diff --git a/404.html b/404.html index 02970fa1..ebc26eaf 100644 --- a/404.html +++ b/404.html @@ -54,9 +54,36 @@ Lightning module + + + +
  • + + Trainer +
    • diff --git a/Pytorch-Lightning/LightningModule/index.html b/Pytorch-Lightning/LightningModule/index.html index bda91e12..e6dd4193 100644 --- a/Pytorch-Lightning/LightningModule/index.html +++ b/Pytorch-Lightning/LightningModule/index.html @@ -93,9 +93,36 @@ + + + +
  • + + Trainer +
    • @@ -142,7 +169,8 @@

    Lightning module

    [Github Code]

    A lightning module is a strict superclass of nn.Module, it provides a standard interface for the trainer to interact with the model.

    -

    To Define a Lightning Module, implement the following methods:

    +

    The easiest thing to do is copy this template and modify accordingly.

    +

    Otherwise, to Define a Lightning Module, implement the following methods:

    Required:

    + + +
  • + + Trainer +
    • @@ -234,5 +261,5 @@ diff --git a/search.html b/search.html index 357a507c..7d73030f 100644 --- a/search.html +++ b/search.html @@ -54,9 +54,36 @@ Lightning module + + + +
  • + + Trainer +
    • diff --git a/search/search_index.json b/search/search_index.json index 3ea7627b..d3ddb95c 100644 --- a/search/search_index.json +++ b/search/search_index.json @@ -1 +1 @@ -{"config":{"lang":["en"],"prebuild_index":false,"separator":"[\\s\\-]+"},"docs":[{"location":"","text":"PYTORCH-LIGHTNING DOCUMENTATION Quick start Define a LightningModule Set up the trainer Quick start examples CPU example Single GPU example Multi-gpu example SLURM cluster grid search example Training loop Accumulate gradients Check GPU usage Check which gradients are nan Check validation every n epochs Display metrics in progress bar Force training for min or max epochs Inspect gradient norms Hooks Learning rate annealing Make model overfit on subset of data Multiple optimizers (like GANs) Set how much of the training set to check (1-100%) training_step function Validation loop Display metrics in progress bar hooks Set how much of the validation set to check (1-100%) Set validation check frequency within 1 training epoch (1-100%) validation_step function Why does validation run first for 5 steps? Distributed training Single-gpu Multi-gpu Multi-node 16-bit mixed precision Checkpointing Model saving Model loading Computing cluster (SLURM) Automatic checkpointing Automatic saving, loading Running grid search on a cluster Walltime auto-resubmit","title":"PYTORCH-LIGHTNING DOCUMENTATION"},{"location":"#pytorch-lightning-documentation","text":"","title":"PYTORCH-LIGHTNING DOCUMENTATION"},{"location":"#quick-start","text":"Define a LightningModule Set up the trainer","title":"Quick start"},{"location":"#quick-start-examples","text":"CPU example Single GPU example Multi-gpu example SLURM cluster grid search example","title":"Quick start examples"},{"location":"#training-loop","text":"Accumulate gradients Check GPU usage Check which gradients are nan Check validation every n epochs Display metrics in progress bar Force training for min or max epochs Inspect gradient norms Hooks Learning rate annealing Make model overfit on subset of data Multiple optimizers (like GANs) Set how much of the training set to check (1-100%) training_step function","title":"Training loop"},{"location":"#validation-loop","text":"Display metrics in progress bar hooks Set how much of the validation set to check (1-100%) Set validation check frequency within 1 training epoch (1-100%) validation_step function Why does validation run first for 5 steps?","title":"Validation loop"},{"location":"#distributed-training","text":"Single-gpu Multi-gpu Multi-node 16-bit mixed precision","title":"Distributed training"},{"location":"#checkpointing","text":"Model saving Model loading","title":"Checkpointing"},{"location":"#computing-cluster-slurm","text":"Automatic checkpointing Automatic saving, loading Running grid search on a cluster Walltime auto-resubmit","title":"Computing cluster (SLURM)"},{"location":"Pytorch-Lightning/LightningModule/","text":"Lightning module [ Github Code ] A lightning module is a strict superclass of nn.Module, it provides a standard interface for the trainer to interact with the model. To Define a Lightning Module, implement the following methods: Required : training_step validation_step validation_end configure_optimizers get_save_dict load_model_specific tng_dataloader tng_dataloader test_dataloader Optional : update_tng_log_metrics add_model_specific_args training_step def training_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. Params Param description data_batch The output of your dataloader. A tensor, tuple or list batch_nb Integer displaying which batch this is Return Dictionary or OrderedDict key value is required loss tensor scalar Y prog Dict for progress bar display. Must have only tensors N Example def training_step(self, data_batch, batch_nb): x, y, z = data_batch # implement your own out = self.forward(x) loss = self.loss(out, x) output = { 'loss': loss, # required 'prog': {'tng_loss': loss, 'batch_nb': batch_nb} # optional } # return a dict return output validation_step 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. Params Param description data_batch The output of your dataloader. A tensor, tuple or list batch_nb Integer displaying which batch this is Return Return description optional dict Dict of OrderedDict with metrics to display in progress bar. All keys must be tensors. Y Example def validation_step(self, data_batch, batch_nb): x, y, z = data_batch # implement your own out = self.forward(x) loss = self.loss(out, x) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # all optional... # return whatever you need for the collation function validation_end output = OrderedDict({ 'val_loss': loss_val, 'val_acc': torch.tensor(val_acc), # everything