Model Hooks

There are cases when you might want to do something different at different parts of the training/validation loop. To enable a hook, simply override the method in your LightningModule and the trainer will call it at the correct time.

Contributing If there’s a hook you’d like to add, simply:

1. Fork PyTorchLightning.

2. Add the hook to pytorch_lightning.core.hooks.ModelHooks.

3. Add it in the correct place in pytorch_lightning.trainer where it should be called.

---

Hooks lifecycle

Training set-up

• prepare_data()

• setup()

• init_ddp_connection()

• init_optimizers()

• configure_apex()

• configure_ddp()

• train_dataloader()

• test_dataloader()

• val_dataloader()

• summarize()

• restore_weights()

Warning

prepare_data is only called from global_rank=0. Don’t assign state (self.something), use setup for that

---

Training loop

• on_epoch_start()

• on_batch_start()

• tbptt_split_batch()

• training_step()

• training_step_end() (optional)

• on_before_zero_grad()

• backward()

• on_after_backward()

• optimizer.step()

• on_batch_end()

• training_epoch_end()

• on_epoch_end()

---

Validation loop

• model.zero_grad()

• model.eval()

• torch.set_grad_enabled(False)

• validation_step()

• validation_step_end()

• validation_epoch_end()

• model.train()

• torch.set_grad_enabled(True)

• on_post_performance_check()

---

Test loop

• model.zero_grad()

• model.eval()

• torch.set_grad_enabled(False)

• test_step()

• test_step_end()

• test_epoch_end()

• model.train()

• torch.set_grad_enabled(True)

• on_post_performance_check()

---

General hooks

class pytorch_lightning.core.hooks.ModelHooks(*args, **kwargs)

Bases: torch.nn.Module

backward(trainer, loss, optimizer, optimizer_idx)

Override backward with your own implementation if you need to.

Parameters

• trainer – Pointer to the trainer

• loss (Tensor) – Loss is already scaled by accumulated grads

• optimizer (Optimizer) – Current optimizer being used

• optimizer_idx (int) – Index of the current optimizer being used

Called to perform backward step. Feel free to override as needed.

The loss passed in has already been scaled for accumulated gradients if requested.

Example:

def backward(self, trainer, loss, optimizer, optimizer_idx):
    loss.backward()

Return type

None

on_after_backward()

Called in the training loop after loss.backward() and before optimizers do anything. This is the ideal place to inspect or log gradient information.

Example:

def on_after_backward(self):
    # example to inspect gradient information in tensorboard
    if self.trainer.global_step % 25 == 0:  # don't make the tf file huge
        params = self.state_dict()
        for k, v in params.items():
            grads = v
            name = k
            self.logger.experiment.add_histogram(tag=name, values=grads,
                                                 global_step=self.trainer.global_step)

Return type

None

on_batch_end()

Called in the training loop after the batch.

Return type

None

on_batch_start(batch)

Called in the training loop before anything happens for that batch.

If you return -1 here, you will skip training for the rest of the current epoch.

Parameters

batch (Any) – The batched data as it is returned by the training DataLoader.

Return type

None

on_before_zero_grad(optimizer)

Called after optimizer.step() and before optimizer.zero_grad().

Called in the training loop after taking an optimizer step and before zeroing grads. Good place to inspect weight information with weights updated.

This is where it is called:

for optimizer in optimizers:
    optimizer.step()
    model.on_before_zero_grad(optimizer) # < ---- called here
    optimizer.zero_grad

Parameters

optimizer (Optimizer) – The optimizer for which grads should be zeroed.

Return type

None

on_epoch_end()

Called in the training loop at the very end of the epoch.

Return type

None

on_epoch_start()

Called in the training loop at the very beginning of the epoch.

Return type

None

on_fit_end()

Called at the very end of fit. If on DDP it is called on every process

on_fit_start()

Called at the very beginning of fit. If on DDP it is called on every process

on_post_performance_check()

Called at the very end of the validation loop.

Return type

None

on_pre_performance_check()

Called at the very beginning of the validation loop.

Return type

None

on_sanity_check_start()

Called before starting evaluation.

Warning

Deprecated. Will be removed in v0.9.0.

on_train_end()

Called at the end of training before logger experiment is closed.

Return type

None

on_train_start()

Called at the beginning of training before sanity check.

Return type

None

setup(stage)

Called at the beginning of fit and test. This is a good hook when you need to build models dynamically or adjust something about them. This hook is called on every process when using DDP.

Parameters

stage (str) – either ‘fit’ or ‘test’

Example:

class LitModel(...):
    def __init__(self):
        self.l1 = None

    def prepare_data(self):
        download_data()
        tokenize()

        # don't do this
        self.something = else

    def setup(stage):
        data = Load_data(...)
        self.l1 = nn.Linear(28, data.num_classes)

teardown(stage)

Called at the end of fit and test.

Parameters

stage (str) – either ‘fit’ or ‘test’

transfer_batch_to_device(batch, device)

Override this hook if your DataLoader returns tensors wrapped in a custom data structure.

The data types listed below (and any arbitrary nesting of them) are supported out of the box:

• torch.Tensor or anything that implements .to(…)

• list

• dict

• tuple

• torchtext.data.batch.Batch

For anything else, you need to define how the data is moved to the target device (CPU, GPU, TPU, …).

Example:

def transfer_batch_to_device(self, batch, device)
    if isinstance(batch, CustomBatch):
        # move all tensors in your custom data structure to the device
        batch.samples = batch.samples.to(device)
        batch.targets = batch.targets.to(device)
    else:
        batch = super().transfer_batch_to_device(data, device)
    return batch

Parameters

• batch (Any) – A batch of data that needs to be transferred to a new device.

• device (device) – The target device as defined in PyTorch.

Return type

Any

Returns

A reference to the data on the new device.

Note

This hook should only transfer the data and not modify it, nor should it move the data to any other device than the one passed in as argument (unless you know what you are doing). The Trainer already takes care of splitting the batch and determines the target devices.

See also

• move_data_to_device()

• apply_to_collection()