release v0.2.6

add a hook for on_tng_metrics so that users get access to the grad_no…

README.md

@@ -116,8 +116,8 @@ def validation_end(self, outputs):
     return tqdm_dic
 ```
    
-## TensorboardX    
-Lightning is fully integrated with tensorboardX.   
+## Tensorboard    
+Lightning is fully integrated with tensorboard.   
 
 <p align="center">
   <a href="https://williamfalcon.github.io/pytorch-lightning/">

docs/Trainer/Training Loop.md

@@ -19,7 +19,7 @@ Cut the learning rate by 10 at every epoch listed in this list.
 trainer = Trainer(lr_scheduler_milestones=None)
 
 # cut LR by 10 at 100, 200, and 300 epochs 
-trainer = Trainer(lr_scheduler_milestones=[100, 200, 300])
+trainer = Trainer(lr_scheduler_milestones='100, 200, 300')
 ```
 
 ---

examples/new_project_templates/lightning_module_template.py

@@ -7,6 +7,8 @@ import torch
 import torch.nn.functional as F
 from test_tube import HyperOptArgumentParser
 from torch import optim
+from torch.utils.data import DataLoader
+from torch.utils.data.distributed import DistributedSampler
 
 from pytorch_lightning.root_module.root_module import LightningModule
 
@@ -82,9 +84,10 @@ class LightningTemplateModel(LightningModule):
         loss_val = self.loss(y, y_hat)
 
         output = OrderedDict({
-            'loss': loss_val,
-            'tqdm_metrics': {}
+            'loss': loss_val
         })
+
+        # can also return just a scalar instead of a dict (return loss_val)
         return output
 
     def validation_step(self, data_batch, batch_i):
@@ -105,8 +108,10 @@ class LightningTemplateModel(LightningModule):
 
         output = OrderedDict({
             'val_loss': loss_val,
-            'val_acc': torch.tensor(val_acc),
+            'val_acc': torch.tensor(val_acc).cuda(loss_val.device.index),
         })
+
+        # can also return just a scalar instead of a dict (return loss_val)
         return output
 
     def validation_end(self, outputs):
@@ -115,6 +120,10 @@ class LightningTemplateModel(LightningModule):
         :param outputs: list of individual outputs of each validation step
         :return:
         """
+        # if returned a scalar from validation_step, outputs is a list of tensor scalars
+        # we return just the average in this case (if we want)
+        # return torch.stack(outputs).mean()
+
         val_loss_mean = 0
         val_acc_mean = 0
         for output in outputs:
@@ -154,13 +163,25 @@ class LightningTemplateModel(LightningModule):
     def __dataloader(self, train):
         # init data generators
         transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
-
         dataset = MNIST(root=self.hparams.data_root, train=train, transform=transform, download=True)
 
-        loader = torch.utils.data.DataLoader(
+        # when using multi-node we need to add the datasampler
+        train_sampler = None
+        batch_size = self.hparams.batch_size
+
+        try:
+            if self.on_gpu:
+                train_sampler = DistributedSampler(dataset, rank=self.trainer.proc_rank)
+                batch_size = batch_size // self.trainer.world_size  # scale batch size
+        except Exception as e:
+            pass
+
+        should_shuffle = train_sampler is None
+        loader = DataLoader(
             dataset=dataset,
-            batch_size=self.hparams.batch_size,
-            shuffle=True
+            batch_size=batch_size,
+            shuffle=should_shuffle,
+            sampler=train_sampler
         )
 
         return loader
@@ -210,16 +231,19 @@ class LightningTemplateModel(LightningModule):
 
         # network params
         parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=False)
-        parser.add_argument('--in_features', default=28*28)
-        parser.add_argument('--out_features', default=10)
-        parser.add_argument('--hidden_dim', default=50000) # use 500 for CPU, 50000 for GPU to see speed difference
+        parser.add_argument('--in_features', default=28*28, type=int)
+        parser.add_argument('--out_features', default=10, type=int)
+        parser.add_argument('--hidden_dim', default=50000, type=int) # use 500 for CPU, 50000 for GPU to see speed difference
 
         # data
         parser.add_argument('--data_root', default=os.path.join(root_dir, 'mnist'), type=str)
 
         # training params (opt)
-        parser.opt_list('--learning_rate', default=0.001, type=float, options=[0.0001, 0.0005, 0.001, 0.005],
+        parser.opt_list('--learning_rate', default=0.001*8, type=float, options=[0.0001, 0.0005, 0.001, 0.005],
                         tunable=False)
-        parser.opt_list('--batch_size', default=256, type=int, options=[32, 64, 128, 256], tunable=False)
         parser.opt_list('--optimizer_name', default='adam', type=str, options=['adam'], tunable=False)
+
+        # if using 2 nodes with 4 gpus each the batch size here (256) will be 256 / (2*8) = 16 per gpu
+        parser.opt_list('--batch_size', default=256*8, type=int, options=[32, 64, 128, 256], tunable=False,
+                        help='batch size will be divided over all the gpus being used across all nodes')
         return parser

