release v0.4.9

Metrics load (#228 )
* load from metrics defaults to CPU * load from metrics defaults to CPU * load from metrics defaults to CPU
2026-09-12 12:40:20 +08:00 · 2019-09-16 10:50:59 -04:00 · 2019-09-16 10:47:19 -04:00 · 2019-09-16 10:21:00 -04:00 · 2019-09-14 09:55:42 -04:00 · 2019-09-14 02:23:49 -04:00
61 changed files with 4650 additions and 1898 deletions
@@ -1,4 +1,7 @@
-#see https://github.com/codecov/support/wiki/Codecov-Yaml
+# see https://docs.codecov.io/docs/codecov-yaml
+# Validation check:
+# $ curl --data-binary @.codecov.yml https://codecov.io/validate
+
 codecov:
  notify:
    require_ci_to_pass: yes
@@ -41,4 +44,3 @@ comment:
  behavior: default  # update if exists else create new
  # branches: *

-
@@ -0,0 +1,76 @@
+# Contributor Covenant Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to making participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+ advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+ address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+ professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies both within project spaces and in public spaces
+when an individual is representing the project or its community. Examples of
+representing a project or community include using an official project e-mail
+address, posting via an official social media account, or acting as an appointed
+representative at an online or offline event. Representation of a project may be
+further defined and clarified by project maintainers.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at waf2107@columbia.edu. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq
@@ -1,5 +1,10 @@
 # Contributing    
-Welcome to the PyTorch Lightning community! We're building the most advanced research platform on the planet to implement the latest, best practices that the amazing PyTorch team rolls out!  
+Welcome to the PyTorch Lightning community! We're building the most advanced research platform on the planet to implement the latest, best practices that the amazing PyTorch team rolls out!   
+
+## One less thing to remember    
+Simplify the API as much as possible from the user perspective. Any additions or improvements should minimize things the user needs to remember.   
+
+For example: One benefit of the validation_step is that the user doesn't have to remember to set the model to .eval(). This avoids all sorts of subtle errors the user could make.   

 ## Lightning Design Principles   
 We encourage all sorts of contributions you're interested in adding! When coding for lightning, please follow these principles.    
@@ -0,0 +1,26 @@
+---
+name: How to question
+about: Asking how-to questions
+title: ''
+labels: question
+assignees: ''
+
+---
+
+### Before asking:   
+1. search the issues.   
+2. search the docs.    
+
+If you still can't find what you need:     
+#### What is your question?    
+
+#### Code    
+Please paste a code snippet if your question requires it!   
+
+#### What have you tried?    
+
+#### What's your environment?   
+- conda version (no venv)   
+- PyTorch version   
+- Lightning version   
+- Test-tube version
@@ -0,0 +1,16 @@
+# Before submitting
+
+- Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
+- Did you read the [contributor guideline](https://github.com/williamFalcon/pytorch-lightning/blob/master/.github/CONTRIBUTING.md)?
+- Did you make sure to update the docs?   
+- Did you write any new necessary tests?  
+
+## What does this PR do?
+Fixes # (issue).
+
+## PR review    
+Anyone in the community is free to review the PR once the tests have passed.     
+If we didn't discuss your PR in Github issues there's a high chance it will not be merged.   
+
+## Did you have fun?
+Make sure you had fun coding 🙃
@@ -11,6 +11,7 @@ pip-wheel-metadata/
 test_tube_exp/
 tests/tests_tt_dir/
 tests/save_dir
+default/

 # Byte-compiled / optimized / DLL files
 __pycache__/
@@ -8,8 +8,6 @@
 # this file is *not* meant to cover or endorse the use of travis, but rather to
 # help confirm pull requests to this project.

-dist: xenial  # Ubuntu 16.04
-
 env:
  global:
    - DISPLAY=""
@@ -18,10 +16,36 @@ language: python

 matrix:
  include:
-    - python: 3.6
+    - os: linux
+      dist: xenial  # Ubuntu 16.04
+      python: 3.6
      env: TOXENV=py36
-    - python: 3.7
+    - os: linux
+      dist: bionic  # Ubuntu 18.04
+      python: 3.6
+      env: TOXENV=py36
+    - os: linux
+      dist: bionic  # Ubuntu 18.04
+      python: 3.7
      env: TOXENV=py37
+    - os: osx
+      osx_image: xcode9.4
+      language: generic
+      env: TOXENV=py36
+      addons:
+        homebrew:
+          # update: true
+          packages: python3
+      before_install:
+        - pip3 install virtualenv
+        - virtualenv -p python3 ~/venv
+        - source ~/venv/bin/activate
+#    - os: windows
+#      language: minimal
+#      before_install:
+#        - choco install python3
+#        - export PATH="/c/Python37:/c/Python37/Scripts:$PATH"
+#      env: TOXENV=py37

 # See http://docs.travis-ci.com/user/caching/#pip-cache
 cache: pip
@@ -13,11 +13,16 @@ include LICENSE
 exclude *.sh
 exclude *.toml
 exclude *.svg    
-recursive-include examples *.py
 recursive-include pytorch_lightning *.py

+# include examples
+recursive-include examples *.py
+recursive-include examples *.md
+recursive-include examples *.sh
+
 # exclude tests from package
 recursive-exclude tests *
+recursive-exclude site *
 exclude tests

 # Exclude the documentation files
@@ -4,23 +4,23 @@

 # PyTorch Lightning

-**The PyTorch Keras for ML researchers. More control. Less boilerplate.**
+**The lightweight PyTorch wrapper for ML researchers. Scale your models. Write less boilerplate.**


 [![PyPI Status](https://badge.fury.io/py/pytorch-lightning.svg)](https://badge.fury.io/py/pytorch-lightning)
 [![PyPI Status](https://pepy.tech/badge/pytorch-lightning)](https://pepy.tech/project/pytorch-lightning)
 [![Build Status](https://travis-ci.org/williamFalcon/pytorch-lightning.svg?branch=master)](https://travis-ci.org/williamFalcon/pytorch-lightning)
-<!-- 
-removed until windows install issues resolved.
-[![Build status](https://ci.appveyor.com/api/projects/status/rum89d7hq8l1kfye?svg=true)](https://ci.appveyor.com/project/Borda/pytorch-lightning) -->
+[![Build status](https://ci.appveyor.com/api/projects/status/rum89d7hq8l1kfye?svg=true)](https://ci.appveyor.com/project/Borda/pytorch-lightning)
+[![Coverage](https://github.com/williamFalcon/pytorch-lightning/blob/master/docs/source/_static/coverage.svg)](https://github.com/williamFalcon/pytorch-lightning/tree/master/tests#running-coverage)
+[![CodeFactor](https://www.codefactor.io/repository/github/borda/pytorch-lightning/badge)](https://www.codefactor.io/repository/github/borda/pytorch-lightning)    
+
+[![ReadTheDocs](https://readthedocs.org/projects/pytorch-lightning/badge/?version=latest)](https://pytorch-lightning.readthedocs.io/en/latest)
+[![Gitter](https://badges.gitter.im/PyTorch-Lightning/community.svg)](https://gitter.im/PyTorch-Lightning/community?utm_source=badge&utm_medium=badge&utm_campaign=pr-badge)
+[![license](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://github.com/williamFalcon/pytorch-lightning/blob/master/LICENSE)
 <!-- 
 removed until codecov badge isn't empy. likely a config error showing nothing on master.
 [![codecov](https://codecov.io/gh/Borda/pytorch-lightning/branch/master/graph/badge.svg)](https://codecov.io/gh/Borda/pytorch-lightning)
 -->
-[![Coverage](https://github.com/williamFalcon/pytorch-lightning/blob/master/docs/source/_static/coverage.svg)](https://github.com/williamFalcon/pytorch-lightning/tree/master/tests#running-coverage)
-[![CodeFactor](https://www.codefactor.io/repository/github/borda/pytorch-lightning/badge)](https://www.codefactor.io/repository/github/borda/pytorch-lightning)
-[![ReadTheDocs](https://readthedocs.org/projects/pytorch-lightning/badge/?version=latest)](https://pytorch-lightning.readthedocs.io/en/latest)
-[![license](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://github.com/williamFalcon/pytorch-lightning/blob/master/LICENSE)

 </div>

@@ -33,17 +33,18 @@ pip install pytorch-lightning
 **[View the docs here](https://williamfalcon.github.io/pytorch-lightning/)**

 ## What is it?  
-Lightning is a very lightweight wrapper on PyTorch. This means you don't have to learn a new library. It defers core training and validation logic to you and automates the rest. It guarantees tested, correct, modern best practices for the automated parts.
+Lightning is a very lightweight wrapper on PyTorch. This means you don't have to learn a new library. To use Lightning, simply refactor your research code into the [LightningModule](https://github.com/williamFalcon/pytorch-lightning#how-do-i-do-use-it) format and Lightning will automate the rest. Lightning guarantees tested, correct, modern best practices for the automated parts.

+## Starting a new project?   
+[Use our seed-project aimed at reproducibility!](https://github.com/williamFalcon/pytorch-lightning-conference-seed)     

 ## Why do I want to use lightning?
-When starting a new project the last thing you want to do is recode a training loop, multi-cluster training, 16-bit precision, early-stopping, model loading/saving, 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.
+Every research project starts the same, a model, a training loop, validation loop, etc. As your research advances, you're likely to need distributed training, 16-bit precision, checkpointing, gradient accumulation, etc.   

-With lightning, you guarantee those parts of your code work so you can focus on what the meat of the research: The data and the training/validation loop logic. 
+Lightning sets up all the boilerplate state-of-the-art training for you so you can focus on the research.   

-Don't worry about training on multiple gpus or speeding up your code, lightning will do that for you!
-
---   
+---
+ 
 ## README Table of Contents        
 - [How do I use it](https://github.com/williamFalcon/pytorch-lightning#how-do-i-do-use-it)     
 - [What lightning automates](https://github.com/williamFalcon/pytorch-lightning#what-does-lightning-control-for-me)    
@@ -53,14 +54,19 @@ Don't worry about training on multiple gpus or speeding up your code, lightning
 - [Tutorials](https://github.com/williamFalcon/pytorch-lightning#tutorials)
 - [Contributing](https://github.com/williamFalcon/pytorch-lightning/blob/master/CONTRIBUTING.md)    
 - [Bleeding edge install](https://github.com/williamFalcon/pytorch-lightning#bleeding-edge)   
- [Lightning Design Principles](https://github.com/williamFalcon/pytorch-lightning#lightning-design-principles)
+- [Lightning Design Principles](https://github.com/williamFalcon/pytorch-lightning#lightning-design-principles)   
+- [Asking for help](https://github.com/williamFalcon/pytorch-lightning#asking-for-help)
 - [FAQ](https://github.com/williamFalcon/pytorch-lightning#faq)    

---   
+---
+
 ## How do I do use it?   
+Think about Lightning as refactoring your research code instead of using a new framework. The research code goes into a [LightningModule]((https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/)) which you fit using a Trainer.    
+
+The LightningModule defines a *system* such as seq-2-seq, GAN, etc... It can ALSO define a simple classifier such as the example below.     

 To use lightning do 2 things:  
-1. [Define a LightningModel](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/)         
+1. [Define a LightningModule](https://williamfalcon.github.io/pytorch-lightning/LightningModule/RequiredTrainerInterface/)         
 ```python
 import os
 import torch
@@ -71,61 +77,65 @@ import torchvision.transforms as transforms

 import pytorch_lightning as pl

-class CoolModel(pl.LightningModule):
+class CoolSystem(pl.LightningModule):

    def __init__(self):
-        super(CoolModel, self).__init__()
+        super(CoolSystem, 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):
+        # REQUIRED
        x, y = batch
        y_hat = self.forward(x)
-        return {'loss': self.my_loss(y_hat, y)}
+        return {'loss': F.cross_entropy(y_hat, y)}

    def validation_step(self, batch, batch_nb):
+        # OPTIONAL
        x, y = batch
        y_hat = self.forward(x)
-        return {'val_loss': self.my_loss(y_hat, y)}
+        return {'val_loss': F.cross_entropy(y_hat, y)}

    def validation_end(self, outputs):
+        # OPTIONAL
        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)]
+        # REQUIRED
+        # can return multiple optimizers and learning_rate schedulers
+        return torch.optim.Adam(self.parameters(), lr=0.02)

    @pl.data_loader
    def tng_dataloader(self):
+        # REQUIRED
        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)

    @pl.data_loader
    def val_dataloader(self):
+        # OPTIONAL
        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)

    @pl.data_loader
    def test_dataloader(self):
+        # OPTIONAL
        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

-model = CoolModel()
+model = CoolSystem()

 # most basic trainer, uses good defaults
 trainer = Trainer()    
 trainer.fit(model)   
 ```     

-Or with tensorboard logger and some options turned on such as multi-gpu, etc...    
+Or with tensorboard logger and some options turned on such as multi-gpu, etc...
 ```python   
 from test_tube import Experiment    

@@ -133,7 +143,7 @@ from test_tube import Experiment
 exp = Experiment(save_dir=os.getcwd())

 # train on cpu using only 10% of the data (for demo purposes)
-# pass in experi
+# pass in experiment for automatic tensorboard logging.    
 trainer = Trainer(experiment=exp, max_nb_epochs=1, train_percent_check=0.1)

 # train on 4 gpus
@@ -150,6 +160,11 @@ print('View tensorboard logs by running\ntensorboard --logdir %s' % os.getcwd())
 print('and going to http://localhost:6006 on your browser')
 ```    

+When you're all done you can even run the test set separately.   
+```python
+trainer.test()
+```
+
 ## What does lightning control for me?   

