import logging
import time
from collections import defaultdict
from pathlib import Path
from typing import Any, DefaultDict, Dict, List, Optional
import jax
import jax.numpy as jnp
from flax import jax_utils
from flax.training import checkpoints
from flax.training.train_state import TrainState
from tango.common.dataset_dict import DatasetDictBase
from tango.common.exceptions import ConfigurationError
from tango.common.lazy import Lazy
from tango.common.tqdm import Tqdm
from tango.format import Format
from tango.step import Step
from tango.workspace import Workspace
from .data import FlaxDataLoader
from .format import FlaxFormat
from .model import Model
from .optim import LRScheduler, Optimizer
from .train_callback import TrainCallback
from .train_config import TrainConfig
from .util import get_multiple_keys, get_PRNGkey
from .wrapper import FlaxWrapper
PyTree = Any
[docs]@Step.register("flax::train")
class FlaxTrainStep(Step):
"""
A Flax training step that supports distributed training with configurable dataloaders, callbacks,
optimizer.
.. tip::
Registered as a :class:`~tango.step.Step` under the name "flax::train".
.. important::
To train on GPUs and TPUs, installation of jax[cuda] or jax[tpu] is required. Follow the
instructions here: https://github.com/google/jax to set up jax for GPUs and TPUs.
Note: CUDA and cuDNN installation is required to run jax on NVidia GPUs. It is recommended to
install cuDNN in your conda environment using: ``conda install -c anaconda cudnn``.
Distributed data parallel training is activated when the ``device_count`` is greater than 1.
You can control which CUDA devices to use with the environment variable ``CUDA_VISIBLE_DEVICES``.
For example, to only use the GPUs with IDs 0 and 1, set ``CUDA_VISIBLE_DEVICES=0,1``
(and ``device_count`` to 2).
.. warning::
During validation, the validation metric (specified by the ``val_metric_name`` parameter)
is aggregated by simply averaging across validation batches and distributed processes.
This behavior is usually correct when your validation metric is "loss" or "accuracy",
for example, but may not be correct for other metrics like "F1".
If this is not correct for your metric you will need to handle the aggregation
internally in your model or with a :class:`TrainCallback`
using the :meth:`TrainCallback.post_val_batch()` method.
Then set the parameter ``auto_aggregate_val_metric`` to ``False``.
Jax pre-allocates 90% of GPU memory. If you run into out-of-memory (OOM) issues, please refer
to this: https://jax.readthedocs.io/en/latest/gpu_memory_allocation.html.
"""
DETERMINISTIC = True
CACHEABLE = True
FORMAT: Format = FlaxFormat()
SKIP_ID_ARGUMENTS = {"log_every"}
METADATA = {"artifact_kind": "model"}
[docs] def run( # type: ignore[override]
self,
model: Model,
dataset: DatasetDictBase,
optimizer: Lazy[Optimizer],
train_dataloader: Lazy[FlaxDataLoader],
*,
wrapper: FlaxWrapper,
seed: int = 42,
keep_checkpoints: int = 5,
lr_scheduler: Optional[Lazy[LRScheduler]] = None,
train_split: str = "train",
validation_dataloader: Optional[Lazy[FlaxDataLoader]] = None,
validation_split: Optional[str] = None,
train_steps: Optional[int] = None,
train_epoch: Optional[int] = None,
validation_steps: Optional[int] = None,
log_every: int = 10,
checkpoint_every: int = 100,
validate_every: Optional[int] = None,
val_metric_name: str = "loss",
minimize_val_metric: bool = True,
auto_aggregate_val_metric: bool = True,
callbacks: Optional[List[Lazy[TrainCallback]]] = None,
remove_stale_checkpoints: bool = True,
) -> PyTree:
"""
Run a basic training loop to train the ``model``.
:param model:
The flax model to train. It should define ``__call__()``. Defining ``setup()`` is Optional.
:param dataset:
The train and optional validation dataset.
:param optimizer:
The name of the optax Optimizer to use for training.