must be a tensor }) # return an optional dict return output validation_end def validation_end(self, outputs) Called at the end of the validation loop with the output of each validation_step. Params Param description outputs List of outputs you defined in validation_step Return Return description optional dict Dict of OrderedDict with metrics to display in progress bar Y Example def validation_end(self, outputs): \"\"\" Called at the end of validation to aggregate outputs :param outputs: list of individual outputs of each validation step :return: \"\"\" val_loss_mean = 0 val_acc_mean = 0 for output in outputs: val_loss_mean += output['val_loss'] val_acc_mean += output['val_acc'] val_loss_mean /= len(outputs) val_acc_mean /= len(outputs) tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()} return tqdm_dic configure_optimizers def configure_optimizers(self) Set up as many optimizers as you need. Normally you'd need one. But in the case of GANs or something more esoteric you might have multiple. Lightning will call .backward() and .step() on each one. If you use 16 bit precision it will also handle that. Return List - List of optimizers Example # most cases def configure_optimizers(self): opt = Adam(lr=0.01) return [opt] # gan example def configure_optimizers(self): generator_opt = Adam(lr=0.01) disriminator_opt = Adam(lr=0.02) return [generator_opt, disriminator_opt] get_save_dict def get_save_dict(self) Called by lightning to checkpoint your model. Lightning saves current epoch, current batch nb, etc... All you have to return is what specifically about your lightning model you want to checkpoint. Return Dictionary - No required keys. Most of the time as described in this example. Example def get_save_dict(self): # 99% of use cases this is all you need to return checkpoint = {'state_dict': self.state_dict()} return checkpoint load_model_specific def load_model_specific(self, checkpoint) Called by lightning to restore your model. This is your chance to restore your model using the keys you added in get_save_dict. Lightning will automatically restore current epoch, batch nb, etc. Return Nothing Example def load_model_specific(self, checkpoint): # you defined 'state_dict' in get_save_dict() self.load_state_dict(checkpoint['state_dict']) tng_dataloader @property def tng_dataloader(self) Called by lightning during training loop. Define it as a property. Return Pytorch DataLoader Example @property def tng_dataloader(self): if self._tng_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=True, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._tng_dataloader = loader except Exception as e: raise e return self._tng_dataloader val_dataloader @property def tng_dataloader(self) Called by lightning during validation loop. Define it as a property. Return Pytorch DataLoader Example @property def val_dataloader(self): if self._val_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._val_dataloader = loader except Exception as e: raise e return self._val_dataloader test_dataloader @property def test_dataloader(self) Called by lightning during test loop. Define it as a property. Return Pytorch DataLoader Example @property def test_dataloader(self): if self._test_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._test_dataloader = loader except Exception as e: raise e return self._test_dataloader update_tng_log_metrics 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. Return Dict Example def update_tng_log_metrics(self, logs): # modify or add to logs return logs add_model_specific_args @staticmethod def add_model_specific_args(parent_parser, root_dir) Lightning has a list of default argparse commands. This method is your chance to add or modify commands specific to your model. The argument parser is available anywhere in your model by calling self.hparams Return An argument parser Example @staticmethod def add_model_specific_args(parent_parser, root_dir): parser = HyperOptArgumentParser(strategy=parent_parser.strategy, parents=[parent_parser]) # param overwrites # parser.set_defaults(gradient_clip=5.0) # network params parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=False) parser.add_argument('--in_features', default=28*28) parser.add_argument('--out_features', default=10) parser.add_argument('--hidden_dim', default=50000) # use 500 for CPU, 50000 for GPU to see speed difference # data parser.add_argument('--data_root', default=os.path.join(root_dir, 'mnist'), type=str) # training params (opt) parser.opt_list('--learning_rate', default=0.001, type=float, options=[0.0001, 0.0005, 0.001, 0.005], tunable=False) parser.opt_list('--batch_size', default=256, type=int, options=[32, 64, 128, 256], tunable=False) parser.opt_list('--optimizer_name', default='adam', type=str, options=['adam'], tunable=False) return parser","title":"Lightning module"},{"location":"Pytorch-Lightning/LightningModule/#lightning-module","text":"[ Github Code ] A lightning module is a strict superclass of nn.Module, it provides a standard interface for the trainer to interact with the model. To Define a Lightning Module, implement the following methods: Required : training_step validation_step validation_end configure_optimizers get_save_dict load_model_specific tng_dataloader tng_dataloader test_dataloader Optional : update_tng_log_metrics add_model_specific_args","title":"Lightning module"},{"location":"Pytorch-Lightning/LightningModule/#training_step","text":"def training_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. Params Param description data_batch The output of your dataloader. A tensor, tuple or list batch_nb Integer displaying which batch this is Return Dictionary or OrderedDict key value is required loss tensor scalar Y prog Dict for progress bar display. Must have only tensors N Example def training_step(self, data_batch, batch_nb): x, y, z = data_batch # implement your own out = self.forward(x) loss = self.loss(out, x) output = { 'loss': loss, # required 'prog': {'tng_loss': loss, 'batch_nb': batch_nb} # optional } # return a dict return output","title":"training_step"},{"location":"Pytorch-Lightning/LightningModule/#validation_step","text":"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. Params Param description data_batch The output of your dataloader. A tensor, tuple or list batch_nb Integer displaying which batch this is Return Return description optional dict Dict of OrderedDict with metrics to display in progress bar. All keys must be tensors. Y Example def validation_step(self, data_batch, batch_nb): x, y, z = data_batch # implement your own out = self.forward(x) loss = self.loss(out, x) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # all optional... # return whatever you need for the collation function validation_end output = OrderedDict({ 'val_loss': loss_val, 'val_acc': torch.tensor(val_acc), # everything must be a tensor }) # return an optional dict return output","title":"validation_step"},{"location":"Pytorch-Lightning/LightningModule/#validation_end","text":"def validation_end(self, outputs) Called at the end of the validation loop with the output of each validation_step. Params Param description outputs List of outputs you defined in validation_step Return Return description optional dict Dict of OrderedDict with metrics to display in progress bar Y Example def validation_end(self, outputs): \"\"\" Called at the end of validation to aggregate outputs :param outputs: list of individual outputs of each validation step :return: \"\"\" val_loss_mean = 0 val_acc_mean = 0 for output in outputs: val_loss_mean += output['val_loss'] val_acc_mean += output['val_acc'] val_loss_mean /= len(outputs) val_acc_mean /= len(outputs) tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()} return tqdm_dic","title":"validation_end"},{"location":"Pytorch-Lightning/LightningModule/#configure_optimizers","text":"def configure_optimizers(self) Set up as many optimizers as you need. Normally you'd need one. But in the case of GANs or something more esoteric you might have multiple. Lightning will call .backward() and .step() on each one. If you use 16 bit precision it will also handle that.","title":"configure_optimizers"},{"location":"Pytorch-Lightning/LightningModule/#return","text":"List - List of optimizers Example # most cases def configure_optimizers(self): opt = Adam(lr=0.01) return [opt] # gan example def configure_optimizers(self): generator_opt = Adam(lr=0.01) disriminator_opt = Adam(lr=0.02) return [generator_opt, disriminator_opt]","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#get_save_dict","text":"def get_save_dict(self) Called by lightning to checkpoint your model. Lightning saves current epoch, current batch nb, etc... All you have to return is what specifically about your lightning model you want to checkpoint.","title":"get_save_dict"},{"location":"Pytorch-Lightning/LightningModule/#return_1","text":"Dictionary - No required keys. Most of the time as described in this example. Example def get_save_dict(self): # 99% of use cases this is all you need to return checkpoint = {'state_dict': self.state_dict()} return checkpoint","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#load_model_specific","text":"def load_model_specific(self, checkpoint) Called by lightning to restore your model. This is your chance to restore your model using the keys you added in get_save_dict. Lightning will automatically restore current epoch, batch nb, etc.","title":"load_model_specific"},{"location":"Pytorch-Lightning/LightningModule/#return_2","text":"Nothing Example def load_model_specific(self, checkpoint): # you defined 'state_dict' in get_save_dict() self.load_state_dict(checkpoint['state_dict'])","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#tng_dataloader","text":"@property def tng_dataloader(self) Called by lightning during training loop. Define it as a property.","title":"tng_dataloader"},{"location":"Pytorch-Lightning/LightningModule/#return_3","text":"Pytorch DataLoader Example @property def tng_dataloader(self): if self._tng_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=True, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._tng_dataloader = loader except Exception as e: raise e return self._tng_dataloader","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#val_dataloader","text":"@property def tng_dataloader(self) Called by lightning during validation loop. Define it as a property.","title":"val_dataloader"},{"location":"Pytorch-Lightning/LightningModule/#return_4","text":"Pytorch DataLoader Example @property def val_dataloader(self): if self._val_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._val_dataloader = loader except Exception as e: raise e return self._val_dataloader","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#test_dataloader","text":"@property def test_dataloader(self) Called by lightning during test loop. Define it as a property.","title":"test_dataloader"},{"location":"Pytorch-Lightning/LightningModule/#return_5","text":"Pytorch DataLoader Example @property def test_dataloader(self): if self._test_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._test_dataloader = loader except Exception as e: raise e return self._test_dataloader","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#update_tng_log_metrics","text":"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.","title":"update_tng_log_metrics"},{"location":"Pytorch-Lightning/LightningModule/#return_6","text":"Dict Example def update_tng_log_metrics(self, logs): # modify or add to logs return logs","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#add_model_specific_args","text":"@staticmethod def add_model_specific_args(parent_parser, root_dir) Lightning