPE YOUR RUN COMMAND HERE
+"""
+
+
+def main_local(hparams):
+    main(hparams, None, None)
+
+
+def main(hparams, cluster, results_dict):
+    """
+    Main training routine specific for this project
+    :param hparams:
+    :return:
+    """
+    # ------------------------
+    # 1 INIT LIGHTNING MODEL
+    # ------------------------
+    print('loading model...')
+    model = LightningTemplateModel(hparams)
+    print('model built')
+
+    # ------------------------
+    # 2 INIT TEST TUBE EXP
+    # ------------------------
+    # when using grid search, it's possible for all models to start at once
+    # and use the same test tube experiment version
+    relative_node_id = int(os.environ['SLURM_NODEID'])
+    sleep(relative_node_id + 1)
+
+    # init experiment
+    exp = Experiment(
+        name=hyperparams.experiment_name,
+        save_dir=hyperparams.test_tube_save_path,
+        autosave=False,
+        description='test demo'
+    )
+
+    exp.argparse(hparams)
+    exp.save()
+
+    # ------------------------
+    # 3 DEFINE CALLBACKS
+    # ------------------------
+    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
+    early_stop = EarlyStopping(
+        monitor='val_acc',
+        patience=3,
+        verbose=True,
+        mode='max'
+    )
+
+    checkpoint = ModelCheckpoint(
+        filepath=model_save_path,
+        save_best_only=True,
+        verbose=True,
+        monitor='val_loss',
+        mode='min'
+    )
+
+    # ------------------------
+    # 4 INIT TRAINER
+    # ------------------------
+    trainer = Trainer(
+        experiment=exp,
+        cluster=cluster,
+        checkpoint_callback=checkpoint,
+        early_stop_callback=early_stop,
+        gpus=hparams.gpus,
+        nb_gpu_nodes=hyperparams.nb_gpu_nodes
+    )
+
+    # ------------------------
+    # 5 START TRAINING
+    # ------------------------
+    trainer.fit(model)
+
+
+def optimize_on_cluster(hyperparams):
+    # enable cluster training
+    # log all scripts to the test tube folder
+    cluster = SlurmCluster(
+        hyperparam_optimizer=hyperparams,
+        log_path=hyperparams.slurm_log_path,
+    )
+
+    # email for cluster coms
+    cluster.notify_job_status(email='add_email_here', on_done=True, on_fail=True)
+
+    # configure cluster
+    cluster.per_experiment_nb_gpus = hyperparams.per_experiment_nb_gpus
+    cluster.per_experiment_nb_nodes = hyperparams.nb_gpu_nodes
+    cluster.job_time = '2:00:00'
+    cluster.gpu_type = 'volta'
+    cluster.memory_mb_per_node = 0
+
+    # any modules for code to run in env
+    cluster.add_command('source activate lightning')
+
+    # run only on 32GB voltas
+    cluster.add_slurm_cmd(cmd='constraint', value='volta32gb', comment='use 32gb gpus')
+    cluster.add_slurm_cmd(cmd='partition', value=hyperparams.gpu_partition, comment='use 32gb gpus')
+
+    # run hopt
+    # creates and submits jobs to slurm
+    cluster.optimize_parallel_cluster_gpu(
+        main,
+        nb_trials=hyperparams.nb_hopt_trials,
+        job_name=hyperparams.experiment_name
+    )
+
+
+if __name__ == '__main__':
+
+    # use default args
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    demo_log_dir = os.path.join(root_dir, 'pt_lightning_demo_logs')
+
+    checkpoint_dir = os.path.join(demo_log_dir, 'model_weights')
+    test_tube_dir = os.path.join(demo_log_dir, 'test_tube_data')
+    slurm_out_dir = os.path.join(demo_log_dir, 'slurm_scripts')
+
+    parent_parser = HyperOptArgumentParser(strategy='grid_search', add_help=False)
+
+    # cluster args not defined inside the model
+    parent_parser.add_argument('--gpu_partition', type=str, help='consult your cluster manual')
+
+    # TODO: make 1 param
+    parent_parser.add_argument('--per_experiment_nb_gpus', type=int, help='how many gpus to use in a node')
+    parent_parser.add_argument('--gpus', type=str, default='-1', help='how many gpus to use in the node')
+
+    parent_parser.add_argument('--nb_gpu_nodes', type=int, default=1, help='how many nodes to use in a cluster')
+    parent_parser.add_argument('--test_tube_save_path', type=str, default=test_tube_dir, help='where to save logs')
+    parent_parser.add_argument('--slurm_log_path', type=str, default=slurm_out_dir, help='where to save slurm meta')
+    parent_parser.add_argument('--model_save_path', type=str, default=checkpoint_dir, help='where to save model')
+    parent_parser.add_argument('--experiment_name', type=str, default='pt_lightning_exp_a', help='test tube exp name')
+    parent_parser.add_argument('--nb_hopt_trials', type=int, default=1, help='how many grid search trials to run')
+
+    # allow model to overwrite or extend args
+    parser = LightningTemplateModel.add_model_specific_args(parent_parser, root_dir)
+    hyperparams = parser.parse_args()
+
+    # ---------------------
+    # RUN TRAINING
+    # ---------------------
+    # run on HPC cluster
+    print('RUNNING ON SLURM CLUSTER')
+    optimize_on_cluster(hyperparams)

pytorch_lightning/examples/new_project_templates/single_cpu_template.py

@@ -0,0 +1,110 @@
+"""
+Runs a model on a single node across N-gpus.
+"""
+import os
+import sys
+import numpy as np
+from time import sleep
+import torch
+
+from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
+from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning.utils.arg_parse import add_default_args
+
+from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
+
+SEED = 2334
+torch.manual_seed(SEED)
+np.random.seed(SEED)
+
+from lightning_module_template import LightningTemplateModel
+
+
+def main(hparams):
+    """
+    Main training routine specific for this project
+    :param hparams:
+    :return:
+    """
+    # ------------------------
+    # 1 INIT LIGHTNING MODEL
+    # ------------------------
+    print('loading model...')
+    model = LightningTemplateModel(hparams)
+    print('model built')
+
+    # ------------------------
+    # 2 INIT TEST TUBE EXP
+    # ------------------------
+
+    # init experiment
+    exp = Experiment(
+        name=hyperparams.experiment_name,
+        save_dir=hyperparams.test_tube_save_path,
+        autosave=False,
+        description='test demo'
+    )
+
+    exp.argparse(hparams)
+    exp.save()
+
+    # ------------------------
+    # 3 DEFINE CALLBACKS
+    # ------------------------
+    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
+    early_stop = EarlyStopping(
+        monitor='val_acc',
+        patience=3,
+        verbose=True,
+        mode='max'
+    )
+
+    checkpoint = ModelCheckpoint(
+        filepath=model_save_path,
+        save_best_only=True,
+        verbose=True,
+        monitor='val_loss',
+        mode='min'
+    )
+
+    # ------------------------
+    # 4 INIT TRAINER
+    # ------------------------
+    trainer = Trainer(
+        experiment=exp,
+        checkpoint_callback=checkpoint,
+        early_stop_callback=early_stop,
+    )
+
+    # ------------------------
+    # 5 START TRAINING
+    # ------------------------
+    trainer.fit(model)
+
+
+if __name__ == '__main__':
+
+    # dirs
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    demo_log_dir = os.path.join(root_dir, 'pt_lightning_demo_logs')
+    checkpoint_dir = os.path.join(demo_log_dir, 'model_weights')
+    test_tube_dir = os.path.join(demo_log_dir, 'test_tube_data')
+
+    # although we user hyperOptParser, we are using it only as argparse right now
+    parent_parser = HyperOptArgumentParser(strategy='grid_search', add_help=False)
+
+    # gpu args
+    parent_parser.add_argument('--test_tube_save_path', type=str, default=test_tube_dir, help='where to save logs')
+    parent_parser.add_argument('--model_save_path', type=str, default=checkpoint_dir, help='where to save model')
+    parent_parser.add_argument('--experiment_name', type=str, default='pt_lightning_exp_a', help='test tube exp name')
+
+    # allow model to overwrite or extend args
+    parser = LightningTemplateModel.add_model_specific_args(parent_parser, root_dir)
+    hyperparams = parser.parse_args()
+
+    # ---------------------
+    # RUN TRAINING
+    # ---------------------
+    # run on HPC cluster
+    print(f'RUNNING ON CPU')
+    main(hyperparams)

pytorch_lightning/examples/new_project_templates/single_gpu_node_16bit_template.py

@@ -0,0 +1,113 @@
+"""
+Runs a model on a single node across N-gpus.
+"""
+import os
+import sys
+import numpy as np
+from time import sleep
+import torch
+
+from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
+from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning.utils.arg_parse import add_default_args
+
+from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
+
+SEED = 2334
+torch.manual_seed(SEED)
+np.random.seed(SEED)
+
+from lightning_module_template import LightningTemplateModel
+
+
+def main(hparams):
+    """
+    Main training routine specific for this project
+    :param hparams:
+    :return:
+    """
+    # ------------------------
+    # 1 INIT LIGHTNING MODEL
+    # ------------------------
+    print('loading model...')
+    model = LightningTemplateModel(hparams)
+    print('model built')
+
+    # ------------------------
+    # 2 INIT TEST TUBE EXP
+    # ------------------------
+
+    # init experiment
+    exp = Experiment(
+        name=hyperparams.experiment_name,
+        save_dir=hyperparams.test_tube_save_path,
+        autosave=False,
+        description='test demo'
+    )
+
+    exp.argparse(hparams)
+    exp.save()
+
+    # ------------------------
+    # 3 DEFINE CALLBACKS
+    # ------------------------
+    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
+    early_stop = EarlyStopping(
+        monitor='val_acc',
+        patience=3,
+        verbose=True,
+        mode='max'
+    )
+
+    checkpoint = ModelCheckpoint(
+        filepath=model_save_path,
+        save_best_only=True,
+        verbose=True,
+        monitor='val_loss',
+        mode='min'
+    )
+
+    # ------------------------
+    # 4 INIT TRAINER
+    # ------------------------
+    trainer = Trainer(
+        experiment=exp,
+        checkpoint_callback=checkpoint,
+        early_stop_callback=early_stop,
+        gpus=hparams.gpus,
+        use_amp=True
+    )
+
+    # ------------------------
+    # 5 START TRAINING
+    # ------------------------
+    trainer.fit(model)
+
+
+if __name__ == '__main__':
+
+    # dirs
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    demo_log_dir = os.path.join(root_dir, 'pt_lightning_demo_logs')
+    checkpoint_dir = os.path.join(demo_log_dir, 'model_weights')
+    test_tube_dir = os.path.join(demo_log_dir, 'test_tube_data')
+
+    # although we user hyperOptParser, we are using it only as argparse right now
+    parent_parser = HyperOptArgumentParser(strategy='grid_search', add_help=False)
+
+    # gpu args
+    parent_parser.add_argument('--gpus', type=str, default='-1', help='how many gpus to use in the node. -1 uses all the gpus on the node')
+    parent_parser.add_argument('--test_tube_save_path', type=str, default=test_tube_dir, help='where to save logs')
+    parent_parser.add_argument('--model_save_path', type=str, default=checkpoint_dir, help='where to save model')
+    parent_parser.add_argument('--experiment_name', type=str, default='pt_lightning_exp_a', help='test tube exp name')
+
+    # allow model to overwrite or extend args
+    parser = LightningTemplateModel.add_model_specific_args(parent_parser, root_dir)
+    hyperparams = parser.parse_args()
+
+    # ---------------------
+    # RUN TRAINING
+    # ---------------------
+    # run on HPC cluster
+    print(f'RUNNING INTERACTIVE MODE ON GPUS. gpu ids: {hyperparams.gpus}')
+    main(hyperparams)

pytorch_lightning/examples/new_project_templates/single_gpu_node_dp_template.py

@@ -0,0 +1,112 @@
+"""
+Runs a model on a single node across N-gpus.
+"""
+import os
+import sys
+import numpy as np
+from time import sleep
+import torch
+
+from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
+from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning.utils.arg_parse import add_default_args
+
+from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
+
+SEED = 2334
+torch.manual_seed(SEED)
+np.random.seed(SEED)
+
+from lightning_module_template import LightningTemplateModel
+
+
+def main(hparams):
+    """
+    Main training routine specific for this project
+    :param hparams:
+    :return:
+    """
+    # ------------------------
+    # 1 INIT LIGHTNING MODEL
+    # ------------------------
+    print('loading model...')
+    model = LightningTemplateModel(hparams)
+    print('model built')
+
+    # ------------------------
+    # 2 INIT TEST TUBE EXP
+    # ------------------------
+
+    # init experiment
+    exp = Experiment(
+        name=hyperparams.experiment_name,
+        save_dir=hyperparams.test_tube_save_path,
+        autosave=False,
+        description='test demo'
+    )
+
+    exp.argparse(hparams)
+    exp.save()
+
+    # ------------------------
+    # 3 DEFINE CALLBACKS
+    # ------------------------
+    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
+    early_stop = EarlyStopping(
+        monitor='val_acc',
+        patience=3,
+        verbose=True,
+        mode='max'
+    )
+
+    checkpoint = ModelCheckpoint(
+        filepath=model_save_path,
+        save_best_only=True,
+        verbose=True,
+        monitor='val_loss',
+        mode='min'
+    )
+
+    # ------------------------
+    # 4 INIT TRAINER
+    # ------------------------
+    trainer = Trainer(
+        experiment=exp,
+        checkpoint_callback=checkpoint,
+        early_stop_callback=early_stop,
+        gpus=hparams.gpus,
+    )
+
+    # ------------------------
+    # 5 START TRAINING
+    # ------------------------
+    trainer.fit(model)
+
+
+if __name__ == '__main__':
+
+    # dirs
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    demo_log_dir = os.path.join(root_dir, 'pt_lightning_demo_logs')
+    checkpoint_dir = os.path.join(demo_log_dir, 'model_weights')
+    test_tube_dir = os.path.join(demo_log_dir, 'test_tube_data')
+
+    # although we user hyperOptParser, we are using it only as argparse right now
+    parent_parser = HyperOptArgumentParser(strategy='grid_search', add_help=False)
+
+    # gpu args
+    parent_parser.add_argument('--gpus', type=str, default='-1', help='how many gpus to use in the node. -1 uses all the gpus on the node')
+    parent_parser.add_argument('--test_tube_save_path', type=str, default=test_tube_dir, help='where to save logs')
+    parent_parser.add_argument('--model_save_path', type=str, default=checkpoint_dir, help='where to save model')
+    parent_parser.add_argument('--experiment_name', type=str, default='pt_lightning_exp_a', help='test tube exp name')
+
+    # allow model to overwrite or extend args
+    parser = LightningTemplateModel.add_model_specific_args(parent_parser, root_dir)
+    hyperparams = parser.parse_args()
+
+    # ---------------------
+    # RUN TRAINING
+    # ---------------------
+    # run on HPC cluster
+    print(f'RUNNING INTERACTIVE MODE ON GPUS. gpu ids: {hyperparams.gpus}')
+    main(hyperparams)