examples/new_project_templates/multi_node_cluster_template.py

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

examples/new_project_templates/single_cpu_template.py

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

examples/new_project_templates/single_gpu_node_16bit_template.py

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

examples/new_project_templates/single_gpu_node_dp_template.py

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

examples/new_project_templates/single_gpu_node_template.py

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

examples/new_project_templates/single_gpu_template.py

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

examples/new_project_templates/trainer_gpu_cluster_template.py

@@ -1,208 +0,0 @@
-import os
-import sys
-import numpy as np
-from time import sleep
-import torch
-
-from test_tube import HyperOptArgumentParser, Experiment, SlurmCluster
-from pytorch_lightning.models.trainer import Trainer
-from pytorch_lightning.utils.arg_parse import add_default_args
-
-from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
-
-SEED = 2334
-torch.manual_seed(SEED)
-np.random.seed(SEED)
-
-# ---------------------
-# DEFINE MODEL HERE
-# ---------------------
-from lightning_module_template import LightningTemplateModel
-# ---------------------
-
-AVAILABLE_MODELS = {
-    'model_template': LightningTemplateModel
-}
-
-
-"""
-Allows training by using command line arguments
-Run by:
-# TYPE YOUR RUN COMMAND HERE
-"""
-
-
-def main_local(hparams):
-    main(hparams, None, None)
-
-
-def main(hparams, cluster, results_dict):
-    """
-    Main training routine specific for this project
-    :param hparams:
-    :return:
-    """
-    on_gpu = hparams.gpus is not None and torch.cuda.is_available()
-
-    device = 'cuda' if on_gpu else 'cpu'
-    hparams.__setattr__('device', device)
-    hparams.__setattr__('on_gpu', on_gpu)
-    hparams.__setattr__('nb_gpus', torch.cuda.device_count())
-    hparams.__setattr__('inference_mode', hparams.model_load_weights_path is not None)
-
-    # delay each training start to not overwrite logs
-    process_position, current_gpu = TRAINING_MODEL.get_process_position(hparams.gpus)
-    sleep(process_position + 1)
-
-    # init experiment
-    log_dir = os.path.dirname(os.path.realpath(__file__))
-    log_dir = os.path.join(log_dir, 'test_tube_demo_logs')
-    exp = Experiment(
-        name='test_tube_exp',
-        save_dir=log_dir,
-        autosave=False,
-        description='test demo'
-    )
-
-    exp.argparse(hparams)
-    exp.save()
-
-    # build model
-    print('loading model...')
-    model = TRAINING_MODEL(hparams)
-    print('model built')
-
-    # callbacks
-    early_stop = EarlyStopping(
-        monitor=hparams.early_stop_metric,
-        patience=hparams.early_stop_patience,
-        verbose=True,
-        mode=hparams.early_stop_mode
-    )
-
-    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
-    checkpoint = ModelCheckpoint(
-        filepath=model_save_path,
-        save_best_only=True,
-        verbose=True,
-        monitor=hparams.model_save_monitor_value,
-        mode=hparams.model_save_monitor_mode
-    )
-
-    # gpus are ; separated for inside a node and , within nodes
-    gpu_list = None
-    if hparams.gpus is not None:
-        gpu_list = [int(x) for x in hparams.gpus.split(';')]
-
-    # configure trainer
-    trainer = Trainer(
-        experiment=exp,
-        cluster=cluster,
-        checkpoint_callback=checkpoint,
-        early_stop_callback=early_stop,
-        gpus=gpu_list,
-    )
-
-    # train model
-    trainer.fit(model)
-
-
-def get_default_parser(strategy, root_dir):
-
-    possible_model_names = list(AVAILABLE_MODELS.keys())
-    parser = HyperOptArgumentParser(strategy=strategy, add_help=False)
-    add_default_args(parser, root_dir, possible_model_names=possible_model_names, rand_seed=SEED)
-    return parser
-
-
-def get_model_name(args):
-    for i, arg in enumerate(args):
-        if 'model_name' in arg:
-            return args[i+1]
-
-
-def optimize_on_cluster(hyperparams):
-    # enable cluster training
-    cluster = SlurmCluster(
-        hyperparam_optimizer=hyperparams,
-        log_path=hyperparams.tt_save_path,
-        test_tube_exp_name=hyperparams.tt_name
-    )
-
-    # email for cluster coms
-    cluster.notify_job_status(email='add_email_here', on_done=True, on_fail=True)
-
-    # configure cluster
-    cluster.per_experiment_nb_gpus = hyperparams.per_experiment_nb_gpus
-    cluster.job_time = '48:00:00'
-    cluster.gpu_type = '1080ti'
-    cluster.memory_mb_per_node = 48000
-
-    # any modules for code to run in env
-    cluster.add_command('source activate pytorch_lightning')
-
-    # name of exp
-    job_display_name = hyperparams.tt_name.split('_')[0]
-    job_display_name = job_display_name[0:3]
-
-    # run hopt
-    print('submitting jobs...')
-    cluster.optimize_parallel_cluster_gpu(
-        main,
-        nb_trials=hyperparams.nb_hopt_trials,
-        job_name=job_display_name
-    )
-
-
-if __name__ == '__main__':
-
-    model_name = get_model_name(sys.argv)
-    if model_name is None:
-        model_name = 'model_template'
-
-    # use default args
-    root_dir = os.path.dirname(os.path.realpath(__file__))
-    parent_parser = get_default_parser(strategy='random_search', root_dir=root_dir)
-
-    # allow model to overwrite or extend args
-    TRAINING_MODEL = AVAILABLE_MODELS[model_name]
-    parser = TRAINING_MODEL.add_model_specific_args(parent_parser, root_dir)
-    hyperparams = parser.parse_args()
-
-    # format GPU layout
-    os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
-
-    # ---------------------
-    # RUN TRAINING
-    # ---------------------
-
-    # cluster and CPU
-    if hyperparams.on_cluster:
-        # run on HPC cluster
-        print('RUNNING ON SLURM CLUSTER')
-        gpu_ids = hyperparams.gpus.split(';')
-        os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(gpu_ids)
-        optimize_on_cluster(hyperparams)
-
-    elif hyperparams.gpus is None:
-        # run on cpu
-        print('RUNNING ON CPU')
-        main(hyperparams, None, None)
-
-    # single or multiple GPUs on same machine
-    gpu_ids = hyperparams.gpus.split(';')
-    if hyperparams.interactive:
-        # run on 1 gpu
-        print(f'RUNNING INTERACTIVE MODE ON GPUS. gpu ids: {gpu_ids}')
-        os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(gpu_ids)
-        main(hyperparams, None, None)
-
-    else:
-        # multiple GPUs on same machine
-        print(f'RUNNING MULTI GPU. GPU ids: {gpu_ids}')
-        hyperparams.optimize_parallel_gpu(
-            main_local,
-            gpu_ids=gpu_ids,
-            nb_trials=hyperparams.nb_hopt_trials,
-            nb_workers=len(gpu_ids)
-        )

pytorch_lightning/__init__.py

@@ -1 +1,2 @@
-from .models import Trainer
+from .models import Trainer
+from .root_module.root_module import LightningModule