has a list of default argparse commands. This method is your chance to add or modify commands specific to your model. The argument parser is available anywhere in your model by calling self.hparams","title":"add_model_specific_args"},{"location":"Pytorch-Lightning/LightningModule/#return_7","text":"An argument parser Example @staticmethod def add_model_specific_args(parent_parser, root_dir): parser = HyperOptArgumentParser(strategy=parent_parser.strategy, parents=[parent_parser]) # param overwrites # parser.set_defaults(gradient_clip=5.0) # network params parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=False) parser.add_argument('--in_features', default=28*28) parser.add_argument('--out_features', default=10) parser.add_argument('--hidden_dim', default=50000) # use 500 for CPU, 50000 for GPU to see speed difference # data parser.add_argument('--data_root', default=os.path.join(root_dir, 'mnist'), type=str) # training params (opt) parser.opt_list('--learning_rate', default=0.001, type=float, options=[0.0001, 0.0005, 0.001, 0.005], tunable=False) parser.opt_list('--batch_size', default=256, type=int, options=[32, 64, 128, 256], tunable=False) parser.opt_list('--optimizer_name', default='adam', type=str, options=['adam'], tunable=False) return parser","title":"Return"},{"location":"Pytorch-Lightning/Trainer/","text":"Trainer","title":"Trainer"},{"location":"Pytorch-Lightning/Trainer/#trainer","text":"","title":"Trainer"}]} \ No newline at end of file +{"config":{"lang":["en"],"prebuild_index":false,"separator":"[\\s\\-]+"},"docs":[{"location":"","text":"PYTORCH-LIGHTNING DOCUMENTATION Quick start Define a LightningModule Set up the trainer Quick start examples CPU example Single GPU example Multi-gpu example SLURM cluster grid search example Training loop Accumulate gradients Check GPU usage Check which gradients are nan Check validation every n epochs Display metrics in progress bar Force training for min or max epochs Inspect gradient norms Hooks Learning rate annealing Make model overfit on subset of data Multiple optimizers (like GANs) Set how much of the training set to check (1-100%) training_step function Validation loop Display metrics in progress bar hooks Set how much of the validation set to check (1-100%) Set validation check frequency within 1 training epoch (1-100%) validation_step function Why does validation run first for 5 steps? Distributed training Single-gpu Multi-gpu Multi-node 16-bit mixed precision Checkpointing Model saving Model loading Computing cluster (SLURM) Automatic checkpointing Automatic saving, loading Running grid search on a cluster Walltime auto-resubmit","title":"PYTORCH-LIGHTNING DOCUMENTATION"},{"location":"#pytorch-lightning-documentation","text":"","title":"PYTORCH-LIGHTNING DOCUMENTATION"},{"location":"#quick-start","text":"Define a LightningModule Set up the trainer","title":"Quick start"},{"location":"#quick-start-examples","text":"CPU example Single GPU example Multi-gpu example SLURM cluster grid search example","title":"Quick start examples"},{"location":"#training-loop","text":"Accumulate gradients Check GPU usage Check which gradients are nan Check validation every n epochs Display metrics in progress bar Force training for min or max epochs Inspect gradient norms Hooks Learning rate annealing Make model overfit on subset of data Multiple optimizers (like GANs) Set how much of the training set to check (1-100%) training_step function","title":"Training loop"},{"location":"#validation-loop","text":"Display metrics in progress bar hooks Set how much of the validation set to check (1-100%) Set validation check frequency within 1 training epoch (1-100%) validation_step function Why does validation run first for 5 steps?","title":"Validation loop"},{"location":"#distributed-training","text":"Single-gpu Multi-gpu Multi-node 16-bit mixed precision","title":"Distributed training"},{"location":"#checkpointing","text":"Model saving Model loading","title":"Checkpointing"},{"location":"#computing-cluster-slurm","text":"Automatic checkpointing Automatic saving, loading Running grid search on a cluster Walltime auto-resubmit","title":"Computing cluster (SLURM)"},{"location":"Pytorch-Lightning/LightningModule/","text":"Lightning module [ Github Code ] A lightning module is a strict superclass of nn.Module, it provides a standard interface for the trainer to interact with the model. The easiest thing to do is copy this template and modify accordingly. Otherwise, to Define a Lightning Module, implement the following methods: Required : training_step validation_step validation_end configure_optimizers get_save_dict load_model_specific tng_dataloader tng_dataloader test_dataloader Optional : update_tng_log_metrics add_model_specific_args training_step def training_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. Params Param description data_batch The output of your dataloader. A tensor, tuple or list batch_nb Integer displaying which batch this is Return Dictionary or OrderedDict key value is required loss tensor scalar Y prog Dict for progress bar display. Must have only tensors N Example def training_step(self, data_batch, batch_nb): x, y, z = data_batch # implement your own out = self.forward(x) loss = self.loss(out, x) output = { 'loss': loss, # required 'prog': {'tng_loss': loss, 'batch_nb': batch_nb} # optional } # return a dict return output validation_step 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. Params Param description data_batch The output of your dataloader. A tensor, tuple or list batch_nb Integer displaying which batch this is Return Return description optional dict Dict of OrderedDict with metrics to display in progress bar. All keys must be tensors. Y Example def validation_step(self, data_batch, batch_nb): x, y, z = data_batch # implement your own out = self.forward(x) loss = self.loss(out, x) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # all optional... # return whatever you need for the collation function validation_end output = OrderedDict({ 'val_loss': loss_val, 'val_acc': torch.tensor(val_acc), # everything must be a tensor }) # return an optional dict return output validation_end def validation_end(self, outputs) Called at the end of the validation loop with the output of each validation_step. Params Param description outputs List of outputs you defined in validation_step Return Return description optional dict Dict of OrderedDict with metrics to display in progress bar Y Example def validation_end(self, outputs): \"\"\" Called at the end of validation to aggregate outputs :param outputs: list of individual outputs of each validation step :return: \"\"\" val_loss_mean = 0 val_acc_mean = 0 for output in outputs: val_loss_mean += output['val_loss'] val_acc_mean += output['val_acc'] val_loss_mean /= len(outputs) val_acc_mean /= len(outputs) tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()} return tqdm_dic configure_optimizers def configure_optimizers(self) Set up as many optimizers as you need. Normally you'd need one. But in the case of GANs or something more esoteric you might have multiple. Lightning will call .backward() and .step() on each one. If you use 16 bit precision it will also handle that. Return List - List of optimizers Example # most cases def configure_optimizers(self): opt = Adam(lr=0.01) return [opt] # gan example def configure_optimizers(self): generator_opt = Adam(lr=0.01) disriminator_opt = Adam(lr=0.02) return [generator_opt, disriminator_opt] get_save_dict def get_save_dict(self) Called by lightning to checkpoint your model. Lightning saves current epoch, current batch nb, etc... All you have to return is what specifically about your lightning model you want to checkpoint. Return Dictionary - No required keys. Most of the time as described in this example. Example def get_save_dict(self): # 99% of use cases this is all you need to return checkpoint = {'state_dict': self.state_dict()} return checkpoint load_model_specific def load_model_specific(self, checkpoint) Called by lightning to restore your model. This is your chance to restore your model using the keys you added in get_save_dict. Lightning will automatically restore current epoch, batch nb, etc. Return Nothing Example def load_model_specific(self, checkpoint): # you defined 'state_dict' in get_save_dict() self.load_state_dict(checkpoint['state_dict']) tng_dataloader @property def tng_dataloader(self) Called by lightning during training loop. Define it as a property. Return Pytorch DataLoader Example @property def tng_dataloader(self): if self._tng_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=True, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._tng_dataloader = loader except Exception as e: raise e return self._tng_dataloader val_dataloader @property def tng_dataloader(self) Called by lightning during validation loop. Define it as a property. Return Pytorch DataLoader Example @property def val_dataloader(self): if self._val_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._val_dataloader = loader except Exception as e: raise e return self._val_dataloader test_dataloader @property def test_dataloader(self) Called by lightning during test loop. Define it as a property. Return Pytorch DataLoader Example @property def test_dataloader(self): if self._test_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._test_dataloader = loader except Exception as e: raise e return self._test_dataloader update_tng_log_metrics 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. Return Dict Example def update_tng_log_metrics(self, logs): # modify or add to logs return logs add_model_specific_args @staticmethod def add_model_specific_args(parent_parser, root_dir) Lightning has a list of default argparse commands. This method is your chance to add or modify commands specific to your model. The argument parser is available anywhere in your model by calling self.hparams Return An argument parser Example @staticmethod def add_model_specific_args(parent_parser, root_dir): parser = HyperOptArgumentParser(strategy=parent_parser.strategy, parents=[parent_parser]) # param overwrites # parser.set_defaults(gradient_clip=5.0) # network params parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=False) parser.add_argument('--in_features', default=28*28) parser.add_argument('--out_features', default=10) parser.add_argument('--hidden_dim', default=50000) # use 500 for CPU, 50000 for GPU to see speed difference # data parser.add_argument('--data_root', default=os.path.join(root_dir, 'mnist'), type=str) # training params (opt) parser.opt_list('--learning_rate', default=0.001, type=float, options=[0.0001, 0.0005, 0.001, 0.005], tunable=False) parser.opt_list('--batch_size', default=256, type=int, options=[32, 64, 128, 256], tunable=False) parser.opt_list('--optimizer_name', default='adam', type=str, options=['adam'], tunable=False) return parser","title":"Lightning module"},{"location":"Pytorch-Lightning/LightningModule/#lightning-module","text":"[ Github Code ] A lightning module is a strict superclass of nn.Module, it provides a standard interface for the trainer to interact with the model. The easiest thing to do is copy this template and modify accordingly. Otherwise, to Define a Lightning Module, implement the following methods: Required : training_step validation_step validation_end configure_optimizers get_save_dict load_model_specific tng_dataloader tng_dataloader test_dataloader Optional : update_tng_log_metrics add_model_specific_args","title":"Lightning module"},{"location":"Pytorch-Lightning/LightningModule/#training_step","text":"def training_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. Params Param