lightning/pytorch_lightning/core/lightning.py

1749 lines
65 KiB
Python

import collections
import inspect
import os
import re
from abc import ABC, abstractmethod
from argparse import Namespace
from typing import Any, Callable, Dict, List, Optional, Tuple, Union, Sequence
import torch
import torch.distributed as torch_distrib
from torch import Tensor
from torch.nn import Module
from torch.nn.parallel import DistributedDataParallel
from torch.optim.optimizer import Optimizer
from torch.utils.data import DataLoader
from pytorch_lightning import _logger as log
from pytorch_lightning.core.grads import GradInformation
from pytorch_lightning.core.hooks import ModelHooks
from pytorch_lightning.core.memory import ModelSummary
from pytorch_lightning.core.saving import ModelIO, PRIMITIVE_TYPES, ALLOWED_CONFIG_TYPES
from pytorch_lightning.utilities.device_dtype_mixin import DeviceDtypeModuleMixin
from pytorch_lightning.overrides.data_parallel import LightningDistributedDataParallel
from pytorch_lightning.utilities import rank_zero_warn
from pytorch_lightning.utilities.parsing import AttributeDict, collect_init_args, get_init_args
try:
import torch_xla.core.xla_model as xm
except ImportError:
XLA_AVAILABLE = False
else:
XLA_AVAILABLE = True
class LightningModule(ABC, DeviceDtypeModuleMixin, GradInformation, ModelIO, ModelHooks, Module):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.exp_save_path = None
#: The current epoch
self.current_epoch = 0
#: Total training batches seen across all epochs
self.global_step = 0
self.loaded_optimizer_states_dict = {}
#: Pointer to the trainer object
self.trainer = None
#: Pointer to the logger object
self.logger = None
#: True if using dp
self.use_dp = False
#: True if using ddp
self.use_ddp = False
#: True if using ddp2
self.use_ddp2 = False
# True if on tpu
self.use_tpu = False
#: True if using amp
self.use_amp = False
#: Current dtype
self._dtype = torch.float
#: device reference
self._device = torch.device('cpu')
# optionally can be set by user
self._example_input_array = None
@property
def example_input_array(self) -> Any:
return self._example_input_array
@example_input_array.setter
def example_input_array(self, example: Any) -> None:
self._example_input_array = example
@property
def on_gpu(self):
"""
True if your model is currently running on GPUs.
Useful to set flags around the LightningModule for different CPU vs GPU behavior.
"""
return self.device.type == 'cuda'
def print(self, *args, **kwargs) -> None:
r"""
Prints only from process 0. Use this in any distributed mode to log only once.
Args:
*args: The thing to print. Will be passed to Python's built-in print function.
**kwargs: Will be passed to Python's built-in print function.
Example:
.. code-block:: python
def forward(self, x):
self.print(x, 'in forward')
"""
if self.trainer.is_global_zero:
print(*args, **kwargs)
@abstractmethod
def forward(self, *args, **kwargs):
r"""
Same as :meth:`torch.nn.Module.forward()`, however in Lightning you want this to define
the operations you want to use for prediction (i.e.: on a server or as a feature extractor).
Normally you'd call ``self()`` from your :meth:`training_step` method.
This makes it easy to write a complex system for training with the outputs
you'd want in a prediction setting.
You may also find the :func:`~pytorch_lightning.core.decorators.auto_move_data` decorator useful
when using the module outside Lightning in a production setting.
Args:
*args: Whatever you decide to pass into the forward method.
**kwargs: Keyword arguments are also possible.
Return:
Predicted output
Examples:
.. code-block:: python
# example if we were using this model as a feature extractor
def forward(self, x):
feature_maps = self.convnet(x)
return feature_maps
def training_step(self, batch, batch_idx):
x, y = batch
feature_maps = self(x)
logits = self.classifier(feature_maps)
# ...
return loss
# splitting it this way allows model to be used a feature extractor
model = MyModelAbove()
inputs = server.get_request()
results = model(inputs)
server.write_results(results)
# -------------
# This is in stark contrast to torch.nn.Module where normally you would have this:
def forward(self, batch):
x, y = batch
feature_maps = self.convnet(x)
logits = self.classifier(feature_maps)
return logits
"""
def training_step(self, *args, **kwargs) -> Union[
int, Dict[str, Union[Tensor, Dict[str, Tensor]]]
]:
r"""
Here you compute and return the training loss and some additional metrics for e.g.
the progress bar or logger.
Args:
batch (:class:`~torch.Tensor` | (:class:`~torch.Tensor`, ...) | [:class:`~torch.Tensor`, ...]):
The output of your :class:`~torch.utils.data.DataLoader`. A tensor, tuple or list.
batch_idx (int): Integer displaying index of this batch
optimizer_idx (int): When using multiple optimizers, this argument will also be present.
hiddens(:class:`~torch.Tensor`): Passed in if
:paramref:`~pytorch_lightning.trainer.trainer.Trainer.truncated_bptt_steps` > 0.
Return:
Dict with loss key and optional log or progress bar keys.
When implementing :meth:`training_step`, return whatever you need in that step:
- loss -> tensor scalar **REQUIRED**
- progress_bar -> Dict for progress bar display. Must have only tensors
- log -> Dict of metrics to add to logger. Must have only tensors (no images, etc)
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 model specific.
Examples:
.. code-block:: python
def training_step(self, batch, batch_idx):
x, y, z = batch
# implement your own
out = self(x)
loss = self.loss(out, x)
logger_logs = {'training_loss': loss} # optional (MUST ALL BE TENSORS)
# if using TestTubeLogger or TensorBoardLogger you can nest scalars
logger_logs = {'losses': logger_logs} # optional (MUST ALL BE TENSORS)
output = {
'loss': loss, # required
'progress_bar': {'training_loss': loss}, # optional (MUST ALL BE TENSORS)
'log': logger_logs
}
# return a dict
return output
If you define multiple optimizers, this step will be called with an additional
``optimizer_idx`` parameter.