pytorch_lightning/callbacks/pt_callbacks.py

@@ -1,6 +1,6 @@
 import numpy as np
 import os, shutil
-from pytorch_lightning.pt_overrides.override_data_parallel import LightningDataParallel
+from pytorch_lightning.pt_overrides.override_data_parallel import LightningDistributedDataParallel
 
 
 class Callback(object):
@@ -34,7 +34,7 @@ class Callback(object):
         self.params = params
 
     def set_model(self, model):
-        if type(model) is LightningDataParallel:
+        if type(model) is LightningDistributedDataParallel:
             model = model.module
         self.model = model
 

pytorch_lightning/models/trainer.py

@@ -1,12 +1,25 @@
-import torch
-import tqdm
-import numpy as np
-from pytorch_lightning.root_module.memory import get_gpu_memory_map
+"""
+The trainer handles all the logic for running a val loop, training loop, distributing, etc...
+"""
+import subprocess
 import traceback
-from pytorch_lightning.root_module.model_saving import TrainerIO
-from torch.optim.lr_scheduler import MultiStepLR
-from pytorch_lightning.pt_overrides.override_data_parallel import LightningDataParallel
+import warnings
+import os
 import pdb
+import re
+
+import torch
+from torch.utils.data.distributed import DistributedSampler
+from torch.optim.lr_scheduler import MultiStepLR
+import torch.multiprocessing as mp
+import torch.distributed as dist
+import numpy as np
+import tqdm
+
+from pytorch_lightning.root_module.memory import get_gpu_memory_map
+from pytorch_lightning.root_module.model_saving import TrainerIO
+from pytorch_lightning.pt_overrides.override_data_parallel import LightningDistributedDataParallel, LightningDataParallel
+
 
 try:
     from apex import amp
@@ -16,6 +29,14 @@ except ModuleNotFoundError:
 
 
 def reduce_distributed_output(output, nb_gpus):
+    if nb_gpus <= 1:
+        return output
+
+    # when using DP, we get one output per gpu
+    # average outputs and return
+    if type(output) is torch.Tensor:
+        return output.mean()
+
     for k, v in output.items():
         # recurse on nested dics
         if isinstance(output[k], dict):
@@ -32,11 +53,13 @@ class Trainer(TrainerIO):
 
     def __init__(self,
                  experiment,
-                 checkpoint_callback, early_stop_callback,
+                 early_stop_callback=None,
+                 checkpoint_callback=None,
                  gradient_clip=0,
                  cluster=None,
                  process_position=0,
                  current_gpu_name=0,
+                 nb_gpu_nodes=1,
                  gpus=None,
                  progress_bar=True,
                  overfit_pct=0.0,
@@ -44,19 +67,57 @@ class Trainer(TrainerIO):
                  check_val_every_n_epoch=1,
                  fast_dev_run=False,
                  accumulate_grad_batches=1,
-                 enable_early_stop=True, max_nb_epochs=1000, min_nb_epochs=1,
-                 train_percent_check=1.0, val_percent_check=1.0, test_percent_check=1.0, val_check_interval=0.95,
+                 max_nb_epochs=1000, min_nb_epochs=1,
+                 train_percent_check=1.0, val_percent_check=1.0, test_percent_check=1.0,
+                 val_check_interval=0.95,
                  log_save_interval=100, add_log_row_interval=10,
                  lr_scheduler_milestones=None,
+                 distributed_backend='dp',
                  use_amp=False,
                  print_nan_grads=False,
+                 print_weights_summary=True,
                  amp_level='O2',
                  nb_sanity_val_steps=5):
 
+        """
+
+        :param experiment: Test-tube experiment
+        :param early_stop_callback: from pytorch_lightning import EarlyStopping
+        :param checkpoint_callback: from pytorch_lightning import Checkpoint
+        :param gradient_clip:
+        :param cluster:
+        :param process_position:
+        :param current_gpu_name:
+        :param nb_gpu_nodes:
+        :param gpus:
+        :param progress_bar:
+        :param overfit_pct:
+        :param track_grad_norm:
+        :param check_val_every_n_epoch:
+        :param fast_dev_run:
+        :param accumulate_grad_batches:
+        :param max_nb_epochs:
+        :param min_nb_epochs:
+        :param train_percent_check:
+        :param val_percent_check:
+        :param test_percent_check:
+        :param val_check_interval:
+        :param log_save_interval:
+        :param add_log_row_interval:
+        :param lr_scheduler_milestones:
+        :param distributed_backend: 'np' to use DistributedParallel, 'ddp' to use DistributedDataParallel
+        :param use_amp:
+        :param print_nan_grads:
+        :param print_weights_summary:
+        :param amp_level:
+        :param nb_sanity_val_steps:
+        """
+
         # Transfer params
+        self.nb_gpu_nodes = nb_gpu_nodes
         self.gradient_clip = gradient_clip
         self.check_val_every_n_epoch = check_val_every_n_epoch
-        self.enable_early_stop = enable_early_stop
+        self.enable_early_stop = early_stop_callback is not None
         self.track_grad_norm = track_grad_norm
         self.fast_dev_run = fast_dev_run
         self.on_gpu = gpus is not None and torch.cuda.is_available()
@@ -66,8 +127,12 @@ class Trainer(TrainerIO):
         self.cluster = cluster
         self.process_position = process_position
         self.current_gpu_name = current_gpu_name
+        self.print_weights_summary = print_weights_summary
         self.checkpoint_callback = checkpoint_callback
-        self.checkpoint_callback.save_function = self.save_checkpoint
+
+        if self.checkpoint_callback is not None:
+            self.checkpoint_callback.save_function = self.save_checkpoint
+
         self.early_stop = early_stop_callback
         self.model = None
         self.max_nb_epochs = max_nb_epochs
@@ -79,8 +144,35 @@ class Trainer(TrainerIO):
         self.lr_schedulers = []
         self.amp_level = amp_level
         self.print_nan_grads = print_nan_grads
-        self.data_parallel_device_ids = gpus
-        self.data_parallel = gpus is not None and len(gpus) > 0
+        self.data_parallel_device_ids = None
+        self.world_size = 1
+        self.node_rank = 0
+        self.use_ddp = False
+        self.use_dp = False
+
+        # gpus come in as a string.
+        # if gpus = -1 then use all available devices
+        # otherwise, split the string using commas
+        if gpus is not None:
+            if gpus == '-1':
+                self.data_parallel_device_ids = list(range(0, torch.cuda.device_count()))
+            else:
+                self.data_parallel_device_ids = [int(x.strip()) for x in gpus.split(',')]
+
+            # set the correct cuda visible devices (using pci order)
+            os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+            os.environ["CUDA_VISIBLE_DEVICES"] = ','.join([str(x) for x in self.data_parallel_device_ids])
+            print(f'VISIBLE GPUS: {os.environ["CUDA_VISIBLE_DEVICES"]}')
+
+        # make DP and DDP mutually exclusive
+        # single GPU will also use DP with devices=[0]
+        have_gpus = self.data_parallel_device_ids is not None and len(self.data_parallel_device_ids) > 0
+        if have_gpus:
+            self.use_dp = distributed_backend == 'dp'
+            self.use_ddp = distributed_backend == 'ddp'
+
+        # process info
+        self.proc_rank = 0
 