description data_batch The output of your dataloader. A tensor, tuple or list batch_nb Integer displaying which batch this is Return Dictionary or OrderedDict key value is required loss tensor scalar Y prog Dict for progress bar display. Must have only tensors N Example def training_step(self, data_batch, batch_nb): x, y, z = data_batch # implement your own out = self.forward(x) loss = self.loss(out, x) output = { 'loss': loss, # required 'prog': {'tng_loss': loss, 'batch_nb': batch_nb} # optional } # return a dict return output","title":"training_step"},{"location":"Pytorch-Lightning/LightningModule/#validation_step","text":"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. Params Param description data_batch The output of your dataloader. A tensor, tuple or list batch_nb Integer displaying which batch this is Return Return description optional dict Dict of OrderedDict with metrics to display in progress bar. All keys must be tensors. Y Example def validation_step(self, data_batch, batch_nb): x, y, z = data_batch # implement your own out = self.forward(x) loss = self.loss(out, x) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # all optional... # return whatever you need for the collation function validation_end output = OrderedDict({ 'val_loss': loss_val, 'val_acc': torch.tensor(val_acc), # everything must be a tensor }) # return an optional dict return output","title":"validation_step"},{"location":"Pytorch-Lightning/LightningModule/#validation_end","text":"def validation_end(self, outputs) Called at the end of the validation loop with the output of each validation_step. Params Param description outputs List of outputs you defined in validation_step Return Return description optional dict Dict of OrderedDict with metrics to display in progress bar Y Example def validation_end(self, outputs): \"\"\" Called at the end of validation to aggregate outputs :param outputs: list of individual outputs of each validation step :return: \"\"\" val_loss_mean = 0 val_acc_mean = 0 for output in outputs: val_loss_mean += output['val_loss'] val_acc_mean += output['val_acc'] val_loss_mean /= len(outputs) val_acc_mean /= len(outputs) tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()} return tqdm_dic","title":"validation_end"},{"location":"Pytorch-Lightning/LightningModule/#configure_optimizers","text":"def configure_optimizers(self) Set up as many optimizers as you need. Normally you'd need one. But in the case of GANs or something more esoteric you might have multiple. Lightning will call .backward() and .step() on each one. If you use 16 bit precision it will also handle that.","title":"configure_optimizers"},{"location":"Pytorch-Lightning/LightningModule/#return","text":"List - List of optimizers Example # most cases def configure_optimizers(self): opt = Adam(lr=0.01) return [opt] # gan example def configure_optimizers(self): generator_opt = Adam(lr=0.01) disriminator_opt = Adam(lr=0.02) return [generator_opt, disriminator_opt]","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#get_save_dict","text":"def get_save_dict(self) Called by lightning to checkpoint your model. Lightning saves current epoch, current batch nb, etc... All you have to return is what specifically about your lightning model you want to checkpoint.","title":"get_save_dict"},{"location":"Pytorch-Lightning/LightningModule/#return_1","text":"Dictionary - No required keys. Most of the time as described in this example. Example def get_save_dict(self): # 99% of use cases this is all you need to return checkpoint = {'state_dict': self.state_dict()} return checkpoint","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#load_model_specific","text":"def load_model_specific(self, checkpoint) Called by lightning to restore your model. This is your chance to restore your model using the keys you added in get_save_dict. Lightning will automatically restore current epoch, batch nb, etc.","title":"load_model_specific"},{"location":"Pytorch-Lightning/LightningModule/#return_2","text":"Nothing Example def load_model_specific(self, checkpoint): # you defined 'state_dict' in get_save_dict() self.load_state_dict(checkpoint['state_dict'])","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#tng_dataloader","text":"@property def tng_dataloader(self) Called by lightning during training loop. Define it as a property.","title":"tng_dataloader"},{"location":"Pytorch-Lightning/LightningModule/#return_3","text":"Pytorch DataLoader Example @property def tng_dataloader(self): if self._tng_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=True, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._tng_dataloader = loader except Exception as e: raise e return self._tng_dataloader","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#val_dataloader","text":"@property def tng_dataloader(self) Called by lightning during validation loop. Define it as a property.","title":"val_dataloader"},{"location":"Pytorch-Lightning/LightningModule/#return_4","text":"Pytorch DataLoader Example @property def val_dataloader(self): if self._val_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._val_dataloader = loader except Exception as e: raise e return self._val_dataloader","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#test_dataloader","text":"@property def test_dataloader(self) Called by lightning during test loop. Define it as a property.","title":"test_dataloader"},{"location":"Pytorch-Lightning/LightningModule/#return_5","text":"Pytorch DataLoader Example @property def test_dataloader(self): if self._test_dataloader is None: try: transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))]) dataset = MNIST(root='/path/to/mnist/', train=False, transform=transform, download=True) loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.hparams.batch_size, shuffle=True ) self._test_dataloader = loader except Exception as e: raise e return self._test_dataloader","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#update_tng_log_metrics","text":"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.","title":"update_tng_log_metrics"},{"location":"Pytorch-Lightning/LightningModule/#return_6","text":"Dict Example def update_tng_log_metrics(self, logs): # modify or add to logs return logs","title":"Return"},{"location":"Pytorch-Lightning/LightningModule/#add_model_specific_args","text":"@staticmethod def add_model_specific_args(parent_parser, root_dir) Lightning has a list of default argparse commands. This method is your chance to add or modify commands specific to your model. The argument parser is available anywhere in your model by calling self.hparams","title":"add_model_specific_args"},{"location":"Pytorch-Lightning/LightningModule/#return_7","text":"An argument parser Example @staticmethod def add_model_specific_args(parent_parser, root_dir): parser = HyperOptArgumentParser(strategy=parent_parser.strategy, parents=[parent_parser]) # param overwrites # parser.set_defaults(gradient_clip=5.0) # network params parser.opt_list('--drop_prob', default=0.2, options=[0.2, 0.5], type=float, tunable=False) parser.add_argument('--in_features', default=28*28) parser.add_argument('--out_features', default=10) parser.add_argument('--hidden_dim', default=50000) # use 500 for CPU, 50000 for GPU to see speed difference # data parser.add_argument('--data_root', default=os.path.join(root_dir, 'mnist'), type=str) # training params (opt) parser.opt_list('--learning_rate', default=0.001, type=float, options=[0.0001, 0.0005, 0.001, 0.005], tunable=False) parser.opt_list('--batch_size', default=256, type=int, options=[32, 64, 128, 256], tunable=False) parser.opt_list('--optimizer_name', default='adam', type=str, options=['adam'], tunable=False) return parser","title":"Return"},{"location":"Trainer/","text":"Trainer [ Github Code ] The lightning trainer abstracts best practices for running a training, val, test routine. It calls parts of your model when it wants to hand over full control and otherwise makes training assumptions which are now standard practice in AI research. This is the basic use of the trainer: from pytorch_lightning import Trainer model = LightningTemplate() trainer = Trainer() trainer.fit(model) But of course the fun is in all the advanced things it can do: Training loop Accumulate gradients Anneal Learning rate Check GPU usage Check which gradients are nan Check validation every n epochs Display metrics in progress bar Display the parameter count by layer Force training for min or max epochs Inspect gradient norms Make model overfit on subset of data Use multiple optimizers (like GANs) Set how much of the training set to check (1-100%) Validation loop Display metrics in progress bar Set how much of the validation set to check (1-100%) Set validation check frequency within 1 training epoch (1-100%) validation_step function Why does validation run first for 5 steps? Distributed training Single-gpu Multi-gpu Multi-node 16-bit mixed precision Checkpointing Model saving Model loading Computing cluster (SLURM) Automatic checkpointing Automatic saving, loading Running grid search on a cluster Walltime auto-resubmit","title":"Trainer"},{"location":"Trainer/#trainer","text":"[ Github Code ] The lightning trainer abstracts best practices for running a training, val, test routine. It calls parts of your model when it wants to hand over full control and otherwise makes training assumptions which are now standard practice in AI research. This is the basic use of the trainer: from pytorch_lightning import Trainer model = LightningTemplate() trainer = Trainer() trainer.fit(model) But of course the fun is in all the advanced things it can do: Training loop Accumulate gradients Anneal Learning rate Check GPU usage Check which gradients are nan Check validation every n epochs Display metrics in progress bar Display the parameter count by layer Force training for min or max epochs Inspect gradient norms Make model overfit on subset of data Use multiple optimizers (like GANs) Set how much of the training set to check (1-100%) Validation loop Display metrics in progress bar Set how much of the validation set to check (1-100%) Set validation check frequency within 1 training epoch (1-100%) validation_step function Why does validation run first for 5 steps? Distributed training Single-gpu Multi-gpu Multi-node 16-bit mixed precision Checkpointing Model saving Model loading Computing cluster (SLURM) Automatic checkpointing Automatic saving, loading Running grid search on a cluster Walltime auto-resubmit","title":"Trainer"},{"location":"Trainer/Checkpointing/","text":"","title":"Checkpointing"},{"location":"Trainer/Distributed training/","text":"","title":"Distributed training"},{"location":"Trainer/SLURM Managed Cluster/","text":"","title":"SLURM Managed Cluster"},{"location":"Trainer/Training Loop/","text":"The asdf Accumulated gradients Accumulated gradients runs K small batches of size N before doing a backwards pass. The effect is a large effective batch size of size KxN. # DEFAULT (ie: no accumulated grads) trainer = Trainer(accumulate_grad_batches=1) Anneal Learning rate Cut the learning rate by 10 at every epoch listed in this list. # DEFAULT (don't anneal) trainer = Trainer(lr_scheduler_milestones=None) # cut LR by 10 at 100, 200, and 300 epochs trainer = Trainer(lr_scheduler_milestones=[100, 200, 300]) Check GPU usage Lightning automatically logs gpu usage to the test tube logs. It'll only do it at the metric logging interval, so it doesn't slow down training. Check which gradients are nan This option prints a list of tensors with nan gradients. # DEFAULT trainer = Trainer(print_nan_grads=False) Check validation every n epochs If you have a small dataset