.. code-block:: python
# Multiple optimizers (e.g.: GANs)
def training_step(self, batch, batch_idx, optimizer_idx):
if optimizer_idx == 0:
# do training_step with encoder
if optimizer_idx == 1:
# do training_step with decoder
If you add truncated back propagation through time you will also get an additional
argument with the hidden states of the previous step.
.. code-block:: python
# Truncated back-propagation through time
def training_step(self, batch, batch_idx, hiddens):
# hiddens are the hidden states from the previous truncated backprop step
...
out, hiddens = self.lstm(data, hiddens)
...
return {
"loss": ...,
"hiddens": hiddens # remember to detach() this
}
Notes:
The loss value shown in the progress bar is smoothed (averaged) over the last values,
so it differs from the actual loss returned in train/validation step.
"""
rank_zero_warn('`training_step` must be implemented to be used with the Lightning Trainer')
def training_end(self, *args, **kwargs):
"""
Warnings:
Deprecated in v0.7.0. Use :meth:`training_step_end` instead.
"""
def training_epoch_end(
self,
outputs: Union[List[Dict[str, Tensor]], List[List[Dict[str, Tensor]]]]
) -> Dict[str, Dict[str, Tensor]]:
"""Called at the end of the training epoch with the outputs of all training steps.
.. code-block:: python
# the pseudocode for these calls
train_outs = []
for train_batch in train_data:
out = training_step(train_batch)
train_outs.append(out)
training_epoch_end(train_outs)
Args:
outputs: List of outputs you defined in :meth:`training_step`, or if there are
multiple dataloaders, a list containing a list of outputs for each dataloader.
Return:
Dict or OrderedDict.
May contain the following optional keys:
- log (metrics to be added to the logger; only tensors)
- progress_bar (dict for progress bar display)
- any metric used in a callback (e.g. early stopping).
Note:
If this method is not overridden, this won't be called.
- The outputs here are strictly for logging or progress bar.
- If you don't need to display anything, don't return anything.
- If you want to manually set current step, you can specify the 'step' key in the 'log' dict.
Examples:
With a single dataloader:
.. code-block:: python
def training_epoch_end(self, outputs):
train_acc_mean = 0
for output in outputs:
train_acc_mean += output['train_acc']
train_acc_mean /= len(outputs)
# log training accuracy at the end of an epoch
results = {
'log': {'train_acc': train_acc_mean.item()},
'progress_bar': {'train_acc': train_acc_mean},
}
return results
With multiple dataloaders, ``outputs`` will be a list of lists. The outer list contains
one entry per dataloader, while the inner list contains the individual outputs of
each training step for that dataloader.
.. code-block:: python
def training_epoch_end(self, outputs):
train_acc_mean = 0
i = 0
for dataloader_outputs in outputs:
for output in dataloader_outputs:
train_acc_mean += output['train_acc']
i += 1
train_acc_mean /= i
# log training accuracy at the end of an epoch
results = {
'log': {'train_acc': train_acc_mean.item(), 'step': self.current_epoch}
'progress_bar': {'train_acc': train_acc_mean},
}
return results
"""
def training_step_end(self, *args, **kwargs) -> Dict[
str, Union[Tensor, Dict[str, Tensor]]
]:
"""
Use this when training with dp or ddp2 because :meth:`training_step`
will operate on only part of the batch. However, this is still optional
and only needed for things like softmax or NCE loss.
Note:
If you later switch to ddp or some other mode, this will still be called
so that you don't have to change your code
.. code-block:: python
# pseudocode
sub_batches = split_batches_for_dp(batch)
batch_parts_outputs = [training_step(sub_batch) for sub_batch in sub_batches]
training_step_end(batch_parts_outputs)
Args:
batch_parts_outputs: What you return in `training_step` for each batch part.
Return:
Dict with loss key and optional log or progress bar keys.
- loss -> tensor scalar **REQUIRED**
- progress_bar -> Dict for progress bar display. Must have only tensors
- log -> Dict of metrics to add to logger. Must have only tensors (no images, etc)
Examples:
.. code-block:: python
# WITHOUT training_step_end
# if used in DP or DDP2, this batch is 1/num_gpus large
def training_step(self, batch, batch_idx):
# batch is 1/num_gpus big
x, y = batch
out = self(x)
loss = self.softmax(out)
loss = nce_loss(loss)
return {'loss': loss}
# --------------
# with training_step_end to do softmax over the full batch
def training_step(self, batch, batch_idx):
# batch is 1/num_gpus big
x, y = batch
out = self(x)
return {'out': out}
def training_step_end(self, outputs):
# this out is now the full size of the batch
out = outputs['out']
# this softmax now uses the full batch size
loss = nce_loss(loss)
return {'loss': loss}
See Also:
See the :ref:`multi-gpu-training` guide for more details.
"""
def validation_step(self, *args, **kwargs) -> Dict[str, Tensor]:
r"""
Operates on a single batch of data from the validation set.
In this step you'd might generate examples or calculate anything of interest like accuracy.
.. code-block:: python
# the pseudocode for these calls
val_outs = []
for val_batch in val_data:
out = validation_step(train_batch)
val_outs.append(out)
validation_epoch_end(val_outs)
Args:
batch (:class:`~torch.Tensor` | (:class:`~torch.Tensor`, ...) | [:class:`~torch.Tensor`, ...]):
The output of your :class:`~torch.utils.data.DataLoader`. A tensor, tuple or list.
batch_idx (int): The index of this batch
dataloader_idx (int): The index of the dataloader that produced this batch
(only if multiple val datasets used)
Return:
Dict or OrderedDict - passed to :meth:`validation_epoch_end`.
If you defined :meth:`validation_step_end` it will go to that first.
.. code-block:: python
# pseudocode of order
out = validation_step()
if defined('validation_step_end'):
out = validation_step_end(out)
out = validation_epoch_end(out)
.. code-block:: python
# if you have one val dataloader:
def validation_step(self, batch, batch_idx)
# if you have multiple val dataloaders:
def validation_step(self, batch, batch_idx, dataloader_idx)
Examples:
.. code-block:: python
# CASE 1: A single validation dataset
def validation_step(self, batch, batch_idx):
x, y = batch
# implement your own
out = self(x)
loss = self.loss(out, y)
# log 6 example images
# or generated text... or whatever
sample_imgs = x[:6]
grid = torchvision.utils.make_grid(sample_imgs)
self.logger.experiment.add_image('example_images', grid, 0)
# 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_epoch_end
output = OrderedDict({
'val_loss': loss_val,
'val_acc': torch.tensor(val_acc), # everything must be a tensor
})
# return an optional dict
return output
If you pass in multiple val datasets, validation_step will have an additional argument.