         # training state
         self.optimizers = None
@@ -103,11 +195,25 @@ class Trainer(TrainerIO):
         self.__determine_data_use_amount(train_percent_check, val_percent_check, test_percent_check, overfit_pct)
         print('gpu available: {}, used: {}'.format(torch.cuda.is_available(), self.on_gpu))
 
-        # apex test
+        # 16 bit mixed precision training using apex
         self.use_amp = use_amp and APEX_AVAILABLE
         if self.use_amp:
             print('using 16bit precision')
 
+        if use_amp and not APEX_AVAILABLE:
+            msg = '''
+            You set use_amp=True but do not have apex installed.
+            Install apex first using this guide and rerun with use_amp=True: 
+            https://github.com/NVIDIA/apex#linux
+            
+            this run will NOT use 16 bit precision
+            '''
+            warnings.warn(msg)
+
+    @property
+    def data_parallel(self):
+        return self.use_dp or self.use_ddp
+
     def __determine_data_use_amount(self, train_percent_check, val_percent_check, test_percent_check, overfit_pct):
         """
         Use less data for debugging purposes
@@ -120,8 +226,12 @@ class Trainer(TrainerIO):
             self.val_percent_check = overfit_pct
             self.test_percent_check = overfit_pct
 
+    def __get_model(self):
+        return self.model.module if self.data_parallel else self.model
+
     def __is_function_implemented(self, f_name):
-        f_op = getattr(self.model, f_name, None)
+        model = self.__get_model()
+        f_op = getattr(model, f_name, None)
         return callable(f_op)
 
     @property
@@ -139,7 +249,7 @@ class Trainer(TrainerIO):
 
         return tqdm_dic
 
-    def __layout_bookeeping(self, model):
+    def __layout_bookeeping(self):
         # training bookeeping
         self.total_batch_nb = 0
         self.running_loss = []
@@ -148,17 +258,17 @@ class Trainer(TrainerIO):
         self.tqdm_metrics = {}
 
         # determine number of training batches
-        self.nb_tng_batches = model.nb_batches(self.tng_dataloader)
+        self.nb_tng_batches = len(self.tng_dataloader)
         self.nb_tng_batches = int(self.nb_tng_batches * self.train_percent_check)
 
         # determine number of validation batches
-        self.nb_val_batches = model.nb_batches(self.val_dataloader)
+        self.nb_val_batches = len(self.val_dataloader)
         self.nb_val_batches = int(self.nb_val_batches * self.val_percent_check)
         self.nb_val_batches = max(1, self.nb_val_batches)
         self.nb_val_batches = self.nb_val_batches
 
         # determine number of test batches
-        self.nb_test_batches = model.nb_batches(self.test_dataloader)
+        self.nb_test_batches = len(self.test_dataloader)
         self.nb_test_batches = int(self.nb_test_batches * self.test_percent_check)
 
         # determine when to check validation
@@ -179,12 +289,9 @@ class Trainer(TrainerIO):
         :param max_batches: Scalar
         :return:
         """
-        print('validating...')
-
         # enable eval mode
         model.zero_grad()
         model.eval()
-        model.from_lightning = True
 
         # disable gradients to save memory
         torch.set_grad_enabled(False)
@@ -205,9 +312,12 @@ class Trainer(TrainerIO):
             # -----------------
             # RUN VALIDATION STEP
             # -----------------
-            if self.data_parallel:
+            if self.use_ddp:
+                output = model(data_batch, batch_i)
+            elif self.use_dp:
                 output = model(data_batch, batch_i)
                 output = reduce_distributed_output(output, len(self.data_parallel_device_ids))
+
             else:
                 output = model.validation_step(data_batch, batch_i)
 
@@ -241,32 +351,208 @@ class Trainer(TrainerIO):
         self.test_dataloader = model.test_dataloader
         self.val_dataloader = model.val_dataloader
 
+        if self.use_ddp and not isinstance(self.tng_dataloader.sampler, DistributedSampler):
+            msg = '''
+            when using multiple gpus and multiple nodes you must pass a DistributedSampler to DataLoader(sampler).
+            
+            ie: this:
+            dataset = myDataset()
+            dataloader = Dataloader(dataset)
+            
+            becomes:
+            dataset = myDataset()
+            dist_sampler = torch.utils.data.distributed.DistributedSampler(dataset)
+            dataloader = Dataloader(dataset, sampler=dist_sampler)
+            '''
+            raise Exception(msg)
+
     # -----------------------------
     # MODEL TRAINING
     # -----------------------------
     def fit(self, model):
 
-        # give model convenience properties
-        model.trainer = self
-        model.experiment = self.experiment
+        # when using multi-node or DDP within a node start each module in a separate process
+        if self.use_ddp:
+            # must copy only the meta of the exp so it survives pickle/unpickle when going to new process
+            self.experiment = self.experiment.get_meta_copy()
 