 Everything in gray!    
@@ -157,7 +172,7 @@ You define the blue parts using the LightningModule interface:

 ![Ouverview](./docs/source/_static/overview_flat.jpg)

-```{.python}
+```python
 # what to do in the training loop
 def training_step(self, data_batch, batch_nb):

@@ -243,10 +258,10 @@ Lightning also adds a text column with all the hyperparameters for this experime

 ![tensorboard-support](./docs/source/_static/tf_tags.png)

-Simply note the path you set for the Experiment    
-``` {.python}   
+Simply note the path you set for the [Experiment](https://williamfalcon.github.io/test-tube/experiment_tracking/experiment/) from [test_tube](https://github.com/williamFalcon/test-tube)
+```python
 from test_tube import Experiment
-from pytorch-lightning import  Trainer
+from pytorch_lightning import Trainer

 exp = Experiment(save_dir='/some/path')
 trainer = Trainer(experiment=exp)
@@ -261,18 +276,18 @@ tensorboard --logdir /some/path
 ## Lightning automates all of the following ([each is also configurable](https://williamfalcon.github.io/pytorch-lightning/Trainer/)):


-###### Checkpointing    
+#### 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) 
 - [Restoring training session](https://williamfalcon.github.io/pytorch-lightning/Trainer/Checkpointing/#restoring-training-session)

-###### Computing cluster (SLURM)    
+#### 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  
+#### 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)
@@ -283,7 +298,7 @@ tensorboard --logdir /some/path
 - [Print input and output size of every module in system](https://williamfalcon.github.io/pytorch-lightning/LightningModule/properties/#example_input_array)


-###### 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)
@@ -292,7 +307,7 @@ tensorboard --logdir /some/path
 - [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 metric row every k batches](https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#log-metric-row-every-k-batches)
@@ -302,7 +317,7 @@ tensorboard --logdir /some/path
 - [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)
 - [Force training for min or max epochs](https://williamfalcon.github.io/pytorch-lightning/Trainer/Training%20Loop/#force-training-for-min-or-max-epochs)
@@ -312,8 +327,9 @@ tensorboard --logdir /some/path
 - [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)
+- [Step optimizers at arbitrary intervals](https://williamfalcon.github.io/pytorch-lightning/Trainer/hooks/#optimizer_step)

-###### 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/)
@@ -322,11 +338,13 @@ tensorboard --logdir /some/path
 - [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)

+#### Testing loop  
+- [Run test set](https://williamfalcon.github.io/pytorch-lightning/Trainer/Testing%20loop/)  

 ## Demo
 ```bash
 # install lightning
-pip install pytorch-lightning
+pip install pytorch_lightning

 # clone lightning for the demo
 git clone https://github.com/williamFalcon/pytorch-lightning.git
@@ -350,8 +368,24 @@ python multi_node_cluster_template.py --nb_gpu_nodes 4 --gpus '0,1,2,3,4,5,6,7'
 - [9 key speed features in Pytorch-Lightning](https://towardsdatascience.com/9-tips-for-training-lightning-fast-neural-networks-in-pytorch-8e63a502f565)    
 - [SLURM, multi-node training with Lightning](https://towardsdatascience.com/trivial-multi-node-training-with-pytorch-lightning-ff75dfb809bd)     

+---    
+## Asking for help    
+Welcome to the Lightning community!   
+
+If you have any questions, feel free to:   
+1. [read the docs](https://williamfalcon.github.io/pytorch-lightning/).     
+2. [Search through the issues](https://github.com/williamFalcon/pytorch-lightning/issues?utf8=%E2%9C%93&q=my++question).      
+3. [Ask on stackoverflow](https://stackoverflow.com/questions/ask?guided=false) with the tag pytorch-lightning.   
+
+If no one replies to you quickly enough, feel free to post the stackoverflow link to our Gitter chat!   
+
+To chat with the rest of us visit our [gitter channel](https://gitter.im/PyTorch-Lightning/community)!     
+
 ---   
 ## FAQ    
+**How do I use Lightning for rapid research?**   
+[Here's a walk-through](https://williamfalcon.github.io/pytorch-lightning/)  
+
 **Why was Lightning created?**     
 Lightning has 3 goals in mind:
 1. Maximal flexibility while abstracting out the common boilerplate across research projects.   
@@ -371,11 +405,36 @@ Nope.
 Nope. Please use anaconda or miniconda.    

 **Which PyTorch versions do you support?**    
-Lightning 0.4.2+ supports PyTorch 1.2.0.    
-For PyTorch 1.1.0 install Lightning 0.4.0 with test-tube=0.6.7.6.    
+- **PyTorch 1.1.0**       
+    ```bash    
+    # install pytorch 1.1.0 using the official instructions   
+    
+    # install test-tube 0.6.7.6 which supports 1.1.0   
+    pip install test-tube==0.6.7.6   
+    
+    # install latest Lightning version without upgrading deps    
+    pip install -U --no-deps pytorch-lightning
+    ```     
+- **PyTorch 1.2.0**
+    Install via pip as normal   

-## Bleeding edge
-If you can't wait for the next release, install the most up to date code with:  
+## Custom installation
+
+### Bleeding edge
+
+If you can't wait for the next release, install the most up to date code with:
+* using GIT (locally clone whole repo with full history)
+    ```bash
+    pip install git+https://github.com/williamFalcon/pytorch-lightning.git@master --upgrade
+    ```
+* using instant zip (last state of the repo without git history)
+    ```bash
+    pip install https://github.com/williamFalcon/pytorch-lightning/archive/master.zip --upgrade
+    ```
+
+### Any release installation
+
+You can also install any past release from this repository:
 ```bash
-pip install git+https://github.com/williamFalcon/pytorch-lightning.git@master --upgrade
+pip install https://github.com/williamFalcon/pytorch-lightning/archive/0.4.4.zip --upgrade
 ```
@@ -45,9 +45,7 @@ install:
  # directly to master instead of just PR builds (or the converse).
  - SET PATH=%PYTHON%;%PYTHON%\\Scripts;%path%
  - pip install -U --user pip
-  - pip install "https://download.pytorch.org/whl/cu90/torch-1.1.0-cp%PIP_PYVER%-cp%PIP_PYVER%m-win_amd%PYTHON_ARCH%.whl"
-    pip install "https://download.pytorch.org/whl/cu90/torchvision-0.3.0-cp%PIP_PYVER%-cp%PIP_PYVER%m-win_amd%PYTHON_ARCH%.whl"
-  - pip install -r requirements.txt
+  - pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html
  - pip install -r ./tests/requirements.txt

 # scripts to run before tests (working directory and environment changes are persisted from the previous steps such as "before_build")
@@ -9,22 +9,22 @@ Otherwise, to Define a Lightning Module, implement the following methods:

 **Required**:  

- [training_step](RequiredTrainerInterface.md#training_step)   
- [validation_step](RequiredTrainerInterface.md#validation_step)
- [validation_end](RequiredTrainerInterface.md#validation_end)
-    
- [configure_optimizers](RequiredTrainerInterface.md#configure_optimizers)
-
- [tng_dataloader](RequiredTrainerInterface.md#tng_dataloader)
- [tng_dataloader](RequiredTrainerInterface.md#tng_dataloader)
- [test_dataloader](RequiredTrainerInterface.md#test_dataloader)
+- [training_step](RequiredTrainerInterface.md#training_step)      
+- [tng_dataloader](RequiredTrainerInterface.md#tng_dataloader)    
+- [configure_optimizers](RequiredTrainerInterface.md#configure_optimizers)    

 **Optional**:   

- [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)
+- [validation_step](RequiredTrainerInterface.md#validation_step)    
+- [validation_end](RequiredTrainerInterface.md#validation_end) 
+- [test_step](RequiredTrainerInterface.md#test_step)    
+- [test_end](RequiredTrainerInterface.md#test_end) 
+- [val_dataloader](RequiredTrainerInterface.md#val_dataloader)    
+- [test_dataloader](RequiredTrainerInterface.md#test_dataloader)    
+- [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   
@@ -48,24 +48,36 @@ class CoolModel(pl.LightningModule):
    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):
+        # REQUIRED
        x, y = batch
        y_hat = self.forward(x)
-        return {'loss': self.my_loss(y_hat, y)}
+        return {'loss': F.cross_entropy(y_hat, y)}

    def validation_step(self, batch, batch_nb):
+        # OPTIONAL
        x, y = batch
        y_hat = self.forward(x)
-        return {'val_loss': self.my_loss(y_hat, y)}
+        return {'val_loss': F.cross_entropy(y_hat, y)}

    def validation_end(self, outputs):
+        # OPTIONAL
        avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
        return {'avg_val_loss': avg_loss}

+    def test_step(self, batch, batch_nb):
+        # OPTIONAL
+        x, y = batch
+        y_hat = self.forward(x)
+        return {'test_loss': F.cross_entropy(y_hat, y)}
+
+    def test_end(self, outputs):
+        # OPTIONAL
+        avg_loss = torch.stack([x['test_loss'] for x in outputs]).mean()
+        return {'avg_test_loss': avg_loss}
+
    def configure_optimizers(self):
+        # REQUIRED
        return [torch.optim.Adam(self.parameters(), lr=0.02)]

    @pl.data_loader
@@ -74,11 +86,15 @@ class CoolModel(pl.LightningModule):

    @pl.data_loader
    def val_dataloader(self):
+        # OPTIONAL
+        # can also return a list of val dataloaders
        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)

    @pl.data_loader
    def test_dataloader(self):
-        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)
+        # OPTIONAL
+        # can also return a list of test dataloaders
+        return DataLoader(MNIST(os.getcwd(), train=False, download=True, transform=transforms.ToTensor()), batch_size=32)
 ```
 ---   
 ### How do these methods fit into the broader training?     
@@ -90,7 +106,7 @@ The LightningModule interface is on the right. Each method corresponds to a part
  </a>
 </p>   

---
+## Required Methods    

 ### training_step

@@ -134,18 +150,107 @@ def training_step(self, data_batch, batch_nb):
    
    # return a dict
    return output
+```    
+
+If you define multiple optimizers, this step will also be called with an additional ```optimizer_idx``` param.    
+``` {.python}
+# Multiple optimizers (ie: GANs)     
+def training_step(self, data_batch, batch_nb, optimizer_idx):
+    if optimizer_idx == 0:
+        # do training_step with encoder
+    if optimizer_idx == 1:
+        # do training_step with decoder    
+```    
+
+--- 
+### tng_dataloader 
+
+``` {.python}
+@pl.data_loader
+def tng_dataloader(self)
+```
+Called by lightning during training loop. Make sure to use the @pl.data_loader decorator, this ensures not calling this function until the data are needed.
+
+##### Return
+PyTorch DataLoader
+
+**Example**
+
+``` {.python}
+@pl.data_loader
+def tng_dataloader(self):
+    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
 ```

---
+--- 
+### configure_optimizers 
+
+``` {.python}
+def configure_optimizers(self)
+```
+
+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.
+
+**Note:** If you use multiple optimizers, training_step will have an additional ```optimizer_idx``` parameter.    
+
+
+
+##### Return    
+Return any of these 3 options:   
+Single optimizer   
+List or Tuple - List of optimizers    
+Two lists - The first list has multiple optimizers, the second a list of learning-rate schedulers
+
+**Example**
+
+``` {.python}
+# most cases
+def configure_optimizers(self):
+    opt = Adam(self.parameters(), lr=0.01)
+    return opt
+
+# multiple optimizer case (eg: GAN)
+def configure_optimizers(self):
+    generator_opt = Adam(self.model_gen.parameters(), lr=0.01)
+    disriminator_opt = Adam(self.model_disc.parameters(), lr=0.02)
+    return generator_opt, disriminator_opt
+    
+# example with learning_rate schedulers  
+def configure_optimizers(self):
+    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]
+```  
+
+If you need to control how often those optimizers step or override the default .step() schedule, override 
+the [optimizer_step](https://williamfalcon.github.io/pytorch-lightning/Trainer/hooks/#optimizer_step) hook.   
+
+## Optional Methods    

 ### validation_step

 ``` {.python}
-def validation_step(self, data_batch, batch_nb)
-```
+# if you have one val dataloader:
+def validation_step(self, data_batch, batch_nb)   

-In this step you'd normally do the forward pass and calculate the loss for a batch. You can also do fancier things like multiple forward passes or something specific to your model.
-This is most likely the same as your training_step. But unlike training step, the outputs from here will go to validation_end for collation.
+# if you have multiple val dataloaders:  
+def validation_step(self, data_batch, batch_nb, dataloader_idx)
+```
+**OPTIONAL**    
+If you don't need to validate you don't need to implement this method. In this step you'd normally generate examples or calculate anything of interest such as accuracy. 
+
+When the validation_step is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, model goes back to training mode and gradients are enabled.
+
+The dict you return here will be available in the `validation_end` method. 

 **Params**   

@@ -153,22 +258,30 @@ This is most likely the same as your training_step. But unlike training step, th
 |---|---|
 |  data_batch | The output of your dataloader. A tensor, tuple or list  |
 |  batch_nb | Integer displaying which batch this is  |
+|  dataloader_i | Integer displaying which dataloader this is (only if multiple val datasets used)  |

 **Return**   

 | Return  | description  | optional |
 |---|---|---|   
-|  dict | Dict of OrderedDict with metrics to display in progress bar. All keys must be tensors. | Y |
+|  dict | Dict or OrderedDict with metrics to display in progress bar. All keys must be tensors. | Y |

 **Example**

 ``` {.python}
+# CASE 1: A single validation dataset
 def validation_step(self, data_batch, batch_nb):
-    x, y, z = data_batch
+    x, y = data_batch
    
    # implement your own
    out = self.forward(x)
-    loss = self.loss(out, x)
+    loss = self.loss(out, y)
+    
+    # log 6 example images
+    # or generated text... or whatever
+    sample_imgs = x[:6]
+    grid = torchvision.utils.make_grid(sample_imgs)
+    self.experiment.add_image('example_images', grid, 0) 
    
    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
@@ -183,22 +296,35 @@ def validation_step(self, data_batch, batch_nb):
    
    # return an optional dict
    return output
-```
+```   
+
+If you pass in multiple validation datasets, validation_step will have an additional argument.
+
+```python
+# CASE 2: multiple validation datasets
+def validation_step(self, data_batch, batch_nb, dataset_idx):
+    # dataset_idx tells you which dataset this is.   
+```   
+
+The ```dataset_idx``` corresponds to the order of datasets returned in ```val_dataloader```.    

 --- 
 ### validation_end

 ``` {.python}
 def validation_end(self, outputs)
-```
+```   
+If you didn't define a validation_step, this won't be called.       

-Called at the end of the validation loop with the output of each validation_step.
+Called at the end of the validation loop with the outputs of validation_step.
+
+The outputs here are strictly for the progress bar. If you don't need to display anything, don't return anything.    