.. code-block:: python
# CASE 2: multiple validation datasets
def validation_step(self, batch, batch_idx, dataset_idx):
# dataset_idx tells you which dataset this is.
Note:
If you don't need to validate you don't need to implement this method.
Note:
When the :meth:`validation_step` is called, the model has been put in eval mode
and PyTorch gradients have been disabled. At the end of validation,
the model goes back to training mode and gradients are enabled.
"""
def validation_step_end(self, *args, **kwargs) -> Dict[str, Tensor]:
"""
Use this when validating with dp or ddp2 because :meth:`validation_step`
will operate on only part of the batch. However, this is still optional
and only needed for things like softmax or NCE loss.
Note:
If you later switch to ddp or some other mode, this will still be called
so that you don't have to change your code.
.. code-block:: python
# pseudocode
sub_batches = split_batches_for_dp(batch)
batch_parts_outputs = [validation_step(sub_batch) for sub_batch in sub_batches]
validation_step_end(batch_parts_outputs)
Args:
batch_parts_outputs: What you return in :meth:`validation_step`
for each batch part.
Return:
Dict or OrderedDict - passed to the :meth:`validation_epoch_end` method.
Examples:
.. code-block:: python
# WITHOUT validation_step_end
# if used in DP or DDP2, this batch is 1/num_gpus large
def validation_step(self, batch, batch_idx):
# batch is 1/num_gpus big
x, y = batch
out = self(x)
loss = self.softmax(out)
loss = nce_loss(loss)
return {'loss': loss}
# --------------
# with validation_step_end to do softmax over the full batch
def validation_step(self, batch, batch_idx):
# batch is 1/num_gpus big
x, y = batch
out = self(x)
return {'out': out}
def validation_epoch_end(self, outputs):
# this out is now the full size of the batch
out = outputs['out']
# this softmax now uses the full batch size
loss = nce_loss(loss)
return {'loss': loss}
See Also:
See the :ref:`multi-gpu-training` guide for more details.
"""
def validation_end(self, outputs):
"""
Warnings:
Deprecated in v0.7.0. Use :meth:`validation_epoch_end` instead.
Will be removed in 1.0.0.
"""
def validation_epoch_end(
self,
outputs: Union[List[Dict[str, Tensor]], List[List[Dict[str, Tensor]]]]
) -> Dict[str, Dict[str, Tensor]]:
"""
Called at the end of the validation epoch with the outputs of all validation steps.
.. code-block:: python
# the pseudocode for these calls
val_outs = []
for val_batch in val_data:
out = validation_step(val_batch)
val_outs.append(out)
validation_epoch_end(val_outs)
Args:
outputs: List of outputs you defined in :meth:`validation_step`, or if there
are multiple dataloaders, a list containing a list of outputs for each dataloader.
Return:
Dict or OrderedDict.
May have the following optional keys:
- progress_bar (dict for progress bar display; only tensors)
- log (dict of metrics to add to logger; only tensors).
Note:
If you didn't define a :meth:`validation_step`, this won't be called.
- The outputs here are strictly for logging or progress bar.
- If you don't need to display anything, don't return anything.
- If you want to manually set current step, you can specify the 'step' key in the 'log' dict.
Examples:
With a single dataloader:
.. code-block:: python
def validation_epoch_end(self, outputs):
val_acc_mean = 0
for output in outputs:
val_acc_mean += output['val_acc']
val_acc_mean /= len(outputs)
tqdm_dict = {'val_acc': val_acc_mean.item()}
# show val_acc in progress bar but only log val_loss
results = {
'progress_bar': tqdm_dict,
'log': {'val_acc': val_acc_mean.item()}
}
return results
With multiple dataloaders, `outputs` will be a list of lists. The outer list contains
one entry per dataloader, while the inner list contains the individual outputs of
each validation step for that dataloader.
.. code-block:: python
def validation_epoch_end(self, outputs):
val_acc_mean = 0
i = 0
for dataloader_outputs in outputs:
for output in dataloader_outputs:
val_acc_mean += output['val_acc']
i += 1
val_acc_mean /= i
tqdm_dict = {'val_acc': val_acc_mean.item()}
# show val_loss and val_acc in progress bar but only log val_loss
results = {
'progress_bar': tqdm_dict,
'log': {'val_acc': val_acc_mean.item(), 'step': self.current_epoch}
}
return results
"""
def test_step(self, *args, **kwargs) -> Dict[str, Tensor]:
r"""
Operates on a single batch of data from the test set.
In this step you'd normally generate examples or calculate anything of interest
such as accuracy.
.. code-block:: python
# the pseudocode for these calls
test_outs = []
for test_batch in test_data:
out = test_step(test_batch)
test_outs.append(out)
test_epoch_end(test_outs)
Args:
batch (:class:`~torch.Tensor` | (:class:`~torch.Tensor`, ...) | [:class:`~torch.Tensor`, ...]):
The output of your :class:`~torch.utils.data.DataLoader`. A tensor, tuple or list.
batch_idx (int): The index of this batch.
dataloader_idx (int): The index of the dataloader that produced this batch
(only if multiple test datasets used).
Return:
Dict or OrderedDict - passed to the :meth:`test_epoch_end` method.
If you defined :meth:`test_step_end` it will go to that first.
.. code-block:: python
# if you have one test dataloader:
def test_step(self, batch, batch_idx)
# if you have multiple test dataloaders:
def test_step(self, batch, batch_idx, dataloader_idx)
Examples:
.. code-block:: python
# CASE 1: A single test dataset
def test_step(self, batch, batch_idx):
x, y = batch
# implement your own
out = self(x)
loss = self.loss(out, y)
# log 6 example images
# or generated text... or whatever
sample_imgs = x[:6]
grid = torchvision.utils.make_grid(sample_imgs)
self.logger.experiment.add_image('example_images', grid, 0)
# 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 test_epoch_end
output = OrderedDict({
'val_loss': loss_val,
'val_acc': torch.tensor(val_acc), # everything must be a tensor
})
# return an optional dict
return output
If you pass in multiple validation datasets, :meth:`test_step` will have an additional
argument.