-        # transfer data loaders from model
-        self.__get_dataloaders(model)
+            # whenever we have the correct number of tasks, we let slurm manage processes
+            # otherwise we launch the required number of processes
+            try:
+                nb_slurm_tasks = int(os.environ['SLURM_NTASKS'])
+                nb_requested_gpus = len(self.data_parallel_device_ids) * self.nb_gpu_nodes
+                is_slurm_managing_tasks = nb_slurm_tasks == nb_requested_gpus
+            except Exception as e:
+                # likely not on slurm, so set the slurm managed flag to false
+                is_slurm_managing_tasks = False
 
-        # init training constants
-        self.__layout_bookeeping(model)
+            if is_slurm_managing_tasks:
+                task = int(os.environ['SLURM_LOCALID'])
+                self.ddp_train(task, model)
+            else:
+                msg = f"""
+                You requested {nb_requested_gpus} GPUs but launched {nb_slurm_tasks} slurm tasks. 
+                We will launch {nb_requested_gpus} processes for you. 
+                We recommend you let slurm manage the processes by setting: --ntasks-per-node={nb_requested_gpus}
+                If you're not using SLURM, ignore this message!
+                """
+                warnings.warn(msg)
+                mp.spawn(self.ddp_train, nprocs=len(self.data_parallel_device_ids), args=(model, ))
+
+        # 1 gpu or dp option triggers training using DP module
+        # easier to avoid NCCL issues
+        elif self.use_dp:
+            self.dp_train(model)
+
+        # ON CPU
+        else:
+            # CHOOSE OPTIMIZER
+            # filter out the weights that were done on gpu so we can load on good old cpus
+            self.optimizers = model.configure_optimizers()
+
+            # run through amp wrapper
+            if self.use_amp:
+                # An example
+                model, optimizers = amp.initialize(
+                    model, self.optimizers, opt_level=self.amp_level,
+                )
+                self.optimizers = optimizers
+
+            self.__run_pretrain_routine(model)
+
+    def dp_train(self, model):
 
         # CHOOSE OPTIMIZER
         # filter out the weights that were done on gpu so we can load on good old cpus
         self.optimizers = model.configure_optimizers()
 
+        model.cuda(self.data_parallel_device_ids[0])
+        model = LightningDataParallel(model, device_ids=self.data_parallel_device_ids)
+
+        # run through amp wrapper
         if self.use_amp:
             # An example
-            model, optimizer = amp.initialize(
-                model, self.optimizers[0], opt_level=self.amp_level,
+            model, optimizers = amp.initialize(
+                model, self.optimizers, opt_level=self.amp_level,
             )
-            self.optimizers[0] = optimizer
-            model.trainer = self
+            self.optimizers = optimizers
+
+        self.__run_pretrain_routine(model)
+
+    def ddp_train(self, gpu_nb, model):
+        """
+        Entry point into a DP thread
+        :param gpu_nb:
+        :param model:
+        :param cluster_obj:
+        :return:
+        """
+        # node rank using relative slurm id
+        # otherwise default to node rank 0
+        try:
+            node_id = os.environ['SLURM_NODEID']
+            self.node_rank = int(node_id)
+        except Exception as e:
+            self.node_rank = 0
+
+        # recover original exp before went into process
+        # init in write mode only on proc 0
+        self.experiment.debug = self.proc_rank > 0
+        self.experiment = self.experiment.get_non_ddp_exp()
+
+        # show progbar only on prog_rank 0
+        self.prog_bar = self.prog_bar and self.node_rank == 0 and gpu_nb == 0
+
+        # determine which process we are and world size
+        self.proc_rank = self.node_rank * len(self.data_parallel_device_ids) + gpu_nb
+        self.world_size = self.nb_gpu_nodes * len(self.data_parallel_device_ids)
+
+        # set up server using proc 0's ip address
+        # try to init for 20 times at max in case ports are taken
+        # where to store ip_table
+        self.__init_tcp_connection()
+
+        # CHOOSE OPTIMIZER
+        # filter out the weights that were done on gpu so we can load on good old cpus
+        self.optimizers = model.configure_optimizers()
+
+        # MODEL
+        # copy model to each gpu
+        torch.cuda.set_device(gpu_nb)
+        model.cuda(gpu_nb)
+
+        # AMP
+        # run through amp wrapper before going to distributed DP
+        if self.use_amp:
+            # An example
+            model, optimizers = amp.initialize(
+                model, self.optimizers, opt_level=self.amp_level,
+            )
+            self.optimizers = optimizers
+
+        model = LightningDistributedDataParallel(model, device_ids=[gpu_nb])
+
+        # continue training routine
+        self.__run_pretrain_routine(model)
+
+    def __init_tcp_connection(self):
+        """
+        Connect all procs in the world using the env:// init
+        Use the first node as the root address
+        :param port:
+        :param tries:
+        :return:
+        """
+        try:
+            port = os.environ['MASTER_PORT']
+        except Exception as e:
+            port = 12910
+            os.environ['MASTER_PORT'] = f'{port}'
+
+        root_node = self.__resolve_root_node_address()
+        os.environ['MASTER_ADDR'] = root_node
+
+        dist.init_process_group("nccl", rank=self.proc_rank, world_size=self.world_size)
+
+
+    def __resolve_root_node_address(self):
+        try:
+            root_node = os.environ['SLURM_NODELIST'].split(' ')[0]
+
+            if '[' in root_node:
+                name = root_node.split('[')[0]
+                number = root_node.split(',')[0]
+                if '-' in number:
+                    number = number.split('-')[0]
+
+                number = re.sub('[^0-9]', '', number)
+                root_node = name + number
+
+        except Exception as e:
+            root_node = '127.0.0.2'
+
+        return root_node
+
+    def __run_pretrain_routine(self, model):
+        """
+        Sanity check a few things before starting actual training
+        :param model:
+        :return:
+        """
+        ref_model = model
+        if self.data_parallel:
+            ref_model = model.module
+
+        ref_model.trainer = self
+
+        # set local properties on the model
+        ref_model.on_gpu = self.on_gpu
+
+        # transfer data loaders from model
+        self.__get_dataloaders(ref_model)
+
+        # init training constants
+        self.__layout_bookeeping()
 
         # add lr schedulers
         if self.lr_scheduler_milestones is not None:
@@ -275,17 +561,19 @@ class Trainer(TrainerIO):
                 self.lr_schedulers.append(scheduler)
 