 **Params**    

 | Param  | description  |
 |---|---|
-|  outputs | List of outputs you defined in validation_step |
+|  outputs | List of outputs you defined in validation_step, or if there are multiple dataloaders, a list containing a list of outputs for each dataloader |

 **Return**   

@@ -208,6 +334,8 @@ Called at the end of the validation loop with the output of each validation_step

 **Example**

+With a single dataloader
+
 ``` {.python}
 def validation_end(self, outputs):
    """
@@ -227,34 +355,169 @@ def validation_end(self, outputs):
    return tqdm_dic
 ```

--- 
-### configure_optimizers 
+With multiple dataloaders, `outputs` will be a list of lists. The outer list contains
+one entry per dataloader, while the inner list contains the individual outputs of 
+each validation step for that dataloader.

 ``` {.python}
-def configure_optimizers(self)
+def validation_end(self, outputs):
+    """
+    Called at the end of validation to aggregate outputs
+    :param outputs: list of list of individual outputs of each validation step
+    :return:
+    """
+    val_loss_mean = 0
+    val_acc_mean = 0
+    i = 0
+    for dataloader_outputs in outputs:
+        for output in dataloader_outputs:
+            val_loss_mean += output['val_loss']
+            val_acc_mean += output['val_acc']
+            i += 1
+
+    val_loss_mean /= i
+    val_acc_mean /= i
+    tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()}
+    return tqdm_dic
 ```

-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.
+### test_step

+``` {.python}
+# if you have one test dataloader:
+def test_step(self, data_batch, batch_nb)   

-##### Return
-List or Tuple - List of optimizers with an optional second list of learning-rate schedulers
+# if you have multiple test dataloaders:  
+def test_step(self, data_batch, batch_nb, dataloader_idx)
+```
+**OPTIONAL**    
+If you don't need to test you don't need to implement this method. In this step you'd normally generate examples or calculate anything of interest such as accuracy.   
+
+When the validation_step is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, model goes back to training mode and gradients are enabled.
+
+The dict you return here will be available in the `test_end` method. 
+
+This function is used when you execute `trainer.test()`.
+
+**Params**   
+
+| Param  | description  |
+|---|---|
+|  data_batch | The output of your dataloader. A tensor, tuple or list  |
+|  batch_nb | Integer displaying which batch this is  |
+|  dataloader_i | Integer displaying which dataloader this is (only if multiple test datasets used)  |
+
+**Return**   
+
+| Return  | description  | optional |
+|---|---|---|   
+|  dict | Dict or OrderedDict with metrics to display in progress bar. All keys must be tensors. | Y |

 **Example**

 ``` {.python}
-# most cases
-def configure_optimizers(self):
-    opt = Adam(self.parameters(), lr=0.01)
-    return [opt]
+# CASE 1: A single test dataset
+def test_step(self, data_batch, batch_nb):
+    x, y = data_batch
    
-# gan example, with scheduler for discriminator
-def configure_optimizers(self):
-    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]
+    # implement your own
+    out = self.forward(x)
+    loss = self.loss(out, y)
+    
+    # calculate acc
+    labels_hat = torch.argmax(out, dim=1)
+    test_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)
+    
+    # all optional...
+    # return whatever you need for the collation function test_end
+    output = OrderedDict({
+        'test_loss': loss_test,
+        'test_acc': torch.tensor(test_acc), # everything must be a tensor
+    })
+    
+    # return an optional dict
+    return output
+```   
+
+If you pass in multiple test datasets, test_step will have an additional argument.
+
+```python
+# CASE 2: multiple test datasets
+def test_step(self, data_batch, batch_nb, dataset_idx):
+    # dataset_idx tells you which dataset this is.   
+```   
+
+The ```dataset_idx``` corresponds to the order of datasets returned in ```test_dataloader```.    
+
+--- 
+### test_end
+
+``` {.python}
+def test_end(self, outputs)
+```   
+If you didn't define a test_step, this won't be called.       
+
+Called at the end of the test step with the output of each test_step.
+
+The outputs here are strictly for the progress bar. If you don't need to display anything, don't return anything.    
+
+**Params**    
+
+| Param  | description  |
+|---|---|
+|  outputs | List of outputs you defined in test_step, or if there are multiple dataloaders, a list containing a list of outputs for each dataloader |
+
+**Return**   
+
+| Return  | description  | optional |
+|---|---|---|   
+|  dict | Dict of OrderedDict with metrics to display in progress bar | Y |
+
+**Example**
+
+``` {.python}
+def test_end(self, outputs):
+    """
+    Called at the end of test to aggregate outputs
+    :param outputs: list of individual outputs of each test step
+    :return:
+    """
+    test_loss_mean = 0
+    test_acc_mean = 0
+    for output in outputs:
+        test_loss_mean += output['test_loss']
+        test_acc_mean += output['test_acc']
+
+    test_loss_mean /= len(outputs)
+    test_acc_mean /= len(outputs)
+    tqdm_dic = {'test_loss': test_loss_mean.item(), 'test_acc': test_acc_mean.item()}
+    return tqdm_dic
+```
+
+With multiple dataloaders, `outputs` will be a list of lists. The outer list contains
+one entry per dataloader, while the inner list contains the individual outputs of 
+each validation step for that dataloader.
+
+``` {.python}
+def test_end(self, outputs):
+    """
+    Called at the end of test to aggregate outputs
+    :param outputs: list of individual outputs of each test step
+    :return:
+    """
+    test_loss_mean = 0
+    test_acc_mean = 0
+    i = 0
+    for dataloader_outputs in outputs:
+        for output in dataloader_outputs:
+            test_loss_mean += output['test_loss']
+            test_acc_mean += output['test_acc']
+            i += 1
+
+    test_loss_mean /= i 
+    test_acc_mean /= i
+    tqdm_dic = {'test_loss': test_loss_mean.item(), 'test_acc': test_acc_mean.item()}
+    return tqdm_dic
 ```

 --- 
@@ -299,44 +562,20 @@ def on_load_checkpoint(self, checkpoint):
    self.something_cool_i_want_to_save = checkpoint['something_cool_i_want_to_save']
 ```

--- 
-### tng_dataloader 
-
-``` {.python}
-@pl.data_loader
-def tng_dataloader(self)
-```
-Called by lightning during training loop. Make sure to use the @pl.data_loader decorator, this ensures not calling this function until the data are needed.
-
-##### Return
-PyTorch DataLoader
-
-**Example**
-
-``` {.python}
-@pl.data_loader
-def tng_dataloader(self):
-    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}
@pl.data_loader
-def tng_dataloader(self)
+def val_dataloader(self)
 ```
-Called by lightning during validation loop. Make sure to use the @pl.data_loader decorator, this ensures not calling this function until the data are needed.
+**OPTIONAL**    
+If you don't need a validation dataset and a validation_step, you don't need to implement this method.    
+
+Called by lightning during validation loop. Make sure to use the @pl.data_loader decorator, this ensures not calling this function until the data are needed.   

 ##### Return
-PyTorch DataLoader
+PyTorch DataLoader or list of PyTorch Dataloaders.    

 **Example**

@@ -352,8 +591,16 @@ def val_dataloader(self):
    )
    
    return loader
+
+# can also return multiple dataloaders   
+@pl.data_loader
+def val_dataloader(self):
+    return [loader_a, loader_b, ..., loader_n]   
 ```

+In the case where you return multiple val_dataloaders, the validation_step will have an arguement ```dataset_idx```
+which matches the order here.    
+
 --- 
 ### test_dataloader 

@@ -361,6 +608,9 @@ def val_dataloader(self):
@pl.data_loader
 def test_dataloader(self)
 ```
+**OPTIONAL**    
+If you don't need a test dataset and a test_step, you don't need to implement this method.    
+
 Called by lightning during test loop. Make sure to use the @pl.data_loader decorator, this ensures not calling this function until the data are needed.

 ##### Return
@@ -389,7 +639,7 @@ def test_dataloader(self):
 def update_tng_log_metrics(self, logs)
 ```
 Called by lightning right before it logs metrics for this batch.
-This is a chance to ammend or add to the metrics about to be logged.
+This is a chance to amend or add to the metrics about to be logged.

 ##### Return
 Dict 
@@ -10,7 +10,7 @@ Current dtype

 ---
 #### experiment    
-An instance of test-tube Experiment which you can use to log anything for tensorboarX.   
+An instance of test-tube Experiment which you can use to log anything for tensorboard (subclass of [PyTorch SummaryWriter](https://pytorch.org/docs/stable/tensorboard.html)).   
 ```{.python}
 self.experiment.add_embedding(...)   
 self.experiment.log({'val_loss': 0.9})   
@@ -1,4 +1,4 @@
-i Lightning can automate saving and loading checkpoints.
+Lightning can automate saving and loading checkpoints.

 ---
 ### Model saving
@@ -38,6 +38,7 @@ trainer.fit(model)
 ```

 The trainer restores:  
+
 - global_step    
 - current_epoch    
 - All optimizers    
@@ -10,10 +10,13 @@ For multi-node training you must use DistributedDataParallel.

 You can toggle between each mode by setting this flag.
 ``` {.python}
-# DEFAULT uses DataParallel
+# DEFAULT (when using single GPU or no GPUs)
+trainer = Trainer(distributed_backend=None)
+
+# Change to DataParallel (gpus > 1)
 trainer = Trainer(distributed_backend='dp')

-# change to distributed data parallel
+# change to distributed data parallel (gpus > 1)
 trainer = Trainer(distributed_backend='ddp')
 ```

@@ -32,12 +35,24 @@ Below are the possible configurations we support.

 | 1 GPU  | 1+ GPUs  | DP  | DDP  | 16-bit  | command |
 |---|---|---|---|---|---|
-| Y  |   |   |   |  | ```Trainer(gpus=[0])``` |
-| Y  |   |   |   | Y | ```Trainer(gpus=[0], use_amp=True)``` |
-|   | Y | Y |   |   | ```Trainer(gpus=[0, ...])``` |
-|   | Y |  | Y  |  | ```Trainer(gpus=[0, ...], distributed_backend='ddp')``` |
-|   | Y |  | Y  | Y | ```Trainer(gpus=[0, ...], distributed_backend='ddp', use_amp=True)``` |
+| Y  |   |   |   |  | ```Trainer(gpus=1)``` |
+| Y  |   |   |   | Y | ```Trainer(gpus=1, use_amp=True)``` |
+|   | Y | Y |   |   | ```Trainer(gpus=k)``` |
+|   | Y |  | Y  |  | ```Trainer(gpus=k, distributed_backend='ddp')``` |
+|   | Y |  | Y  | Y | ```Trainer(gpus=k, distributed_backend='ddp', use_amp=True)``` |

+You also have the option of specifying which GPUs to use by passing a list:   
+
+```python
+# DEFAULT (int)
+Trainer(gpus=k)  
+
+# You specify which GPUs (don't use if running on cluster)  
+Trainer(gpus=[0, 1])  
+
+# can also be a string
+Trainer(gpus='0, 1')
+```

 ---
 #### CUDA flags   
@@ -49,6 +64,9 @@ Lightning sets these for you automatically, there's NO NEED to do this yourself.
 # os.environ["CUDA_VISIBLE_DEVICES"] = "0"
 ```

+However, when using a cluster, Lightning will NOT set these flags (and you should not either). 
+SLURM will set these for you.   
+
 ---
 #### 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.    
@@ -70,7 +88,7 @@ trainer = Trainer(amp_level='O2', use_amp=False)
 Make sure you're on a GPU machine. 
 ```python
 # DEFAULT
-trainer = Trainer(gpus=[0])
+trainer = Trainer(gpus=1)
 ```

 ---
@@ -78,11 +96,11 @@ 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
-# to use DataParallel (default)
-trainer = Trainer(gpus=[0,1,2,3,4,5,6,7], distributed_backend='dp')
+# to use DataParallel
+trainer = Trainer(gpus=8, distributed_backend='dp')

 # RECOMMENDED use DistributedDataParallel
-trainer = Trainer(gpus=[0,1,2,3,4,5,6,7], distributed_backend='ddp')
+trainer = Trainer(gpus=8, distributed_backend='ddp')
 ```

 ---
@@ -90,7 +108,7 @@ trainer = Trainer(gpus=[0,1,2,3,4,5,6,7], distributed_backend='ddp')
 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)
+trainer = Trainer(gpus=8, 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.
@@ -5,7 +5,7 @@ Lighting offers a few options for logging information about model, gpu usage, et
 #### Display metrics in progress bar 
 ``` {.python}
 # DEFAULT
-trainer = Trainer(progress_bar=True)
+trainer = Trainer(show_progress_bar=True)
 ```

 ---
@@ -14,6 +14,14 @@ Every k batches lightning will make an entry in the metrics log
 ``` {.python}
 # DEFAULT (ie: save a .csv log file every 10 batches)
 trainer = Trainer(add_log_row_interval=10)
+```   
+
+---
+#### Log metric row every k batches 
+Logs GPU memory when metrics are logged.   
+``` {.python}
+# DEFAULT
+trainer = Trainer(log_gpu_memory=False)
 ```

 ---
@@ -1,8 +1,10 @@
 Lightning supports model training on a cluster managed by SLURM in the following cases:    

-1. Training on single or multi-cpus only.
-2. Training on single or multi-gpus on the same node.
-3. Coming SOON: Training across multiple nodes.
+1. Training on a single cpu or single GPU.
+2. Train on multiple GPUs on the same node using DataParallel or DistributedDataParallel
+3. Training across multiple GPUs on multiple different nodes via DistributedDataParallel.
+
+**Note: A node means a machine with multiple GPUs**

 ---
 #### Running grid search on a cluster
@@ -55,8 +57,8 @@ cluster.memory_mb_per_node = 10000
 cluster.job_time = '10:00'
 ```

-(3). Give trainer the cluster_manager in your main function:    
-
+(3). Make a main function with your model and trainer. Each job will call this function with a particular
+hparams configuration.    
 ```{.python}
 from pytorch_lightning import Trainer

@@ -66,12 +68,12 @@ def train_fx(trial_hparams, cluster_manager, _):
    my_model = MyLightningModel()
    
    # give the trainer the cluster object
-    trainer = Trainer(cluster=cluster_manager)
+    trainer = Trainer()
    trainer.fit(my_model)

 ```

-(4). Start the grid search     
+(3). Start the grid/random search     
 ```{.python}
 # run the models on the cluster
 cluster.optimize_parallel_cluster_gpu(
@@ -81,24 +83,22 @@ cluster.optimize_parallel_cluster_gpu(
    job_display_name='my_exp')
 ```

-That's it! The SlurmCluster object will automatically checkpoint the lightning model and resubmit if it runs into the walltime!
-
-
 ---
 #### Walltime auto-resubmit
-Lightning automatically resubmits jobs when they reach the walltime. You get this behavior for free if you give lightning
-a slurm cluster object.
+Lightning automatically resubmits jobs when they reach the walltime. Make sure to set the SIGUSR1 signal in 
+your SLURM script.   