pytorch_lightning/examples/new_project_templates/single_gpu_node_template.py

@@ -0,0 +1,112 @@
+"""
+Runs a model on a single node across N-gpus.
+"""
+import os
+import sys
+import numpy as np
+from time import sleep
+import torch
+
+from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
+from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning.utils.arg_parse import add_default_args
+
+from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
+
+SEED = 2334
+torch.manual_seed(SEED)
+np.random.seed(SEED)
+
+from lightning_module_template import LightningTemplateModel
+
+
+def main(hparams):
+    """
+    Main training routine specific for this project
+    :param hparams:
+    :return:
+    """
+    # ------------------------
+    # 1 INIT LIGHTNING MODEL
+    # ------------------------
+    print('loading model...')
+    model = LightningTemplateModel(hparams)
+    print('model built')
+
+    # ------------------------
+    # 2 INIT TEST TUBE EXP
+    # ------------------------
+
+    # init experiment
+    exp = Experiment(
+        name=hyperparams.experiment_name,
+        save_dir=hyperparams.test_tube_save_path,
+        autosave=False,
+        description='test demo'
+    )
+
+    exp.argparse(hparams)
+    exp.save()
+
+    # ------------------------
+    # 3 DEFINE CALLBACKS
+    # ------------------------
+    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
+    early_stop = EarlyStopping(
+        monitor='val_acc',
+        patience=3,
+        verbose=True,
+        mode='max'
+    )
+
+    checkpoint = ModelCheckpoint(
+        filepath=model_save_path,
+        save_best_only=True,
+        verbose=True,
+        monitor='val_loss',
+        mode='min'
+    )
+
+    # ------------------------
+    # 4 INIT TRAINER
+    # ------------------------
+    trainer = Trainer(
+        experiment=exp,
+        checkpoint_callback=checkpoint,
+        early_stop_callback=early_stop,
+        gpus=hparams.gpus,
+    )
+
+    # ------------------------
+    # 5 START TRAINING
+    # ------------------------
+    trainer.fit(model)
+
+
+if __name__ == '__main__':
+
+    # dirs
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    demo_log_dir = os.path.join(root_dir, 'pt_lightning_demo_logs')
+    checkpoint_dir = os.path.join(demo_log_dir, 'model_weights')
+    test_tube_dir = os.path.join(demo_log_dir, 'test_tube_data')
+
+    # although we user hyperOptParser, we are using it only as argparse right now
+    parent_parser = HyperOptArgumentParser(strategy='grid_search', add_help=False)
+
+    # gpu args
+    parent_parser.add_argument('--gpus', type=str, default='-1', help='how many gpus to use in the node. -1 uses all the gpus on the node')
+    parent_parser.add_argument('--test_tube_save_path', type=str, default=test_tube_dir, help='where to save logs')
+    parent_parser.add_argument('--model_save_path', type=str, default=checkpoint_dir, help='where to save model')
+    parent_parser.add_argument('--experiment_name', type=str, default='pt_lightning_exp_a', help='test tube exp name')
+
+    # allow model to overwrite or extend args
+    parser = LightningTemplateModel.add_model_specific_args(parent_parser, root_dir)
+    hyperparams = parser.parse_args()
+
+    # ---------------------
+    # RUN TRAINING
+    # ---------------------
+    # run on HPC cluster
+    print(f'RUNNING INTERACTIVE MODE ON GPUS. gpu ids: {hyperparams.gpus}')
+    main(hyperparams)

pytorch_lightning/examples/new_project_templates/single_gpu_template.py

@@ -0,0 +1,112 @@
+"""
+Runs a model on a single node across N-gpus.
+"""
+import os
+import sys
+import numpy as np
+from time import sleep
+import torch
+
+from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
+from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning.utils.arg_parse import add_default_args
+
+from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
+
+SEED = 2334
+torch.manual_seed(SEED)
+np.random.seed(SEED)
+
+from lightning_module_template import LightningTemplateModel
+
+
+def main(hparams):
+    """
+    Main training routine specific for this project
+    :param hparams:
+    :return:
+    """
+    # ------------------------
+    # 1 INIT LIGHTNING MODEL
+    # ------------------------
+    print('loading model...')
+    model = LightningTemplateModel(hparams)
+    print('model built')
+
+    # ------------------------
+    # 2 INIT TEST TUBE EXP
+    # ------------------------
+
+    # init experiment
+    exp = Experiment(
+        name=hyperparams.experiment_name,
+        save_dir=hyperparams.test_tube_save_path,
+        autosave=False,
+        description='test demo'
+    )
+
+    exp.argparse(hparams)
+    exp.save()
+
+    # ------------------------
+    # 3 DEFINE CALLBACKS
+    # ------------------------
+    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
+    early_stop = EarlyStopping(
+        monitor='val_acc',
+        patience=3,
+        verbose=True,
+        mode='max'
+    )
+
+    checkpoint = ModelCheckpoint(
+        filepath=model_save_path,
+        save_best_only=True,
+        verbose=True,
+        monitor='val_loss',
+        mode='min'
+    )
+
+    # ------------------------
+    # 4 INIT TRAINER
+    # ------------------------
+    trainer = Trainer(
+        experiment=exp,
+        checkpoint_callback=checkpoint,
+        early_stop_callback=early_stop,
+        gpus=hparams.gpus,
+    )
+
+    # ------------------------
+    # 5 START TRAINING
+    # ------------------------
+    trainer.fit(model)
+
+
+if __name__ == '__main__':
+
+    # dirs
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    demo_log_dir = os.path.join(root_dir, 'pt_lightning_demo_logs')
+    checkpoint_dir = os.path.join(demo_log_dir, 'model_weights')
+    test_tube_dir = os.path.join(demo_log_dir, 'test_tube_data')
+
+    # although we user hyperOptParser, we are using it only as argparse right now
+    parent_parser = HyperOptArgumentParser(strategy='grid_search', add_help=False)
+
+    # gpu args
+    parent_parser.add_argument('--gpus', type=str, default='0', help='how many gpus to use in the node. -1 uses all the gpus on the node')
+    parent_parser.add_argument('--test_tube_save_path', type=str, default=test_tube_dir, help='where to save logs')
+    parent_parser.add_argument('--model_save_path', type=str, default=checkpoint_dir, help='where to save model')
+    parent_parser.add_argument('--experiment_name', type=str, default='pt_lightning_exp_a', help='test tube exp name')
+
+    # allow model to overwrite or extend args
+    parser = LightningTemplateModel.add_model_specific_args(parent_parser, root_dir)
+    hyperparams = parser.parse_args()
+
+    # ---------------------
+    # RUN TRAINING
+    # ---------------------
+    # run on HPC cluster
+    print(f'RUNNING INTERACTIVE MODE ON GPUS. gpu ids: {hyperparams.gpus}')
+    main(hyperparams)

pytorch_lightning/examples/new_project_templates/trainer_cpu_template.py

@@ -41,7 +41,6 @@ def main(hparams):
     model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
     checkpoint = ModelCheckpoint(
         filepath=model_save_path,
-        save_function=None,
         save_best_only=True,
         verbose=True,
         monitor='val_acc',

pytorch_lightning/models/sample_model_template/model_template.py

@@ -1,203 +0,0 @@
-import torch.nn as nn
-import numpy as np
-from pytorch_lightning.root_module.root_module import LightningModule
-from test_tube import HyperOptArgumentParser
-from torchvision.datasets import MNIST
-import torchvision.transforms as transforms
-import torch
-import torch.nn.functional as F
-
-
-class ExampleModel1(LightningModule):
-    """
-    Sample model to show how to define a template
-    """
-
-    def __init__(self, hparams):
-        # init superclass
-        super(ExampleModel1, self).__init__(hparams)
-
-        self.batch_size = hparams.batch_size
-
-        # build model
-        self.__build_model()
-
-    # ---------------------
-    # MODEL SETUP
-    # ---------------------
-    def __build_model(self):
-        """
-        Layout model
-        :return:
-        """
-        self.c_d1 = nn.Linear(in_features=self.hparams.in_features, out_features=self.hparams.hidden_dim)
-        self.c_d1_bn = nn.BatchNorm1d(self.hparams.hidden_dim)
-        self.c_d1_drop = nn.Dropout(self.hparams.drop_prob)
-
-        self.c_d2 = nn.Linear(in_features=self.hparams.hidden_dim, out_features=self.hparams.out_features)
-
-    # ---------------------
-    # TRAINING
-    # ---------------------
-    def forward(self, x):
-        x = self.c_d1(x)
-        x = F.tanh(x)
-        x = self.c_d1_bn(x)
-        x = self.c_d1_drop(x)
-
-        x = self.c_d2(x)
-        logits = F.log_softmax(x, dim=1)
-
-        return logits
-
-    def loss(self, labels, logits):
-        nll = F.nll_loss(logits, labels)
-        return nll
-
-    def training_step(self, data_batch):
-        """
-        Called inside the training loop
-        :param data_batch:
-        :return:
-        """
-        # forward pass
-        x, y = data_batch
-        x = x.view(x.size(0), -1)
-        y_hat = self.forward(x)
-
-        # calculate loss
-        loss_val = self.loss(y, y_hat)
-
-        tqdm_dic = {'jefe': 1}
-        return loss_val, tqdm_dic
-
-    def validation_step(self, data_batch):
-        """
-        Called inside the validation loop
-        :param data_batch:
-        :return:
-        """
-        x, y = data_batch
-        x = x.view(x.size(0), -1)
-        y_hat = self.forward(x)
-
-        loss_val = self.loss(y, y_hat)
-
-        # acc
-        labels_hat = torch.argmax(y_hat, dim=1)
-        val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)
-
-        output = {'y_hat': y_hat, 'val_loss': loss_val.item(), 'val_acc': val_acc}
-        return output
-
-    def validation_end(self, outputs):
-        """
-        Called at the end of validation to aggregate outputs
-        :param outputs: list of individual outputs of each validation step
-        :return:
-        """
-        val_loss_mean = 0
-        accs = []
-        for output in outputs:
-            val_loss_mean += output['val_loss']
-            accs.append(output['val_acc'])
-
-        val_loss_mean /= len(outputs)
-        tqdm_dic = {'val_loss': val_loss_mean, 'val_acc': np.mean(accs)}
-        return tqdm_dic
-
-    def update_tng_log_metrics(self, logs):
-        return logs
-
-    # ---------------------
-    # MODEL SAVING
-    # ---------------------
-    def get_save_dict(self):
-        checkpoint = {
-            'state_dict': self.state_dict(),
-        }
-
-        return checkpoint
-
-    def load_model_specific(self, checkpoint):
-        self.load_state_dict(checkpoint['state_dict'])
-        pass
-
-    # ---------------------
-    # TRAINING SETUP
-    # ---------------------
-    def configure_optimizers(self):
-        """
-        return whatever optimizers we want here
-        :return: list of optimizers
-        """
-        optimizer = self.choose_optimizer(self.hparams.optimizer_name, self.parameters(), {'lr': self.hparams.learning_rate}, 'optimizer')
-        self.optimizers = [optimizer]
-        return self.optimizers
-
-    def __dataloader(self, train):
-        # init data generators
-        transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
-
-        dataset = MNIST(root=self.hparams.data_root, train=train, transform=transform, download=True)
-
-        loader = torch.utils.data.DataLoader(
-            dataset=dataset,
-            batch_size=self.hparams.batch_size,
-            shuffle=True
-        )
-
-        return loader
-
-    @property
-    def tng_dataloader(self):
-        if self._tng_dataloader is None:
-            try:
-                self._tng_dataloader = self.__dataloader(train=True)
-            except Exception as e:
-                print(e)
-                raise e
-        return self._tng_dataloader
-
-    @property
-    def val_dataloader(self):
-        if self._val_dataloader is None:
-            try:
-                self._val_dataloader = self.__dataloader(train=False)
-            except Exception as e:
-                print(e)
-                raise e
-        return self._val_dataloader
-
-    @property
-    def test_dataloader(self):
-        if self._test_dataloader is None:
-            try:
-                self._test_dataloader = self.__dataloader(train=False)
-            except Exception as e:
-                print(e)
-                raise e
-        return self._test_dataloader
-
-    @staticmethod
-    def add_model_specific_args(parent_parser):
-        parser = HyperOptArgumentParser(strategy=parent_parser.strategy, parents=[parent_parser])
-
-        # param overwrites
-        # parser.set_defaults(gradient_clip=5.0)
-
-        # network params
-        parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=False)
-        parser.add_argument('--in_features', default=28*28)
-        parser.add_argument('--hidden_dim', default=500)
-        parser.add_argument('--out_features', default=10)
-
-        # data
-        parser.add_argument('--data_root', default='/Users/williamfalcon/Developer/personal/research_lib/research_proj/datasets/mnist', type=str)
-
-        # training params (opt)
-        parser.opt_list('--learning_rate', default=0.001, type=float, options=[0.0001, 0.0005, 0.001, 0.005],
-                        tunable=False)
-        parser.opt_list('--batch_size', default=256, type=int, options=[32, 64, 128, 256], tunable=False)
-        parser.opt_list('--optimizer_name', default='adam', type=str, options=['adam'], tunable=False)
-        return parser

pytorch_lightning/models/trainer.py

@@ -1,21 +1,41 @@
-import torch
-import tqdm
-import numpy as np
-from pytorch_lightning.root_module.memory import get_gpu_memory_map
+"""
+The trainer handles all the logic for running a val loop, training loop, distributing, etc...
+"""
+import subprocess
 import traceback
-from pytorch_lightning.root_module.model_saving import TrainerIO
-from torch.optim.lr_scheduler import MultiStepLR
-from pytorch_lightning.pt_overrides.override_data_parallel import LightningDataParallel
+import warnings
+import os
 import pdb
+import re
+
+import torch
+from torch.utils.data.distributed import DistributedSampler
+import torch.multiprocessing as mp
+import torch.distributed as dist
+import numpy as np
+import tqdm
+
+from pytorch_lightning.root_module.memory import get_gpu_memory_map
+from pytorch_lightning.root_module.model_saving import TrainerIO
+from pytorch_lightning.pt_overrides.override_data_parallel import LightningDistributedDataParallel, LightningDataParallel
+from pytorch_lightning.utils.debugging import MisconfigurationException
 
 try:
     from apex import amp
     APEX_AVAILABLE = True
-except ModuleNotFoundError:
+except Exception:
     APEX_AVAILABLE = False
 
 
 def reduce_distributed_output(output, nb_gpus):
+    if nb_gpus <= 1:
+        return output
+
+    # when using DP, we get one output per gpu
+    # average outputs and return
+    if type(output) is torch.Tensor:
+        return output.mean()
+
     for k, v in output.items():
         # recurse on nested dics
         if isinstance(output[k], dict):
@@ -32,10 +52,13 @@ class Trainer(TrainerIO):
 
     def __init__(self,
                  experiment,
-                 checkpoint_callback, early_stop_callback,
+                 early_stop_callback=None,
+                 checkpoint_callback=None,
+                 gradient_clip=0,
                  cluster=None,
                  process_position=0,
                  current_gpu_name=0,
+                 nb_gpu_nodes=1,
                  gpus=None,
                  progress_bar=True,
                  overfit_pct=0.0,
@@ -43,18 +66,56 @@ class Trainer(TrainerIO):
                  check_val_every_n_epoch=1,
                  fast_dev_run=False,
                  accumulate_grad_batches=1,
-                 enable_early_stop=True, max_nb_epochs=1000, min_nb_epochs=1,
-                 train_percent_check=1.0, val_percent_check=1.0, test_percent_check=1.0, val_check_interval=0.95,
+                 max_nb_epochs=1000, min_nb_epochs=1,
+                 train_percent_check=1.0, val_percent_check=1.0, test_percent_check=1.0,
+                 val_check_interval=0.95,
                  log_save_interval=100, add_log_row_interval=10,
-                 lr_scheduler_milestones=None,
+                 distributed_backend='dp',
                  use_amp=False,
                  print_nan_grads=False,
+                 print_weights_summary=True,
                  amp_level='O2',
                  nb_sanity_val_steps=5):
 