.. code-block:: python
# CASE 2: multiple test datasets
def test_step(self, batch, batch_idx, dataset_idx):
# dataset_idx tells you which dataset this is.
Note:
If you don't need to validate you don't need to implement this method.
Note:
When the :meth:`test_step` is called, the model has been put in eval mode and
PyTorch gradients have been disabled. At the end of the test epoch, the model goes back
to training mode and gradients are enabled.
"""
def test_step_end(self, *args, **kwargs) -> Dict[str, Tensor]:
"""
Use this when testing with dp or ddp2 because :meth:`test_step` will operate
on only part of the batch. However, this is still optional
and only needed for things like softmax or NCE loss.
Note:
If you later switch to ddp or some other mode, this will still be called
so that you don't have to change your code.
.. code-block:: python
# pseudocode
sub_batches = split_batches_for_dp(batch)
batch_parts_outputs = [test_step(sub_batch) for sub_batch in sub_batches]
test_step_end(batch_parts_outputs)
Args:
batch_parts_outputs: What you return in :meth:`test_step` for each batch part.
Return:
Dict or OrderedDict - passed to the :meth:`test_epoch_end`.
Examples:
.. code-block:: python
# WITHOUT test_step_end
# if used in DP or DDP2, this batch is 1/num_gpus large
def test_step(self, batch, batch_idx):
# batch is 1/num_gpus big
x, y = batch
out = self(x)
loss = self.softmax(out)
loss = nce_loss(loss)
return {'loss': loss}
# --------------
# with test_step_end to do softmax over the full batch
def test_step(self, batch, batch_idx):
# batch is 1/num_gpus big
x, y = batch
out = self(x)
return {'out': out}
def test_step_end(self, outputs):
# this out is now the full size of the batch
out = outputs['out']
# this softmax now uses the full batch size
loss = nce_loss(loss)
return {'loss': loss}
See Also:
See the :ref:`multi-gpu-training` guide for more details.
"""
def test_end(self, outputs):
"""
Warnings:
Deprecated in v0.7.0. Use :meth:`test_epoch_end` instead.
Will be removed in 1.0.0.
"""
def test_epoch_end(
self,
outputs: Union[List[Dict[str, Tensor]], List[List[Dict[str, Tensor]]]]
) -> Dict[str, Dict[str, Tensor]]:
"""
Called at the end of a test epoch with the output of all test steps.
.. code-block:: python
# the pseudocode for these calls
test_outs = []
for test_batch in test_data:
out = test_step(test_batch)
test_outs.append(out)
test_epoch_end(test_outs)
Args:
outputs: List of outputs you defined in :meth:`test_step_end`, or if there
are multiple dataloaders, a list containing a list of outputs for each dataloader
Return:
Dict or OrderedDict: Dict has the following optional keys:
- progress_bar -> Dict for progress bar display. Must have only tensors.
- log -> Dict of metrics to add to logger. Must have only tensors (no images, etc).
Note:
If you didn't define a :meth:`test_step`, this won't be called.
- The outputs here are strictly for logging or progress bar.
- If you don't need to display anything, don't return anything.
- If you want to manually set current step, specify it with the 'step' key in the 'log' Dict
Examples:
With a single dataloader:
.. code-block:: python
def test_epoch_end(self, outputs):
test_acc_mean = 0
for output in outputs:
test_acc_mean += output['test_acc']
test_acc_mean /= len(outputs)
tqdm_dict = {'test_acc': test_acc_mean.item()}
# show test_loss and test_acc in progress bar but only log test_loss
results = {
'progress_bar': tqdm_dict,
'log': {'test_acc': test_acc_mean.item()}
}
return results
With multiple dataloaders, `outputs` will be a list of lists. The outer list contains
one entry per dataloader, while the inner list contains the individual outputs of
each test step for that dataloader.
.. code-block:: python
def test_epoch_end(self, outputs):
test_acc_mean = 0
i = 0
for dataloader_outputs in outputs:
for output in dataloader_outputs:
test_acc_mean += output['test_acc']
i += 1
test_acc_mean /= i
tqdm_dict = {'test_acc': test_acc_mean.item()}
# show test_loss and test_acc in progress bar but only log test_loss
results = {
'progress_bar': tqdm_dict,
'log': {'test_acc': test_acc_mean.item(), 'step': self.current_epoch}
}
return results
"""
def configure_ddp(
self,
model: 'LightningModule',
device_ids: List[int]
) -> DistributedDataParallel:
r"""
Override to init DDP in your own way or with your own wrapper.
The only requirements are that:
1. On a validation batch the call goes to ``model.validation_step``.
2. On a training batch the call goes to ``model.training_step``.
3. On a testing batch, the call goes to ``model.test_step``.+
Args:
model: the :class:`LightningModule` currently being optimized.
device_ids: the list of GPU ids.
Return:
DDP wrapped model
Examples:
.. code-block:: python
# default implementation used in Trainer
def configure_ddp(self, model, device_ids):
# Lightning DDP simply routes to test_step, val_step, etc...
model = LightningDistributedDataParallel(
model,
device_ids=device_ids,
find_unused_parameters=True
)
return model
"""
model = LightningDistributedDataParallel(
model,
device_ids=device_ids,
find_unused_parameters=True
)
return model
def _init_slurm_connection(self) -> None:
"""
Sets up environment variables necessary for pytorch distributed communications
based on slurm environment.