-```{.python}
-def my_main_fx(hparams, slurm_manager, _):
-    trainer = Trainer(cluster=slurm_manager)
+```bash
+# 90 seconds before training ends
+#SBATCH --signal=SIGUSR1@90
 ``` 

-(See the grid search example above for cluster configuration).
-With this feature lightning will:    
+When lightning receives the SIGUSR1 signal it will:
+1. save a checkpoint with 'hpc_ckpt' in the name.
+2. resubmit the job using the SLURM_JOB_ID  
+
+When the script starts again, Lightning will:
+1. search for a 'hpc_ckpt' checkpoint. 
+2. restore the model, optimizers, schedulers, epoch, etc...   

-1. automatically checkpoint the model
-2. checkpoint the trainer session
-3. resubmit a continuation job.
-4. load the checkpoint and trainer session in the new model

@@ -0,0 +1,31 @@
+To ensure you don't accidentally use test data to guide training decisions Lightning makes running the test set deliberate.   
+
+---
+#### test
+You have two options to run the test set. 
+First case is where you test right after a full training routine.
+``` {.python}
+# run full training
+trainer.fit(model)
+
+# run test set
+trainer.test()
+```   
+
+Second case is where you load a model and run the test set   
+```{.python}
+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
+)
+
+# init trainer with whatever options
+trainer = Trainer(...)
+    
+# test (pass in the model)
+trainer.test(model)
+```
+In this second case, the options you pass to trainer will be used when running the test set (ie: 16-bit, dp, ddp, etc...)  
+ 
@@ -54,7 +54,10 @@ trainer = Trainer(track_grad_norm=2)

 ---
 #### Set how much of the training set to check
-If you don't want to check 100% of the training set (for debugging or if it's huge), set this flag
+If you don't want to check 100% of the training set (for debugging or if it's huge), set this flag.
+
+train_percent_check will be overwritten by overfit_pct if `overfit_pct > 0`
+
 ``` {.python}
 # DEFAULT
 trainer = Trainer(train_percent_check=1.0)
@@ -5,8 +5,6 @@ Below are all the things lightning automates for you in the validation loop.
 Lightning will run 5 steps of validation in the beginning of training as a sanity check so you don't have to wait until a full epoch to catch possible validation issues.


-
-
 ---
 #### Check validation every n epochs
 If you have a small dataset you might want to check validation every n epochs
@@ -18,6 +16,9 @@ trainer = Trainer(check_val_every_n_epoch=1)
 ---
 #### Set how much of the validation set to check 
 If you don't want to check 100% of the validation set (for debugging or if it's huge), set this flag
+
+val_percent_check will be overwritten by overfit_pct if `overfit_pct > 0`
+
 ``` {.python}
 # DEFAULT
 trainer = Trainer(val_percent_check=1.0)
@@ -29,6 +30,9 @@ trainer = Trainer(val_percent_check=0.1)
 ---
 #### Set how much of the test set to check 
 If you don't want to check 100% of the test set (for debugging or if it's huge), set this flag
+
+test_percent_check will be overwritten by overfit_pct if `overfit_pct > 0`
+
 ``` {.python}
 # DEFAULT
 trainer = Trainer(test_percent_check=1.0)
@@ -54,4 +58,6 @@ Lightning runs a few steps of validation in the beginning of training. This avoi
 ``` {.python}
 # DEFAULT
 trainer = Trainer(nb_sanity_val_steps=5)
-```
+```
+
+You can use `Trainer(nb_sanity_val_steps=0)` to skip the sanity check.
@@ -23,6 +23,9 @@ trainer = Trainer(track_grad_norm=2)
 ---
 #### Make model overfit on subset of data
 A useful debugging trick is to make your model overfit a tiny fraction of the data.
+
+setting `overfit_pct > 0` will overwrite train_percent_check, val_percent_check, test_percent_check
+
 ``` {.python}
 # DEFAULT don't overfit (ie: normal training)
 trainer = Trainer(overfit_pct=0.0)
@@ -67,6 +67,53 @@ def on_tng_metrics(self, metrics):
    # do something before validation end
 ```

+---   
+#### optimizer_step 
+Calls .step() and .zero_grad for each optimizer.  
+You can override this method to adjust how you do the optimizer step for each optimizer
+
+Called once per optimizer
+```python
+# DEFAULT
+def optimizer_step(self, current_epoch, batch_nb, optimizer, optimizer_i):
+    optimizer.step()   
+    optimizer.zero_grad()   
+    
+# Alternating schedule for optimizer steps (ie: GANs)    
+def optimizer_step(self, current_epoch, batch_nb, optimizer, optimizer_i):
+    # update generator opt every 2 steps
+    if optimizer_i == 0:
+        if batch_nb % 2 == 0 :
+            optimizer.step()
+            optimizer.zero_grad()
+   
+    # update discriminator opt every 4 steps
+    if optimizer_i == 1:
+        if batch_nb % 4 == 0 :
+            optimizer.step()
+            optimizer.zero_grad()    
+    
+    # ...
+    # add as many optimizers as you want 
+```
+
+This step allows you to do a lot of non-standard training tricks such as learning-rate warm-up:   
+
+```python
+# learning rate warm-up
+def optimizer_step(self, current_epoch, batch_nb, optimizer, optimizer_i):
+    # warm up lr
+    if self.trainer.global_step < 500:
+        lr_scale = min(1., float(self.trainer.global_step + 1) / 500.)
+        for pg in optimizer.param_groups:
+            pg['lr'] = lr_scale * self.hparams.learning_rate
+    
+    # update params
+    optimizer.step()
+    optimizer.zero_grad() 
+```
+
+
 ---
 #### on_before_zero_grad
 Called in the training loop after taking an optimizer step and before zeroing grads.
@@ -68,6 +68,7 @@ But of course the fun is in all the advanced things it can do:
 - [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)
+- [Step optimizers at arbitrary intervals](https://williamfalcon.github.io/pytorch-lightning/Trainer/hooks/#optimizer_step)

 **Validation loop**    

@@ -77,3 +78,8 @@ But of course the fun is in all the advanced things it can do:
 - [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)
+
+
+**Testing loop**    
+
+- [Run test set](https://williamfalcon.github.io/pytorch-lightning/Trainer/Testing%20loop/)  
@@ -3,7 +3,7 @@ In 99% of cases you want to just copy [this template](https://github.com/william