         # print model summary
-        model.summarize()
+        if self.proc_rank == 0 and self.print_weights_summary:
+            ref_model.summarize()
 
-        # put on gpu if needed
-        if self.on_gpu:
-            model = LightningDataParallel(model, device_ids=self.data_parallel_device_ids)
+        # give model convenience properties
+        ref_model.trainer = self
+        ref_model.experiment = self.experiment
 
         # run tiny validation to make sure program won't crash during val
         _ = self.validate(model, self.val_dataloader, max_batches=self.nb_sanity_val_steps)
 
         # save exp to get started
-        self.experiment.save()
+        if self.proc_rank == 0:
+            self.experiment.save()
 
         # enable cluster checkpointing
         if self.cluster is not None:
@@ -304,12 +592,12 @@ class Trainer(TrainerIO):
             for lr_scheduler in self.lr_schedulers:
                 lr_scheduler.step()
 
-            model = self.model.module if self.data_parallel else self.model
+            model = self.__get_model()
             model.current_epoch = epoch_nb
 
             # hook
             if self.__is_function_implemented('on_epoch_start'):
-                model = self.model.module if self.data_parallel else self.model
+                model = self.__get_model()
                 model.on_epoch_start()
 
             self.current_epoch = epoch_nb
@@ -324,7 +612,7 @@ class Trainer(TrainerIO):
                 self.batch_nb = batch_nb
                 self.global_step += 1
 
-                model = self.model.module if self.data_parallel else self.model
+                model = self.__get_model()
                 model.global_step = self.global_step
 
                 # stop when the flag is changed or we've gone past the amount requested in the batches
@@ -348,17 +636,16 @@ class Trainer(TrainerIO):
 
                 # when batch should be saved
                 if (batch_nb + 1) % self.log_save_interval == 0 or early_stop_epoch:
-                    self.experiment.save()
+                    if self.proc_rank == 0:
+                        self.experiment.save()
 
                 # when metrics should be logged
                 if batch_nb % self.add_log_row_interval == 0 or early_stop_epoch:
                     # count items in memory
                     # nb_params, nb_tensors = count_mem_items()
 
-                    if self.data_parallel:
-                        metrics = self.model.module.update_tng_log_metrics(self.__tng_tqdm_dic)
-                    else:
-                        metrics = self.model.update_tng_log_metrics(self.__tng_tqdm_dic)
+                    model = self.__get_model()
+                    metrics = model.update_tng_log_metrics(self.__tng_tqdm_dic)
 
                     # add gpu memory
                     if self.on_gpu:
@@ -367,19 +654,22 @@ class Trainer(TrainerIO):
 
                     # add norms
                     if self.track_grad_norm > 0:
-                        model = self.model.module if self.data_parallel else self.model
+                        model = self.__get_model()
                         grad_norm_dic = model.grad_norm(self.track_grad_norm)
-
                         metrics.update(grad_norm_dic)
 
+                    if self.__is_function_implemented('on_tng_metrics'):
+                        model.on_tng_metrics(metrics)
+
                     # log metrics
                     scalar_metrics = self.__metrics_to_scalars(metrics, blacklist=self.__log_vals_blacklist())
-                    self.experiment.log(scalar_metrics, global_step=self.global_step)
-                    self.experiment.save()
+                    if self.proc_rank == 0:
+                        self.experiment.log(scalar_metrics, global_step=self.global_step)
+                        self.experiment.save()
 
                 # hook
                 if self.__is_function_implemented('on_batch_end'):
-                    model = self.model.module if self.data_parallel else self.model
+                    model = self.__get_model()
                     model.on_batch_end()
 
                 # end epoch early
@@ -388,13 +678,13 @@ class Trainer(TrainerIO):
 
             # hook
             if self.__is_function_implemented('on_epoch_end'):
-                model = self.model.module if self.data_parallel else self.model
+                model = self.__get_model()
                 model.on_epoch_end()
 
             # early stopping
-            if self.enable_early_stop:
+            met_min_epochs = epoch_nb > self.min_nb_epochs
+            if self.enable_early_stop and met_min_epochs:
                 should_stop = self.early_stop_callback.on_epoch_end(epoch=epoch_nb, logs=self.__tng_tqdm_dic)
-                met_min_epochs = epoch_nb > self.min_nb_epochs
 
                 # stop training
                 stop = should_stop and met_min_epochs
@@ -417,7 +707,7 @@ class Trainer(TrainerIO):
 
     def __log_vals_blacklist(self):
         """avoid logging some vals lightning uses to maintain state"""
-        blacklist = {'batch_nb', 'v_nb', 'epoch', 'gpu'}
+        blacklist = {'batch_nb', 'v_nb', 'gpu'}
         return blacklist
 
     def __run_tng_batch(self, data_batch, batch_nb):
@@ -426,8 +716,8 @@ class Trainer(TrainerIO):
 
         # hook
         if self.__is_function_implemented('on_batch_start'):
-            model = self.model.module if self.data_parallel else self.model
-            response = model.on_batch_start(data_batch)
+            model_ref = self.__get_model()
+            response = model_ref.on_batch_start(data_batch)
 
             if response == -1:
                 return -1
@@ -437,19 +727,32 @@ class Trainer(TrainerIO):
 
         # forward pass
         # return a scalar value and a dic with tqdm metrics
-        if self.data_parallel:
+        if self.use_ddp:
+            output = self.model(data_batch, batch_nb)
+        elif self.use_dp:
             output = self.model(data_batch, batch_nb)
             output = reduce_distributed_output(output, len(self.data_parallel_device_ids))
         else:
             output = self.model.training_step(data_batch, batch_nb)
 
-        model_specific_tqdm_metrics_dic = output['tqdm_metrics']
-        loss = output['loss']
+        try:
+            model_specific_tqdm_metrics_dic = output['tqdm_metrics']
+        except Exception as e:
+            model_specific_tqdm_metrics_dic = {}
+
+        # if output dict doesn't have the keyword loss
+        # then assume the output=loss if scalar
+        try:
+            loss = output['loss']
+        except Exception as e:
+            if type(output) is torch.Tensor:
+                loss = output
 
         self.__add_tqdm_metrics(model_specific_tqdm_metrics_dic)
 