+        """
+
+        :param experiment: Test-tube experiment
+        :param early_stop_callback: from pytorch_lightning import EarlyStopping
+        :param checkpoint_callback: from pytorch_lightning import Checkpoint
+        :param gradient_clip:
+        :param cluster:
+        :param process_position:
+        :param current_gpu_name:
+        :param nb_gpu_nodes:
+        :param gpus:
+        :param progress_bar:
+        :param overfit_pct:
+        :param track_grad_norm:
+        :param check_val_every_n_epoch:
+        :param fast_dev_run:
+        :param accumulate_grad_batches:
+        :param max_nb_epochs:
+        :param min_nb_epochs:
+        :param train_percent_check:
+        :param val_percent_check:
+        :param test_percent_check:
+        :param val_check_interval:
+        :param log_save_interval:
+        :param add_log_row_interval:
+        :param distributed_backend: 'np' to use DistributedParallel, 'ddp' to use DistributedDataParallel
+        :param use_amp:
+        :param print_nan_grads:
+        :param print_weights_summary:
+        :param amp_level:
+        :param nb_sanity_val_steps:
+        """
+
         # Transfer params
+
+        self.nb_gpu_nodes = nb_gpu_nodes
+        self.gradient_clip = gradient_clip
         self.check_val_every_n_epoch = check_val_every_n_epoch
-        self.enable_early_stop = enable_early_stop
+        self.enable_early_stop = early_stop_callback is not None
         self.track_grad_norm = track_grad_norm
         self.fast_dev_run = fast_dev_run
         self.on_gpu = gpus is not None and torch.cuda.is_available()
@@ -64,8 +125,12 @@ class Trainer(TrainerIO):
         self.cluster = cluster
         self.process_position = process_position
         self.current_gpu_name = current_gpu_name
+        self.print_weights_summary = print_weights_summary
         self.checkpoint_callback = checkpoint_callback
-        self.checkpoint_callback.save_function = self.save_checkpoint
+
+        if self.checkpoint_callback is not None:
+            self.checkpoint_callback.save_function = self.save_checkpoint
+
         self.early_stop = early_stop_callback
         self.model = None
         self.max_nb_epochs = max_nb_epochs
@@ -73,12 +138,75 @@ class Trainer(TrainerIO):
         self.early_stop_callback = early_stop_callback
         self.min_nb_epochs = min_nb_epochs
         self.nb_sanity_val_steps = nb_sanity_val_steps
-        self.lr_scheduler_milestones = [] if lr_scheduler_milestones is None else [int(x.strip()) for x in lr_scheduler_milestones.split(',')]
         self.lr_schedulers = []
         self.amp_level = amp_level
         self.print_nan_grads = print_nan_grads
-        self.data_parallel_device_ids = gpus
-        self.data_parallel = gpus is not None and len(gpus) > 0
+        self.data_parallel_device_ids = None
+        self.world_size = 1
+        self.node_rank = 0
+        self.use_ddp = False
+        self.use_dp = False
+
+        # training bookeeping
+        self.total_batch_nb = 0
+        self.running_loss = []
+        self.avg_loss = 0
+        self.batch_nb = 0
+        self.tqdm_metrics = {}
+        self.nb_val_batches = None
+        self.nb_tng_batches = None
+        self.nb_test_batches = None
+
+        # gpus come in as a string.
+        # if gpus = -1 then use all available devices
+        # otherwise, split the string using commas
+        if gpus is not None:
+            if type(gpus) is list:
+                self.data_parallel_device_ids = gpus
+            elif type(gpus) is str:
+                if gpus == '-1':
+                    self.data_parallel_device_ids = list(range(0, torch.cuda.device_count()))
+                else:
+                    self.data_parallel_device_ids = [int(x.strip()) for x in gpus.split(',')]
+            else:
+                raise Exception('gpus has to be a string or list of ids')
+
+            # set the correct cuda visible devices (using pci order)
+            os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+            os.environ["CUDA_VISIBLE_DEVICES"] = ','.join([str(x) for x in self.data_parallel_device_ids])
+            print(f'VISIBLE GPUS: {os.environ["CUDA_VISIBLE_DEVICES"]}')
+
+        # make DP and DDP mutually exclusive
+        # single GPU will also use DP with devices=[0]
+        have_gpus = self.data_parallel_device_ids is not None and len(self.data_parallel_device_ids) > 0
+        if have_gpus:
+            self.use_dp = distributed_backend == 'dp'
+            self.use_ddp = distributed_backend == 'ddp'
+
+            # use ddp automatically if nb_gpu_nodes > 1
+            if nb_gpu_nodes > 1 and self.use_dp:  # pragma: no cover
+                self.use_ddp = True
+                self.use_dp = False
+                w = 'DataParallel does not support nb_gpu_nodes > 1. ' \
+                    'Switching to DistributedDataParallel for you. ' \
+                    'To silence this warning set distributed_backend=ddp'
+                warnings.warn(w)
+
+        # extract SLURM flag vars
+        # whenever we have the correct number of tasks, we let slurm manage processes
+        # otherwise we launch the required number of processes
+        if self.use_ddp:
+            self.nb_requested_gpus = len(self.data_parallel_device_ids) * self.nb_gpu_nodes
+            self.nb_slurm_tasks = 0
+            try:
+                self.nb_slurm_tasks = int(os.environ['SLURM_NTASKS'])
+                self.is_slurm_managing_tasks = self.nb_slurm_tasks == self.nb_requested_gpus
+            except Exception as e:
+                # likely not on slurm, so set the slurm managed flag to false
+                self.is_slurm_managing_tasks = False
+
+        # process info
+        self.proc_rank = 0
 
         # training state
         self.optimizers = None
@@ -101,11 +229,25 @@ class Trainer(TrainerIO):
         self.__determine_data_use_amount(train_percent_check, val_percent_check, test_percent_check, overfit_pct)
         print('gpu available: {}, used: {}'.format(torch.cuda.is_available(), self.on_gpu))
 
-        # apex test
+        # 16 bit mixed precision training using apex
         self.use_amp = use_amp and APEX_AVAILABLE
         if self.use_amp:
             print('using 16bit precision')
 
+        if use_amp and not APEX_AVAILABLE:  # pragma: no cover
+            msg = '''
+            You set use_amp=True but do not have apex installed.
+            Install apex first using this guide and rerun with use_amp=True: 
+            https://github.com/NVIDIA/apex#linux
+            
+            this run will NOT use 16 bit precision
+            '''
+            raise ModuleNotFoundError(msg)
+
+    @property
+    def data_parallel(self):
+        return self.use_dp or self.use_ddp
+
     def __determine_data_use_amount(self, train_percent_check, val_percent_check, test_percent_check, overfit_pct):
         """
         Use less data for debugging purposes
@@ -118,43 +260,52 @@ class Trainer(TrainerIO):
             self.val_percent_check = overfit_pct
             self.test_percent_check = overfit_pct
 
+    def __get_model(self):
+        return self.model.module if self.data_parallel else self.model
+
     def __is_function_implemented(self, f_name):
-        f_op = getattr(self.model, f_name, None)
+        model = self.__get_model()
+        f_op = getattr(model, f_name, None)
         return callable(f_op)
 
     @property
     def __tng_tqdm_dic(self):
+        # ForkedPdb().set_trace()
         tqdm_dic = {
             'tng_loss': '{0:.3f}'.format(self.avg_loss),
-            'gpu': '{}'.format(self.current_gpu_name),
             'v_nb': '{}'.format(self.experiment.version),
             'epoch': '{}'.format(self.current_epoch),
             'batch_nb':'{}'.format(self.batch_nb),
         }
         tqdm_dic.update(self.tqdm_metrics)
 
+        if self.on_gpu:
+            tqdm_dic['gpu'] = '{}'.format(self.current_gpu_name)
+
         return tqdm_dic
 
-    def __layout_bookeeping(self, model):
-        # training bookeeping
-        self.total_batch_nb = 0
-        self.running_loss = []
-        self.avg_loss = 0
-        self.batch_nb = 0
-        self.tqdm_metrics = {}
+    @property
+    def tng_tqdm_dic(self):
+        """
+        Read-only for tqdm metrics
+        :return:
+        """
+        return self.__tng_tqdm_dic
+
+    def __layout_bookeeping(self):
 
         # determine number of training batches
-        self.nb_tng_batches = model.nb_batches(self.tng_dataloader)
+        self.nb_tng_batches = len(self.tng_dataloader)
         self.nb_tng_batches = int(self.nb_tng_batches * self.train_percent_check)
 
         # determine number of validation batches
-        self.nb_val_batches = model.nb_batches(self.val_dataloader)
+        self.nb_val_batches = len(self.val_dataloader)
         self.nb_val_batches = int(self.nb_val_batches * self.val_percent_check)
         self.nb_val_batches = max(1, self.nb_val_batches)
         self.nb_val_batches = self.nb_val_batches
 
         # determine number of test batches
-        self.nb_test_batches = model.nb_batches(self.test_dataloader)
+        self.nb_test_batches = len(self.test_dataloader)
         self.nb_test_batches = int(self.nb_test_batches * self.test_percent_check)
 
         # determine when to check validation
@@ -175,12 +326,9 @@ class Trainer(TrainerIO):
         :param max_batches: Scalar
         :return:
         """
-        print('validating...')
-
         # enable eval mode
         model.zero_grad()
         model.eval()
-        model.from_lightning = True
 
         # disable gradients to save memory
         torch.set_grad_enabled(False)
@@ -191,7 +339,7 @@ class Trainer(TrainerIO):
         # run training
         for batch_i, data_batch in enumerate(dataloader):
 
-            if data_batch is None:
+            if data_batch is None:  # pragma: no cover
                 continue
 
             # stop short when on fast dev run
@@ -201,9 +349,12 @@ class Trainer(TrainerIO):
             # -----------------
             # RUN VALIDATION STEP
             # -----------------
-            if self.data_parallel:
+            if self.use_ddp:
+                output = model(data_batch, batch_i)
+            elif self.use_dp:
                 output = model(data_batch, batch_i)
                 output = reduce_distributed_output(output, len(self.data_parallel_device_ids))
+
             else:
                 output = model.validation_step(data_batch, batch_i)
 
@@ -227,7 +378,7 @@ class Trainer(TrainerIO):
 
         return val_results
 
-    def __get_dataloaders(self, model):
+    def get_dataloaders(self, model):
         """
         Dataloaders are provided by the model
         :param model:
@@ -237,73 +388,252 @@ class Trainer(TrainerIO):
         self.test_dataloader = model.test_dataloader
         self.val_dataloader = model.val_dataloader
 
+        if self.use_ddp and not isinstance(self.tng_dataloader.sampler, DistributedSampler):
+            msg = '''
+            when using multiple gpus and multiple nodes you must pass a DistributedSampler to DataLoader(sampler).
+            
+            ie: this:
+            dataset = myDataset()
+            dataloader = Dataloader(dataset)
+            
+            becomes:
+            dataset = myDataset()
+            dist_sampler = torch.utils.data.distributed.DistributedSampler(dataset)
+            dataloader = Dataloader(dataset, sampler=dist_sampler)
+            '''
+            raise MisconfigurationException(msg)
+
     # -----------------------------
     # MODEL TRAINING
     # -----------------------------
     def fit(self, model):
 
-        model.trainer = self
+        # when using multi-node or DDP within a node start each module in a separate process
+        if self.use_ddp:
+            # must copy only the meta of the exp so it survives pickle/unpickle when going to new process
+            self.experiment = self.experiment.get_meta_copy()
 
-        # transfer data loaders from model
-        self.__get_dataloaders(model)
+            if self.is_slurm_managing_tasks:
+                task = int(os.environ['SLURM_LOCALID'])
+                self.ddp_train(task, model)
+            else:
+                msg = f"""
+                You requested {self.nb_requested_gpus} GPUs but launched {self.nb_slurm_tasks} slurm tasks. 
+                We will launch {self.nb_requested_gpus} processes for you. 
+                We recommend you let slurm manage the processes by setting: --ntasks-per-node={self.nb_requested_gpus}
+                If you're not using SLURM, ignore this message!
+                """
+                warnings.warn(msg)
+                mp.spawn(self.ddp_train, nprocs=len(self.data_parallel_device_ids), args=(model, ))
 
-        # init training constants
-        self.__layout_bookeeping(model)
+        # 1 gpu or dp option triggers training using DP module
+        # easier to avoid NCCL issues
+        elif self.use_dp:
+            self.__dp_train(model)
+
+        # ON CPU
+        else:
+            # run through amp wrapper
+            if self.use_amp:
+                raise MisconfigurationException('amp + cpu is not supported. Please use a GPU option')
+
+            # CHOOSE OPTIMIZER
+            # allow for lr schedulers as well
+            self.optimizers = model.configure_optimizers()
+            if len(self.optimizers) == 2:
+                self.optimizers, self.lr_schedulers = self.optimizers
+
+            self.__run_pretrain_routine(model)
+
+        # return 1 when finished
+        # used for testing or when we need to know that training succeeded
+        return 1
+
+    def __dp_train(self, model):
 
         # CHOOSE OPTIMIZER
-        # filter out the weights that were done on gpu so we can load on good old cpus
+        # allow for lr schedulers as well
         self.optimizers = model.configure_optimizers()
+        if len(self.optimizers) == 2:
+            self.optimizers, self.lr_schedulers = self.optimizers
 
+        model.cuda(self.data_parallel_device_ids[0])
+
+        # check for this bug (amp + dp + !01 doesn't work)
+        # https://github.com/NVIDIA/apex/issues/227
+        if self.use_dp and self.use_amp:
+            m = f'amp level {self.amp_level} with DataParallel is not supported. ' \
+                f'See this note from NVIDIA for more info: https://github.com/NVIDIA/apex/issues/227. ' \
+                f'We recommend you switch to ddp if you want to use amp'
+            raise MisconfigurationException(m)
+
+        model = LightningDataParallel(model, device_ids=self.data_parallel_device_ids)
+
+        self.__run_pretrain_routine(model)
+
+    def ddp_train(self, gpu_nb, model):
+        """
+        Entry point into a DP thread
+        :param gpu_nb:
+        :param model:
+        :param cluster_obj:
+        :return:
+        """
+        # node rank using relative slurm id
+        # otherwise default to node rank 0
+        try:
+            node_id = os.environ['SLURM_NODEID']
+            self.node_rank = int(node_id)
+        except Exception as e:
+            self.node_rank = 0
+
+        # recover original exp before went into process
+        # init in write mode only on proc 0
+        self.experiment.debug = self.proc_rank > 0
+        self.experiment = self.experiment.get_non_ddp_exp()
+
+        # show progbar only on prog_rank 0
+        self.prog_bar = self.prog_bar and self.node_rank == 0 and gpu_nb == 0
+
+        # determine which process we are and world size
+        self.proc_rank = self.node_rank * len(self.data_parallel_device_ids) + gpu_nb
+        self.world_size = self.nb_gpu_nodes * len(self.data_parallel_device_ids)
+
+        # let the exp know the rank to avoid overwriting logs
+        self.experiment.rank = self.proc_rank
+
+        # set up server using proc 0's ip address
+        # try to init for 20 times at max in case ports are taken
+        # where to store ip_table
+        self.__init_tcp_connection()
+
+        # CHOOSE OPTIMIZER
+        # allow for lr schedulers as well
+        self.optimizers = model.configure_optimizers()
+        if len(self.optimizers) == 2:
+            self.optimizers, self.lr_schedulers = self.optimizers
+
+        # MODEL
+        # copy model to each gpu
+        torch.cuda.set_device(gpu_nb)
+        model.cuda(gpu_nb)
+
+        # AMP
+        # run through amp wrapper before going to distributed DP
         if self.use_amp:
             # An example
-            model, optimizer = amp.initialize(
-                model, self.optimizers[0], opt_level=self.amp_level,
+            model, optimizers = amp.initialize(
+                model, self.optimizers, opt_level=self.amp_level,
             )
-            self.optimizers[0] = optimizer
-            model.trainer = self
+            self.optimizers = optimizers
 
-        # add lr schedulers
-        if self.lr_scheduler_milestones is not None:
-            for optimizer in self.optimizers:
-                scheduler = MultiStepLR(optimizer, self.lr_scheduler_milestones)
-                self.lr_schedulers.append(scheduler)
+        model = LightningDistributedDataParallel(model, device_ids=[gpu_nb], find_unused_parameters=True)
+
+        # continue training routine
+        self.__run_pretrain_routine(model)
+
+    def __init_tcp_connection(self):
+        """
+        Connect all procs in the world using the env:// init
+        Use the first node as the root address
+        :param port:
+        :param tries:
+        :return:
+        """
+        # sets the appropriate port
+        try:
+            port = os.environ['MASTER_PORT']
+        except Exception as e:
+            port = 12910
+            os.environ['MASTER_PORT'] = f'{port}'
+
+        # figure out the root node addr
+        try:
+            root_node = os.environ['SLURM_NODELIST'].split(' ')[0]
+        except Exception as e:
+            root_node = '127.0.0.2'
+
+        root_node = self.resolve_root_node_address(root_node)
+        os.environ['MASTER_ADDR'] = root_node
+
+        dist.init_process_group("nccl", rank=self.proc_rank, world_size=self.world_size)
+
+    def resolve_root_node_address(self, root_node):
+        if '[' in root_node:
+            name = root_node.split('[')[0]
+            number = root_node.split(',')[0]
+            if '-' in number:
+                number = number.split('-')[0]
+
+            number = re.sub('[^0-9]', '', number)
+            root_node = name + number
+
+        return root_node
+
+    def __run_pretrain_routine(self, model):
+        """
+        Sanity check a few things before starting actual training
+        :param model:
+        :return:
+        """
+        ref_model = model
+        if self.data_parallel:
+            ref_model = model.module
+
+        ref_model.trainer = self
+
+        # set local properties on the model
+        ref_model.on_gpu = self.on_gpu
+
+        # transfer data loaders from model
+        self.get_dataloaders(ref_model)
+
+        # init training constants
+        self.__layout_bookeeping()
 
         # print model summary
-        model.summarize()
+        if self.proc_rank == 0 and self.print_weights_summary:
+            ref_model.summarize()
 
-        # put on gpu if needed
-        if self.on_gpu:
-            model = LightningDataParallel(model, device_ids=self.data_parallel_device_ids)
+        # give model convenience properties
+        ref_model.trainer = self
+        ref_model.experiment = self.experiment
 