 ```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
+wget https://raw.githubusercontent.com/williamFalcon/pytorch-lightning/master/examples/new_project_templates/lightning_module_template.py
 ```

 ---    
@@ -40,7 +40,7 @@ The main function should have 3 arguments:
 - 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 _)    

-```{}
+```python
 def main(hparams, cluster, results_dict):
    """
    Main training routine specific for this project
@@ -1,12 +1,63 @@
 ###### New project Quick Start    
-To start a new project you define two files, a LightningModule and a Trainer file.    
+To start a new project define two files, a LightningModule and a Trainer file.    
+To illustrate Lightning power and simplicity, here's an example of a typical research flow.    

-A separate trainer file allows to run many LightningModules. Each LightningModule has the core
-logic to a particular research project.    
+###### Case 1: BERT    
+Let's say you're working on something like BERT but want to try different ways of training or even different networks.  
+You would define a single LightningModule and use flags to switch between your different ideas.   
+```python
+class BERT(pl.LightningModule):
+    def __init__(self, model_name, task):
+        self.task = task
+    
+        if model_name == 'transformer':
+            self.net = Transformer()
+        elif model_name == 'my_cool_version':
+            self.net = MyCoolVersion()
+            
+    def training_step(self, batch, batch_nb):
+        if self.task == 'standard_bert':
+            # do standard bert training with self.net...
+            # return loss
+            
+        if self.task == 'my_cool_task':
+            # do my own version with self.net
+            # return loss
+```   

-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.
+###### Case 2: COOLER NOT BERT    
+But if you wanted to try something **completely** different, you'd define a new module for that.    
+```python

+class CoolerNotBERT(pl.LightningModule):
+    def __init__(self):
+        self.net = ...
+        
+    def training_step(self, batch, batch_nb):
+        # do some other cool task
+        # return loss   
+```   
+
+###### Rapid research flow    
+Then you could do rapid research by switching between these two and using the same trainer.   
+```python
+
+if use_bert:
+    model = BERT()
+else:
+    model = CoolerNotBERT()
+    
+trainer = Trainer(gpus=[0, 1, 2, 3], use_amp=True)
+trainer.fit(model)
+```
+
+Notice a few things about this flow:   
+1. You're writing pure PyTorch... no unnecessary abstractions or new libraries to learn.   
+2. You get free GPU and 16-bit support without writing any of that code in your model.   
+3. You also get all of the capabilities below (without coding or testing yourself).     
+
+---    
+###### Templates 
 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/examples/new_project_templates/single_cpu_template.py) 
@@ -75,6 +126,7 @@ one could be a seq-2-seq model, both (optionally) ran by the same trainer file.
 - [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)
+- [Step optimizers at arbitrary intervals](https://williamfalcon.github.io/pytorch-lightning/Trainer/hooks/#optimizer_step)

 ###### Validation loop    

@@ -85,3 +137,5 @@ one could be a seq-2-seq model, both (optionally) ran by the same trainer file.
 - [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)

+###### Testing loop  
+- [Run test set](https://williamfalcon.github.io/pytorch-lightning/Trainer/Testing%20loop/)  
@@ -161,9 +161,9 @@ class LightningTemplateModel(LightningModule):
            # reduce manually when using dp
            val_acc = output['val_acc']
            if self.trainer.use_dp:
-                val_acc_mean = torch.mean(val_acc)
+                val_acc = torch.mean(val_acc)

-            val_acc_mean += val_acc_mean
+            val_acc_mean += val_acc

        val_loss_mean /= len(outputs)
        val_acc_mean /= len(outputs)
@@ -189,16 +189,13 @@ class LightningTemplateModel(LightningModule):
        dataset = MNIST(root=self.hparams.data_root, train=train,
                        transform=transform, download=True)

-        # when using multi-node we need to add the datasampler
+        # when using multi-node (ddp) 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:
-            pass
+        if self.use_ddp:
+            train_sampler = DistributedSampler(dataset, rank=self.trainer.proc_rank)
+            batch_size = batch_size // self.trainer.world_size  # scale batch size

        should_shuffle = train_sampler is None
        loader = DataLoader(
@@ -243,15 +240,15 @@ class LightningTemplateModel(LightningModule):
        parser.add_argument('--out_features', default=10, type=int)
        # use 500 for CPU, 50000 for GPU to see speed difference
        parser.add_argument('--hidden_dim', default=50000, type=int)
-        parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=False)
+        parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=True)
+        parser.opt_list('--learning_rate', default=0.001 * 8, type=float,
+                        options=[0.0001, 0.0005, 0.001],
+                        tunable=True)

        # 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)

@@ -0,0 +1,107 @@
+# Multi-node examples
+Use these templates for multi-node training. 
+The main complexity around cluster training is how you submit the SLURM jobs.  
+
+## Test-tube   
+Lightning uses test-tube to submit SLURM jobs and to run hyperparameter searches on a cluster.  
+
+To run a hyperparameter search, we normally add the values to search to the Hyperparameter optimizer 
+```python
+from test_tube import HyperOptArgumentParser
+
+parser = HyperOptArgumentParser(strategy='grid_search')
+parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=True)
+parser.opt_list('--learning_rate', default=0.001, type=float,
+                        options=[0.0001, 0.0005, 0.001],
+                        tunable=True)
+                        
+# give your model a chance to add its own parameters
+parser = LightningTemplateModel.add_model_specific_args(parent_parser, root_dir)
+
+# parse args
+hyperparams = parser.parse_args()
+```
+
+The above sets up a grid search on learning rate and drop probability. You can now add this object to the 
+cluster object to perform the grid search:   
+```python
+cluster = SlurmCluster(
+    hyperparam_optimizer=hyperparams,
+    log_path='/path/to/log/slurm/files',
+)
+
+# ... configure cluster options
+
+# run grid search on cluster
+nb_trials = 6   # (2 drop probs * 3 lrs)
+cluster.optimize_parallel_cluster_gpu(
+    YourMainFunction,
+    nb_trials=nb_trials,
+    job_name=hyperparams.experiment_name
+)
+```
+
+Running the above will launch 6 jobs, each with a different drop prob and learning rate combination.   
+The ```tunable``` parameter must be set to True to add that argument to the space of options, otherwise
+Test-Tube will use the ```default=value```.    
+
+
+## SLURM Flags   
+However you decide to submit your jobs, debugging requires a few flags. Without these flags, you'll
+see a nccl error instead of the actual error which caused the bug.   
+
+```sh
+export NCCL_DEBUG=INFO
+export PYTHONFAULTHANDLER=1
+```
+
+On some clusters you might need to set the network interface with this flag.   
+```sh
+export NCCL_SOCKET_IFNAME=^docker0,lo
+```   
+
+You might also need to load the latest version of NCCL  
+```sh
+module load NCCL/2.4.7-1-cuda.10.0
+```
+
+Finally, you must set the master port (usually a random number between 12k and 20k).   
+```sh
+# random port between 12k and 20k
+export MASTER_PORT=$((12000 + RANDOM % 20000))$   
+```
+
+## Simplest example.   
+1. Modify this script with your CoolModel file.   
+2. Update and submit [this bash script](https://github.com/williamFalcon/pytorch-lightning/blob/master/examples/new_project_templates/multi_node_examples/minimal_multi_node_demo_script.sh)   
+```bash
+squeue minimal_multi_node_demo_script.sh
+```
+
+## Grid search on a cluster   
+
+#### Option 1: Run on cluster using your own SLURM script    
+The trainer and model will work on a cluster if you configure your SLURM script correctly.   
+
+1. Update [this demo slurm script](https://github.com/williamFalcon/pytorch-lightning/blob/master/examples/new_project_templates/multi_node_examples/demo_script.sh).  
+2. Submit the script   
+```bash
+$ squeue demo_script.sh
+```
+
+Most people have some way they automatically generate their own scripts.  
+To run a grid search this way, you'd need a way to automatically generate scripts using all the combinations of 
+hyperparameters to search over.   
+
+#### Option 2: Use test-tube for SLURM script
+With test tube we can automatically generate slurm scripts for different hyperparameter options.   
+
+To run this demo:    
+```bash
+source activate YourCondaEnv
+
+python multi_node_cluster_auto_slurm.py --email your@email.com --gpu_partition your_partition --conda_env YourCondaEnv
+```
+
+That will submit 6 jobs. Each job will have a specific combination of hyperparams. Each job will also run on 2 nodes
+where each node has 8 gpus.   
@@ -0,0 +1,66 @@
+#!/bin/bash
+#
+# Auto-generated by test-tube (https://github.com/williamFalcon/test-tube)
+#################
+
+# set a job name
+#SBATCH --job-name=lightning_test
+#################
+
+# a file for job output, you can check job progress
+#SBATCH --output=/slurm_output_%j.out
+#################
+
+# a file for errors
+#SBATCH --error=/slurm_output_%j.err
+#################
+
+# time needed for job
+#SBATCH --time=01:00:00
+#################
+
+# gpus per node
+#SBATCH --gres=gpu:8
+#################
+
+# cpus per job
+#SBATCH --cpus-per-task=10
+#################
+
+# number of requested nodes
+#SBATCH --nodes=2
+#################
+
+# memory per node (0 means all)
+#SBATCH --mem=0
+#################
+
+# slurm will send a signal this far out before it kills the job
+#SBATCH --signal=USR1@300
+#################
+
+# comment
+#SBATCH --comment=lightning_demo
+#################
+
+# 1 task per gpu
+#SBATCH --ntasks-per-node=8
+#################
+
+source activate YourEnv
+
+# debugging flags (optional)
+export NCCL_DEBUG=INFO
+export PYTHONFAULTHANDLER=1
+
+# on your cluster you might need these:
+# set the network interface
+export NCCL_SOCKET_IFNAME=^docker0,lo
+
+# might need the latest cuda
+module load NCCL/2.4.7-1-cuda.10.0
+
+# random port between 12k and 20k
+export MASTER_PORT=$((12000 + RANDOM % 20000))$
+
+srun python multi_node_own_slurm_script.py
@@ -0,0 +1,24 @@
+from pytorch_lightning import Trainer
+from test_tube import Experiment
+import os
+
+
+def main():
+    # use the cool model from the main README.md
+    model = CoolModel()  # noqa: F821
+    exp = Experiment(save_dir=os.getcwd())
+
+    # train on 4 GPUs across 4 nodes
+    trainer = Trainer(
+        experiment=exp,
+        distributed_backend='ddp',
+        max_nb_epochs=10,
+        gpus=4,
+        nb_gpu_nodes=4
+    )
+
+    trainer.fit(model)
+
+
+if __name__ == '__main__':
+    main()
@@ -0,0 +1,30 @@
+#!/bin/bash -l
+
+# SLURM SUBMIT SCRIPT
+#SBATCH --nodes=4
+#SBATCH --gres=gpu:4
+#SBATCH --ntasks-per-node=4
+#SBATCH --mem=0
+#SBATCH --time=0-02:00:00
+
+# activate conda env
+conda activate my_env
+
+# -------------------------
+# debugging flags (optional)
+# export NCCL_DEBUG=INFO
+# export PYTHONFAULTHANDLER=1
+
+# on your cluster you might need these:
+# set the network interface
+# export NCCL_SOCKET_IFNAME=^docker0,lo
+
+# might need the latest cuda
+# module load NCCL/2.4.7-1-cuda.10.0
+# -------------------------
+
+# random port between 12k and 20k
+export MASTER_PORT=$((12000 + RANDOM % 20000))
+
+# run script from above
+python minimal_multi_node_demo.py
@@ -7,11 +7,12 @@ from time import sleep
 import torch

 from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
-from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint

 from examples.new_project_templates.lightning_module_template import LightningTemplateModel

+PORT = np.random.randint(12000, 20000, 1)[0]
 SEED = 2334
 torch.manual_seed(SEED)
 np.random.seed(SEED)
@@ -21,7 +22,7 @@ def main_local(hparams):
    main(hparams, None, None)


-def main(hparams, cluster, results_dict):
+def main(hparams, cluster):
    """
    Main training routine specific for this project
    :param hparams:
@@ -47,6 +48,7 @@ def main(hparams, cluster, results_dict):
        name=hyperparams.experiment_name,
        save_dir=hyperparams.test_tube_save_path,
        autosave=False,
+        version=hparams.hpc_exp_number,  # match the slurm job version number
        description='test demo'
    )

@@ -77,10 +79,9 @@ def main(hparams, cluster, results_dict):
    # ------------------------
    trainer = Trainer(
        experiment=exp,
-        cluster=cluster,
        checkpoint_callback=checkpoint,
        early_stop_callback=early_stop,
-        gpus=hparams.gpus,
+        gpus=hparams.per_experiment_nb_gpus,
        nb_gpu_nodes=hyperparams.nb_gpu_nodes
    )

@@ -99,7 +100,7 @@ def optimize_on_cluster(hyperparams):
    )

    # email for cluster coms
-    cluster.notify_job_status(email='add_email_here', on_done=True, on_fail=True)
+    cluster.notify_job_status(email=hyperparams.email, on_done=True, on_fail=True)

    # configure cluster
    cluster.per_experiment_nb_gpus = hyperparams.per_experiment_nb_gpus
@@ -109,7 +110,24 @@ def optimize_on_cluster(hyperparams):
    cluster.memory_mb_per_node = 0

    # any modules for code to run in env
-    cluster.add_command('source activate lightning')
+    cluster.add_command(f'source activate {hyperparams.conda_env}')
+
+    # set DDP master port
+    cluster.add_command(f'export MASTER_PORT={PORT}')
+
+    # OPTIONAL for debugging
+    # without these flags errors in your code will 
+    # appear to be nccl errors
+    cluster.add_command('export NCCL_DEBUG=INFO')
+    cluster.add_command('export PYTHONFAULTHANDLER=1')
+
+    # depending on your cluster config, you probably want
+    # to limit the wired connection device
+    # cluster.add_command('export NCCL_SOCKET_IFNAME=^docker0,lo')
+
+    # depending on your cluster, you might need to load
+    # the latest NCCL version
+    # cluster.load_modules(['NCCL/2.4.7-1-cuda.10.0'])

    # run only on 32GB voltas
    cluster.add_slurm_cmd(cmd='constraint', value='volta32gb',
@@ -121,7 +139,7 @@ def optimize_on_cluster(hyperparams):
    # creates and submits jobs to slurm
    cluster.optimize_parallel_cluster_gpu(
        main,
-        nb_trials=hyperparams.nb_hopt_trials,
+        nb_trials=hyperparams.num_hyperparam_trials,
        job_name=hyperparams.experiment_name
    )

@@ -139,15 +157,10 @@ if __name__ == '__main__':
    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,
+                               default=8, help='how many gpus to use in a node')
+    parent_parser.add_argument('--nb_gpu_nodes', type=int, default=2,
                               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')
@@ -157,9 +170,15 @@ if __name__ == '__main__':
                               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,
+    parent_parser.add_argument('--num_hyperparam_trials', type=int, default=6,
                               help='how many grid search trials to run')

+    parent_parser.add_argument('--email', type=str, default='add@email.com',
+                               help='email for jobs')
+    parent_parser.add_argument('--conda_env', type=str, default='base',
+                               help='email for jobs')
+    parent_parser.add_argument('--gpu_partition', type=str, help='consult your cluster manual')
+
    # allow model to overwrite or extend args
    parser = LightningTemplateModel.add_model_specific_args(parent_parser, root_dir)
    hyperparams = parser.parse_args()
@@ -0,0 +1,70 @@
+"""
+Multi-node example (GPU)
+"""
+import os
+import numpy as np
+import torch
+
+from test_tube import HyperOptArgumentParser, Experiment
+from pytorch_lightning import Trainer
+from examples.new_project_templates.lightning_module_template import LightningTemplateModel
+
+SEED = 2334
+torch.manual_seed(SEED)
+np.random.seed(SEED)
+
+
+def main(hparams):
+    """
+    Main training routine specific for this project
+    :param hparams:
+    :return:
+    """
+    # ------------------------
+    # 1 INIT LIGHTNING MODEL
+    # ------------------------
+    model = LightningTemplateModel(hparams)
+
+    # ------------------------
+    # 2 INIT TEST TUBE EXP
+    # ------------------------
+    # init experiment
+    exp = Experiment(
+        name='test_exp',
+        save_dir=hyperparams.log_dir,
+        autosave=False,
+        description='test demo'
+    )
+
+    # ------------------------
+    # 2 INIT TRAINER
+    # ------------------------
+    trainer = Trainer(
+        experiment=exp,
+        gpus=8,
+        nb_gpu_nodes=2
+    )
+
+    # ------------------------
+    # 5 START TRAINING
+    # ------------------------
+    trainer.fit(model)
+
+
+if __name__ == '__main__':
+    # use current dir for logging
+    root_dir = os.path.dirname(os.path.realpath(__file__))
+    log_dir = os.path.join(root_dir, 'pt_lightning_demo_logs')
+