:param train_dataloader:
The dataloader object that generates training batches.
:param wrapper:
A Wrapper class that defines ``loss_fn()``, ``eval_fn()`` and ``compute_metrics()``
:param seed:
Used to set the PRNG state. By default, ``seed=42``
:param keep_checkpoints:
An integer which denotes how many previous checkpoints should be stored while training.
By default, ``keep_checkpoints=5``
:param lr_scheduler:
The name of the learning rate scheduler.
:param train_split:
The name of the data split used for training in the ``dataset_dict``.
Default is "train".
:param validation_dataloader:
An optional data loader for generating validation batches. The batches should be
:class:`dict` objects. If not specified, but ``validation_split`` is given,
the validation ``DataLoader`` will be constructed from the same parameters
as the train ``DataLoader``.
:param validation_split:
Optional name of the validation split in the ``dataset_dict``. Default is ``None``,
which means no validation.
:param train_steps:
The number of steps to train for. If not specified training will
stop after a complete iteration through the ``train_dataloader``.
:param train_epoch:
The number of epochs to train for. You cannot specify ``train_steps`` and ``train_epochs``
at the same time.
:param validation_steps:
The number of steps to validate for. If not specified validation
will stop after a complete iteration through the ``validation_dataloader``.
:param log_every:
Log every this many steps.
:param checkpoint_every:
Save a checkpoint every this many steps.
:param validate_every:
Run the validation loop every this many steps.
:param val_metric_name:
The name of the validation metric, i.e. the key of the metric in the dictionary
returned by the forward pass of the model. Default is "loss".
:param minimize_val_metric:
Whether the validation metric is meant to be minimized (such as the loss).
Default is ``True``. When using a metric such as accuracy, you should set
this to ``False``.
:param auto_aggregate_val_metric:
If ``True`` (the default), the validation metric will be averaged across
validation batches and distributed processes. This may not be the correct
behavior for some metrics (such as F1), in which you should set this to
``False`` and handle the aggregation internally in your model
or with a :class:`TrainCallback` (using :meth:`TrainCallback.post_val_batch()`).
:param callbacks:
A list of :class: `TrainCallback`.
:param remove_stale_checkpoints:
If ``True`` (the default), stale checkpoints will be removed throughout training so that
only the latest and best checkpoints are kept.
:returns:
The trained model with the last checkpoint loaded.
"""
return self._train(
dataset=dataset,
model=model,
optimizer=optimizer,
train_dataloader=train_dataloader,
train_wrapper=wrapper,
seed=seed,
keep_checkpoints=keep_checkpoints,
lr_scheduler=lr_scheduler,
train_split=train_split,
validation_split=validation_split,
validation_dataloader=validation_dataloader,
train_steps=train_steps,
train_epochs=train_epoch,
validation_steps=validation_steps,
log_every=log_every,
checkpoint_every=checkpoint_every,
validate_every=validate_every,
val_metric_name=val_metric_name,
minimize_val_metric=minimize_val_metric,
auto_aggregate_val_metric=auto_aggregate_val_metric,
callbacks=callbacks,
remove_stale_checkpoints=remove_stale_checkpoints,
)
def _train(
self,
model: Model,
optimizer: Lazy[Optimizer],
dataset: DatasetDictBase,
train_dataloader: Lazy[FlaxDataLoader],
*,
train_wrapper: FlaxWrapper,
seed: int = 42,
keep_checkpoints: int = 5,
lr_scheduler: Optional[Lazy[LRScheduler]],
train_split: str = "train",
validation_split: Optional[str] = None,
validation_dataloader: Optional[Lazy[FlaxDataLoader]] = None,
train_steps: Optional[int] = None,
train_epochs: Optional[int] = None,
validation_steps: Optional[int] = None,
log_every: int = 10,
checkpoint_every: int = 100,
validate_every: Optional[int] = None,
val_metric_name: str = "loss",
minimize_val_metric: bool = True,
auto_aggregate_val_metric: bool = True,
callbacks: Optional[List[Lazy[TrainCallback]]] = None,
remove_stale_checkpoints: bool = True,
) -> PyTree:
if validate_every is not None and validation_split is None:
raise ConfigurationError(
"You have set a validation interval, but no validation split. "
"That's probably unintentional."