         # run tiny validation to make sure program won't crash during val
         _ = self.validate(model, self.val_dataloader, max_batches=self.nb_sanity_val_steps)
 
         # save exp to get started
-        self.experiment.save()
+        if self.proc_rank == 0:
+            self.experiment.save()
+
+        # track model now.
+        # if cluster resets state, the model will update with the saved weights
+        self.model = model
 
         # enable cluster checkpointing
-        if self.cluster is not None:
+        # also restores training state
+        if self.cluster is not None:  # pragma: no cover
             self.enable_auto_hpc_walltime_manager()
 
         # ---------------------------
         # CORE TRAINING LOOP
         # ---------------------------
-        self.model = model
         self.__train()
 
     def __train(self):
         # run all epochs
         for epoch_nb in range(self.current_epoch, self.max_nb_epochs):
             # update the lr scheduler
-            for lr_scheduler in self.lr_schedulers:
-                lr_scheduler.step()
+            if self.lr_schedulers is not None:
+                for lr_scheduler in self.lr_schedulers:
+                    lr_scheduler.step()
 
-            model = self.model.module if self.data_parallel else self.model
+            model = self.__get_model()
             model.current_epoch = epoch_nb
 
             # hook
             if self.__is_function_implemented('on_epoch_start'):
-                model = self.model.module if self.data_parallel else self.model
+                model = self.__get_model()
                 model.on_epoch_start()
 
             self.current_epoch = epoch_nb
@@ -318,7 +648,7 @@ class Trainer(TrainerIO):
                 self.batch_nb = batch_nb
                 self.global_step += 1
 
-                model = self.model.module if self.data_parallel else self.model
+                model = self.__get_model()
                 model.global_step = self.global_step
 
                 # stop when the flag is changed or we've gone past the amount requested in the batches
@@ -342,17 +672,16 @@ class Trainer(TrainerIO):
 
                 # when batch should be saved
                 if (batch_nb + 1) % self.log_save_interval == 0 or early_stop_epoch:
-                    self.experiment.save()
+                    if self.proc_rank == 0:
+                        self.experiment.save()
 
                 # when metrics should be logged
                 if batch_nb % self.add_log_row_interval == 0 or early_stop_epoch:
                     # count items in memory
                     # nb_params, nb_tensors = count_mem_items()
 
-                    if self.data_parallel:
-                        metrics = self.model.module.update_tng_log_metrics(self.__tng_tqdm_dic)
-                    else:
-                        metrics = self.model.update_tng_log_metrics(self.__tng_tqdm_dic)
+                    model = self.__get_model()
+                    metrics = self.__tng_tqdm_dic
 
                     # add gpu memory
                     if self.on_gpu:
@@ -361,18 +690,22 @@ class Trainer(TrainerIO):
 
                     # add norms
                     if self.track_grad_norm > 0:
-                        model = self.model.module if self.data_parallel else self.model
+                        model = self.__get_model()
                         grad_norm_dic = model.grad_norm(self.track_grad_norm)
-
                         metrics.update(grad_norm_dic)
 
+                    if self.__is_function_implemented('on_tng_metrics'):
+                        model.on_tng_metrics(metrics)
+
                     # log metrics
-                    self.experiment.log(metrics)
-                    self.experiment.save()
+                    scalar_metrics = self.__metrics_to_scalars(metrics, blacklist=self.__log_vals_blacklist())
+                    if self.proc_rank == 0:
+                        self.experiment.log(scalar_metrics, global_step=self.global_step)
+                        self.experiment.save()
 
                 # hook
                 if self.__is_function_implemented('on_batch_end'):
-                    model = self.model.module if self.data_parallel else self.model
+                    model = self.__get_model()
                     model.on_batch_end()
 
                 # end epoch early
@@ -381,19 +714,37 @@ class Trainer(TrainerIO):
 
             # hook
             if self.__is_function_implemented('on_epoch_end'):
-                model = self.model.module if self.data_parallel else self.model
+                model = self.__get_model()
                 model.on_epoch_end()
 
             # early stopping
-            if self.enable_early_stop:
+            met_min_epochs = epoch_nb > self.min_nb_epochs
+            if self.enable_early_stop and met_min_epochs:
                 should_stop = self.early_stop_callback.on_epoch_end(epoch=epoch_nb, logs=self.__tng_tqdm_dic)
-                met_min_epochs = epoch_nb > self.min_nb_epochs
 
                 # stop training
                 stop = should_stop and met_min_epochs
                 if stop:
                     return
 
+    def __metrics_to_scalars(self, metrics, blacklist=[]):
+        new_metrics = {}
+        for k, v in metrics.items():
+            if type(v) is torch.Tensor:
+                v = v.item()
+
+            if type(v) is dict:
+                v = self.__metrics_to_scalars(v)
+
+            if k not in blacklist:
+                new_metrics[k] = float(v)
+
+        return new_metrics
+
+    def __log_vals_blacklist(self):
+        """avoid logging some vals lightning uses to maintain state"""
+        blacklist = {'batch_nb', 'v_nb', 'gpu'}
+        return blacklist
 
     def __run_tng_batch(self, data_batch, batch_nb):
         if data_batch is None:
@@ -401,8 +752,8 @@ class Trainer(TrainerIO):
 
         # hook
         if self.__is_function_implemented('on_batch_start'):
-            model = self.model.module if self.data_parallel else self.model
-            response = model.on_batch_start(data_batch)
+            model_ref = self.__get_model()
+            response = model_ref.on_batch_start(data_batch)
 
             if response == -1:
                 return -1
@@ -412,39 +763,68 @@ class Trainer(TrainerIO):
 
         # forward pass
         # return a scalar value and a dic with tqdm metrics
-        if self.data_parallel:
+        if self.use_ddp:
+            output = self.model(data_batch, batch_nb)
+        elif self.use_dp:
             output = self.model(data_batch, batch_nb)
             output = reduce_distributed_output(output, len(self.data_parallel_device_ids))
         else:
             output = self.model.training_step(data_batch, batch_nb)
 
-        model_specific_tqdm_metrics_dic = output['tqdm_metrics']
-        loss = output['loss']
+        try:
+            model_specific_tqdm_metrics_dic = output['prog']
+        except Exception as e:
+            model_specific_tqdm_metrics_dic = {}
+
+        # if output dict doesn't have the keyword loss
+        # then assume the output=loss if scalar
+        try:
+            loss = output['loss']
+        except Exception as e:
+            if type(output) is torch.Tensor:
+                loss = output
 
         self.__add_tqdm_metrics(model_specific_tqdm_metrics_dic)
 
         # backward pass
         if self.use_amp:
+            # scale loss when using amp
             for optimizer in self.optimizers:
                 with amp.scale_loss(loss, optimizer) as scaled_loss:
                     scaled_loss.backward()
         else:
             loss.backward()
 
+        # insert after step hook
+        if self.__is_function_implemented('on_after_backward'):
+            model_ref = self.__get_model()
+            response = model_ref.on_after_backward()
+
         if self.print_nan_grads:
-            model = self.model.module if self.data_parallel else self.model
+            model = self.__get_model()
             for param in model.parameters():
                 print(param.grad.float().sum())
 
+        # avoid memory leaks
         self.batch_loss_value += loss.item()
 
         # gradient update with accumulated gradients
         if (self.batch_nb + 1) % self.accumulate_grad_batches == 0:
 
+            # clip gradients
+            if self.gradient_clip > 0:
+                model = self.__get_model()
+                torch.nn.utils.clip_grad_norm(model.parameters(), self.gradient_clip)
+
             # update gradients across all optimizers
             for optimizer in self.optimizers:
                 optimizer.step()
 
+                # insert after step hook
+                if self.__is_function_implemented('on_before_zero_grad'):
+                    model_ref = self.__get_model()
+                    response = model_ref.on_before_zero_grad(optimizer)
+
                 # clear gradients
                 optimizer.zero_grad()
 
@@ -464,7 +844,8 @@ class Trainer(TrainerIO):
 
         # activate batch end hook
         if self.__is_function_implemented('on_batch_end'):
-            self.model.on_batch_end()
+            model = self.__get_model()
+            model.on_batch_end()
 
         return 0
 
@@ -476,28 +857,25 @@ class Trainer(TrainerIO):
         elif not can_check_epoch:
             return
 
-        try:
-            # hook
-            if self.__is_function_implemented('on_pre_performance_check'):
-                self.model.on_pre_performance_check()
+        # hook
+        if self.__is_function_implemented('on_pre_performance_check'):
+            model = self.__get_model()
+            model.on_pre_performance_check()
 
-            # use full val set on end of epoch
-            # use a small portion otherwise
-            max_batches = None if not self.fast_dev_run else 1
-            model_specific_tqdm_metrics_dic = self.validate(
-                self.model,
-                self.val_dataloader,
-                max_batches
-            )
-            self.__add_tqdm_metrics(model_specific_tqdm_metrics_dic)
+        # use full val set on end of epoch
+        # use a small portion otherwise
+        max_batches = None if not self.fast_dev_run else 1
+        model_specific_tqdm_metrics_dic = self.validate(
+            self.model,
+            self.val_dataloader,
+            max_batches
+        )
+        self.__add_tqdm_metrics(model_specific_tqdm_metrics_dic)
 
-            # hook
-            if self.__is_function_implemented('on_post_performance_check'):
-                self.model.on_post_performance_check()
-
-        except Exception as e:
-            print(e)
-            print(traceback.print_exc())
+        # hook
+        if self.__is_function_implemented('on_post_performance_check'):
+            model = self.__get_model()
+            model.on_post_performance_check()
 
         if self.progress_bar:
             # add model specific metrics
@@ -505,5 +883,6 @@ class Trainer(TrainerIO):
             self.prog_bar.set_postfix(**tqdm_metrics)
 
         # model checkpointing
-        print('save callback...')
-        self.checkpoint_callback.on_epoch_end(epoch=self.current_epoch, logs=self.__tng_tqdm_dic)
+        if self.proc_rank == 0 and self.checkpoint_callback is not None:
+            print('save callback...')
+            self.checkpoint_callback.on_epoch_end(epoch=self.current_epoch, logs=self.__tng_tqdm_dic)

pytorch_lightning/pt_overrides/override_data_parallel.py

@@ -1,4 +1,7 @@
 from torch.nn import DataParallel
+from torch.nn.parallel import DistributedDataParallel
+import itertools
+from itertools import chain
 
 import threading
 import torch
@@ -6,7 +9,20 @@ from torch.cuda._utils import _get_device_index
 import pdb
 
 
-def get_a_var(obj):
+def _find_tensors(obj):  # pragma: no cover
+    r"""
+    Recursively find all tensors contained in the specified object.
+    """
+    if isinstance(obj, torch.Tensor):
+        return [obj]
+    if isinstance(obj, (list, tuple)):
+        return itertools.chain(*map(_find_tensors, obj))
+    if isinstance(obj, dict):
+        return itertools.chain(*map(_find_tensors, obj.values()))
+    return []
+
+
+def get_a_var(obj):  # pragma: no cover
     if isinstance(obj, torch.Tensor):
         return obj
 
@@ -26,11 +42,77 @@ class LightningDataParallel(DataParallel):
     Override the forward call in lightning so it goes to training and validation step respectively
     """
 
+    def forward(self, *inputs, **kwargs):
+        if not self.device_ids:
+            return self.module(*inputs, **kwargs)
+
+        for t in chain(self.module.parameters(), self.module.buffers()):
+            if t.device != self.src_device_obj:
+                raise RuntimeError("module must have its parameters and buffers "
+                                   "on device {} (device_ids[0]) but found one of "
+                                   "them on device: {}".format(self.src_device_obj, t.device))
+
+        inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids)
+        if len(self.device_ids) == 1:
+            # lightning
+            if self.module.training:
+                return self.module.training_step(*inputs[0], **kwargs[0])
+            else:
+                return self.module.validation_step(*inputs[0], **kwargs[0])
+
+        replicas = self.replicate(self.module, self.device_ids[:len(inputs)])
+        outputs = self.parallel_apply(replicas, inputs, kwargs)
+        return self.gather(outputs, self.output_device)
+
+
     def parallel_apply(self, replicas, inputs, kwargs):
         return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])
 
 
-def parallel_apply(modules, inputs, kwargs_tup=None, devices=None):
+class LightningDistributedDataParallel(DistributedDataParallel):
+    """
+    Override the forward call in lightning so it goes to training and validation step respectively
+    """
+
+    def parallel_apply(self, replicas, inputs, kwargs):
+        return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])
+
+    def forward(self, *inputs, **kwargs):  # pragma: no cover
+        self._sync_params()
+        if self.device_ids:
+            inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids)
+            if len(self.device_ids) == 1:
+                # --------------
+                # LIGHTNING MOD
+                # --------------
+                # normal
+                # output = self.module(*inputs[0], **kwargs[0])
+
+                # lightning
+                if self.module.training:
+                    output = self.module.training_step(*inputs[0], **kwargs[0])
+                else:
+                    output = self.module.validation_step(*inputs[0], **kwargs[0])
+            else:
+                outputs = self.parallel_apply(self._module_copies[:len(inputs)], inputs, kwargs)
+                output = self.gather(outputs, self.output_device)
+        else:
+            output = self.module(*inputs, **kwargs)
+
+        if torch.is_grad_enabled():
+            # We'll return the output object verbatim since it is a freeform
+            # object. We need to find any tensors in this object, though,
+            # because we need to figure out which parameters were used during
+            # this forward pass, to ensure we short circuit reduction for any
+            # unused parameters. Only if `find_unused_parameters` is set.
+            if self.find_unused_parameters:
+                self.reducer.prepare_for_backward(list(_find_tensors(output)))
+            else:
+                self.reducer.prepare_for_backward([])
+        return output
+
+
+def parallel_apply(modules, inputs, kwargs_tup=None, devices=None):  # pragma: no cover
     r"""Applies each `module` in :attr:`modules` in parallel on arguments
     contained in :attr:`inputs` (positional) and :attr:`kwargs_tup` (keyword)
     on each of :attr:`devices`.
@@ -102,4 +184,4 @@ def parallel_apply(modules, inputs, kwargs_tup=None, devices=None):
         if isinstance(output, Exception):
             raise output
         outputs.append(output)
-    return outputs
+    return outputs

pytorch_lightning/root_module/decorators.py

@@ -0,0 +1,17 @@
+
+def data_loader(fn):
+    """
+    Decorator to make any fx with this use the lazy property
+    :param fn:
+    :return:
+    """
+
+    attr_name = '_lazy_' + fn.__name__
+
+    @property
+    def _data_loader(self):
+        if not hasattr(self, attr_name):
+            setattr(self, attr_name, fn(self))
+        return getattr(self, attr_name)
+
+    return _data_loader

pytorch_lightning/root_module/grads.py

@@ -27,14 +27,3 @@ class GradInformation(nn.Module):
         results['grad_{}_norm_total'.format(norm_type)] = round(total_norm.data.cpu().numpy().flatten()[0], 3)
         return results
 
-
-    def describe_grads(self):
-        for p in self.parameters():
-            g = p.grad.data.numpy().flatten()
-            print(np.max(g), np.min(g), np.mean(g))
-
-
-    def describe_params(self):
-        for p in self.parameters():
-            g = p.data.numpy().flatten()
-            print(np.max(g), np.min(g), np.mean(g))

pytorch_lightning/root_module/hooks.py

@@ -19,3 +19,27 @@ class ModelHooks(torch.nn.Module):
     def on_post_performance_check(self):
         pass
 
+    def on_tng_metrics(self, metrics):
+        pass
+
+    def on_before_zero_grad(self, optimizer):
+        """
+        Called after optimizer.step() and before optimizer.zero_grad()
+
+        for optimizer in optimizers:
+            optimizer.step()
+            model.on_before_zero_grad(optimizer) # < ---- called here
+            optimizer.zero_grad
+
+        :param optimizer:
+        :return:
+        """
+        pass
+
+    def on_after_backward(self):
+        """
+        Called after loss.backward() and before optimizers do anything
+        :return:
+        """
+        pass
+

pytorch_lightning/root_module/memory.py

@@ -33,33 +33,42 @@ class ModelSummary(object):
         mods = list(self.model.modules())
         in_sizes = []
         out_sizes = []
-        input_ = self.example_input_array
-        for i in range(1, len(mods)):
-            m = mods[i]
-            if type(input_) is list or type(input_) is tuple:
-                out = m(*input_)
-            else:
-                out = m(input_)
+        input_ = self.model.example_input_array
 