+    parent_parser = HyperOptArgumentParser(strategy='grid_search', add_help=False)
+    parent_parser.add_argument('--log_dir', type=str, default=log_dir,
+                               help='where to save logs')
+
+    # allow model to overwrite or extend args
+    parser = LightningTemplateModel.add_model_specific_args(parent_parser, root_dir)
+    hyperparams = parser.parse_args()
+
+    # ---------------------
+    # RUN TRAINING
+    # ---------------------
+    main(hyperparams)
@@ -6,7 +6,7 @@ import numpy as np
 import torch

 from test_tube import HyperOptArgumentParser, Experiment
-from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint

 from examples.new_project_templates.lightning_module_template import LightningTemplateModel
@@ -25,14 +25,11 @@ def main(hparams):
    # ------------------------
    # 1 INIT LIGHTNING MODEL
    # ------------------------
-    print('loading model...')
    model = LightningTemplateModel(hparams)
-    print('model built')

    # ------------------------
-    # 2 INIT TEST TUBE EXP
+    # 2 INIT EXP
    # ------------------------
-
    # init experiment
    exp = Experiment(
        name=hyperparams.experiment_name,
@@ -6,7 +6,7 @@ import numpy as np
 import torch

 from test_tube import HyperOptArgumentParser, Experiment
-from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint

 from examples.new_project_templates.lightning_module_template import LightningTemplateModel
@@ -110,5 +110,5 @@ if __name__ == '__main__':
    # RUN TRAINING
    # ---------------------
    # run on HPC cluster
-    print('RUNNING INTERACTIVE MODE ON GPUS. gpu ids: %i' % hyperparams.gpus)
+    print(f'RUNNING INTERACTIVE MODE ON GPUS. gpu ids: {hyperparams.gpus}')
    main(hyperparams)
@@ -6,7 +6,7 @@ import numpy as np
 import torch

 from test_tube import HyperOptArgumentParser, Experiment
-from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint

 from examples.new_project_templates.lightning_module_template import LightningTemplateModel
@@ -110,5 +110,5 @@ if __name__ == '__main__':
    # RUN TRAINING
    # ---------------------
    # run on HPC cluster
-    print('RUNNING INTERACTIVE MODE ON GPUS. gpu ids: %i' % hyperparams.gpus)
+    print(f'RUNNING INTERACTIVE MODE ON GPUS. gpu ids: {hyperparams.gpus}')
    main(hyperparams)
@@ -6,7 +6,7 @@ import numpy as np
 import torch

 from test_tube import HyperOptArgumentParser, Experiment
-from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint

 from examples.new_project_templates.lightning_module_template import LightningTemplateModel
@@ -109,5 +109,5 @@ if __name__ == '__main__':
    # RUN TRAINING
    # ---------------------
    # run on HPC cluster
-    print('RUNNING INTERACTIVE MODE ON GPUS. gpu ids: %i' % hyperparams.gpus)
+    print(f'RUNNING INTERACTIVE MODE ON GPUS. gpu ids: {hyperparams.gpus}')
    main(hyperparams)
@@ -2,7 +2,7 @@ import os
 import sys

 from test_tube import HyperOptArgumentParser, Experiment
-from pytorch_lightning.models.trainer import Trainer
+from pytorch_lightning import Trainer
 from pytorch_lightning.utilities.arg_parse import add_default_args
 from pytorch_lightning.callbacks.pt_callbacks import EarlyStopping, ModelCheckpoint

@@ -0,0 +1,178 @@
+"""
+To run this template just do:  
+python gan.py   
+
+After a few epochs, launch tensorboard to see the images being generated at every batch.   
+
+tensorboard --logdir default
+"""
+from argparse import ArgumentParser
+import os
+import numpy as np
+
+import torchvision
+import torchvision.transforms as transforms
+from torchvision.datasets import MNIST
+
+from torch.utils.data import DataLoader
+
+import torch.nn as nn
+import torch.nn.functional as F
+import torch
+
+import pytorch_lightning as pl
+from test_tube import Experiment
+
+
+class Generator(nn.Module):
+    def __init__(self, latent_dim, img_shape):
+        super(Generator, self).__init__()
+        self.img_shape = img_shape
+
+        def block(in_feat, out_feat, normalize=True):
+            layers = [nn.Linear(in_feat, out_feat)]
+            if normalize:
+                layers.append(nn.BatchNorm1d(out_feat, 0.8))
+            layers.append(nn.LeakyReLU(0.2, inplace=True))
+            return layers
+
+        self.model = nn.Sequential(
+            *block(latent_dim, 128, normalize=False),
+            *block(128, 256),
+            *block(256, 512),
+            *block(512, 1024),
+            nn.Linear(1024, int(np.prod(img_shape))),
+            nn.Tanh()
+        )
+
+    def forward(self, z):
+        img = self.model(z)
+        img = img.view(img.size(0), *self.img_shape)
+        return img
+
+
+class Discriminator(nn.Module):
+    def __init__(self, img_shape):
+        super(Discriminator, self).__init__()
+
+        self.model = nn.Sequential(
+            nn.Linear(int(np.prod(img_shape)), 512),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.Linear(512, 256),
+            nn.LeakyReLU(0.2, inplace=True),
+            nn.Linear(256, 1),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, img):
+        img_flat = img.view(img.size(0), -1)
+        validity = self.model(img_flat)
+
+        return validity
+
+
+class GAN(pl.LightningModule):
+
+    def __init__(self, hparams):
+        super(GAN, self).__init__()
+        self.hparams = hparams
+
+        # networks
+        mnist_shape = (1, 28, 28)
+        self.generator = Generator(latent_dim=hparams.latent_dim, img_shape=mnist_shape)
+        self.discriminator = Discriminator(img_shape=mnist_shape)
+
+        # cache for generated images
+        self.generated_imgs = None
+
+    def forward(self, z):
+        return self.generator(z)
+
+    def adversarial_loss(self, y_hat, y):
+        return F.binary_cross_entropy(y_hat, y)
+
+    def training_step(self, batch, batch_nb, optimizer_i):
+        imgs, _ = batch
+
+        # train generator
+        if optimizer_i == 0:
+            # sample noise
+            z = torch.randn(imgs.shape[0], self.hparams.latent_dim)
+
+            # match gpu device (or keep as cpu)
+            if self.on_gpu:
+                z = z.cuda(imgs.device.index)
+
+            # generate images
+            self.generated_imgs = self.forward(z)
+
+            # log sampled images
+            sample_imgs = self.generated_imgs[:6]
+            grid = torchvision.utils.make_grid(sample_imgs)
+            self.experiment.add_image('generated_images', grid, 0)
+
+            # ground truth result (ie: all fake)
+            valid = torch.ones(imgs.size(0), 1)
+
+            # adversarial loss is binary cross-entropy
+            g_loss = self.adversarial_loss(self.discriminator(self.generated_imgs), valid)
+
+            return g_loss
+
+        # train discriminator
+        if optimizer_i == 1:
+            # Measure discriminator's ability to classify real from generated samples
+
+            # how well can it label as real?
+            valid = torch.ones(imgs.size(0), 1)
+            real_loss = self.adversarial_loss(self.discriminator(imgs), valid)
+
+            # how well can it label as fake?
+            fake = torch.zeros(imgs.size(0), 1)
+            fake_loss = self.adversarial_loss(self.discriminator(self.generated_imgs.detach()), fake)
+
+            # discriminator loss is the average of these
+            d_loss = (real_loss + fake_loss) / 2
+
+            return d_loss
+
+    def configure_optimizers(self):
+        lr = self.hparams.lr
+        b1 = self.hparams.b1
+        b2 = self.hparams.b2
+
+        opt_g = torch.optim.Adam(self.generator.parameters(), lr=lr, betas=(b1, b2))
+        opt_d = torch.optim.Adam(self.discriminator.parameters(), lr=lr, betas=(b1, b2))
+        return [opt_g, opt_d], []
+
+    @pl.data_loader
+    def tng_dataloader(self):
+        transform = transforms.Compose([transforms.ToTensor(),
+                                        transforms.Normalize([0.5], [0.5])])
+        dataset = MNIST(os.getcwd(), train=True, download=True, transform=transform)
+        return DataLoader(dataset, batch_size=self.hparams.batch_size)
+
+
+def main(hparams):
+    # save tensorboard logs
+    exp = Experiment(save_dir=os.getcwd())
+
+    # init model
+    model = GAN(hparams)
+
+    # fit trainer on CPU
+    trainer = pl.Trainer(experiment=exp, max_nb_epochs=200)
+    trainer.fit(model)
+
+
+if __name__ == '__main__':
+    parser = ArgumentParser()
+    parser.add_argument("--batch_size", type=int, default=64, help="size of the batches")
+    parser.add_argument("--lr", type=float, default=0.0002, help="adam: learning rate")
+    parser.add_argument("--b1", type=float, default=0.5, help="adam: decay of first order momentum of gradient")
+    parser.add_argument("--b2", type=float, default=0.999, help="adam: decay of first order momentum of gradient")
+    parser.add_argument("--latent_dim", type=int, default=100, help="dimensionality of the latent space")
+
+    hparams = parser.parse_args()
+
+    main(hparams)
@@ -2,9 +2,15 @@ site_name: PyTorch lightning Documentation
 theme:
  name: 'material'
 docs_dir: docs
+repo_name: 'williamFalcon/pytorch-lightning'
 repo_url: https://github.com/williamFalcon/pytorch-lightning
 site_dir: 'site'
 site_description: 'Documentation for PyTorch LightningModule, the researcher version of keras.'

 dev_addr: '0.0.0.0:8000'
 #google_analytics: ['UA-aasd', 'sitename']
+
+markdown_extensions:
+  - codehilite:
+      guess_lang: false
+      linenums: true
@@ -1,4 +1,4 @@
-from .models.trainer import Trainer
+from .trainer.trainer import Trainer
 from .root_module.root_module import LightningModule
 from .root_module.decorators import data_loader

@@ -1,6 +1,7 @@
-from .pt_callbacks import EarlyStopping, ModelCheckpoint
+from .pt_callbacks import EarlyStopping, ModelCheckpoint, GradientAccumulationScheduler

 __all__ = [
    'EarlyStopping',
    'ModelCheckpoint',
+    'GradientAccumulationScheduler',
 ]
@@ -1,5 +1,6 @@
 import os
 import shutil
+import warnings

 import numpy as np

@@ -11,7 +12,6 @@ class Callback(object):
    # Properties
        params: dict. Training parameters
            (eg. verbosity, batch size, number of epochs...).
-        model: instance of `keras.models.Model`.
            Reference of the model being trained.
    The `logs` dictionary that callback methods
    take as argument will contain keys for quantities relevant to
@@ -254,6 +254,37 @@ class ModelCheckpoint(Callback):
                self.save_model(filepath, overwrite=False)


+class GradientAccumulationScheduler(Callback):
+    """Change gradient accumulation factor according to scheduling.
+    # Arguments
+        scheduling: dict, scheduling in format {epoch: accumulation_factor}
+    """
+    def __init__(self, scheduling: dict):
+        if scheduling == {}:  # empty dict error
+            raise TypeError("Empty dict cannot be interpreted correct")
+
+        for key in scheduling.keys():
+            if not isinstance(key, int) or not isinstance(scheduling[key], int):
+                raise TypeError("All epoches and accumulation factor must be integers")
+
+        minimal_epoch = min(scheduling.keys())
+        if minimal_epoch < 1:
+            msg = f"Epochs indexing from 1, epoch {minimal_epoch} cannot be interpreted correct"
+            raise IndexError(msg)
+        elif minimal_epoch != 1:  # if user didnt define first epoch accumulation factor
+            scheduling.update({1: 1})
+
+        self.scheduling = scheduling
+        self.epochs = sorted(scheduling.keys())
+
+    def on_epoch_begin(self, epoch, trainer):
+        epoch += 1  # indexing epochs from 1
+        for i in reversed(range(len(self.epochs))):
+            if epoch >= self.epochs[i]:
+                trainer.accumulate_grad_batches = self.scheduling.get(self.epochs[i])
+                break
+
+
 if __name__ == '__main__':
    c = EarlyStopping(min_delta=0.9, patience=2, verbose=True)
    losses = [10, 9, 8, 8, 6, 4.3, 5, 4.4, 2.8, 2.5]
@@ -56,6 +56,8 @@ class LightningDataParallel(DataParallel):
            # lightning
            if self.module.training:
                return self.module.training_step(*inputs[0], **kwargs[0])
+            elif self.module.testing:
+                return self.module.test_step(*inputs[0], **kwargs[0])
            else:
                return self.module.validation_step(*inputs[0], **kwargs[0])

@@ -89,6 +91,8 @@ class LightningDistributedDataParallel(DistributedDataParallel):
                # lightning
                if self.module.training:
                    output = self.module.training_step(*inputs[0], **kwargs[0])
+                elif self.module.testing:
+                    output = self.module.test_step(*inputs[0], **kwargs[0])
                else:
                    output = self.module.validation_step(*inputs[0], **kwargs[0])
            else:
@@ -153,6 +157,10 @@ def parallel_apply(modules, inputs, kwargs_tup=None, devices=None):  # pragma: n
                # CHANGE
                if module.training:
                    output = module.training_step(*input, **kwargs)
+
+                elif module.testing:
+                    output = module.test_step(*input, **kwargs)
+
                else:
                    output = module.validation_step(*input, **kwargs)
                # ---------------
@@ -1,3 +1,5 @@
+import traceback
+

 def data_loader(fn):
    """
@@ -10,8 +12,17 @@ def data_loader(fn):

    @property
    def _data_loader(self):
-        if not hasattr(self, attr_name):
-            setattr(self, attr_name, fn(self))
-        return getattr(self, attr_name)
+        try:
+            value = getattr(self, attr_name)
+        except AttributeError:
+            try:
+                value = fn(self)  # Lazy evaluation, done only once.
+            except AttributeError as e:
+                # Guard against AttributeError suppression. (Issue #142)
+                traceback.print_exc()
+                error = f'{fn.__name__}: An AttributeError was encountered: ' + str(e)
+                raise RuntimeError(error) from e
+            setattr(self, attr_name, value)  # Memoize evaluation.
+        return value

    return _data_loader
@@ -5,7 +5,7 @@ class ModelHooks(torch.nn.Module):

    def on_sanity_check_start(self):
        """
-        Called before starting validate
+        Called before starting evaluate
        :return:
        """
        pass
@@ -1,12 +1,3 @@
-import os
-import re
-
-import torch
-
-from pytorch_lightning.pt_overrides.override_data_parallel import (
-    LightningDistributedDataParallel, LightningDataParallel)
-
-
 class ModelIO(object):

    def on_load_checkpoint(self, checkpoint):
@@ -41,221 +32,3 @@ class ModelIO(object):
        :return:
        """
        pass
-
-
-class TrainerIO(object):
-
-    def __get_model(self):
-        is_dp_module = isinstance(self.model, (LightningDistributedDataParallel,
-                                               LightningDataParallel))
-        model = self.model.module if is_dp_module else self.model
-        return model
-
-    # --------------------
-    # MODEL SAVE CHECKPOINT
-    # --------------------
-    def save_checkpoint(self, filepath):
-        checkpoint = self.dump_checkpoint()
-
-        # do the actual save
-        torch.save(checkpoint, filepath)
-
-    def restore(self, checkpoint_path, on_gpu):
-
-        if on_gpu:
-            checkpoint = torch.load(checkpoint_path)
-        else:
-            checkpoint = torch.load(checkpoint_path, map_location=lambda storage, loc: storage)
-
-        # load training state (affects trainer only)
-        self.restore_training_state(checkpoint)
-
-        # load model state
-        model = self.__get_model()
-
-        # load the state_dict on the model automatically
-        model.load_state_dict(checkpoint['state_dict'])
-
-    def dump_checkpoint(self):
-
-        checkpoint = {
-            'epoch': self.current_epoch,
-            '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())
-
-        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)
-
-        return checkpoint
-
-    # --------------------
-    # HPC IO
-    # --------------------
-    def enable_auto_hpc_walltime_manager(self):
-        if self.cluster is None:
-            return
-
-        # allow test tube to handle model check pointing automatically
-        # only if proc 0 so we don't trigger world_size resubmits
-        if self.proc_rank == 0:
-            self.cluster.set_checkpoint_save_function(
-                self.hpc_save,
-                kwargs={
-                    'folderpath': self.checkpoint_callback.filepath,
-                    'experiment': self.experiment
-                }
-            )
-
-        self.cluster.set_checkpoint_load_function(
-            self.hpc_load,
-            kwargs={
-                'folderpath': self.checkpoint_callback.filepath,
-                'on_gpu': self.on_gpu
-            }
-        )
-
-    def restore_training_state(self, checkpoint):
-        """
-        Restore trainer state.
-        Model will get its change to update
-        :param checkpoint:
-        :return:
-        """
-        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']
-
-        # restore the optimizers
-        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