you might want to check validation every n epochs # DEFAULT trainer = Trainer(check_val_every_n_epoch=1) Display metrics in progress bar # DEFAULT trainer = Trainer(progress_bar=True) Display the parameter count by layer By default lightning prints a list of parameters and submodules when it starts training. Force training for min or max epochs It can be useful to force training for a minimum number of epochs or limit to a max number # DEFAULT trainer = Trainer(min_nb_epochs=1, max_nb_epochs=1000) Inspect gradient norms Looking at grad norms can help you figure out where training might be going wrong. # DEFAULT (-1 doesn't track norms) trainer = Trainer(track_grad_norm=-1) # track the LP norm (P=2 here) 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. # DEFAULT don't overfit (ie: normal training) trainer = Trainer(overfit_pct=0.0) # overfit on 1% of data trainer = Trainer(overfit_pct=0.01) Set how much of the training 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 # DEFAULT trainer = Trainer(train_percent_check=1.0) # check 10% only trainer = Trainer(train_percent_check=0.1)","title":"Training Loop"},{"location":"Trainer/Training Loop/#accumulated-gradients","text":"Accumulated gradients runs K small batches of size N before doing a backwards pass. The effect is a large effective batch size of size KxN. # DEFAULT (ie: no accumulated grads) trainer = Trainer(accumulate_grad_batches=1)","title":"Accumulated gradients"},{"location":"Trainer/Training Loop/#anneal-learning-rate","text":"Cut the learning rate by 10 at every epoch listed in this list. # DEFAULT (don't anneal) trainer = Trainer(lr_scheduler_milestones=None) # cut LR by 10 at 100, 200, and 300 epochs trainer = Trainer(lr_scheduler_milestones=[100, 200, 300])","title":"Anneal Learning rate"},{"location":"Trainer/Training Loop/#check-gpu-usage","text":"Lightning automatically logs gpu usage to the test tube logs. It'll only do it at the metric logging interval, so it doesn't slow down training.","title":"Check GPU usage"},{"location":"Trainer/Training Loop/#check-which-gradients-are-nan","text":"This option prints a list of tensors with nan gradients. # DEFAULT trainer = Trainer(print_nan_grads=False)","title":"Check which gradients are nan"},{"location":"Trainer/Training Loop/#check-validation-every-n-epochs","text":"If you have a small dataset you might want to check validation every n epochs # DEFAULT trainer = Trainer(check_val_every_n_epoch=1)","title":"Check validation every n epochs"},{"location":"Trainer/Training Loop/#display-metrics-in-progress-bar","text":"# DEFAULT trainer = Trainer(progress_bar=True)","title":"Display metrics in progress bar"},{"location":"Trainer/Training Loop/#display-the-parameter-count-by-layer","text":"By default lightning prints a list of parameters and submodules when it starts training.","title":"Display the parameter count by layer"},{"location":"Trainer/Training Loop/#force-training-for-min-or-max-epochs","text":"It can be useful to force training for a minimum number of epochs or limit to a max number # DEFAULT trainer = Trainer(min_nb_epochs=1, max_nb_epochs=1000)","title":"Force training for min or max epochs"},{"location":"Trainer/Training Loop/#inspect-gradient-norms","text":"Looking at grad norms can help you figure out where training might be going wrong. # DEFAULT (-1 doesn't track norms) trainer = Trainer(track_grad_norm=-1) # track the LP norm (P=2 here) trainer = Trainer(track_grad_norm=2)","title":"Inspect gradient norms"},{"location":"Trainer/Training Loop/#make-model-overfit-on-subset-of-data","text":"A useful debugging trick is to make your model overfit a tiny fraction of the data. # DEFAULT don't overfit (ie: normal training) trainer = Trainer(overfit_pct=0.0) # overfit on 1% of data trainer = Trainer(overfit_pct=0.01)","title":"Make model overfit on subset of data"},{"location":"Trainer/Training Loop/#set-how-much-of-the-training-set-to-check","text":"If you don't want to check 100% of the validation set (for debugging or if it's huge), set this flag # DEFAULT trainer = Trainer(train_percent_check=1.0) # check 10% only trainer = Trainer(train_percent_check=0.1)","title":"Set how much of the training set to check"},{"location":"Trainer/Vaildation loop/","text":"","title":"Vaildation loop"}]} \ No newline at end of file diff --git a/sitemap.xml b/sitemap.xml index 014f5456..4ef71170 100644 --- a/sitemap.xml +++ b/sitemap.xml @@ -15,4 +15,29 @@ 2019-06-27 daily + + None + 2019-06-27 + daily + + + None + 2019-06-27 + daily + + + None + 2019-06-27 + daily + + + None + 2019-06-27 + daily + + + None + 2019-06-27 + daily + \ No newline at end of file diff --git a/sitemap.xml.gz b/sitemap.xml.gz index 7e0704ce..ec79a08e 100644 Binary files a/sitemap.xml.gz and b/sitemap.xml.gz differ diff --git a/source/examples/basic_trainer.py b/source/examples/basic_trainer.py index 308f24eb..a1f6d851 100644 --- a/source/examples/basic_trainer.py +++ b/source/examples/basic_trainer.py @@ -4,7 +4,7 @@ import sys from test_tube import HyperOptArgumentParser, Experiment from pytorch_lightning.models.trainer import Trainer from pytorch_lightning.utils.arg_parse import add_default_args -from pytorch_lightning.utils.pt_callbacks import EarlyStopping, ModelCheckpoint +from pytorch_lightning.callbacks.pt_callbacks import EarlyStopping, ModelCheckpoint from docs.source.examples.example_model import ExampleModel diff --git a/source/examples/fully_featured_trainer.py b/source/examples/fully_featured_trainer.py index 0ef013df..bbf05c5f 100644 --- a/source/examples/fully_featured_trainer.py +++ b/source/examples/fully_featured_trainer.py @@ -8,7 +8,7 @@ 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.utils.pt_callbacks import EarlyStopping, ModelCheckpoint +from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint SEED = 2334 torch.manual_seed(SEED)