-            if type(input_) is tuple or type(input_) is list:
-                in_size = []
-                for x in input_:
-                    if type(x) is list:
-                        in_size.append(len(x))
-                    else:
-                        in_size.append(x.size())
-            else:
-                in_size = np.array(input_.size())
+        if self.model.on_gpu:
+            input_ = input_.cuda(0)
 
-            in_sizes.append(in_size)
+        if self.model.trainer.use_amp:
+            input_ = input_.half()
 
-            if type(out) is tuple or type(out) is list:
-                out_size = np.asarray([x.size() for x in out])
-            else:
-                out_size = np.array(out.size())
+        with torch.no_grad():
 
-            out_sizes.append(out_size)
-            input_ = out
+            for i in range(1, len(mods)):
+                m = mods[i]
+                if type(input_) is list or type(input_) is tuple:  # pragma: no cover
+                    out = m(*input_)
+                else:
+                    out = m(input_)
+
+                if type(input_) is tuple or type(input_) is list:  # pragma: no cover
+                    in_size = []
+                    for x in input_:
+                        if type(x) is list:
+                            in_size.append(len(x))
+                        else:
+                            in_size.append(x.size())
+                else:
+                    in_size = np.array(input_.size())
+
+                in_sizes.append(in_size)
+
+                if type(out) is tuple or type(out) is list:  # pragma: no cover
+                    out_size = np.asarray([x.size() for x in out])
+                else:
+                    out_size = np.array(out.size())
+
+                out_sizes.append(out_size)
+                input_ = out
 
         self.in_sizes = in_sizes
         self.out_sizes = out_sizes
@@ -114,13 +123,22 @@ class ModelSummary(object):
         Layer Name, Layer Type, Input Size, Output Size, Number of Parameters
         '''
 
-        df = pd.DataFrame( np.zeros( (len(self.layer_names), 3) ) )
-        df.columns = ['Name', 'Type', 'Params']
+        cols = ['Name', 'Type', 'Params']
+        if self.model.example_input_array is not None:
+            cols.extend(['In_sizes', 'Out_sizes'])
+
+        df = pd.DataFrame(np.zeros( (len(self.layer_names), len(cols))))
+        df.columns = cols
 
         df['Name'] = self.layer_names
         df['Type'] = self.layer_types
         df['Params'] = self.param_nums
 
+        if self.model.example_input_array is not None:
+
+            df['In_sizes'] = self.in_sizes
+            df['Out_sizes'] = self.out_sizes
+
         self.summary = df
         return
 
@@ -128,10 +146,13 @@ class ModelSummary(object):
         self.get_layer_names()
         self.get_parameter_sizes()
         self.get_parameter_nums()
+
+        if self.model.example_input_array is not None:
+            self.get_variable_sizes()
         self.make_summary()
 
 
-def print_mem_stack():
+def print_mem_stack(): # pragma: no cover
     for obj in gc.get_objects():
         try:
             if torch.is_tensor(obj) or (hasattr(obj, 'data') and torch.is_tensor(obj.data)):
@@ -140,7 +161,7 @@ def print_mem_stack():
             pass
 
 
-def count_mem_items():
+def count_mem_items(): # pragma: no cover
     nb_params = 0
     nb_tensors = 0
     for obj in gc.get_objects():

pytorch_lightning/root_module/model_saving.py

@@ -2,28 +2,52 @@ import torch
 import os
 import re
 import pdb
-from pytorch_lightning.pt_overrides.override_data_parallel import LightningDataParallel
+from pytorch_lightning.pt_overrides.override_data_parallel import LightningDistributedDataParallel, LightningDataParallel
+
 
 class ModelIO(object):
 
-    def load_model_specific(self, checkpoint):
+    def on_load_checkpoint(self, checkpoint):
         """
         Do something with the checkpoint
+        Gives model a chance to load something before state_dict is restored
         :param checkpoint:
         :return:
         """
-        raise NotImplementedError
+        pass
 
-    def get_save_dict(self):
+    def on_save_checkpoint(self, checkpoint):
         """
-        Return specific things for the model
+        Give the model a chance to add something to the checkpoint.
+        state_dict is already there
+        """
+        pass
+
+    # -------------------------
+    # OPTIONAL HOOKS
+    # -------------------------
+    def on_hpc_save(self, checkpoint):
+        """
+        Hook to do whatever you need right before Slurm manager saves the model
         :return:
         """
-        raise NotImplementedError
+        pass
+
+    def on_hpc_load(self, checkpoint):
+        """
+        Hook to do whatever you need right before Slurm manager loads the model
+        :return:
+        """
+        pass
 
 
 class TrainerIO(object):
 
+    def __get_model(self):
+        is_dp_module = type(self.model) is LightningDistributedDataParallel or type(self.model) is LightningDataParallel
+        model = self.model.module if is_dp_module else self.model
+        return model
+
     # --------------------
     # MODEL SAVE CHECKPOINT
     # --------------------
@@ -34,26 +58,40 @@ class TrainerIO(object):
         torch.save(checkpoint, filepath)
 
     def dump_checkpoint(self):
+
         checkpoint = {
             'epoch': self.current_epoch,
-            'checkpoint_callback_best': self.checkpoint_callback.best,
-            'early_stop_callback_wait': self.early_stop_callback.wait,
-            'early_stop_callback_patience': self.early_stop_callback.patience,
             'global_step': self.global_step
         }
 
+        if self.checkpoint_callback is not None:
+            checkpoint['checkpoint_callback_best'] = self.checkpoint_callback.best
+
+        if self.early_stop_callback is not None:
+            checkpoint['early_stop_callback_wait'] = self.early_stop_callback.wait
+            checkpoint['early_stop_callback_patience'] = self.early_stop_callback.patience
+
+        # save optimizers
         optimizer_states = []
         for i, optimizer in enumerate(self.optimizers):
             optimizer_states.append(optimizer.state_dict())
 
         checkpoint['optimizer_states'] = optimizer_states
+        
+        # save lr schedulers
+        lr_schedulers = []
+        for i, scheduler in enumerate(self.lr_schedulers):
+            lr_schedulers.append(scheduler.state_dict())
 
-        # request what to save from the model
-        model = self.model.module if type(self.model) is LightningDataParallel else self.model
-        checkpoint_dict = model.get_save_dict()
+        checkpoint['lr_schedulers'] = lr_schedulers
+
+        # add the state_dict from the model
+        model = self.__get_model()
+        checkpoint['state_dict'] = model.state_dict()
+
+        # give the model a chance to add a few things
+        model.on_save_checkpoint(checkpoint)
 
-        # merge trainer and model saving items
-        checkpoint.update(checkpoint_dict)
         return checkpoint
 
     # --------------------
@@ -86,9 +124,13 @@ class TrainerIO(object):
         :param checkpoint:
         :return:
         """
-        self.checkpoint_callback.best = checkpoint['checkpoint_callback_best']
-        self.early_stop_callback.wait = checkpoint['early_stop_callback_wait']
-        self.early_stop_callback.patience = checkpoint['early_stop_callback_patience']
+        if self.checkpoint_callback is not None:
+            self.checkpoint_callback.best = checkpoint['checkpoint_callback_best']
+
+        if self.early_stop_callback is not None:
+            self.early_stop_callback.wait = checkpoint['early_stop_callback_wait']
+            self.early_stop_callback.patience = checkpoint['early_stop_callback_patience']
+
         self.global_step = checkpoint['global_step']
         self.current_epoch = checkpoint['epoch']
 
@@ -96,6 +138,11 @@ class TrainerIO(object):
         optimizer_states = checkpoint['optimizer_states']
         for optimizer, opt_state in zip(self.optimizers, optimizer_states):
             optimizer.load_state_dict(opt_state)
+        
+        # restore the lr schedulers
+        lr_schedulers = checkpoint['lr_schedulers']
+        for scheduler, lrs_state in zip(self.lr_schedulers, lr_schedulers):
+            scheduler.load_state_dict(lrs_state)
 
     # ----------------------------------
     # PRIVATE OPS
@@ -107,17 +154,25 @@ class TrainerIO(object):
         # save exp to make sure we get all the metrics
         experiment.save()
 
+        # close experiment to avoid issues
+        experiment.close()
+
         ckpt_number = self.max_ckpt_in_folder(folderpath) + 1
 
         if not os.path.exists(folderpath):
             os.makedirs(folderpath, exist_ok=True)
         filepath = '{}/hpc_ckpt_{}.ckpt'.format(folderpath, ckpt_number)
 
-        # request what to save from the model
-        checkpoint_dict = self.dump_checkpoint()
+        # give model a chance to do something on hpc_save
+        model = self.__get_model()
+        checkpoint = self.dump_checkpoint()
+
+        model.on_hpc_save(checkpoint)
 
         # do the actual save
-        torch.save(checkpoint_dict, filepath)
+        torch.save(checkpoint, filepath)
+
+        return filepath
 
     def hpc_load(self, folderpath, on_gpu):
         filepath = '{}/hpc_ckpt_{}.ckpt'.format(folderpath, self.max_ckpt_in_folder(folderpath))
@@ -127,12 +182,17 @@ class TrainerIO(object):
         else:
             checkpoint = torch.load(filepath, map_location=lambda storage, loc: storage)
 
-        # load training state
+        # load training state (affects trainer only)
         self.restore_training_state(checkpoint)
 
         # load model state
-        model = self.model.module if type(self.model) is LightningDataParallel else self.model
-        model.load_model_specific(checkpoint)
+        model = self.__get_model()
+
+        # load the state_dict on the model automatically
+        model.load_state_dict(checkpoint['state_dict'])
+
+        # call model hook
+        model.on_hpc_load(checkpoint)
 
     def max_ckpt_in_folder(self, path):
         files = os.listdir(path)

pytorch_lightning/root_module/optimization.py

@@ -1,22 +0,0 @@
-from torch import nn
-from torch import optim
-
-
-class OptimizerConfig(nn.Module):
-
-    def choose_optimizer(self, optimizer, params, optimizer_params, opt_name_key):
-        if optimizer == 'adam':
-            optimizer = optim.Adam(params, **optimizer_params)
-        if optimizer == 'sparse_adam':
-            optimizer = optim.SparseAdam(params, **optimizer_params)
-        if optimizer == 'sgd':
-            optimizer = optim.SGD(params, **optimizer_params)
-        if optimizer == 'adadelta':
-            optimizer = optim.Adadelta(params, **optimizer_params)
-
-        # transfer opt state if loaded
-        if opt_name_key in self.loaded_optimizer_states_dict:
-            state = self.loaded_optimizer_states_dict[opt_name_key]
-            optimizer.load_state_dict(state)
-
-        return optimizer

pytorch_lightning/root_module/root_module.py

@@ -1,48 +1,27 @@
-import os
 import torch
-import math
-
 from pytorch_lightning.root_module.memory import ModelSummary
 from pytorch_lightning.root_module.grads import GradInformation
 from pytorch_lightning.root_module.model_saving import ModelIO, load_hparams_from_tags_csv
-from pytorch_lightning.root_module.optimization import OptimizerConfig
 from pytorch_lightning.root_module.hooks import ModelHooks
+from pytorch_lightning.root_module.decorators import data_loader
 
 
+class LightningModule(GradInformation, ModelIO, ModelHooks):
 
-class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
-
-    def __init__(self, hparams):
-        super(LightningModule, self).__init__()
-        self.hparams = hparams
+    def __init__(self, *args, **kwargs):
+        super(LightningModule, self).__init__(*args, **kwargs)
 
         self.dtype = torch.FloatTensor
         self.exp_save_path = None
         self.current_epoch = 0
         self.global_step = 0
         self.loaded_optimizer_states_dict = {}
-        self.fast_dev_run = hparams.fast_dev_run
-        self.overfit = hparams.overfit
-        self.gradient_clip = hparams.gradient_clip
-        self.num = 2
         self.trainer = None
-        self.from_lightning = True
+        self.experiment = None
+        self.example_input_array = None
 
         # track if gpu was requested for checkpointing
         self.on_gpu = False
-        try:
-            self.on_gpu = hparams.on_gpu
-        except Exception as e:
-            pass
-
-        # computed vars for the dataloaders
-        self._tng_dataloader = None
-        self._val_dataloader = None
-        self._test_dataloader = None
-
-        if self.on_gpu:
-            print('running on gpu...')
-            torch.set_default_tensor_type(hparams.default_tensor_type)
 
     def forward(self, *args, **kwargs):
         """
@@ -79,45 +58,12 @@ class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
 
     def configure_optimizers(self):
         """
-        Return array of optimizers
+        Return a list of optimizers and a list of schedulers (could be empty)
         :return:
         """
         raise NotImplementedError
 
-    def update_tng_log_metrics(self, logs):
-        """
-        Chance to update metrics to be logged for training step.
-        For example, add music, images, etc... to log
-        :param logs:
-        :return:
-        """
-        raise NotImplementedError
-
-    def loss(self, *args, **kwargs):
-        """
-        Expand model_out into your components
-        :param model_out:
-        :return:
-        """
-        raise NotImplementedError
-
-    def summarize(self):
-        model_summary = ModelSummary(self)
-        print(model_summary)
-
-    def nb_batches(self, dataloader):
-        a = math.ceil(float(len(dataloader.dataset) / self.batch_size))
-        return int(a)
-
-    def freeze(self):
-        for param in self.parameters():
-            param.requires_grad = False
-
-    def unfreeze(self):
-        for param in self.parameters():
-            param.requires_grad = True
-
-    @property
+    @data_loader
     def tng_dataloader(self):
         """
         Implement a function to load an h5py of this data
@@ -125,7 +71,7 @@ class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
         """
         raise NotImplementedError
 
-    @property
+    @data_loader
     def test_dataloader(self):
         """
         Implement a function to load an h5py of this data
@@ -133,7 +79,7 @@ class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
         """
         raise NotImplementedError
 
-    @property
+    @data_loader
     def val_dataloader(self):
         """
         Implement a function to load an h5py of this data
@@ -141,16 +87,6 @@ class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
         """
         raise NotImplementedError
 
-    @staticmethod
-    def get_process_position(gpus):
-        try:
-            current_gpu = os.environ["CUDA_VISIBLE_DEVICES"]
-            gpu_ids = gpus.split(';')
-            process_position = gpu_ids.index(current_gpu)
-            return process_position, current_gpu
-        except Exception as e:
-            return 0, 0
-
     @classmethod
     def load_from_metrics(cls, weights_path, tags_csv, on_gpu, map_location=None):
         """
@@ -172,9 +108,26 @@ class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
         else:
             checkpoint = torch.load(weights_path, map_location=lambda storage, loc: storage)
 
+        # load the state_dict on the model automatically
         model = cls(hparams)
+        model.load_state_dict(checkpoint['state_dict'])
+
+        # give model a chance to load something
+        model.on_load_checkpoint(checkpoint)
 