-    # ----------------------------------
-    def hpc_save(self, folderpath, experiment):
-        # make sure the checkpoint folder exists
-        os.makedirs(folderpath, exist_ok=True)
-
-        # 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)
-
-        # 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, filepath)
-
-        return filepath
-
-    def hpc_load(self, folderpath, on_gpu):
-        filepath = '{}/hpc_ckpt_{}.ckpt'.format(folderpath, self.max_ckpt_in_folder(folderpath))
-
-        if on_gpu:
-            checkpoint = torch.load(filepath)
-        else:
-            checkpoint = torch.load(filepath, map_location=lambda storage, loc: storage)
-
-        # load training state (affects trainer only)
-        self.restore_training_state(checkpoint)
-
-        # load model state
-        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, name_key='ckpt_'):
-        files = os.listdir(path)
-        files = [x for x in files if name_key in x]
-        if len(files) == 0:
-            return 0
-
-        ckpt_vs = []
-        for name in files:
-            name = name.split(name_key)[-1]
-            name = re.sub('[^0-9]', '', name)
-            ckpt_vs.append(int(name))
-
-        return max(ckpt_vs)
-
-
-def load_hparams_from_tags_csv(tags_csv):
-    from argparse import Namespace
-    import pandas as pd
-
-    tags_df = pd.read_csv(tags_csv)
-    dic = tags_df.to_dict(orient='records')
-
-    ns_dict = {row['key']: convert(row['value']) for row in dic}
-
-    ns = Namespace(**ns_dict)
-    return ns
-
-
-def convert(val):
-    constructors = [int, float, str]
-
-    if type(val) is str:
-        if val.lower() == 'true':
-            return True
-        if val.lower() == 'false':
-            return False
-
-    for c in constructors:
-        try:
-            return c(val)
-        except ValueError:
-            pass
-    return val
@@ -2,7 +2,8 @@ import torch

 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.trainer.trainer_io import load_hparams_from_tags_csv
+from pytorch_lightning.root_module.model_saving import ModelIO
 from pytorch_lightning.root_module.hooks import ModelHooks
 from pytorch_lightning.root_module.decorators import data_loader

@@ -23,6 +24,9 @@ class LightningModule(GradInformation, ModelIO, ModelHooks):

        # track if gpu was requested for checkpointing
        self.on_gpu = False
+        self.use_dp = False
+        self.use_ddp = False
+        self.use_amp = False

    def forward(self, *args, **kwargs):
        """
@@ -33,29 +37,52 @@ class LightningModule(GradInformation, ModelIO, ModelHooks):
        """
        raise NotImplementedError

-    def validation_step(self, data_batch, batch_nb):
+    def training_step(self, *args, **kwargs):
        """
-        return whatever outputs will need to be aggregated in validation_end
-        :param data_batch:
+        return loss, dict with metrics for tqdm
+        :param called with batch, batch_nb
+        additional: optimizer_i if multiple optimizers used
        :return:
        """
        raise NotImplementedError

+    def validation_step(self, *args, **kwargs):
+        """
+        return whatever outputs will need to be aggregated in validation_end
+        OPTIONAL
+        :param called with batch, batch_nb
+        additional: dataset_i if multiple val datasets used
+        :return:
+        """
+        pass
+
+    def test_step(self, *args, **kwargs):
+        """
+        return whatever outputs will need to be aggregated in test_end
+        OPTIONAL
+        :param called with batch, batch_nb
+        additional: dataset_i if multiple val datasets used
+        :return:
+        """
+        pass
+
    def validation_end(self, outputs):
        """
        Outputs has the appended output after each validation step
+        OPTIONAL
        :param outputs:
        :return: dic_with_metrics for tqdm
        """
-        raise NotImplementedError
+        pass

-    def training_step(self, data_batch, batch_nb):
+    def test_end(self, outputs):
        """
-        return loss, dict with metrics for tqdm
-        :param data_batch:
-        :return:
+        Outputs has the appended output after each test step
+        OPTIONAL
+        :param outputs:
+        :return: dic_with_metrics for tqdm
        """
-        raise NotImplementedError
+        pass

    def configure_optimizers(self):
        """
@@ -64,10 +91,24 @@ class LightningModule(GradInformation, ModelIO, ModelHooks):
        """
        raise NotImplementedError

+    def optimizer_step(self, epoch_nb, batch_nb, optimizer, optimizer_i):
+        """
+        Do something instead of the standard optimizer behavior
+        :param epoch_nb:
+        :param batch_nb:
+        :param optimizer:
+        :param optimizer_i:
+        :return:
+        """
+        optimizer.step()
+
+        # clear gradients
+        optimizer.zero_grad()
+
    @data_loader
    def tng_dataloader(self):
        """
-        Implement a function to load an h5py of this data
+        Implement a PyTorch DataLoader
        :return:
        """
        raise NotImplementedError
@@ -75,21 +116,21 @@ class LightningModule(GradInformation, ModelIO, ModelHooks):
    @data_loader
    def test_dataloader(self):
        """
-        Implement a function to load an h5py of this data
+        Implement a PyTorch DataLoader
        :return:
        """
-        raise NotImplementedError
+        return None

    @data_loader
    def val_dataloader(self):
        """
-        Implement a function to load an h5py of this data
+        Implement a PyTorch DataLoader
        :return:
        """
-        raise NotImplementedError
+        return None

    @classmethod
-    def load_from_metrics(cls, weights_path, tags_csv, on_gpu, map_location=None):
+    def load_from_metrics(cls, weights_path, tags_csv, on_gpu):
        """
        Primary way of loading model from csv weights path
        :param weights_path:
@@ -101,13 +142,9 @@ class LightningModule(GradInformation, ModelIO, ModelHooks):
        hparams = load_hparams_from_tags_csv(tags_csv)
        hparams.__setattr__('on_gpu', on_gpu)

-        if on_gpu:
-            if map_location is not None:
-                checkpoint = torch.load(weights_path, map_location=map_location)
-            else:
-                checkpoint = torch.load(weights_path)
-        else:
-            checkpoint = torch.load(weights_path, map_location=lambda storage, loc: storage)
+        # load on CPU only to avoid OOM issues
+        # then its up to user to put back on GPUs
+        checkpoint = torch.load(weights_path, map_location=lambda storage, loc: storage)

        # load the state_dict on the model automatically
        model = cls(hparams)
@@ -0,0 +1,12 @@
+from .lm_test_module import LightningTestModel
+from .lm_test_module_base import LightningTestModelBase
+from .lm_test_module_mixins import (
+    LightningValidationStepMixin,
+    LightningValidationMixin,
+    LightningValidationStepMultipleDataloadersMixin,
+    LightningValidationMultipleDataloadersMixin,
+    LightningTestStepMixin,
+    LightningTestMixin,
+    LightningTestStepMultipleDataloadersMixin,
+    LightningTestMultipleDataloadersMixin,
+)
@@ -14,256 +14,14 @@ from test_tube import HyperOptArgumentParser
 from pytorch_lightning.root_module.root_module import LightningModule
 from pytorch_lightning import data_loader

+from .lm_test_module_base import LightningTestModelBase
+from .lm_test_module_mixins import LightningValidationMixin, LightningTestMixin

-class LightningTestModel(LightningModule):
+
+class LightningTestModel(LightningValidationMixin, LightningTestMixin, LightningTestModelBase):
    """
-    Sample model to show how to define a template
+    Most common test case. Validation and test dataloaders
    """

-    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 batch_i % 1 == 0:
-            output = OrderedDict({
-                'val_loss': loss_val,
-                'val_acc': val_acc,
-            })
-            return output
-        if batch_i % 2 == 0:
-            return val_acc
-
-        if batch_i % 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:
-            pass
-
-        should_shuffle = train_sampler is None
-        loader = DataLoader(
-            dataset=dataset,
-            batch_size=batch_size,
-            shuffle=should_shuffle,
-            sampler=train_sampler
-        )
-
-        return loader
-
-    @data_loader
-    def tng_dataloader(self):
-        return self.__dataloader(train=True)
-
-    @data_loader
-    def val_dataloader(self):
-        return self.__dataloader(train=False)
-
-    @data_loader
-    def test_dataloader(self):
-        return self.__dataloader(train=False)
-
-    @staticmethod
-    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:
-        :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)
-        # use 500 for CPU, 50000 for GPU to see speed difference
-        parser.add_argument('--hidden_dim', default=50000, type=int)
-
-        # 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 gpus being used across all nodes')
-        return parser
@@ -0,0 +1,194 @@
+import os
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import optim
+from torch.utils.data import DataLoader
+from torch.utils.data.distributed import DistributedSampler
+from torchvision.datasets import MNIST
+from torchvision import transforms
+from test_tube import HyperOptArgumentParser
+
+from pytorch_lightning.root_module.root_module import LightningModule
+from pytorch_lightning import data_loader
+
+
+class LightningTestModelBase(LightningModule):
+    """
+    Base LightningModule for testing. Implements only the required
+    interface
+    """
+
+    def __init__(self, hparams, force_remove_distributed_sampler=False):
+        """
+        Pass in parsed HyperOptArgumentParser to the model
+        :param hparams:
+        """
+        # init superclass
+        super(LightningTestModelBase, 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
+
+    # ---------------------
+    # 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.use_ddp 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:
+            pass
+
+        should_shuffle = train_sampler is None
+        loader = DataLoader(
+            dataset=dataset,
+            batch_size=batch_size,
+            shuffle=should_shuffle,
+            sampler=train_sampler
+        )
+
+        return loader
+
+    @data_loader
+    def tng_dataloader(self):
+        return self._dataloader(train=True)
+
+    @staticmethod
+    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:
+        :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)
+        # use 500 for CPU, 50000 for GPU to see speed difference
+        parser.add_argument('--hidden_dim', default=50000, type=int)
+
+        # 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 gpus being used across all nodes')
+        return parser
@@ -0,0 +1,387 @@
+import os
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import optim
+from torch.utils.data import DataLoader
+from torch.utils.data.distributed import DistributedSampler
+from torchvision.datasets import MNIST
+from torchvision import transforms
+from test_tube import HyperOptArgumentParser
+
+from pytorch_lightning.root_module.root_module import LightningModule
+from pytorch_lightning import data_loader
+
+
+class LightningValidationStepMixin:
+    """
+    Add val_dataloader and validation_step methods for the case
+    when val_dataloader returns a single dataloader
+    """
+
+    @data_loader
+    def val_dataloader(self):
+        return self._dataloader(train=False)
+
+    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 batch_i % 1 == 0:
+            output = OrderedDict({
+                'val_loss': loss_val,
+                'val_acc': val_acc,
+            })
+            return output
+        if batch_i % 2 == 0:
+            return val_acc
+
+        if batch_i % 3 == 0:
+            output = OrderedDict({
+                'val_loss': loss_val,
+                'val_acc': val_acc,
+                'test_dic': {'val_loss_a': loss_val}
+            })
+            return output
+
+
+class LightningValidationMixin(LightningValidationStepMixin):
+    """
+    Add val_dataloader, validation_step, and validation_end methods for the case
+    when val_dataloader returns a single dataloader
+    """
+
+    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 = output['val_loss']
+
+            # reduce manually when using dp
+            if self.trainer.use_dp:
+                val_loss = torch.mean(val_loss)
+            val_loss_mean += val_loss
+
+            # reduce manually when using dp
+            val_acc = output['val_acc']
+            if self.trainer.use_dp:
+                val_acc = torch.mean(val_acc)
+
+            val_acc_mean += 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
+
+
+class LightningValidationStepMultipleDataloadersMixin:
+    """
+    Add val_dataloader and validation_step methods for the case
+    when val_dataloader returns multiple dataloaders
+    """
+
+    @data_loader
+    def val_dataloader(self):
+        return [self._dataloader(train=False), self._dataloader(train=False)]
+
+    def validation_step(self, data_batch, batch_i, dataloader_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 batch_i % 1 == 0:
+            output = OrderedDict({
+                'val_loss': loss_val,
+                'val_acc': val_acc,
+            })
+            return output
+        if batch_i % 2 == 0:
+            return val_acc
+
+        if batch_i % 3 == 0:
+            output = OrderedDict({
+                'val_loss': loss_val,
+                'val_acc': val_acc,
+                'test_dic': {'val_loss_a': loss_val}
+            })
+            return output
+        if batch_i % 5 == 0:
+            output = OrderedDict({
+                f'val_loss_{dataloader_i}': loss_val,
+                f'val_acc_{dataloader_i}': val_acc,
+            })
+            return output
+
+
+class LightningValidationMultipleDataloadersMixin(LightningValidationStepMultipleDataloadersMixin):
+    """
+    Add val_dataloader, validation_step, and validation_end methods for the case
+    when val_dataloader returns multiple dataloaders
+    """
+
+    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
+        i = 0
+        for dl_output in outputs:
+            for output in dl_output:
+                val_loss = output['val_loss']
+
+                # reduce manually when using dp
+                if self.trainer.use_dp:
+                    val_loss = torch.mean(val_loss)
+                val_loss_mean += val_loss
+
+                # reduce manually when using dp
+                val_acc = output['val_acc']
+                if self.trainer.use_dp:
+                    val_acc = torch.mean(val_acc)
+
+                val_acc_mean += val_acc
+                i += 1
+
+        val_loss_mean /= i
+        val_acc_mean /= i
+
+        tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()}
+        return tqdm_dic
+
+
+class LightningTestStepMixin:
+
+    @data_loader
+    def test_dataloader(self):
+        return self._dataloader(train=False)
+
+    def test_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_test = self.loss(y, y_hat)
+
+        # acc
+        labels_hat = torch.argmax(y_hat, dim=1)
+        test_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)
+        test_acc = torch.tensor(test_acc)
+
+        if self.on_gpu:
+            test_acc = test_acc.cuda(loss_test.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_test = loss_test.unsqueeze(0)
+            test_acc = test_acc.unsqueeze(0)
+
+        # alternate possible outputs to test
+        if batch_i % 1 == 0:
+            output = OrderedDict({
+                'test_loss': loss_test,
+                'test_acc': test_acc,
+            })
+            return output
+        if batch_i % 2 == 0:
+            return test_acc
+
+        if batch_i % 3 == 0:
+            output = OrderedDict({
+                'test_loss': loss_test,
+                'test_acc': test_acc,
+                'test_dic': {'test_loss_a': loss_test}
+            })
+            return output
+
+
+class LightningTestMixin(LightningTestStepMixin):
+    def test_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 test_step, outputs is a list of tensor scalars
+        # we return just the average in this case (if we want)
+        # return torch.stack(outputs).mean()
+        test_loss_mean = 0
+        test_acc_mean = 0
+        for output in outputs:
+            test_loss = output['test_loss']
+
+            # reduce manually when using dp
+            if self.trainer.use_dp:
+                test_loss = torch.mean(test_loss)
+            test_loss_mean += test_loss
+
+            # reduce manually when using dp
+            test_acc = output['test_acc']
+            if self.trainer.use_dp:
+                test_acc = torch.mean(test_acc)
+
+            test_acc_mean += test_acc
+
+        test_loss_mean /= len(outputs)
+        test_acc_mean /= len(outputs)
+
+        tqdm_dic = {'test_loss': test_loss_mean.item(), 'test_acc': test_acc_mean.item()}