         # backward pass
         if self.use_amp:
+            # scale loss when using amp
             for optimizer in self.optimizers:
                 with amp.scale_loss(loss, optimizer) as scaled_loss:
                     scaled_loss.backward()
@@ -457,10 +760,11 @@ class Trainer(TrainerIO):
             loss.backward()
 
         if self.print_nan_grads:
-            model = self.model.module if self.data_parallel else self.model
+            model = self.__get_model()
             for param in model.parameters():
                 print(param.grad.float().sum())
 
+        # avoid memory leaks
         self.batch_loss_value += loss.item()
 
         # gradient update with accumulated gradients
@@ -468,7 +772,7 @@ class Trainer(TrainerIO):
 
             # clip gradients
             if self.gradient_clip > 0:
-                model = self.model.module if self.data_parallel else self.model
+                model = self.__get_model()
                 torch.nn.utils.clip_grad_norm(model.parameters(), self.gradient_clip)
 
             # update gradients across all optimizers
@@ -494,7 +798,8 @@ class Trainer(TrainerIO):
 
         # activate batch end hook
         if self.__is_function_implemented('on_batch_end'):
-            self.model.on_batch_end()
+            model = self.__get_model()
+            model.on_batch_end()
 
         return 0
 
@@ -509,7 +814,8 @@ class Trainer(TrainerIO):
         try:
             # hook
             if self.__is_function_implemented('on_pre_performance_check'):
-                self.model.on_pre_performance_check()
+                model = self.__get_model()
+                model.on_pre_performance_check()
 
             # use full val set on end of epoch
             # use a small portion otherwise
@@ -523,7 +829,8 @@ class Trainer(TrainerIO):
 
             # hook
             if self.__is_function_implemented('on_post_performance_check'):
-                self.model.on_post_performance_check()
+                model = self.__get_model()
+                model.on_post_performance_check()
 
         except Exception as e:
             print(e)
@@ -535,5 +842,6 @@ class Trainer(TrainerIO):
             self.prog_bar.set_postfix(**tqdm_metrics)
 
         # model checkpointing
-        print('save callback...')
-        self.checkpoint_callback.on_epoch_end(epoch=self.current_epoch, logs=self.__tng_tqdm_dic)
+        if self.proc_rank == 0 and self.checkpoint_callback:
+            print('save callback...')
+            self.checkpoint_callback.on_epoch_end(epoch=self.current_epoch, logs=self.__tng_tqdm_dic)

pytorch_lightning/pt_overrides/override_data_parallel.py

@@ -1,4 +1,7 @@
 from torch.nn import DataParallel
+from torch.nn.parallel import DistributedDataParallel
+import itertools
+from itertools import chain
 
 import threading
 import torch
@@ -6,6 +9,20 @@ from torch.cuda._utils import _get_device_index
 import pdb
 
 
+def _find_tensors(obj):
+    r"""
+    Recursively find all tensors contained in the specified object.
+    """
+    if isinstance(obj, torch.Tensor):
+        return [obj]
+    if isinstance(obj, (list, tuple)):
+        return itertools.chain(*map(_find_tensors, obj))
+    if isinstance(obj, dict):
+        return itertools.chain(*map(_find_tensors, obj.values()))
+    return []
+
+
+
 def get_a_var(obj):
     if isinstance(obj, torch.Tensor):
         return obj
@@ -26,10 +43,76 @@ class LightningDataParallel(DataParallel):
     Override the forward call in lightning so it goes to training and validation step respectively
     """
 
+    def forward(self, *inputs, **kwargs):
+        if not self.device_ids:
+            return self.module(*inputs, **kwargs)
+
+        for t in chain(self.module.parameters(), self.module.buffers()):
+            if t.device != self.src_device_obj:
+                raise RuntimeError("module must have its parameters and buffers "
+                                   "on device {} (device_ids[0]) but found one of "
+                                   "them on device: {}".format(self.src_device_obj, t.device))
+
+        inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids)
+        if len(self.device_ids) == 1:
+            # lightning
+            if self.module.training:
+                return self.module.training_step(*inputs[0], **kwargs[0])
+            else:
+                return self.module.validation_step(*inputs[0], **kwargs[0])
+
+        replicas = self.replicate(self.module, self.device_ids[:len(inputs)])
+        outputs = self.parallel_apply(replicas, inputs, kwargs)
+        return self.gather(outputs, self.output_device)
+
+
     def parallel_apply(self, replicas, inputs, kwargs):
         return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])
 
 
+class LightningDistributedDataParallel(DistributedDataParallel):
+    """
+    Override the forward call in lightning so it goes to training and validation step respectively
+    """
+
+    def parallel_apply(self, replicas, inputs, kwargs):
+        return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])
+
+    def forward(self, *inputs, **kwargs):
+        self._sync_params()
+        if self.device_ids:
+            inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids)
+            if len(self.device_ids) == 1:
+                # --------------
+                # LIGHTNING MOD
+                # --------------
+                # normal
+                # output = self.module(*inputs[0], **kwargs[0])
+
+                # lightning
+                if self.module.training:
+                    output = self.module.training_step(*inputs[0], **kwargs[0])
+                else:
+                    output = self.module.validation_step(*inputs[0], **kwargs[0])
+            else:
+                outputs = self.parallel_apply(self._module_copies[:len(inputs)], inputs, kwargs)
+                output = self.gather(outputs, self.output_device)
+        else:
+            output = self.module(*inputs, **kwargs)
+
+        if torch.is_grad_enabled():
+            # We'll return the output object verbatim since it is a freeform
+            # object. We need to find any tensors in this object, though,
+            # because we need to figure out which parameters were used during
+            # this forward pass, to ensure we short circuit reduction for any
+            # unused parameters. Only if `find_unused_parameters` is set.
+            if self.find_unused_parameters:
+                self.reducer.prepare_for_backward(list(_find_tensors(output)))
+            else:
+                self.reducer.prepare_for_backward([])
+        return output
+
+
 def parallel_apply(modules, inputs, kwargs_tup=None, devices=None):
     r"""Applies each `module` in :attr:`modules` in parallel on arguments
     contained in :attr:`inputs` (positional) and :attr:`kwargs_tup` (keyword)
@@ -102,4 +185,4 @@ def parallel_apply(modules, inputs, kwargs_tup=None, devices=None):
         if isinstance(output, Exception):
             raise output
         outputs.append(output)
-    return outputs
+    return outputs

pytorch_lightning/root_module/hooks.py

@@ -19,3 +19,6 @@ class ModelHooks(torch.nn.Module):
     def on_post_performance_check(self):
         pass
 