-        # allow model to load
-        model.load_model_specific(checkpoint)
-        model.load_state_dict(checkpoint['state_dict'], strict=False)
         return model
+
+    def summarize(self):
+        model_summary = ModelSummary(self)
+        print(model_summary)
+
+    def freeze(self):
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def unfreeze(self):
+        for param in self.parameters():
+            param.requires_grad = True
+
+
+

pytorch_lightning/testing_models/lm_test_module.py

@@ -0,0 +1,259 @@
+import os
+from collections import OrderedDict
+import torch.nn as nn
+from torchvision.datasets import MNIST
+import torchvision.transforms as transforms
+import torch
+import torch.nn.functional as F
+from test_tube import HyperOptArgumentParser
+from torch import optim
+from torch.utils.data import DataLoader
+from torch.utils.data.distributed import DistributedSampler
+
+from pytorch_lightning.root_module.root_module import LightningModule
+import pytorch_lightning as ptl
+
+
+class LightningTestModel(LightningModule):
+    """
+    Sample model to show how to define a template
+    """
+
+    def __init__(self, hparams, force_remove_distributed_sampler=False):
+        """
+        Pass in parsed HyperOptArgumentParser to the model
+        :param hparams:
+        """
+        # init superclass
+        super(LightningTestModel, self).__init__()
+        self.hparams = hparams
+
+        self.batch_size = hparams.batch_size
+
+        # if you specify an example input, the summary will show input/output for each layer
+        self.example_input_array = torch.rand(5, 28 * 28)
+
+        # remove to test warning for dist sampler
+        self.force_remove_distributed_sampler = force_remove_distributed_sampler
+
+        # build model
+        self.__build_model()
+
+    # ---------------------
+    # MODEL SETUP
+    # ---------------------
+    def __build_model(self):
+        """
+        Layout model
+        :return:
+        """
+        self.c_d1 = nn.Linear(in_features=self.hparams.in_features, out_features=self.hparams.hidden_dim)
+        self.c_d1_bn = nn.BatchNorm1d(self.hparams.hidden_dim)
+        self.c_d1_drop = nn.Dropout(self.hparams.drop_prob)
+
+        self.c_d2 = nn.Linear(in_features=self.hparams.hidden_dim, out_features=self.hparams.out_features)
+
+    # ---------------------
+    # TRAINING
+    # ---------------------
+    def forward(self, x):
+        """
+        No special modification required for lightning, define as you normally would
+        :param x:
+        :return:
+        """
+
+        x = self.c_d1(x)
+        x = torch.tanh(x)
+        x = self.c_d1_bn(x)
+        x = self.c_d1_drop(x)
+
+        x = self.c_d2(x)
+        logits = F.log_softmax(x, dim=1)
+
+        return logits
+
+    def loss(self, labels, logits):
+        nll = F.nll_loss(logits, labels)
+        return nll
+
+    def training_step(self, data_batch, batch_i):
+        """
+        Lightning calls this inside the training loop
+        :param data_batch:
+        :return:
+        """
+        # forward pass
+        x, y = data_batch
+        x = x.view(x.size(0), -1)
+
+        y_hat = self.forward(x)
+
+        # calculate loss
+        loss_val = self.loss(y, y_hat)
+
+        # in DP mode (default) make sure if result is scalar, there's another dim in the beginning
+        if self.trainer.use_dp:
+            loss_val = loss_val.unsqueeze(0)
+
+        # alternate possible outputs to test
+        if self.trainer.batch_nb % 1 == 0:
+            output = OrderedDict({
+                'loss': loss_val,
+                'prog': {'some_val': loss_val * loss_val}
+            })
+            return output
+        if self.trainer.batch_nb % 2 == 0:
+            return loss_val
+
+    def validation_step(self, data_batch, batch_i):
+        """
+        Lightning calls this inside the validation loop
+        :param data_batch:
+        :return:
+        """
+        x, y = data_batch
+        x = x.view(x.size(0), -1)
+        y_hat = self.forward(x)
+
+        loss_val = self.loss(y, y_hat)
+
+        # acc
+        labels_hat = torch.argmax(y_hat, dim=1)
+        val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)
+        val_acc = torch.tensor(val_acc)
+
+        if self.on_gpu:
+            val_acc = val_acc.cuda(loss_val.device.index)
+
+        # in DP mode (default) make sure if result is scalar, there's another dim in the beginning
+        if self.trainer.use_dp:
+            loss_val = loss_val.unsqueeze(0)
+            val_acc = val_acc.unsqueeze(0)
+
+        # alternate possible outputs to test
+        if self.trainer.batch_nb % 1 == 0:
+            output = OrderedDict({
+                'val_loss': loss_val,
+                'val_acc': val_acc,
+            })
+            return output
+        if self.trainer.batch_nb % 2 == 0:
+            return val_acc
+
+        if self.trainer.batch_nb % 3 == 0:
+            output = OrderedDict({
+                'val_loss': loss_val,
+                'val_acc': val_acc,
+                'test_dic': {'val_loss_a': loss_val}
+            })
+            return output
+
+    def validation_end(self, outputs):
+        """
+        Called at the end of validation to aggregate outputs
+        :param outputs: list of individual outputs of each validation step
+        :return:
+        """
+        # if returned a scalar from validation_step, outputs is a list of tensor scalars
+        # we return just the average in this case (if we want)
+        # return torch.stack(outputs).mean()
+
+        val_loss_mean = 0
+        val_acc_mean = 0
+        for output in outputs:
+            val_loss_mean += output['val_loss']
+            val_acc_mean += output['val_acc']
+
+        val_loss_mean /= len(outputs)
+        val_acc_mean /= len(outputs)
+
+        tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()}
+        return tqdm_dic
+
+    def on_tng_metrics(self, logs):
+        logs['some_tensor_to_test'] = torch.rand(1)
+
+    # ---------------------
+    # TRAINING SETUP
+    # ---------------------
+    def configure_optimizers(self):
+        """
+        return whatever optimizers we want here
+        :return: list of optimizers
+        """
+        # try no scheduler for this model (testing purposes)
+        optimizer = optim.Adam(self.parameters(), lr=self.hparams.learning_rate)
+
+        # test returning only 1 list instead of 2
+        return [optimizer]
+
+    def __dataloader(self, train):
+        # init data generators
+        transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
+        dataset = MNIST(root=self.hparams.data_root, train=train, transform=transform, download=True)
+
+        # when using multi-node we need to add the datasampler
+        train_sampler = None
+        batch_size = self.hparams.batch_size
+
+        try:
+            if self.on_gpu and not self.force_remove_distributed_sampler:
+                train_sampler = DistributedSampler(dataset, rank=self.trainer.proc_rank)
+                batch_size = batch_size // self.trainer.world_size  # scale batch size
+        except Exception as e:
+            pass
+
+        should_shuffle = train_sampler is None
+        loader = DataLoader(
+            dataset=dataset,
+            batch_size=batch_size,
+            shuffle=should_shuffle,
+            sampler=train_sampler
+        )
+
+        return loader
+
+    @ptl.data_loader
+    def tng_dataloader(self):
+        return self.__dataloader(train=True)
+
+    @ptl.data_loader
+    def val_dataloader(self):
+        return self.__dataloader(train=False)
+
+    @ptl.data_loader
+    def test_dataloader(self):
+        return self.__dataloader(train=False)
+
+    @staticmethod
+    def add_model_specific_args(parent_parser, root_dir):
+        """
+        Parameters you define here will be available to your model through self.hparams
+        :param parent_parser:
+        :param root_dir:
+        :return:
+        """
+        parser = HyperOptArgumentParser(strategy=parent_parser.strategy, parents=[parent_parser])
+
+        # param overwrites
+        # parser.set_defaults(gradient_clip=5.0)
+
+        # network params
+        parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=False)
+        parser.add_argument('--in_features', default=28*28, type=int)
+        parser.add_argument('--out_features', default=10, type=int)
+        parser.add_argument('--hidden_dim', default=50000, type=int) # use 500 for CPU, 50000 for GPU to see speed difference
+
+        # data
+        parser.add_argument('--data_root', default=os.path.join(root_dir, 'mnist'), type=str)
+
+        # training params (opt)
+        parser.opt_list('--learning_rate', default=0.001*8, type=float, options=[0.0001, 0.0005, 0.001, 0.005],
+                        tunable=False)
+        parser.opt_list('--optimizer_name', default='adam', type=str, options=['adam'], tunable=False)
+
+        # if using 2 nodes with 4 gpus each the batch size here (256) will be 256 / (2*8) = 16 per gpu
+        parser.opt_list('--batch_size', default=256*8, type=int, options=[32, 64, 128, 256], tunable=False,
+                        help='batch size will be divided over all the gpus being used across all nodes')
+        return parser

pytorch_lightning/trainer_main.py

@@ -1,214 +0,0 @@
-import os
-import sys
-
-import torch
-import numpy as np
-from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
-from pytorch_lightning.models.trainer import Trainer
-from pytorch_lightning.utils.arg_parse import add_default_args
-from time import sleep
-
-from pytorch_lightning.callbacks.pt_callbacks import EarlyStopping, ModelCheckpoint
-SEED = 2334
-torch.manual_seed(SEED)
-np.random.seed(SEED)
-
-# ---------------------
-# DEFINE MODEL HERE
-# ---------------------
-from pytorch_lightning.models.sample_model_template.model_template import ExampleModel1
-# ---------------------
-
-AVAILABLE_MODELS = {
-    'model_1': ExampleModel1
-}
-
-
-"""
-Allows training by using command line arguments
-
-Run by: 
-# TYPE YOUR RUN COMMAND HERE
-"""
-
-
-def main_local(hparams):
-    main(hparams, None, None)
-
-
-def main(hparams, cluster, results_dict):
-    """
-    Main training routine specific for this project
-    :param hparams:
-    :return:
-    """
-    on_gpu = torch.cuda.is_available()
-    if hparams.disable_cuda:
-        on_gpu = False
-
-    device = 'cuda' if on_gpu else 'cpu'
-    hparams.__setattr__('device', device)
-    hparams.__setattr__('on_gpu', on_gpu)
-    hparams.__setattr__('nb_gpus', torch.cuda.device_count())
-    hparams.__setattr__('inference_mode', hparams.model_load_weights_path is not None)
-
-    # delay each training start to not overwrite logs
-    process_position, current_gpu = TRAINING_MODEL.get_process_position(hparams.gpus)
-    sleep(process_position + 1)
-
-    # init experiment
-    exp = Experiment(
-        name=hparams.tt_name,
-        debug=hparams.debug,
-        save_dir=hparams.tt_save_path,
-        version=hparams.hpc_exp_number,
-        autosave=False,
-        description=hparams.tt_description
-    )
-
-    exp.argparse(hparams)
-    exp.save()
-
-    # build model
-    print('loading model...')
-    model = TRAINING_MODEL(hparams)
-    print('model built')
-
-    # callbacks
-    early_stop = EarlyStopping(
-        monitor=hparams.early_stop_metric,
-        patience=hparams.early_stop_patience,
-        verbose=True,
-        mode=hparams.early_stop_mode
-    )
-
-    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
-    checkpoint = ModelCheckpoint(
-        filepath=model_save_path,
-        save_function=None,
-        save_best_only=True,
-        verbose=True,
-        monitor=hparams.model_save_monitor_value,
-        mode=hparams.model_save_monitor_mode
-    )
-
-    # configure trainer
-    trainer = Trainer(
-        experiment=exp,
-        on_gpu=on_gpu,
-        cluster=cluster,
-        progress_bar=hparams.enable_tqdm,
-        overfit_pct=hparams.overfit,
-        track_grad_norm=hparams.track_grad_norm,
-        fast_dev_run=hparams.fast_dev_run,
-        check_val_every_n_epoch=hparams.check_val_every_n_epoch,
-        accumulate_grad_batches=hparams.accumulate_grad_batches,
-        process_position=process_position,
-        current_gpu_name=current_gpu,
-        checkpoint_callback=checkpoint,
-        early_stop_callback=early_stop,
-        enable_early_stop=hparams.enable_early_stop,
-        max_nb_epochs=hparams.max_nb_epochs,
-        min_nb_epochs=hparams.min_nb_epochs,
-        train_percent_check=hparams.train_percent_check,
-        val_percent_check=hparams.val_percent_check,
-        test_percent_check=hparams.test_percent_check,
-        val_check_interval=hparams.val_check_interval,
-        log_save_interval=hparams.log_save_interval,
-        add_log_row_interval=hparams.add_log_row_interval,
-        lr_scheduler_milestones=hparams.lr_scheduler_milestones
-    )
-
-    # train model
-    trainer.fit(model)
-
-
-def get_default_parser(strategy, root_dir):
-
-    possible_model_names = list(AVAILABLE_MODELS.keys())
-    parser = HyperOptArgumentParser(strategy=strategy, add_help=False)
-    add_default_args(parser, root_dir, possible_model_names, SEED)
-    return parser
-
-
-def get_model_name(args):
-    for i, arg in enumerate(args):
-        if 'model_name' in arg:
-            return args[i+1]
-
-
-def optimize_on_cluster(hyperparams):
-    # enable cluster training
-    cluster = SlurmCluster(
-        hyperparam_optimizer=hyperparams,
-        log_path=hyperparams.tt_save_path,
-        test_tube_exp_name=hyperparams.tt_name
-    )
-
-    # email for cluster coms
-    cluster.notify_job_status(email='add_email_here', on_done=True, on_fail=True)
-
-    # configure cluster
-    cluster.per_experiment_nb_gpus = hyperparams.per_experiment_nb_gpus
-    cluster.job_time = '48:00:00'
-    cluster.gpu_type = '1080ti'
-    cluster.memory_mb_per_node = 48000
-
-    # any modules for code to run in env
-    cluster.add_command('source activate pytorch_lightning')
-
-    # name of exp
-    job_display_name = hyperparams.tt_name.split('_')[0]
-    job_display_name = job_display_name[0:3]
-
-    # run hopt
-    print('submitting jobs...')
-    cluster.optimize_parallel_cluster_gpu(
-        main,
-        nb_trials=hyperparams.nb_hopt_trials,
-        job_name=job_display_name
-    )
-
-
-if __name__ == '__main__':
-
-    model_name = get_model_name(sys.argv)
-
-    # use default args
-    root_dir = os.path.split(os.path.dirname(sys.modules['__main__'].__file__))[0]
-    parent_parser = get_default_parser(strategy='random_search', root_dir=root_dir)
-
-    # allow model to overwrite or extend args
-    TRAINING_MODEL = AVAILABLE_MODELS[model_name]
-    parser = TRAINING_MODEL.add_model_specific_args(parent_parser)
-    parser.json_config('-c', '--config', default=root_dir + '/run_configs/local.json')
-    hyperparams = parser.parse_args()
-
-    # format GPU layout
-    os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
-    gpu_ids = hyperparams.gpus.split(';')
-
-    # RUN TRAINING
-    if hyperparams.on_cluster:
-        print('RUNNING ON SLURM CLUSTER')
-        os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(gpu_ids)
-        optimize_on_cluster(hyperparams)
-
-    elif hyperparams.single_run_gpu:
-        print(f'RUNNING 1 TRIAL ON GPU. gpu: {gpu_ids[0]}')
-        os.environ["CUDA_VISIBLE_DEVICES"] = gpu_ids[0]
-        main(hyperparams, None, None)
-
-    elif hyperparams.local or hyperparams.single_run:
-        os.environ["CUDA_VISIBLE_DEVICES"] = '0'
-        print('RUNNING LOCALLY')
-        main(hyperparams, None, None)
-
-    else:
-        print(f'RUNNING MULTI GPU. GPU ids: {gpu_ids}')
-        hyperparams.optimize_parallel_gpu(
-            main_local,
-            gpu_ids=gpu_ids,
-            nb_trials=hyperparams.nb_hopt_trials,
-            nb_workers=len(gpu_ids)
-        )

pytorch_lightning/utils/arg_parse.py

@@ -39,8 +39,8 @@ def add_default_args(parser, root_dir, rand_seed=None, possible_model_names=None
         parser.add_argument('--model_name', default='', help=','.join(possible_model_names))
 