+        return tqdm_dic
+
+
+class LightningTestStepMultipleDataloadersMixin:
+
+    @data_loader
+    def test_dataloader(self):
+        return [self._dataloader(train=False), self._dataloader(train=False)]
+
+    def test_step(self, data_batch, batch_i, dataloader_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_test = self.loss(y, y_hat)
+
+        # acc
+        labels_hat = torch.argmax(y_hat, dim=1)
+        test_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)
+        test_acc = torch.tensor(test_acc)
+
+        if self.on_gpu:
+            test_acc = test_acc.cuda(loss_test.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_test = loss_test.unsqueeze(0)
+            test_acc = test_acc.unsqueeze(0)
+
+        # alternate possible outputs to test
+        if batch_i % 1 == 0:
+            output = OrderedDict({
+                'test_loss': loss_test,
+                'test_acc': test_acc,
+            })
+            return output
+        if batch_i % 2 == 0:
+            return test_acc
+
+        if batch_i % 3 == 0:
+            output = OrderedDict({
+                'test_loss': loss_test,
+                'test_acc': test_acc,
+                'test_dic': {'test_loss_a': loss_test}
+            })
+            return output
+        if batch_i % 5 == 0:
+            output = OrderedDict({
+                f'test_loss_{dataloader_i}': loss_test,
+                f'test_acc_{dataloader_i}': test_acc,
+            })
+            return output
+
+
+class LightningTestMultipleDataloadersMixin(LightningTestStepMultipleDataloadersMixin):
+    def test_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 test_step, outputs is a list of tensor scalars
+        # we return just the average in this case (if we want)
+        # return torch.stack(outputs).mean()
+        test_loss_mean = 0
+        test_acc_mean = 0
+        i = 0
+        for dl_output in outputs:
+            for output in dl_output:
+                test_loss = output['test_loss']
+
+                # reduce manually when using dp
+                if self.trainer.use_dp:
+                    test_loss = torch.mean(test_loss)
+                test_loss_mean += test_loss
+
+                # reduce manually when using dp
+                test_acc = output['test_acc']
+                if self.trainer.use_dp:
+                    test_acc = torch.mean(test_acc)
+
+                test_acc_mean += test_acc
+                i += 1
+
+        test_loss_mean /= i
+        test_acc_mean /= i
+
+        tqdm_dic = {'test_loss': test_loss_mean.item(), 'test_acc': test_acc_mean.item()}
+        return tqdm_dic
@@ -0,0 +1,325 @@
+import os
+import re
+import signal
+import pdb
+from subprocess import call
+
+import torch
+
+from pytorch_lightning.pt_overrides.override_data_parallel import (
+    LightningDistributedDataParallel, LightningDataParallel)
+
+
+class TrainerIO(object):
+
+    def __get_model(self):
+        is_dp_module = isinstance(self.model, (LightningDistributedDataParallel,
+                                               LightningDataParallel))
+        model = self.model.module if is_dp_module else self.model
+        return model
+
+    # --------------------
+    # CHECK-POINTING
+    # --------------------
+    def restore_weights(self, model):
+        """
+        To restore weights we have two cases.
+        First, if we use the same experiment version, then restore the latest ckpt.
+        AFTER that, if we find weights from hpc checkpoint, then restore that.
+        :param model:
+        :return:
+        """
+        # restore weights if same exp version
+        self.restore_state_if_checkpoint_exists(model)
+
+        # if script called from hpc resubmit, load weights
+        self.restore_hpc_weights_if_needed(model)
+
+    def restore_state_if_checkpoint_exists(self, model):
+        # do nothing if there's not dir or callback
+        no_ckpt_callback = self.checkpoint_callback is None
+        if no_ckpt_callback or not os.path.exists(self.checkpoint_callback.filepath):
+            return
+
+        # restore trainer state and model if there is a weight for this experiment
+        last_epoch = -1
+        last_ckpt_name = None
+
+        # find last epoch
+        checkpoints = os.listdir(self.checkpoint_callback.filepath)
+        for name in checkpoints:
+            # ignore hpc ckpts
+            if 'hpc_' in name:
+                continue
+
+            if '.ckpt' in name:
+                epoch = name.split('epoch_')[1]
+                epoch = int(re.sub('[^0-9]', '', epoch))
+
+                if epoch > last_epoch:
+                    last_epoch = epoch
+                    last_ckpt_name = name
+
+        # restore last checkpoint
+        if last_ckpt_name is not None:
+            last_ckpt_path = os.path.join(self.checkpoint_callback.filepath, last_ckpt_name)
+            self.restore(last_ckpt_path, self.on_gpu)
+            print(f'model and trainer restored from checkpoint: {last_ckpt_path}')
+
+    # --------------------
+    # HPC SIGNAL HANDLING
+    # --------------------
+    def register_slurm_signal_handlers(self):
+        # see if we're using slurm (not interactive)
+        on_slurm = False
+        try:
+            job_name = os.environ['SLURM_JOB_NAME']
+            if job_name != 'bash':
+                on_slurm = True
+        except Exception as e:
+            pass
+
+        if on_slurm:
+            print('set slurm handle signals')
+            signal.signal(signal.SIGUSR1, self.sig_handler)
+            signal.signal(signal.SIGTERM, self.term_handler)
+
+    def sig_handler(self, signum, frame):
+        if self.proc_rank == 0:
+            # save weights
+            print('handling SIGUSR1')
+            self.hpc_save(self.weights_save_path, self.experiment)
+
+            # find job id
+            job_id = os.environ['SLURM_JOB_ID']
+            cmd = 'scontrol requeue {}'.format(job_id)
+
+            # requeue job
+            print('\nrequeing job {}...'.format(job_id))
+            result = call(cmd, shell=True)
+
+            # print result text
+            if result == 0:
+                print('requeued exp ', job_id)
+            else:
+                print('requeue failed...')
+
+            # close experiment to avoid issues
+            self.experiment.close()
+
+    def term_handler(self, signum, frame):
+        # save
+        print("bypassing sigterm")
+
+    # --------------------
+    # MODEL SAVE CHECKPOINT
+    # --------------------
+    def save_checkpoint(self, filepath):
+        checkpoint = self.dump_checkpoint()
+
+        # do the actual save
+        torch.save(checkpoint, filepath)
+
+    def restore(self, checkpoint_path, on_gpu):
+
+        # if on_gpu:
+        #     checkpoint = torch.load(checkpoint_path)
+        # else:
+        # load on CPU first
+        checkpoint = torch.load(checkpoint_path, map_location=lambda storage, loc: storage)
+
+        # load model state
+        model = self.__get_model()
+
+        # load the state_dict on the model automatically
+        model.load_state_dict(checkpoint['state_dict'])
+        if on_gpu:
+            model.cuda(self.root_gpu)
+
+        # load training state (affects trainer only)
+        self.restore_training_state(checkpoint)
+
+    def dump_checkpoint(self):
+
+        checkpoint = {
+            'epoch': self.current_epoch,
+            '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())
+
+        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)
+
+        return checkpoint
+
+    # --------------------
+    # HPC IO
+    # --------------------
+    def restore_hpc_weights_if_needed(self, model):
+        """
+        If there is a set of hpc weights, use as signal to restore model
+        :param model:
+        :return:
+        """
+        # look for hpc weights
+        folderpath = self.weights_save_path
+        if os.path.exists(folderpath):
+            files = os.listdir(folderpath)
+            hpc_weight_paths = [x for x in files if 'hpc_ckpt' in x]
+
+            # if hpc weights exist restore model
+            if len(hpc_weight_paths) > 0:
+                self.hpc_load(folderpath, self.on_gpu)
+
+    def restore_training_state(self, checkpoint):
+        """
+        Restore trainer state.
+        Model will get its change to update
+        :param checkpoint:
+        :return:
+        """
+        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']
+
+        # restore the optimizers
+        optimizer_states = checkpoint['optimizer_states']
+        for optimizer, opt_state in zip(self.optimizers, optimizer_states):
+            optimizer.load_state_dict(opt_state)
+
+            # move optimizer to GPU 1 weight at a time
+            # avoids OOM
+            if self.root_gpu is not None:
+                for state in optimizer.state.values():
+                    for k, v in state.items():
+                        if isinstance(v, torch.Tensor):
+                            state[k] = v.cuda(self.root_gpu)
+
+        # 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
+    # ----------------------------------
+    def hpc_save(self, folderpath, experiment):
+        # make sure the checkpoint folder exists
+        os.makedirs(folderpath, exist_ok=True)
+
+        # save exp to make sure we get all the metrics
+        experiment.save()
+
+        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)
+
+        # 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, filepath)
+
+        return filepath
+
+    def hpc_load(self, folderpath, on_gpu):
+        filepath = '{}/hpc_ckpt_{}.ckpt'.format(folderpath, self.max_ckpt_in_folder(folderpath))
+
+        # load on CPU first
+        checkpoint = torch.load(filepath, map_location=lambda storage, loc: storage)
+
+        # load model state
+        model = self.__get_model()
+
+        # load the state_dict on the model automatically
+        model.load_state_dict(checkpoint['state_dict'])
+
+        if self.root_gpu is not None:
+            model.cuda(self.root_gpu)
+
+        # load training state (affects trainer only)
+        self.restore_training_state(checkpoint)
+
+        # call model hook
+        model.on_hpc_load(checkpoint)
+
+        print(f'restored hpc model from: {filepath}')
+
+    def max_ckpt_in_folder(self, path, name_key='ckpt_'):
+        files = os.listdir(path)
+        files = [x for x in files if name_key in x]
+        if len(files) == 0:
+            return 0
+
+        ckpt_vs = []
+        for name in files:
+            name = name.split(name_key)[-1]
+            name = re.sub('[^0-9]', '', name)
+            ckpt_vs.append(int(name))
+
+        return max(ckpt_vs)
+
+
+def load_hparams_from_tags_csv(tags_csv):
+    from argparse import Namespace
+    import pandas as pd
+
+    tags_df = pd.read_csv(tags_csv)
+    dic = tags_df.to_dict(orient='records')
+
+    ns_dict = {row['key']: convert(row['value']) for row in dic}
+
+    ns = Namespace(**ns_dict)
+    return ns
+
+
+def convert(val):
+    constructors = [int, float, str]
+
+    if type(val) is str:
+        if val.lower() == 'true':
+            return True
+        if val.lower() == 'false':
+            return False
+
+    for c in constructors:
+        try:
+            return c(val)
+        except ValueError:
+            pass
+    return val
@@ -45,6 +45,7 @@ omit =
    tests/test_models.py
    pytorch_lightning/testing_models/lm_test_module.py
    pytorch_lightning/utilities/arg_parse.py
+    examples/templates

 [flake8]
 ignore = E731,W504,F401,F841
@@ -14,7 +14,7 @@ from setuptools import setup, find_packages
 # engineer specific practices
 setup(
    name='pytorch-lightning',
-    version='0.4.3',
+    version='0.4.9',
    description='The Keras for ML researchers using PyTorch',
    author='William Falcon',
    author_email='waf2107@columbia.edu',
@@ -30,8 +30,8 @@ setup(
    python_requires='>=3.6',
    install_requires=[
        'torch==1.2.0',
-        'tqdm',
-        'test-tube==0.6.8',
+        'tqdm>=4.35.0',
+        'test-tube>=0.6.9',
        'pandas>=0.20.3',
    ],
    classifiers=[
@@ -1,5 +1,6 @@
 from pytorch_lightning import Trainer
 from examples import LightningTemplateModel
+from pytorch_lightning.testing import LightningTestModel
 from argparse import Namespace
 from test_tube import Experiment
 from pytorch_lightning.callbacks import ModelCheckpoint
@@ -11,6 +12,8 @@ import torch
 from torch.nn import functional as F
 from torch.utils.data import DataLoader
 from torchvision.datasets import MNIST
+import numpy as np
+import pdb


 class CoolModel(pl.LightningModule):
@@ -72,10 +75,11 @@ def get_model():
    return model, hparams


-def get_exp(debug=True):
+def get_exp(debug=True, version=None):
    # 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')
+    save_dir = os.path.join(root_dir, 'save_dir')
+    exp = Experiment(debug=debug, save_dir=save_dir, name='tests_tt_dir', version=version)
    return exp


@@ -98,7 +102,7 @@ def clear_save_dir():
        shutil.rmtree(save_dir)


-def load_model(exp, save_dir):
+def load_model(exp, save_dir, on_gpu, map_location=None, module_class=LightningTemplateModel):

    # load trained model
    tags_path = exp.get_data_path(exp.name, exp.version)
@@ -107,8 +111,10 @@ def load_model(exp, save_dir):
    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)
+    trained_model = module_class.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'

@@ -130,45 +136,160 @@ def run_prediction(dataloader, trained_model):
    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, 'this model is expected to get > 0.7 in test set (it got %f)' % val_acc


-def main():
-
+# ------------------------------------------------------------------------
+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)

-    trainer = Trainer(
-        experiment=exp,
-        checkpoint_callback=checkpoint,
-        progress_bar=True,
-        max_nb_epochs=1,
-        gpus=[0, 1],
-        distributed_backend='dp',
-    )
-
-    model = CoolModel()
+    # 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)
+    pretrained_model = load_model(exp, save_dir, on_gpu)

-    # test model preds
+    # test new model accuracy
    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 assert_ok_val_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}'
+
+
+def assert_ok_test_acc(trainer):
+    # this model should get 0.80+ acc
+    acc = trainer.tng_tqdm_dic['test_acc']
+    assert acc > 0.50, f'model failed to get expected 0.50 validation accuracy. Got: {acc}'
+
+
+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 main():
+    """
+    Make sure DDP + AMP continue training correctly
+    :return:
+    """
+    hparams = get_hparams()
+    model = LightningTestModel(hparams)
+
+    trainer_options = dict(
+        show_progress_bar=True,
+        max_nb_epochs=4,
+        gpus=2,
+        distributed_backend='dp',
+    )
+
+    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['experiment'] = exp
+    trainer_options['checkpoint_callback'] = checkpoint
+
+    # fit model
+    trainer = Trainer(**trainer_options)
+    trainer.is_slurm_managing_tasks = True
+    result = trainer.fit(model)
+
+    # track epoch before saving
+    real_global_epoch = trainer.current_epoch
+
+    # correct result and ok accuracy
+    assert result == 1, 'amp + dp model failed to complete'
+
+    # ---------------------------
+    # HPC LOAD/SAVE
+    # ---------------------------
+    # save
+    trainer.hpc_save(save_dir, exp)
+
+    # init new trainer
+    new_exp = get_exp(False, version=exp.version)
+    trainer_options['experiment'] = new_exp
+    trainer_options['checkpoint_callback'] = ModelCheckpoint(save_dir)
+    trainer_options['train_percent_check'] = 0.2
+    trainer_options['val_percent_check'] = 0.2
+    trainer_options['max_nb_epochs'] = 1
+    new_trainer = Trainer(**trainer_options)
+
+    # set the epoch start hook so we can predict before the model does the full training
+    def assert_good_acc():
+        assert trainer.current_epoch == real_global_epoch and trainer.current_epoch > 0
+
+        # if model and state loaded correctly, predictions will be good even though we
+        # haven't trained with the new loaded model
+        dp_model = new_trainer.model
+        dp_model.eval()
+
+        _ = [run_prediction(dataloader, dp_model, dp=True) for dataloader in trainer.val_dataloader]
+
+    # new model
+    model = LightningTestModel(hparams)
+    model.on_sanity_check_start = assert_good_acc
+
+    # fit new model which should load hpc weights
+    new_trainer.fit(model)
+
+    # test freeze on gpu
+    model.freeze()
+    model.unfreeze()
+
    clear_save_dir()


@@ -34,11 +34,11 @@ deps =
 commands =
    check-manifest --ignore tox.ini
    python setup.py check -m -s
-    coverage run --source pytorch_lightning -m py.test pytorch_lightning tests examples -v --doctest-modules
    flake8 .
+    coverage run --source pytorch_lightning -m py.test pytorch_lightning tests examples -v --doctest-modules

 [flake8]
-exclude = .tox,*.egg,build,temp
+exclude = .tox,*.egg,build,temp,examples/*
 select = E,W,F
 doctests = True
 verbose = 2