"""
# use slurm job id for the port number
# guarantees unique ports across jobs from same grid search
try:
# use the last 4 numbers in the job id as the id
default_port = os.environ['SLURM_JOB_ID']
default_port = default_port[-4:]
# all ports should be in the 10k+ range
default_port = int(default_port) + 15000
except Exception:
default_port = 12910
# if user gave a port number, use that one instead
try:
default_port = os.environ['MASTER_PORT']
except Exception:
os.environ['MASTER_PORT'] = str(default_port)
# figure out the root node addr
try:
root_node = os.environ['SLURM_NODELIST'].split(' ')[0]
except Exception:
root_node = '127.0.0.1'
root_node = self.trainer.resolve_root_node_address(root_node)
os.environ['MASTER_ADDR'] = root_node
def init_ddp_connection(
self,
global_rank: int,
world_size: int,
is_slurm_managing_tasks: bool = True
) -> None:
"""
Override to define your custom way of setting up a distributed environment.
Lightning's implementation uses env:// init by default and sets the first node as root
for SLURM managed cluster.
Args:
global_rank: The global process idx.
world_size: Number of GPUs being use across all nodes. (num_nodes * num_gpus).
is_slurm_managing_tasks: is cluster managed by SLURM.
"""
if is_slurm_managing_tasks:
self._init_slurm_connection()
if 'MASTER_ADDR' not in os.environ:
rank_zero_warn("MASTER_ADDR environment variable is not defined. Set as localhost")
os.environ['MASTER_ADDR'] = '127.0.0.1'
log.debug(f"MASTER_ADDR: {os.environ['MASTER_ADDR']}")
if 'MASTER_PORT' not in os.environ:
rank_zero_warn("MASTER_PORT environment variable is not defined. Set as 12910")
os.environ['MASTER_PORT'] = '12910'
log.debug(f"MASTER_PORT: {os.environ['MASTER_PORT']}")
if 'WORLD_SIZE' in os.environ and int(os.environ['WORLD_SIZE']) != world_size:
rank_zero_warn(f"WORLD_SIZE environment variable ({os.environ['WORLD_SIZE']}) "
f"is not equal to the computed world size ({world_size}). Ignored.")
torch_backend = "nccl" if self.trainer.on_gpu else "gloo"
log.info(f"initializing ddp: GLOBAL_RANK: {global_rank}, MEMBER: {global_rank+1}/{world_size}")
torch_distrib.init_process_group(torch_backend, rank=global_rank, world_size=world_size)
def configure_apex(
self,
amp: object,
model: 'LightningModule',
optimizers: List[Optimizer],
amp_level: str
) -> Tuple['LightningModule', List[Optimizer]]:
r"""
Override to init AMP your own way.
Must return a model and list of optimizers.
Args:
amp: pointer to amp library object.
model: pointer to current :class:`LightningModule`.
optimizers: list of optimizers passed in :meth:`configure_optimizers`.
amp_level: AMP mode chosen ('O1', 'O2', etc...)
Return:
Apex wrapped model and optimizers
Examples:
.. code-block:: python
# Default implementation used by Trainer.
def configure_apex(self, amp, model, optimizers, amp_level):
model, optimizers = amp.initialize(
model, optimizers, opt_level=amp_level,
)
return model, optimizers
"""
model, optimizers = amp.initialize(model, optimizers, opt_level=amp_level)
return model, optimizers
def configure_optimizers(self) -> Optional[Union[
Optimizer, Sequence[Optimizer], Dict, Sequence[Dict], Tuple[List, List]
]]:
r"""
Choose what optimizers and learning-rate schedulers to use in your optimization.
Normally you'd need one. But in the case of GANs or similar you might have multiple.
Return:
Any of these 6 options.
- Single optimizer.
- List or Tuple - List of optimizers.
- Two lists - The first list has multiple optimizers, the second a list of LR schedulers (or lr_dict).
- Dictionary, with an 'optimizer' key, and (optionally) a 'lr_scheduler' key which value is a single LR scheduler or lr_dict.
- Tuple of dictionaries as described, with an optional 'frequency' key.
- None - Fit will run without any optimizer.
Note:
The 'frequency' value is an int corresponding to the number of sequential batches
optimized with the specific optimizer. It should be given to none or to all of the optimizers.
There is a difference between passing multiple optimizers in a list,
and passing multiple optimizers in dictionaries with a frequency of 1:
In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
The lr_dict is a dictionary which contains scheduler and its associated configuration.
It has five keys. The default configuration is shown below.
.. code-block:: python
{
'scheduler': lr_scheduler, # The LR schduler
'interval': 'epoch', # The unit of the scheduler's step size
'frequency': 1, # The frequency of the scheduler
'reduce_on_plateau': False, # For ReduceLROnPlateau scheduler
'monitor': 'val_loss' # Metric to monitor
}
If user only provides LR schedulers, then their configuration will set to default as shown above.
Examples:
.. code-block:: python
# most cases
def configure_optimizers(self):
opt = Adam(self.parameters(), lr=1e-3)
return opt
# multiple optimizer case (e.g.: GAN)
def configure_optimizers(self):
generator_opt = Adam(self.model_gen.parameters(), lr=0.01)
disriminator_opt = Adam(self.model_disc.parameters(), lr=0.02)
return generator_opt, disriminator_opt
# example with learning rate schedulers
def configure_optimizers(self):
generator_opt = Adam(self.model_gen.parameters(), lr=0.01)
disriminator_opt = Adam(self.model_disc.parameters(), lr=0.02)
discriminator_sched = CosineAnnealing(discriminator_opt, T_max=10)
return [generator_opt, disriminator_opt], [discriminator_sched]
# example with step-based learning rate schedulers
def configure_optimizers(self):
gen_opt = Adam(self.model_gen.parameters(), lr=0.01)
dis_opt = Adam(self.model_disc.parameters(), lr=0.02)
gen_sched = {'scheduler': ExponentialLR(gen_opt, 0.99),
'interval': 'step'} # called after each training step
dis_sched = CosineAnnealing(discriminator_opt, T_max=10) # called every epoch
return [gen_opt, dis_opt], [gen_sched, dis_sched]
# example with optimizer frequencies
# see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1
# https://arxiv.org/abs/1704.00028
def configure_optimizers(self):
gen_opt = Adam(self.model_gen.parameters(), lr=0.01)
dis_opt = Adam(self.model_disc.parameters(), lr=0.02)
n_critic = 5
return (
{'optimizer': dis_opt, 'frequency': n_critic},
{'optimizer': gen_opt, 'frequency': 1}
)
Note:
Some things to know:
- Lightning calls ``.backward()`` and ``.step()`` on each optimizer
and learning rate scheduler as needed.