     # test_tube settings
-    parser.add_argument('-en', '--tt_name', default='r_lib_')
-    parser.add_argument('-td', '--tt_description', default='test research  lib')
+    parser.add_argument('-en', '--tt_name', default='pt_test')
+    parser.add_argument('-td', '--tt_description', default='pytorch lightning test')
     parser.add_argument('--tt_save_path', default=root_dir + '/test_tube_logs', help='logging dir')
     parser.add_argument('--enable_single_run', dest='single_run', action='store_true')
     parser.add_argument('--nb_hopt_trials', default=1, type=int)

pytorch_lightning/utils/debugging.py

@@ -0,0 +1,5 @@
+import pdb
+import sys
+
+class MisconfigurationException(Exception):
+    pass

pytorch_lightning/utils/embeddings.py

@@ -1,104 +0,0 @@
-import torch
-import numpy as np
-from copy import deepcopy
-
-
-class PretrainedEmbedding(torch.nn.Embedding):
-
-    def __init__(self, embedding_path, embedding_dim, task_vocab, freeze=True, *args, **kwargs):
-        """
-        Loads a prebuilt pytorch embedding from any embedding formated file.
-        Padding=0 by default.
-
-        >>> emb = PretrainedEmbedding(embedding_path='glove.840B.300d.txt',embedding_dim=300, task_vocab={'hello': 1, 'world': 2})
-        >>> data = torch.Tensor([[0, 1], [0, 2]]).long()
-        >>> embedded = emb(data)
-        
-
-
-        :param embedding_path:
-        :param emb_dim:
-        :param task_vocab:
-        :param freeze:
-        :return:
-        """
-        # count the vocab
-        self.vocab_size = max(task_vocab.values()) + 1
-        super(PretrainedEmbedding, self).__init__(self.vocab_size, embedding_dim, padding_idx=0, *args, **kwargs)
-
-        # load pretrained embeddings
-        new_emb = self.__load_task_specific_embeddings(deepcopy(task_vocab), embedding_path, embedding_dim, freeze)
-
-        # transfer weights
-        self.weight = new_emb.weight
-
-        # apply freeze
-        should_freeze = not freeze
-        self.weight.requires_grad = should_freeze
-
-    def __load_task_specific_embeddings(self, vocab_words, embedding_path, emb_dim, freeze):
-        """
-        Iterates embedding file to only pull out task specific embeddings
-        :param vocab_words:
-        :param embedding_path:
-        :param emb_dim:
-        :param freeze:
-        :return:
-        """
-
-        # holds final embeddings for relevant words
-        embeddings = np.zeros(shape=(self.vocab_size, emb_dim))
-
-        # load embedding line by line and extract relevant embeddings
-        with open(embedding_path, encoding='utf-8') as f:
-            for line in f:
-                tokens = line.split(' ')
-                word = tokens[0]
-                embedding = tokens[1:]
-                embedding[-1] = embedding[-1][:-1]  # remove last new line
-
-                if word in vocab_words:
-                    vocab_word_i = vocab_words[word]
-
-                    # skip words that try to overwrite pad idx
-                    if vocab_word_i == 0:
-                        del vocab_words[word]
-                        continue
-
-                    emb_vals = np.asarray([float(x) for x in embedding])
-                    embeddings[vocab_word_i] = emb_vals
-
-                    # remove vocab word to early terminate
-                    del vocab_words[word]
-
-                # early break
-                if len(vocab_words) == 0:
-                    break
-
-        # add random vectors for the non-pretrained words
-        # these are vocab words NOT found in the pretrained embeddings
-        for w, i in vocab_words.items():
-            # skip words that try to overwrite pad idx
-            if i == 0:
-                continue
-
-            embedding = np.random.normal(size=emb_dim)
-            embeddings[i] = embedding
-
-        # turn into pt embedding
-        embeddings = torch.FloatTensor(embeddings)
-        embeddings = torch.nn.Embedding.from_pretrained(embeddings, freeze=freeze)
-
-        return embeddings
-
-
-if __name__ == '__main__':
-    emb = PretrainedEmbedding(
-        embedding_path='/Users/waf/Developer',
-        embedding_dim=300,
-        task_vocab={'hello': 1, 'world': 2}
-    )
-
-    data = torch.Tensor([[0, 1], [0, 2]]).long()
-    embedded = emb(data)
-    print(embedded)

pytorch_lightning/utils/plotting.py

@@ -1,28 +0,0 @@
-from matplotlib import pyplot as plt
-import numpy as np
-np.seterr(divide='ignore', invalid='ignore')
-
-
-def plot_confusion_matrix(cm,
-                          save_path,
-                          normalize=False,
-                          title='Confusion matrix',
-                          ylabel='y',
-                          xlabel='x'):
-    """
-    This function prints and plots the confusion matrix.
-    Normalization can be applied by setting `normalize=True`.
-    """
-    if normalize:
-        cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]
-        print("Normalized confusion matrix")
-    else:
-        print('Confusion matrix, without normalization')
-
-    fig = plt.figure()
-    plt.matshow(cm)
-    plt.title(title)
-    plt.colorbar()
-    plt.ylabel(ylabel)
-    plt.xlabel(xlabel)
-    plt.savefig(save_path)

requirements.txt

@@ -1,27 +1,9 @@
-atomicwrites==1.2.1
-attrs==18.2.0
-certifi==2018.11.29
-cffi==1.11.5
-h5py==2.9.0
-imageio==2.4.1
-mkl-fft==1.0.6
-mkl-random==1.0.2
-more-itertools==5.0.0
-numpy==1.15.4
-olefile==0.46
-pandas==0.23.4
-Pillow==5.3.0
-pluggy==0.8.0
-py==1.7.0
-pycparser==2.19
-pytest==4.0.2
-python-dateutil==2.7.5
-pytz==2018.7
+coverage==4.5.3
+mkdocs==1.0.4
+pytest==5.0.1
 scikit-learn==0.20.2
-scipy==1.2.0
-six==1.12.0
-sklearn==0.0
-test-tube==0.6282
-torch==1.0.0
-torchvision==0.2.1
-tqdm==4.28.1
+tqdm==4.32.1
+twine==1.13.0
+numpy==1.16.4
+torch>=1.1.0
+torchvision==0.3.0

setup.cfg

@@ -16,6 +16,33 @@ markers =
 ignore = E731,W504
 max-line-length = 120
 
+[coverage:report]
+exclude_lines =
+    pragma: no cover
+    def __repr__
+    if self.debug:
+    if settings.DEBUG
+    raise AssertionError
+    raise NotImplementedError
+    if 0:
+    if __name__ == .__main__.:
+    except Exception as e
+    print(e)
+    print(traceback.print_exc())
+    return *
+    raise Exception
+    warnings
+    print
+    raise RuntimeError
+    break
+    pass
+    os.makedirs
+
+omit =
+    pytorch_lightning/callbacks/pt_callbacks.py
+    tests/test_models.py
+    pytorch_lightning/testing_models/lm_test_module.py
+
 [flake8]
 ignore = E731,W504,F401,F841
 max-line-length = 120

setup.py

@@ -7,7 +7,7 @@ from setuptools import setup, find_packages
 # http://blog.ionelmc.ro/2014/05/25/python-packaging/
 setup(
     name="pytorch-lightning",
-    version='0.111',
+    version='0.3.6.6',
     description="The Keras for ML researchers using PyTorch",
     author="William Falcon",
     author_email="waf2107@columbia.edu",
@@ -17,9 +17,9 @@ setup(
     keywords=["deep learning", "pytorch", "AI"],
     python_requires=">=3.5",
     install_requires=[
-        "torch>=1.0.0",
+        "torch>=1.1.0",
         "tqdm",
-        "test-tube",
+        "test-tube>=0.6.7.4",
     ],
     packages=find_packages(),
     long_description=open("README.md", encoding="utf-8").read(),

tests/README.md

@@ -0,0 +1,58 @@
+# Pytorch-Lightning Tests
+
+## Running tests
+The automatic travis tests ONLY run CPU-based tests. Although these cover most of the use cases,
+run on a 2-GPU machine to validate the full test-suite.
+
+
+To run all tests do the following:
+```bash
+git clone https://github.com/williamFalcon/pytorch-lightning
+cd pytorch-lightning
+
+# install module locally
+pip install -e .
+
+# install dev deps
+pip install -r requirements.txt
+
+# run tests
+py.test
+```
+
+To test models that require GPU make sure to run the above command on a GPU machine.
+The GPU machine must have:
+1. At least 2 GPUs.
+2. [NVIDIA-apex](https://github.com/NVIDIA/apex#linux) installed.
+
+
+### test_models.py
+This file fits a tiny model on MNIST using these different set-ups.
+1. CPU only.
+2. Single GPU with DP.
+3. Multiple (2) GPUs using DP.
+3. Multiple (2) GPUs using DDP.
+3. Multiple (2) GPUs using DP + apex (for 16-bit precision).
+3. Multiple (2) GPUs using DDP + apex (for 16-bit precision).   
+
+For each set up it also tests:
+1. model saving.
+2. model loading.
+3. predicting with a loaded model.
+4. simulated save from HPC signal.
+5. simulated load from HPC signal.
+
+## Running Coverage   
+
+```bash
+cd pytorch-lightning
+
+# generate coverage 
+pip install coverage
+coverage run tests/test_models.py   
+
+# print coverage stats
+coverage report -m   
+```
+
+

tests/debug.py

@@ -0,0 +1,180 @@
+import pytest
+from pytorch_lightning import Trainer
+from pytorch_lightning.examples.new_project_templates.lightning_module_template import LightningTemplateModel
+from argparse import Namespace
+from test_tube import Experiment
+from pytorch_lightning.callbacks import ModelCheckpoint
+import numpy as np
+import warnings
+import torch
+import os
+import shutil
+import pdb
+
+import pytorch_lightning as ptl
+import torch
+from torch.nn import functional as F
+from torch.utils.data import DataLoader
+from torchvision.datasets import MNIST
+
+
+class CoolModel(ptl.LightningModule):
+
+    def __init(self):
+        super(CoolModel, self).__init__()
+        # not the best model...
+        self.l1 = torch.nn.Linear(28 * 28, 10)
+
+    def forward(self, x):
+        return torch.relu(self.l1(x))
+
+    def my_loss(self, y_hat, y):
+        return F.cross_entropy(y_hat, y)
+
+    def training_step(self, batch, batch_nb):
+        x, y = batch
+        y_hat = self.forward(x)
+        return {'tng_loss': self.my_loss(y_hat, y)}
+
+    def validation_step(self, batch, batch_nb):
+        x, y = batch
+        y_hat = self.forward(x)
+        return {'val_loss': self.my_loss(y_hat, y)}
+
+    def validation_end(self, outputs):
+        avg_loss = torch.stack([x for x in outputs['val_loss']]).mean()
+        return avg_loss
+
+    def configure_optimizers(self):
+        return [torch.optim.Adam(self.parameters(), lr=0.02)]
+
+    @ptl.data_loader
+    def tng_dataloader(self):
+        return DataLoader(MNIST('path/to/save', train=True), batch_size=32)
+
+    @ptl.data_loader
+    def val_dataloader(self):
+        return DataLoader(MNIST('path/to/save', train=False), batch_size=32)
+
+    @ptl.data_loader
+    def test_dataloader(self):
+        return DataLoader(MNIST('path/to/save', train=False), batch_size=32)
+
+
+def get_model():
+    # set up model with these hyperparams
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    hparams = Namespace(**{'drop_prob': 0.2,
+                           'batch_size': 32,
+                           'in_features': 28*28,
+                           'learning_rate': 0.001*8,
+                           'optimizer_name': 'adam',
+                           'data_root': os.path.join(root_dir, 'mnist'),
+                           'out_features': 10,
+                           'hidden_dim': 1000})
+    model = LightningTemplateModel(hparams)
+
+    return model, hparams
+
+
+def get_exp(debug=True):
+    # set up exp object without actually saving logs
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    exp = Experiment(debug=debug, save_dir=root_dir, name='tests_tt_dir')
+    return exp
+
+
+def init_save_dir():
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    save_dir = os.path.join(root_dir, 'save_dir')
+
+    if os.path.exists(save_dir):
+        shutil.rmtree(save_dir)
+
+    os.makedirs(save_dir, exist_ok=True)
+
+    return save_dir
+
+
+def clear_save_dir():
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    save_dir = os.path.join(root_dir, 'save_dir')
+    if os.path.exists(save_dir):
+        shutil.rmtree(save_dir)
+
+
+def load_model(exp, save_dir):
+
+    # load trained model
+    tags_path = exp.get_data_path(exp.name, exp.version)
+    tags_path = os.path.join(tags_path, 'meta_tags.csv')
+
+    checkpoints = [x for x in os.listdir(save_dir) if '.ckpt' in x]
+    weights_dir = os.path.join(save_dir, checkpoints[0])
+
+    trained_model = LightningTemplateModel.load_from_metrics(weights_path=weights_dir, tags_csv=tags_path, on_gpu=True)
+
+    assert trained_model is not None, 'loading model failed'
+
+    return trained_model
+
+
+def run_prediction(dataloader, trained_model):
+    # run prediction on 1 batch
+    for batch in dataloader:
+        break
+
+    x, y = batch
+    x = x.view(x.size(0), -1)
+
+    y_hat = trained_model(x)
+
+    # acc
+    labels_hat = torch.argmax(y_hat, dim=1)
+    val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)
+    val_acc = torch.tensor(val_acc)
+    val_acc = val_acc.item()
+
+    print(val_acc)
+
+    assert val_acc > 0.70, f'this model is expected to get > 0.7 in test set (it got {val_acc})'
+
+
+def main():
+
+    save_dir = init_save_dir()
+
+    # exp file to get meta
+    exp = get_exp(False)
+    exp.save()
+
+    # exp file to get weights
+    checkpoint = ModelCheckpoint(save_dir)
+
+    trainer = Trainer(
+        experiment=exp,
+        checkpoint_callback=checkpoint,
+        progress_bar=True,
+        max_nb_epochs=1,
+        gpus=[0, 1],
+        distributed_backend='dp',
+    )
+
+    model = CoolModel()
+
+    result = trainer.fit(model)
+
+    # correct result and ok accuracy
+    assert result == 1, 'amp + ddp model failed to complete'
+
+    # test model loading
+    pretrained_model = load_model(exp, save_dir)
+
+    # test model preds
+    run_prediction(model.test_dataloader, pretrained_model)
+
+    clear_save_dir()
+
+
+if __name__ == '__main__':
+    main()