)
if (train_steps is not None) and (train_epochs is not None):
raise ConfigurationError(
"One of 'train_steps' or 'train_epochs' needs to be specified, but not both."
)
if isinstance(dataset, DatasetDictBase) and train_split is None:
raise ConfigurationError("Specify the train split for Datasets object.")
config = TrainConfig(
self.unique_id,
self.work_dir,
step_name=self.name,
train_split=train_split,
validation_split=validation_split,
seed=seed,
train_steps=train_steps,
train_epochs=train_epochs,
log_every=log_every,
checkpoint_every=checkpoint_every,
validate_every=validate_every,
validation_steps=validation_steps,
val_metric_name=val_metric_name,
minimize_val_metric=minimize_val_metric,
auto_aggregate_val_metric=auto_aggregate_val_metric,
remove_stale_checkpoints=remove_stale_checkpoints,
)
optimizer = self._construct_optimizer(optimizer)
lr_scheduler_: Optional[LRScheduler] = None
if lr_scheduler is not None:
lr_scheduler_ = self._construct_lr_scheduler(lr_scheduler)
lr_scheduler = lr_scheduler_
final_model: Model
final_model = self.train_helper(
self.workspace,
config,
model,
optimizer,
keep_checkpoints,
lr_scheduler,
train_wrapper,
dataset,
train_dataloader,
validation_dataloader,
callbacks,
)
assert final_model is not None
return final_model
def train_helper(
self,
workspace: Workspace,
config: TrainConfig,
model: Model,
optimizer: Optimizer,
keep_checkpoints: int,
lr_scheduler: Optional[LRScheduler],
train_wrapper: FlaxWrapper,
dataset: DatasetDictBase,
train_dataloader: Lazy[FlaxDataLoader],
validation_dataloader: Optional[Lazy[FlaxDataLoader]] = None,
callbacks: Optional[List[Lazy[TrainCallback]]] = None,
) -> PyTree:
if lr_scheduler is not None:
raise NotImplementedError(
"Learning rate scheduling is not supported by the flax trainer. "
"Please voice your support for this feature at "
"https://github.com/allenai/tango/issues/477."
)
logger = logging.getLogger(FlaxTrainStep.__name__)
# construct data loaders
validation_dataloader_: Optional[FlaxDataLoader] = None
if config.validation_split is not None:
validation_dataset = dataset[config.validation_split]
validation_dataset.set_format("numpy")
if validation_dataloader is not None:
validation_dataloader_ = validation_dataloader.construct(dataset=validation_dataset)
else:
validation_dataloader_ = train_dataloader.construct(dataset=validation_dataset)
validation_dataloader = validation_dataloader_
train_dataset = dataset[config.train_split]
train_dataset.set_format("numpy") # type:ignore
train_dataloader: FlaxDataLoader = train_dataloader.construct(dataset=train_dataset)
devices = self._get_devices()
do_distributed: bool = False
if len(devices) > 1:
do_distributed = True
if validation_dataloader is not None:
validation_dataloader.batch_size *= len(devices)
train_dataloader.batch_size *= len(devices)
rng = get_PRNGkey(config.seed)
if hasattr(model, "params"):
params = model.params
else:
# TODO: Find better way to init the shape
shape = list(train_dataset["x"].shape)
shape[0] = 1
x = jnp.ones(shape)
params = model.init(rng, x)["params"]
state = TrainState.create(apply_fn=model.__call__, params=params, tx=optimizer)
initial_state: Optional[Dict[str, Any]] = None
if config.state_path.exists():
logger.info("Recovering from previous run at %s" % config.state_path)
state = self.load_checkpoint(config.state_path, state)
if config.train_epochs is None:
assert config.train_steps is not None
try:
train_epochs = len(train_dataloader.dataset) // train_dataloader.batch_size
except TypeError:
raise ConfigurationError(
"You must set train_epochs for streaming/iterable datasets"
)
config.train_epochs = train_epochs
assert config.train_epochs is not None
if validation_dataloader is not None:
if config.validation_steps is None:
try:
config.validation_steps = len(validation_dataloader.dataset)
except TypeError:
raise ConfigurationError(
"You must set 'validation_steps' for streaming/iterable datasets"
)
val_metric: Optional[float] = None
best_val_metric: Optional[float] = None
start_step: int = 0
if initial_state is not None:
val_metric = initial_state[f"val_{config.val_metric_name}"]
best_val_metric = initial_state[f"best_{config.val_metric_name}"]
start_step = initial_state["training_epochs"]
# Initialize callbacks
callbacks: List[TrainCallback] = [
callback.construct(
workspace=workspace,
train_config=config,
dataset=dataset,
train_dataloader=train_dataloader,
model=model,
optimizer=optimizer,
validation_dataloader=validation_dataloader,
)
for callback in (callbacks or [])
]
if initial_state:
for callback, state in zip(callbacks, initial_state["callbacks"]):
callback.load_state_dict(state)
del initial_state
if start_step > 0:
with Tqdm.tqdm(
train_dataloader,
desc=f"Catching dataloader up to step {start_step}",
total=start_step - 1,
) as batch_iter:
for step, batch in enumerate(batch_iter):
del batch
if step >= start_step - 1:
break
def train_step(state, batch, dropout_rng):
# if transformer model
labels = batch.pop("labels")
dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
grad_fn = jax.value_and_grad(train_wrapper.train_loss)
loss, grad = grad_fn(state.params, state, batch, dropout_rng, labels)
if do_distributed:
grad = jax.lax.pmean(grad, "batch")
new_state = state.apply_gradients(grads=grad)
other_metrics = train_wrapper.train_metrics(state, batch, labels=labels)
metrics = {"loss": loss}
metrics.update(other_metrics)
if do_distributed:
metrics = jax.lax.pmean(metrics, axis_name="batch")
return new_state, metrics, new_dropout_rng
def val_step(state, batch):
labels = batch.pop("labels")
logits = state.apply_fn(**batch, params=state.params, train=False)[0]
metrics = train_wrapper.val_metrics(batch, logits, labels)
if do_distributed:
metrics = jax.lax.pmean(metrics, axis_name="batch")
return metrics
if do_distributed:
# NOTE: The trainer currently handles only data parallelism.
state = jax_utils.replicate(state)
dropout_rngs = get_multiple_keys(rng, jax.local_device_count())
parallel_train_step = jax.pmap(train_step, axis_name="batch")
parallel_val_step = jax.pmap(val_step, axis_name="batch")
step_per_epoch = train_dataloader.dataset_size // train_dataloader.batch_size
config.train_steps = step_per_epoch * config.train_epochs
assert config.train_steps is not None # for mypy
for callback in callbacks:
callback.pre_train_loop()
logger.info("***** Running training *****")
logger.info(f" Num examples = {train_dataloader.dataset_size}")
logger.info(f" Num Epochs = {config.train_epochs}")
logger.info(
f" Total train batch size (w. parallel & distributed) = {train_dataloader.batch_size}"
)
logger.info(f" Total optimization steps = {config.train_steps}")
step = start_step
epochs = Tqdm.tqdm(
range(config.train_epochs), desc=f"Epoch (1/{config.train_epochs})", position=0
)
for epoch in epochs:
start = time.time()
train_metrics = []
for callback in callbacks:
callback.pre_epoch(step, epoch)
train_loader = train_dataloader(rng, do_distributed)
for _ in Tqdm.tqdm(range(step_per_epoch), desc="Training", position=1):
batch = next(train_loader)
for callback in callbacks:
callback.pre_batch(step, epoch, batch)
if do_distributed:
state, train_metric, dropout_rngs = parallel_train_step(
state, batch, dropout_rngs
)
else:
state, train_metric, rng = train_step(state, batch, rng)
train_metrics.append(train_metric)
for callback in callbacks:
callback.post_batch(step, epoch, train_metric)
if config.should_log_this_step(step):
for callback in callbacks:
callback.log_batch(step, epoch, train_metric)
if config.should_checkpoint_this_step(step):
self.save_checkpoint(config.state_path, state, step, keep_checkpoints)
step += 1
# check if we need to do validation
if config.validation_split is None:
# If we can't validate, we don't.
should_validate = False
elif step == config.train_steps - 1:
# If we're at the end of the training run, we always validate.
should_validate = True
elif config.validate_every is not None and step % config.validate_every == 0:
# If validate_every is given, we use that to decide.
should_validate = True
else:
# Otherwise, we don't validate.
should_validate = False
if should_validate:
assert validation_dataloader is not None
assert config.validation_steps is not None
val_metrics: DefaultDict = defaultdict(list)
epoch_eval_metrics: DefaultDict = defaultdict(float)
val_dataloader = validation_dataloader(rng, do_distributed)
valid_step = 0
total_val_steps = len(validation_dataset) // validation_dataloader.batch_size
for callback in callbacks:
callback.pre_val_loop(step, valid_step, state)
for _ in Tqdm.tqdm(range(total_val_steps), desc="Evaluating", position=2):
batch = next(val_dataloader)
for callback in callbacks:
callback.pre_val_batch(step, valid_step, epoch, batch)
if do_distributed:
metrics = parallel_val_step(state, batch)
metrics = jax_utils.unreplicate(metrics)
else:
metrics = val_step(state, batch)
for key, value in metrics.items():
val_metrics[key].append(value.item())
for callback in callbacks:
callback.post_val_batch(step, valid_step, epoch, val_metrics)
valid_step += 1
for key, value in val_metrics.items():
if config.auto_aggregate_val_metric:
epoch_eval_metrics[key] = jax.tree_map(
jnp.mean, jnp.array(value)
).item()
else:
epoch_eval_metrics[key] = metrics[key].item()
for key, value in epoch_eval_metrics.items():
print("Validation %s : %.5f" % (key, value))
val_metric = epoch_eval_metrics[config.val_metric_name]
assert val_metric is not None
if best_val_metric is None:
best_val_metric = val_metric
elif config.minimize_val_metric and val_metric <= best_val_metric:
best_val_metric = val_metric
elif not config.minimize_val_metric and val_metric >= best_val_metric:
best_val_metric = val_metric
for callback in callbacks:
callback.post_val_loop(step, epoch, val_metric, best_val_metric)
if do_distributed:
train_metric = jax_utils.unreplicate(train_metric)
for key, value in train_metric.items():
print("Train %s : %.2f" % (key, value))
for callback in callbacks:
callback.post_epoch(step, epoch)
end = time.time()
train_time = (end - start) / 60
desc = f"Epoch... ({epoch + 1}/{config.train_epochs} | Time taken (mins): {train_time})"
epochs.write(desc)
epochs.desc = desc
for callback in callbacks:
callback.post_train_loop(step, epoch)
if do_distributed:
state = jax_utils.unreplicate(state)
return state
def save_checkpoint(self, dir: Path, target: PyTree, step: int, keep_checkpoints: int):
return checkpoints.save_checkpoint(
dir, target, step, prefix="checkpoint_", keep=keep_checkpoints, overwrite=True
)
def load_checkpoint(self, dir: Path, target: PyTree):
return checkpoints.restore_checkpoint(dir, target, prefix="checkpoint_")
def _construct_optimizer(self, optimizer):
self.optimizer = optimizer.construct()
return self.optimizer
def _construct_lr_scheduler(self, scheduler):
self.lr_scheduler = scheduler.construct()
return self.lr_scheduler
def _get_devices(self) -> List[Any]:
device_type = jax.default_backend()
self.devices = jax.devices()
device_count = len(self.devices)
print("Training on %d %s" % (device_count, device_type))
return self.devices