- If you use 16-bit precision (``precision=16``), Lightning will automatically
handle the optimizers for you.
- If you use multiple optimizers, :meth:`training_step` will have an additional
``optimizer_idx`` parameter.
- If you use LBFGS Lightning handles the closure function automatically for you.
- If you use multiple optimizers, gradients will be calculated only
for the parameters of current optimizer at each training step.
- If you need to control how often those optimizers step or override the
default ``.step()`` schedule, override the :meth:`optimizer_step` hook.
- If you only want to call a learning rate scheduler every ``x`` step or epoch,
or want to monitor a custom metric, you can specify these in a lr_dict:
.. code-block:: python
{
'scheduler': lr_scheduler,
'interval': 'step', # or 'epoch'
'monitor': 'val_f1',
'frequency': x,
}
"""
rank_zero_warn('`configure_optimizers` must be implemented to be used with the Lightning Trainer')
def optimizer_step(
self,
epoch: int,
batch_idx: int,
optimizer: Optimizer,
optimizer_idx: int,
second_order_closure: Optional[Callable] = None,
on_tpu: bool = False,
using_native_amp: bool = False,
using_lbfgs: bool = False,
) -> None:
r"""
Override this method to adjust the default way the
:class:`~pytorch_lightning.trainer.trainer.Trainer` calls each optimizer.
By default, Lightning calls ``step()`` and ``zero_grad()`` as shown in the example
once per optimizer.
Args:
epoch: Current epoch
batch_idx: Index of current batch
optimizer: A PyTorch optimizer
optimizer_idx: If you used multiple optimizers this indexes into that list.
second_order_closure: closure for second order methods
on_tpu: true if TPU backward is required
using_native_amp: True if using native amp
using_lbfgs: True if the matching optimizer is lbfgs
Examples:
.. code-block:: python
# DEFAULT
def optimizer_step(self, current_epoch, batch_idx, optimizer, optimizer_idx,
second_order_closure, on_tpu, using_native_amp, using_lbfgs):
optimizer.step()
# Alternating schedule for optimizer steps (i.e.: GANs)
def optimizer_step(self, current_epoch, batch_idx, optimizer, optimizer_idx,
second_order_closure, on_tpu, using_native_amp, using_lbfgs):
# update generator opt every 2 steps
if optimizer_idx == 0:
if batch_idx % 2 == 0 :
optimizer.step()
optimizer.zero_grad()
# update discriminator opt every 4 steps
if optimizer_idx == 1:
if batch_idx % 4 == 0 :
optimizer.step()
optimizer.zero_grad()
# ...
# add as many optimizers as you want
Here's another example showing how to use this for more advanced things such as
learning rate warm-up:
.. code-block:: python
# learning rate warm-up
def optimizer_step(self, current_epoch, batch_idx, optimizer,
optimizer_idx, second_order_closure, on_tpu, using_native_amp, using_lbfgs):
# warm up lr
if self.trainer.global_step < 500:
lr_scale = min(1., float(self.trainer.global_step + 1) / 500.)
for pg in optimizer.param_groups:
pg['lr'] = lr_scale * self.learning_rate
# update params
optimizer.step()
optimizer.zero_grad()
Note:
If you also override the :meth:`~pytorch_lightning.core.hooks.ModelHooks.on_before_zero_grad`
model hook don't forget to add the call to it before ``optimizer.zero_grad()`` yourself.
"""
if on_tpu:
xm.optimizer_step(optimizer)
elif using_native_amp:
self.trainer.scaler.step(optimizer)
elif using_lbfgs:
optimizer.step(second_order_closure)
else:
optimizer.step()
def optimizer_zero_grad(self,
epoch: int,
batch_idx: int,
optimizer: Optimizer,
optimizer_idx: int):
optimizer.zero_grad()
def tbptt_split_batch(self, batch: Tensor, split_size: int) -> list:
r"""
When using truncated backpropagation through time, each batch must be split along the
time dimension. Lightning handles this by default, but for custom behavior override
this function.
Args:
batch: Current batch
split_size: The size of the split
Return:
List of batch splits. Each split will be passed to :meth:`training_step` to enable truncated
back propagation through time. The default implementation splits root level Tensors and
Sequences at dim=1 (i.e. time dim). It assumes that each time dim is the same length.
Examples:
.. code-block:: python
def tbptt_split_batch(self, batch, split_size):
splits = []
for t in range(0, time_dims[0], split_size):
batch_split = []
for i, x in enumerate(batch):
if isinstance(x, torch.Tensor):
split_x = x[:, t:t + split_size]
elif isinstance(x, collections.Sequence):
split_x = [None] * len(x)
for batch_idx in range(len(x)):
split_x[batch_idx] = x[batch_idx][t:t + split_size]
batch_split.append(split_x)
splits.append(batch_split)
return splits
Note:
Called in the training loop after
:meth:`~pytorch_lightning.callbacks.base.Callback.on_batch_start`
if :paramref:`~pytorch_lightning.trainer.Trainer.truncated_bptt_steps` > 0.
Each returned batch split is passed separately to :meth:`training_step`.
"""
time_dims = [len(x[0]) for x in batch if isinstance(x, (torch.Tensor, collections.Sequence))]
assert len(time_dims) >= 1, "Unable to determine batch time dimension"
assert all(x == time_dims[0] for x in time_dims), "Batch time dimension length is ambiguous"
splits = []
for t in range(0, time_dims[0], split_size):
batch_split = []
for i, x in enumerate(batch):
if isinstance(x, torch.Tensor):
split_x = x[:, t:t + split_size]
elif isinstance(x, collections.Sequence):
split_x = [None] * len(x)
for batch_idx in range(len(x)):
split_x[batch_idx] = x[batch_idx][t:t + split_size]
batch_split.append(split_x)
splits.append(batch_split)
return splits
def prepare_data(self) -> None:
"""
Use this to download and prepare data.