tests/test_models.py

@@ -0,0 +1,688 @@
+import pytest
+from pytorch_lightning import Trainer
+from pytorch_lightning.examples.new_project_templates.lightning_module_template import LightningTemplateModel
+from pytorch_lightning.testing_models.lm_test_module import LightningTestModel
+from argparse import Namespace
+from test_tube import Experiment, SlurmCluster
+from pytorch_lightning.callbacks import ModelCheckpoint, EarlyStopping
+from pytorch_lightning.utils.debugging import MisconfigurationException
+from pytorch_lightning.root_module import memory
+from pytorch_lightning.models.trainer import reduce_distributed_output
+from pytorch_lightning.root_module import model_saving
+import numpy as np
+import warnings
+import torch
+import os
+import shutil
+import pdb
+
+SEED = 2334
+torch.manual_seed(SEED)
+np.random.seed(SEED)
+
+
+# ------------------------------------------------------------------------
+# TESTS
+# ------------------------------------------------------------------------
+def test_amp_gpu_ddp():
+    """
+    Make sure DDP + AMP work
+    :return:
+    """
+    if not torch.cuda.is_available():
+        warnings.warn('test_amp_gpu_ddp cannot run. Rerun on a GPU node to run this test')
+        return
+    if not torch.cuda.device_count() > 1:
+        warnings.warn('test_amp_gpu_ddp cannot run. Rerun on a node with 2+ GPUs to run this test')
+        return
+
+    os.environ['MASTER_PORT'] = str(np.random.randint(12000, 19000, 1)[0])
+
+    hparams = get_hparams()
+    model = LightningTestModel(hparams)
+
+    trainer_options = dict(
+        progress_bar=True,
+        max_nb_epochs=1,
+        gpus=[0, 1],
+        distributed_backend='ddp',
+        use_amp=True
+    )
+
+    run_gpu_model_test(trainer_options, model, hparams)
+
+
+def test_cpu_slurm_save_load():
+    """
+    Verify model save/load/checkpoint on CPU
+    :return:
+    """
+    hparams = get_hparams()
+    model = LightningTestModel(hparams)
+
+    save_dir = init_save_dir()
+
+    # exp file to get meta
+    exp = get_exp(False)
+    exp.argparse(hparams)
+    exp.save()
+
+    cluster_a = SlurmCluster()
+    trainer_options = dict(
+        max_nb_epochs=1,
+        cluster=cluster_a,
+        experiment=exp,
+        checkpoint_callback=ModelCheckpoint(save_dir)
+    )
+
+    # fit model
+    trainer = Trainer(**trainer_options)
+    result = trainer.fit(model)
+    real_global_step = trainer.global_step
+
+    # traning complete
+    assert result == 1, 'amp + ddp model failed to complete'
+
+    # predict with trained model before saving
+    # make a prediction
+    for batch in model.test_dataloader:
+        break
+
+    x, y = batch
+    x = x.view(x.size(0), -1)
+
+    model.eval()
+    pred_before_saving = model(x)
+
+    # test registering a save function
+    trainer.enable_auto_hpc_walltime_manager()
+
+    # test HPC saving
+    # simulate snapshot on slurm
+    saved_filepath = trainer.hpc_save(save_dir, exp)
+    assert os.path.exists(saved_filepath)
+
+    # wipe-out trainer and model
+    # retrain with not much data... this simulates picking training back up after slurm
+    # we want to see if the weights come back correctly
+    continue_tng_hparams = get_hparams(continue_training=True, hpc_exp_number=cluster_a.hpc_exp_number)
+    trainer_options = dict(
+        max_nb_epochs=1,
+        cluster=SlurmCluster(continue_tng_hparams),
+        experiment=exp,
+        checkpoint_callback=ModelCheckpoint(save_dir),
+    )
+    trainer = Trainer(**trainer_options)
+    model = LightningTestModel(hparams)
+
+    # set the epoch start hook so we can predict before the model does the full training
+    def assert_pred_same():
+        assert trainer.global_step == real_global_step and trainer.global_step > 0
+
+        # predict with loaded model to make sure answers are the same
+        trainer.model.eval()
+        new_pred = trainer.model(x)
+        assert torch.all(torch.eq(pred_before_saving, new_pred)).item() == 1
+
+    model.on_epoch_start = assert_pred_same
+
+    # by calling fit again, we trigger training, loading weights from the cluster
+    # and our hook to predict using current model before any more weight updates
+    trainer.fit(model)
+
+    clear_save_dir()
+
+
+def test_loading_meta_tags():
+    hparams = get_hparams()
+
+    save_dir = init_save_dir()
+
+    # save tags
+    exp = get_exp(False)
+    exp.tag({'some_str':'a_str', 'an_int': 1, 'a_float': 2.0})
+    exp.argparse(hparams)
+    exp.save()
+
+    # load tags
+    tags_path = exp.get_data_path(exp.name, exp.version) + '/meta_tags.csv'
+    tags = model_saving.load_hparams_from_tags_csv(tags_path)
+
+    assert tags.batch_size == 32 and tags.hidden_dim == 1000
+
+    clear_save_dir()
+
+
+def test_dp_output_reduce():
+
+    # test identity when we have a single gpu
+    out = torch.rand(3, 1)
+    assert reduce_distributed_output(out, nb_gpus=1) is out
+
+    # average when we have multiples
+    assert reduce_distributed_output(out, nb_gpus=2) == out.mean()
+
+    # when we have a dict of vals
+    out = {
+        'a': out,
+        'b': {
+            'c': out
+        }
+    }
+    reduced = reduce_distributed_output(out, nb_gpus=3)
+    assert reduced['a'] == out['a']
+    assert reduced['b']['c'] == out['b']['c']
+
+
+def test_model_saving_loading():
+    """
+    Tests use case where trainer saves the model, and user loads it from tags independently
+    :return:
+    """
+    hparams = get_hparams()
+    model = LightningTestModel(hparams)
+
+    save_dir = init_save_dir()
+
+    # exp file to get meta
+    exp = get_exp(False)
+    exp.argparse(hparams)
+    exp.save()
+
+    trainer_options = dict(
+        max_nb_epochs=1,
+        cluster=SlurmCluster(),
+        experiment=exp,
+        checkpoint_callback=ModelCheckpoint(save_dir)
+    )
+
+    # fit model
+    trainer = Trainer(**trainer_options)
+    result = trainer.fit(model)
+
+    # traning complete
+    assert result == 1, 'amp + ddp model failed to complete'
+
+    # make a prediction
+    for batch in model.test_dataloader:
+        break
+
+    x, y = batch
+    x = x.view(x.size(0), -1)
+
+    # generate preds before saving model
+    model.eval()
+    pred_before_saving = model(x)
+
+    # save model
+    new_weights_path = os.path.join(save_dir, 'save_test.ckpt')
+    trainer.save_checkpoint(new_weights_path)
+
+    # load new model
+    tags_path = exp.get_data_path(exp.name, exp.version)
+    tags_path = os.path.join(tags_path, 'meta_tags.csv')
+    model_2 = LightningTestModel.load_from_metrics(weights_path=new_weights_path, tags_csv=tags_path, on_gpu=False)
+    model_2.eval()
+
+    # make prediction
+    # assert that both predictions are the same
+    new_pred = model_2(x)
+    assert torch.all(torch.eq(pred_before_saving, new_pred)).item() == 1
+
+    clear_save_dir()
+
+
+
+
+def test_model_freeze_unfreeze():
+    hparams = get_hparams()
+    model = LightningTestModel(hparams)
+
+    model.freeze()
+    model.unfreeze()
+
+
+def test_amp_gpu_ddp_slurm_managed():
+    """
+    Make sure DDP + AMP work
+    :return:
+    """
+    if not torch.cuda.is_available():
+        warnings.warn('test_amp_gpu_ddp cannot run. Rerun on a GPU node to run this test')
+        return
+    if not torch.cuda.device_count() > 1:
+        warnings.warn('test_amp_gpu_ddp cannot run. Rerun on a node with 2+ GPUs to run this test')
+        return
+
+    # simulate setting slurm flags
+    os.environ['MASTER_PORT'] = str(np.random.randint(12000, 19000, 1)[0])
+    os.environ['SLURM_LOCALID'] = str(0)
+
+    hparams = get_hparams()
+    model = LightningTestModel(hparams)
+
+    trainer_options = dict(
+        progress_bar=True,
+        max_nb_epochs=1,
+        gpus=[0],
+        distributed_backend='ddp',
+        use_amp=True
+    )
+
+    save_dir = init_save_dir()
+
+    # exp file to get meta
+    exp = get_exp(False)
+    exp.argparse(hparams)
+    exp.save()
+
+    # exp file to get weights
+    checkpoint = ModelCheckpoint(save_dir)
+
+    # add these to the trainer options
+    trainer_options['checkpoint_callback'] = checkpoint
+    trainer_options['experiment'] = exp
+
+    # fit model
+    trainer = Trainer(**trainer_options)
+    trainer.is_slurm_managing_tasks = True
+    result = trainer.fit(model)
+
+    # correct result and ok accuracy
+    assert result == 1, 'amp + ddp model failed to complete'
+
+    # test root model address
+    assert trainer.resolve_root_node_address('abc') == 'abc'
+    assert trainer.resolve_root_node_address('abc[23]') == 'abc23'
+    assert trainer.resolve_root_node_address('abc[23-24]') == 'abc23'
+    assert trainer.resolve_root_node_address('abc[23-24, 45-40, 40]') == 'abc23'
+
+    # test model loading with a map_location
+    map_location = 'cuda:1'
+    pretrained_model = load_model(exp, save_dir, True, map_location)
+
+    # test model preds
+    run_prediction(model.test_dataloader, pretrained_model)
+
+    if trainer.use_ddp:
+        # on hpc this would work fine... but need to hack it for the purpose of the test
+        trainer.model = pretrained_model
+        trainer.optimizers, trainer.lr_schedulers = pretrained_model.configure_optimizers()
+
+    # test HPC loading / saving
+    trainer.hpc_save(save_dir, exp)
+    trainer.hpc_load(save_dir, on_gpu=True)
+
+    # test freeze on gpu
+    model.freeze()
+    model.unfreeze()
+
+    clear_save_dir()
+
+
+def test_early_stopping_cpu_model():
+    """
+    Test each of the trainer options
+    :return:
+    """
+
+    stopping = EarlyStopping()
+    trainer_options = dict(
+        early_stop_callback=stopping,
+        gradient_clip=1.0,
+        overfit_pct=0.20,
+        track_grad_norm=2,
+        print_nan_grads=True,
+        progress_bar=False,
+        experiment=get_exp(),
+        train_percent_check=0.1,
+        val_percent_check=0.1
+    )
+
+    model, hparams = get_model()
+    run_gpu_model_test(trainer_options, model, hparams, on_gpu=False)
+
+    # test freeze on cpu
+    model.freeze()
+    model.unfreeze()
+
+
+def test_cpu_model_with_amp():
+    """
+    Make sure model trains on CPU
+    :return:
+    """
+
+    trainer_options = dict(
+        progress_bar=False,
+        experiment=get_exp(),
+        max_nb_epochs=1,
+        train_percent_check=0.4,
+        val_percent_check=0.4,
+        use_amp=True
+    )
+
+    model, hparams = get_model()
+
+    with pytest.raises((MisconfigurationException, ModuleNotFoundError)):
+        run_gpu_model_test(trainer_options, model, hparams, on_gpu=False)
+
+
+def test_cpu_model():
+    """
+    Make sure model trains on CPU
+    :return:
+    """
+
+    trainer_options = dict(
+        progress_bar=False,
+        experiment=get_exp(),
+        max_nb_epochs=1,
+        train_percent_check=0.4,
+        val_percent_check=0.4
+    )
+
+    model, hparams = get_model()
+
+    run_gpu_model_test(trainer_options, model, hparams, on_gpu=False)
+
+
+def test_all_features_cpu_model():
+    """
+    Test each of the trainer options
+    :return:
+    """
+
+    trainer_options = dict(
+        gradient_clip=1.0,
+        overfit_pct=0.20,
+        track_grad_norm=2,
+        print_nan_grads=True,
+        progress_bar=False,
+        experiment=get_exp(),
+        max_nb_epochs=1,
+        train_percent_check=0.4,
+        val_percent_check=0.4
+    )
+
+    model, hparams = get_model()
+    run_gpu_model_test(trainer_options, model, hparams, on_gpu=False)
+
+
+def test_single_gpu_model():
+    """
+    Make sure single GPU works (DP mode)
+    :return:
+    """
+    if not torch.cuda.is_available():
+        warnings.warn('test_single_gpu_model cannot run. Rerun on a GPU node to run this test')
+        return
+    model, hparams = get_model()
+
+    trainer_options = dict(
+        progress_bar=False,
+        max_nb_epochs=1,
+        train_percent_check=0.1,
+        val_percent_check=0.1,
+        gpus=[0]
+    )
+
+    run_gpu_model_test(trainer_options, model, hparams)
+
+
+def test_multi_gpu_model_dp():
+    """
+    Make sure DP works
+    :return:
+    """
+    if not torch.cuda.is_available():
+        warnings.warn('test_multi_gpu_model_dp cannot run. Rerun on a GPU node to run this test')
+        return
+    if not torch.cuda.device_count() > 1:
+        warnings.warn('test_multi_gpu_model_dp cannot run. Rerun on a node with 2+ GPUs to run this test')
+        return
+    model, hparams = get_model()
+    trainer_options = dict(
+        progress_bar=False,
+        max_nb_epochs=1,
+        train_percent_check=0.1,
+        val_percent_check=0.1,
+        gpus='-1'
+    )
+
+    run_gpu_model_test(trainer_options, model, hparams)
+
+    # test memory helper functions
+    memory.get_gpu_memory_map()
+
+
+def test_amp_gpu_dp():
+    """
+    Make sure DP + AMP work
+    :return:
+    """
+    if not torch.cuda.is_available():
+        warnings.warn('test_amp_gpu_dp cannot run. Rerun on a GPU node to run this test')
+        return
+    if not torch.cuda.device_count() > 1:
+        warnings.warn('test_amp_gpu_dp cannot run. Rerun on a node with 2+ GPUs to run this test')
+        return
+    model, hparams = get_model()
+    trainer_options = dict(
+        max_nb_epochs=1,
+        gpus='0, 1',  # test init with gpu string
+        distributed_backend='dp',
+        use_amp=True
+    )
+    with pytest.raises(MisconfigurationException):
+        run_gpu_model_test(trainer_options, model, hparams)
+
+
+def test_multi_gpu_model_ddp():
+    """
+    Make sure DDP works
+    :return:
+    """
+    if not torch.cuda.is_available():
+        warnings.warn('test_multi_gpu_model_ddp cannot run. Rerun on a GPU node to run this test')
+        return
+    if not torch.cuda.device_count() > 1:
+        warnings.warn('test_multi_gpu_model_ddp cannot run. Rerun on a node with 2+ GPUs to run this test')
+        return
+
+    os.environ['MASTER_PORT'] = str(np.random.randint(12000, 19000, 1)[0])
+    model, hparams = get_model()
+    trainer_options = dict(
+        progress_bar=False,
+        max_nb_epochs=1,
+        train_percent_check=0.4,
+        val_percent_check=0.2,
+        gpus=[0, 1],
+        distributed_backend='ddp'
+    )
+
+    run_gpu_model_test(trainer_options, model, hparams)
+
+
+
+def test_ddp_sampler_error():
+    """
+    Make sure DDP + AMP work
+    :return:
+    """
+    if not torch.cuda.is_available():
+        warnings.warn('test_amp_gpu_ddp cannot run. Rerun on a GPU node to run this test')
+        return
+    if not torch.cuda.device_count() > 1:
+        warnings.warn('test_amp_gpu_ddp cannot run. Rerun on a node with 2+ GPUs to run this test')
+        return
+
+    os.environ['MASTER_PORT'] = str(np.random.randint(12000, 19000, 1)[0])
+
+    hparams = get_hparams()
+    model = LightningTestModel(hparams, force_remove_distributed_sampler=True)
+
+    exp = get_exp(True)
+    exp.save()
+
+    trainer = Trainer(
+        experiment=exp,
+        progress_bar=False,
+        max_nb_epochs=1,
+        gpus=[0, 1],
+        distributed_backend='ddp',
+        use_amp=True
+    )
+
+    with pytest.raises(MisconfigurationException):
+        trainer.get_dataloaders(model)
+
+    clear_save_dir()
+
+
+# ------------------------------------------------------------------------
+# UTILS
+# ------------------------------------------------------------------------
+def run_gpu_model_test(trainer_options, model, hparams, on_gpu=True):
+    save_dir = init_save_dir()
+
+    # exp file to get meta
+    exp = get_exp(False)
+    exp.argparse(hparams)
+    exp.save()
+
+    # exp file to get weights
+    checkpoint = ModelCheckpoint(save_dir)
+
+    # add these to the trainer options
+    trainer_options['checkpoint_callback'] = checkpoint
+    trainer_options['experiment'] = exp
+
+    # fit model
+    trainer = Trainer(**trainer_options)
+    result = trainer.fit(model)
+
+    # correct result and ok accuracy
+    assert result == 1, 'amp + ddp model failed to complete'
+
+    # test model loading
+    pretrained_model = load_model(exp, save_dir, on_gpu)
+
+    # test model preds
+    run_prediction(model.test_dataloader, pretrained_model)
+
+    if trainer.use_ddp:
+        # on hpc this would work fine... but need to hack it for the purpose of the test
+        trainer.model = pretrained_model
+        trainer.optimizers, trainer.lr_schedulers = pretrained_model.configure_optimizers()
+
+    # test HPC loading / saving
+    trainer.hpc_save(save_dir, exp)
+    trainer.hpc_load(save_dir, on_gpu=on_gpu)
+
+    clear_save_dir()
+
+
+def get_hparams(continue_training=False, hpc_exp_number=0):
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+
+    args = {
+        'drop_prob': 0.2,
+        'batch_size': 32,
+        'in_features': 28*28,
+        'learning_rate': 0.001*8,
+        'optimizer_name': 'adam',
+        'data_root': os.path.join(root_dir, 'mnist'),
+        'out_features': 10,
+        'hidden_dim': 1000}
+
+    if continue_training:
+        args['test_tube_do_checkpoint_load'] = True
+        args['hpc_exp_number'] = hpc_exp_number
+
+    hparams = Namespace(**args)
+    return hparams
+
+
+def get_model():
+    # set up model with these hyperparams
+    hparams = get_hparams()
+    model = LightningTemplateModel(hparams)
+
+    return model, hparams
+
+
+def get_exp(debug=True):
+    # set up exp object without actually saving logs
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    exp = Experiment(debug=debug, save_dir=root_dir, name='tests_tt_dir')
+    return exp
+
+
+def init_save_dir():
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    save_dir = os.path.join(root_dir, 'save_dir')
+
+    if os.path.exists(save_dir):
+        shutil.rmtree(save_dir)
+
+    os.makedirs(save_dir, exist_ok=True)
+
+    return save_dir
+
+
+def clear_save_dir():
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    save_dir = os.path.join(root_dir, 'save_dir')
+    if os.path.exists(save_dir):
+        shutil.rmtree(save_dir)
+
+
+def load_model(exp, save_dir, on_gpu, map_location=None):
+
+    # load trained model
+    tags_path = exp.get_data_path(exp.name, exp.version)
+    tags_path = os.path.join(tags_path, 'meta_tags.csv')
+
+    checkpoints = [x for x in os.listdir(save_dir) if '.ckpt' in x]
+    weights_dir = os.path.join(save_dir, checkpoints[0])
+
+    trained_model = LightningTemplateModel.load_from_metrics(weights_path=weights_dir,
+                                                             tags_csv=tags_path,
+                                                             on_gpu=on_gpu,
+                                                             map_location=map_location)
+
+    assert trained_model is not None, 'loading model failed'
+
+    return trained_model
+
+
+def run_prediction(dataloader, trained_model):
+    # run prediction on 1 batch
+    for batch in dataloader:
+        break
+
+    x, y = batch
+    x = x.view(x.size(0), -1)
+
+    y_hat = trained_model(x)
+
+    # acc
+    labels_hat = torch.argmax(y_hat, dim=1)
+    val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)
+    val_acc = torch.tensor(val_acc)
+    val_acc = val_acc.item()
+
+    print(val_acc)
+
+    assert val_acc > 0.50, f'this model is expected to get > 0.50 in test set (it got {val_acc})'
+
+
+def assert_ok_acc(trainer):
+    # this model should get 0.80+ acc
+    acc = trainer.tng_tqdm_dic['val_acc']
+    assert acc > 0.50, f'model failed to get expected 0.50 validation accuracy. Got: {acc}'
+
+
+if __name__ == '__main__':
+    pytest.main([__file__])