Pytorch-lightning

Seed for ML research

Usage

To use lightning, define a model that implements these 10 functions:

Model definition

Name	Description	Input	Return
training_step	Called with a batch of data during training	data from your dataloaders	tuple: scalar, dict
validation_step	Called with a batch of data during validation	data from your dataloaders	tuple: scalar, dict
validation_end	Collate metrics from all validation steps	outputs: array where each item is the output of a validation step	dict: for logging
get_save_dict	called when your model needs to be saved (checkpoints, hpc save, etc...)	None	dict to be saved
load_model_specific

Model training

Name	Description	Input	Return
configure_optimizers	called during training setup	None	list: optimizers you want to use
tng_dataloader	called during training	None	pytorch dataloader
val_dataloader	called during validation	None	pytorch dataloader
test_dataloader	called during testing	None	pytorch dataloader
add_model_specific_args	called with args you defined in your main. This lets you tailor args for each model and keep main the same	argparse	argparse

Model Saving/Loading

Name	Description	Input	Return
get_save_dict	called when your model needs to be saved (checkpoints, hpc save, etc...)	None	dict to be saved
load_model_specific	called when loading a model	checkpoint: dict you created in get_save_dict	dict: modified in whatever way you want

Example

import torch.nn as nn

class ExampleModel(RootModule):
    def __init__(self):
        self.l1 = nn.Linear(100, 20)
    
    # TRAINING
    def training_step(self, data_batch):
        # your dataloader decides what each batch looks like
        x, y = data_batch
        y_hat = self.l1(x)
        loss = some_loss(y_hat)
        
        tqdm_dic = {'train_loss': loss}
        
        # must return scalar, dict for logging
        return loss_val, tqdm_dic
    
    def validation_step(self, data_batch):
        # same as training...
        x, y = data_batch
        y_hat = self.l1(x)
        loss = some_loss(y_hat)
        
        # val specific
        acc = calculate_acc(y_hat, y)
        
        tqdm_dic = {'train_loss': loss, 'val_acc': acc, 'whatever_you_want': 'a'}
        return loss_val, tqdm_dic
 
     def validation_end(self, outputs):
        total_accs = []
        
        # given to you by the framework with all validation outputs.
        # chance to collate
        for output in outputs:
            total_accs.append(output['val_acc'].item())
        
        # return a dict
        return {'total_acc': np.mean(total_accs)}
        
     # SAVING
     def get_save_dict(self):
        # lightning saves for you. Here's your chance to say what you want to save
        checkpoint = {'state_dict': self.state_dict()}

        return checkpoint

    def load_model_specific(self, checkpoint):
        # lightning loads for you. Here's your chance to say what you want to load
        self.load_state_dict(checkpoint['state_dict'])
        pass
    
    # TRAINING CONFIG
    def configure_optimizers(self):
        # give lightning the list of optimizers you want to use.
        # lightning will call automatically
        optimizer = self.choose_optimizer(self.hparams.optimizer_name, self.parameters(), {'lr': self.hparams.learning_rate}, 'optimizer')
        self.optimizers = [optimizer]
        return self.optimizers
    
    # LIGHTING WILL USE THE LOADERS YOU DEFINE HERE
    @property
    def tng_dataloader(self):
        return pytorch_dataloader('train')

    @property
    def val_dataloader(self):
        return pytorch_dataloader('val')

    @property
    def test_dataloader(self):
        return pytorch_dataloader('test')
    
    # MODIFY YOUR COMMAND LINE ARGS
    @staticmethod
    def add_model_specific_args(parent_parser):    
        parser = HyperOptArgumentParser(strategy=parent_parser.strategy, parents=[parent_parser])
        parser.add_argument('--out_features', default=20)
        return parser

Add new model

1. Create a new model under /models.
2. Add model name to trainer_main

AVAILABLE_MODELS = {
    'model_1': ExampleModel1
}

Model methods that can be implemented

Method	Purpose	Input	Output	Required
forward()	Forward pass	model_in tuple with your data	model_out tuple to be passed to loss	Y
loss()	calculate model loss	model_out tuple from forward()	A scalar	Y
check_performance()	run a full loop through val data to check for metrics	dataloader, nb_tests	metrics tuple to be tracked	Y
tng_dataloader	Computed option, used to feed tng data	-	Pytorch DataLoader subclass	Y
val_dataloader	Computed option, used to feed tng data	-	Pytorch DataLoader subclass	Y
test_dataloader	Computed option, used to feed tng data	-	Pytorch DataLoader subclass	Y

Model lifecycle hooks

Use these hooks to customize functionality

Method	Purpose	Input	Output	Required
on_batch_start()	called right before the batch starts	-	-	N
on_batch_end()	called right after the batch ends	-	-	N
on_epoch_start()	called right before the epoch starts	-	-	N
on_epoch_end()	called right afger the epoch ends	-	-	N
on_pre_performance_check()	called right before the performance check starts	-	-	N
on_post_performance_check()	called right after the batch starts	-	-	N