.. warning:: DO NOT set state to the model (use `setup` instead)
since this is NOT called on every GPU in DDP/TPU
Example::
def prepare_data(self):
# good
download_data()
tokenize()
etc()
# bad
self.split = data_split
self.some_state = some_other_state()
In DDP prepare_data can be called in two ways (using Trainer(prepare_data_per_node)):
1. Once per node. This is the default and is only called on LOCAL_RANK=0.
2. Once in total. Only called on GLOBAL_RANK=0.
Example::
# DEFAULT
# called once per node on LOCAL_RANK=0 of that node
Trainer(prepare_data_per_node=True)
# call on GLOBAL_RANK=0 (great for shared file systems)
Trainer(prepare_data_per_node=False)
This is called before requesting the dataloaders:
.. code-block:: python
model.prepare_data()
if ddp/tpu: init()
model.setup(stage)
model.train_dataloader()
model.val_dataloader()
model.test_dataloader()
"""
def train_dataloader(self) -> DataLoader:
"""
Implement a PyTorch DataLoader for training.
Return:
Single PyTorch :class:`~torch.utils.data.DataLoader`.
The dataloader you return will not be called every epoch unless you set
:paramref:`~pytorch_lightning.trainer.Trainer.reload_dataloaders_every_epoch` to ``True``.
For data processing use the following pattern:
- download in :meth:`prepare_data`
- process and split in :meth:`setup`
However, the above are only necessary for distributed processing.
.. warning:: do not assign state in prepare_data
- :meth:`~pytorch_lightning.trainer.Trainer.fit`
- ...
- :meth:`prepare_data`
- :meth:`setup`
- :meth:`train_dataloader`
Note:
Lightning adds the correct sampler for distributed and arbitrary hardware.
There is no need to set it yourself.
Example:
.. code-block:: python
def train_dataloader(self):
transform = transforms.Compose([transforms.ToTensor(),
transforms.Normalize((0.5,), (1.0,))])
dataset = MNIST(root='/path/to/mnist/', train=True, transform=transform,
download=True)
loader = torch.utils.data.DataLoader(
dataset=dataset,
batch_size=self.batch_size,
shuffle=True
)
return loader
"""
rank_zero_warn('`train_dataloader` must be implemented to be used with the Lightning Trainer')
def tng_dataloader(self): # todo: remove in v1.0.0
"""
Warnings:
Deprecated in v0.5.0. Use :meth:`train_dataloader` instead. Will be removed in 1.0.0.
"""
output = self.train_dataloader()
rank_zero_warn("`tng_dataloader` has been renamed to `train_dataloader` since v0.5.0."
" and this method will be removed in v1.0.0", DeprecationWarning)
return output
def test_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
r"""
Implement one or multiple PyTorch DataLoaders for testing.
The dataloader you return will not be called every epoch unless you set
:paramref:`~pytorch_lightning.trainer.Trainer.reload_dataloaders_every_epoch` to ``True``.
For data processing use the following pattern:
- download in :meth:`prepare_data`
- process and split in :meth:`setup`
However, the above are only necessary for distributed processing.
.. warning:: do not assign state in prepare_data
- :meth:`~pytorch_lightning.trainer.Trainer.fit`
- ...
- :meth:`prepare_data`
- :meth:`setup`
- :meth:`train_dataloader`
- :meth:`val_dataloader`
- :meth:`test_dataloader`
Note:
Lightning adds the correct sampler for distributed and arbitrary hardware.
There is no need to set it yourself.
Return:
Single or multiple PyTorch DataLoaders.
Example:
.. code-block:: python
def test_dataloader(self):
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.batch_size,
shuffle=False
)
return loader
Note:
If you don't need a test dataset and a :meth:`test_step`, you don't need to implement
this method.
"""
def val_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
r"""
Implement one or multiple PyTorch DataLoaders for validation.
The dataloader you return will not be called every epoch unless you set
:paramref:`~pytorch_lightning.trainer.Trainer.reload_dataloaders_every_epoch` to ``True``.
It's recommended that all data downloads and preparation happen in :meth:`prepare_data`.
- :meth:`~pytorch_lightning.trainer.Trainer.fit`
- ...
- :meth:`prepare_data`
- :meth:`train_dataloader`
- :meth:`val_dataloader`
- :meth:`test_dataloader`
Note:
Lightning adds the correct sampler for distributed and arbitrary hardware
There is no need to set it yourself.
Return:
Single or multiple PyTorch DataLoaders.
Examples:
.. code-block:: python
def val_dataloader(self):
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.batch_size,
shuffle=False
)
return loader
# can also return multiple dataloaders
def val_dataloader(self):
return [loader_a, loader_b, ..., loader_n]
Note:
If you don't need a validation dataset and a :meth:`validation_step`, you don't need to
implement this method.
Note:
In the case where you return multiple validation dataloaders, the :meth:`validation_step`
will have an argument ``dataset_idx`` which matches the order here.
"""
def summarize(self, mode: str = ModelSummary.MODE_DEFAULT) -> ModelSummary:
model_summary = ModelSummary(self, mode=mode)
log.info('\n' + str(model_summary))
return model_summary
def freeze(self) -> None:
r"""
Freeze all params for inference.
Example:
.. code-block:: python
model = MyLightningModule(...)
model.freeze()
"""
for param in self.parameters():
param.requires_grad = False
self.eval()
def unfreeze(self) -> None:
"""
Unfreeze all parameters for training.
.. code-block:: python
model = MyLightningModule(...)
model.unfreeze()
"""
for param in self.parameters():
param.requires_grad = True
self.train()
def on_load_checkpoint(self, checkpoint: Dict[str, Any]) -> None:
r"""
Called by Lightning to restore your model.
If you saved something with :meth:`on_save_checkpoint` this is your chance to restore this.
Args:
checkpoint: Loaded checkpoint
Example:
.. code-block:: python
def on_load_checkpoint(self, checkpoint):
# 99% of the time you don't need to implement this method
self.something_cool_i_want_to_save = checkpoint['something_cool_i_want_to_save']
Note:
Lightning auto-restores global step, epoch, and train state including amp scaling.
There is no need for you to restore anything regarding training.
"""
def on_save_checkpoint(self, checkpoint: Dict[str, Any]) -> None:
r"""
Called by Lightning when saving a checkpoint to give you a chance to store anything
else you might want to save.