+    def on_tng_metrics(self, metrics):
+        pass
+

pytorch_lightning/root_module/model_saving.py

@@ -2,7 +2,7 @@ import torch
 import os
 import re
 import pdb
-from pytorch_lightning.pt_overrides.override_data_parallel import LightningDataParallel
+from pytorch_lightning.pt_overrides.override_data_parallel import LightningDistributedDataParallel, LightningDataParallel
 
 class ModelIO(object):
 
@@ -21,6 +21,23 @@ class ModelIO(object):
         """
         raise NotImplementedError
 
+    # -------------------------
+    # OPTIONAL HOOKS
+    # -------------------------
+    def on_hpc_save(self):
+        """
+        Hook to do whatever you need right before Slurm manager saves the model
+        :return:
+        """
+        pass
+
+    def on_hpc_load(self):
+        """
+        Hook to do whatever you need right before Slurm manager loads the model
+        :return:
+        """
+        pass
+
 
 class TrainerIO(object):
 
@@ -49,7 +66,8 @@ class TrainerIO(object):
         checkpoint['optimizer_states'] = optimizer_states
 
         # request what to save from the model
-        model = self.model.module if type(self.model) is LightningDataParallel else self.model
+        is_dp_module = type(self.model) is LightningDistributedDataParallel or type(self.model) is LightningDataParallel
+        model = self.model.module if is_dp_module else self.model
         checkpoint_dict = model.get_save_dict()
 
         # merge trainer and model saving items
@@ -116,6 +134,9 @@ class TrainerIO(object):
             os.makedirs(folderpath, exist_ok=True)
         filepath = '{}/hpc_ckpt_{}.ckpt'.format(folderpath, ckpt_number)
 
+        # give model a chance to do something on hpc_save
+        self.on_hpc_save()
+
         # request what to save from the model
         checkpoint_dict = self.dump_checkpoint()
 
@@ -137,6 +158,9 @@ class TrainerIO(object):
         model = self.model.module if type(self.model) is LightningDataParallel else self.model
         model.load_model_specific(checkpoint)
 
+        # call model hook
+        self.on_hpc_load()
+
     def max_ckpt_in_folder(self, path):
         files = os.listdir(path)
         files = [x for x in files if 'ckpt_' in x]

pytorch_lightning/root_module/root_module.py

@@ -9,7 +9,6 @@ from pytorch_lightning.root_module.optimization import OptimizerConfig
 from pytorch_lightning.root_module.hooks import ModelHooks
 
 
-
 class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
 
     def __init__(self, hparams):
@@ -21,29 +20,17 @@ class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
         self.current_epoch = 0
         self.global_step = 0
         self.loaded_optimizer_states_dict = {}
-        self.fast_dev_run = hparams.fast_dev_run
-        self.overfit = hparams.overfit
-        self.gradient_clip = hparams.gradient_clip
         self.trainer = None
-        self.from_lightning = True
         self.experiment = None
 
         # track if gpu was requested for checkpointing
         self.on_gpu = False
-        try:
-            self.on_gpu = hparams.on_gpu
-        except Exception as e:
-            pass
 
         # computed vars for the dataloaders
         self._tng_dataloader = None
         self._val_dataloader = None
         self._test_dataloader = None
 
-        if self.on_gpu:
-            print('running on gpu...')
-            torch.set_default_tensor_type(hparams.default_tensor_type)
-
     def forward(self, *args, **kwargs):
         """
         Expand model in into whatever you need.
@@ -91,7 +78,7 @@ class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
         :param logs:
         :return:
         """
-        raise NotImplementedError
+        return logs
 
     def loss(self, *args, **kwargs):
         """
@@ -105,9 +92,6 @@ class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
         model_summary = ModelSummary(self)
         print(model_summary)
 
-    def nb_batches(self, dataloader):
-        a = math.ceil(float(len(dataloader.dataset) / self.batch_size))
-        return int(a)
 
     def freeze(self):
         for param in self.parameters():
@@ -145,7 +129,7 @@ class LightningModule(GradInformation, ModelIO, OptimizerConfig, ModelHooks):
     def get_process_position(gpus):
         try:
             current_gpu = os.environ["CUDA_VISIBLE_DEVICES"]
-            gpu_ids = gpus.split(';')
+            gpu_ids = gpus.split(',')
             process_position = gpu_ids.index(current_gpu)
             return process_position, current_gpu
         except Exception as e:

pytorch_lightning/utils/arg_parse.py

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

pytorch_lightning/utils/plotting.py

@@ -1,4 +1,3 @@
-from matplotlib import pyplot as plt
 import numpy as np
 np.seterr(divide='ignore', invalid='ignore')
 
@@ -13,6 +12,7 @@ def plot_confusion_matrix(cm,
     This function prints and plots the confusion matrix.
     Normalization can be applied by setting `normalize=True`.
     """
+    from matplotlib import pyplot as plt
     if normalize:
         cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]
         print("Normalized confusion matrix")

setup.py

@@ -7,7 +7,7 @@ from setuptools import setup, find_packages
 # http://blog.ionelmc.ro/2014/05/25/python-packaging/
 setup(
     name="pytorch-lightning",
-    version='0.121',
+    version='0.2.6',
     description="The Keras for ML researchers using PyTorch",
     author="William Falcon",
     author_email="waf2107@columbia.edu",
@@ -19,8 +19,7 @@ setup(
     install_requires=[
         "torch>=1.1.0",
         "tqdm",
-        "test-tube>=0.651",
-        "tensorflow>=1.14.0"
+        "test-tube>=0.6.7.1",
     ],
     packages=find_packages(),
     long_description=open("README.md", encoding="utf-8").read(),