Args:
checkpoint: Checkpoint to be saved
Example:
.. code-block:: python
def on_save_checkpoint(self, checkpoint):
# 99% of use cases you don't need to implement this method
checkpoint['something_cool_i_want_to_save'] = my_cool_pickable_object
Note:
Lightning saves all aspects of training (epoch, global step, etc...)
including amp scaling.
There is no need for you to store anything about training.
"""
def get_progress_bar_dict(self) -> Dict[str, Union[int, str]]:
r"""
Additional items to be displayed in the progress bar.
Return:
Dictionary with the items to be displayed in the progress bar.
"""
# call .item() only once but store elements without graphs
running_train_loss = self.trainer.running_loss.mean()
avg_training_loss = running_train_loss.cpu().item() if running_train_loss is not None else float('NaN')
tqdm_dict = {
'loss': '{:.3f}'.format(avg_training_loss)
}
if self.trainer.truncated_bptt_steps is not None:
tqdm_dict['split_idx'] = self.trainer.split_idx
if self.trainer.logger is not None and self.trainer.logger.version is not None:
tqdm_dict['v_num'] = self.trainer.logger.version
return tqdm_dict
def get_tqdm_dict(self) -> Dict[str, Union[int, str]]:
"""
Additional items to be displayed in the progress bar.
Return:
Dictionary with the items to be displayed in the progress bar.
Warning:
Deprecated since v0.7.3.
Use :meth:`get_progress_bar_dict` instead.
"""
rank_zero_warn("`get_tqdm_dict` was renamed to `get_progress_bar_dict` in v0.7.3"
" and this method will be removed in v1.0.0", DeprecationWarning)
return self.get_progress_bar_dict()
@classmethod
def _auto_collect_arguments(cls, frame=None) -> Tuple[Dict, Dict]:
"""
Collect all module arguments in the current constructor and all child constructors.
The child constructors are all the ``__init__`` methods that reach the current class through
(chained) ``super().__init__()`` calls.
Args:
frame: instance frame
Returns:
self_arguments: arguments dictionary of the first instance
parents_arguments: arguments dictionary of the parent's instances
"""
if not frame:
frame = inspect.currentframe()
frame_args = collect_init_args(frame.f_back, [])
self_arguments = frame_args[-1]
# set module_arguments in child
self_arguments = self_arguments
parents_arguments = {}
# add all arguments from parents
for args in frame_args[:-1]:
parents_arguments.update(args)
return self_arguments, parents_arguments
def save_hyperparameters(self, *args, frame=None) -> None:
"""Save all model arguments.
Args:
args: single object of `dict`, `NameSpace` or `OmegaConf`
or string names or argumenst from class `__init__`
>>> from collections import OrderedDict
>>> class ManuallyArgsModel(LightningModule):
... def __init__(self, arg1, arg2, arg3):
... super().__init__()
... # manually assine arguments
... self.save_hyperparameters('arg1', 'arg3')
... def forward(self, *args, **kwargs):
... ...
>>> model = ManuallyArgsModel(1, 'abc', 3.14)
>>> model.hparams
"arg1": 1
"arg3": 3.14
>>> class AutomaticArgsModel(LightningModule):
... def __init__(self, arg1, arg2, arg3):
... super().__init__()
... # equivalent automatic
... self.save_hyperparameters()
... def forward(self, *args, **kwargs):
... ...
>>> model = AutomaticArgsModel(1, 'abc', 3.14)
>>> model.hparams
"arg1": 1
"arg2": abc
"arg3": 3.14
>>> class SingleArgModel(LightningModule):
... def __init__(self, params):
... super().__init__()
... # manually assign single argument
... self.save_hyperparameters(params)
... def forward(self, *args, **kwargs):
... ...
>>> model = SingleArgModel(Namespace(p1=1, p2='abc', p3=3.14))
>>> model.hparams
"p1": 1
"p2": abc
"p3": 3.14
"""
if not frame:
frame = inspect.currentframe().f_back
init_args = get_init_args(frame)
assert init_args, 'failed to inspect the self init'
if not args:
hp = init_args
self._hparams_name = 'kwargs' if hp else None
else:
isx_non_str = [i for i, arg in enumerate(args) if not isinstance(arg, str)]
if len(isx_non_str) == 1:
hp = args[isx_non_str[0]]
cand_names = [k for k, v in init_args.items() if v == hp]
self._hparams_name = cand_names[0] if cand_names else None
else:
hp = {arg: init_args[arg] for arg in args if isinstance(arg, str)}
self._hparams_name = 'kwargs'
# `hparams` are expected here
if hp:
self._set_hparams(hp)
def _set_hparams(self, hp: Union[dict, Namespace, str]) -> None:
if isinstance(hp, Namespace):
hp = vars(hp)
if isinstance(hp, dict):
hp = AttributeDict(hp)
elif isinstance(hp, PRIMITIVE_TYPES):
raise ValueError(f'Primitives {PRIMITIVE_TYPES} are not allowed.')
elif not isinstance(hp, ALLOWED_CONFIG_TYPES):
raise ValueError(f'Unsupported config type of {type(hp)}.')
if isinstance(hp, dict) and isinstance(self.hparams, dict):
self.hparams.update(hp)
else:
self._hparams = hp
@property
def hparams(self) -> Union[AttributeDict, str]:
if not hasattr(self, '_hparams'):
self._hparams = AttributeDict()
return self._hparams
@hparams.setter
def hparams(self, hp: Union[dict, Namespace, Any]):
hparams_assignment_name = self.__get_hparams_assignment_variable()
self._hparams_name = hparams_assignment_name
self._set_hparams(hp)
def __get_hparams_assignment_variable(self):
"""
looks at the code of the class to figure out what the user named self.hparams
this only happens when the user explicitly sets self.hparams
"""
try:
class_code = inspect.getsource(self.__class__)
lines = class_code.split('\n')
for line in lines:
line = re.sub(r"\s+", "", line, flags=re.UNICODE)
if '.hparams=' in line:
return line.split('=')[1]
except Exception as e:
return 'hparams'
return None