def test_save(self):
state1 = self.get_state(self.model)
mask = Mask.ones_like(self.model)
pruned_model = PrunedModel(self.model, mask)
pruned_model.save(self.root, Step.zero(20))
self.assertTrue(os.path.exists(paths.model(self.root, Step.zero(20))))
self.model.load_state_dict(torch.load(paths.model(self.root, Step.zero(20))))
self.assertStateEqual(state1, self.get_state(self.model))
def test_first_step(self):
init_state = TestStandardCallbacks.get_state(self.model)
train.train(self.hparams.training_hparams,
self.model,
self.train_loader,
self.root,
callbacks=self.callbacks,
end_step=Step.from_epoch(0, 1, len(self.train_loader)))
# Check that the initial state has been saved.
model_state_loc = paths.model(self.root,
Step.zero(len(self.train_loader)))
self.assertTrue(os.path.exists(model_state_loc))
# Check that the model state at init reflects the saved state.
self.model.load_state_dict(torch.load(model_state_loc))
saved_state = TestStandardCallbacks.get_state(self.model)
self.assertStateEqual(init_state, saved_state)
# Check that the checkpoint file exists.
self.assertTrue(os.path.exists(paths.checkpoint(self.root)))
# Check that the logger file doesn't exist.
self.assertFalse(os.path.exists(paths.logger(self.root)))
def lossonly_callbacks(training_hparams: hparams.TrainingHparams,
train_set_loader: DataLoader,
test_set_loader: DataLoader,
eval_on_train: bool = False,
verbose: bool = True,
start_step: Step = None,
evaluate_every_epoch: bool = True):
start = start_step or Step.zero(train_set_loader.iterations_per_epoch)
end = Step.from_str(training_hparams.training_steps,
train_set_loader.iterations_per_epoch)
loss_eval_callback = create_loss_eval_callback(verbose=verbose)
# Basic checkpointing and state saving at the beginning and end.
result = [
run_at_step(start, save_model),
run_at_step(end, save_model),
run_at_step(end, save_logger),
run_every_epoch(checkpointing.save_checkpoint_callback),
]
# Test every epoch if requested.
if evaluate_every_epoch:
result = [run_every_epoch(loss_eval_callback)] + result
elif verbose:
result.append(run_every_epoch(create_timekeeper_callback()))
# Ensure that testing occurs at least at the beginning and end of training.
if start.it != 0 or not evaluate_every_epoch:
result = [run_at_step(start, loss_eval_callback)] + result
if end.it != 0 or not evaluate_every_epoch:
result = [run_at_step(end, loss_eval_callback)] + result
return result
def standard_callbacks(training_hparams: hparams.TrainingHparams,
train_set_loader: DataLoader,
test_set_loader: DataLoader,
eval_on_train: bool = False,
verbose: bool = True,
start_step: Step = None,
evaluate_every_epoch: bool = True,
weight_save_steps: List[Step] = []):
start = start_step or Step.zero(train_set_loader.iterations_per_epoch)
end = Step.from_str(training_hparams.training_steps,
train_set_loader.iterations_per_epoch)
test_eval_callback = create_eval_callback('test',
test_set_loader,
verbose=verbose)
train_eval_callback = create_eval_callback('train',
train_set_loader,
verbose=verbose)
# Basic checkpointing and state saving at the beginning and end.
result = [
run_at_step(start, save_model),
run_at_step(end, save_model),
run_at_step(end, save_logger),
run_every_epoch(checkpointing.save_checkpoint_callback),
]
for s in weight_save_steps:
result.append(run_at_step(s, save_model))
# Test every epoch if requested.
if evaluate_every_epoch:
result = [run_every_epoch(test_eval_callback)] + result
elif verbose:
result.append(run_every_epoch(create_timekeeper_callback()))
# Ensure that testing occurs at least at the beginning and end of training.
if start.it != 0 or not evaluate_every_epoch:
result = [run_at_step(start, test_eval_callback)] + result
if end.it != 0 or not evaluate_every_epoch:
result = [run_at_step(end, test_eval_callback)] + result
# Do the same for the train set if requested.
if eval_on_train:
if evaluate_every_epoch:
result = [run_every_epoch(train_eval_callback)] + result
if start.it != 0 or not evaluate_every_epoch:
result = [run_at_step(start, train_eval_callback)] + result
if end.it != 0 or not evaluate_every_epoch:
result = [run_at_step(end, train_eval_callback)] + result
return result
def test_end_to_end(self):
init_loc = paths.model(self.root, Step.zero(len(self.train_loader)))
end_loc = paths.model(self.root,
Step.from_epoch(3, 0, len(self.train_loader)))
init_state = TestStandardCallbacks.get_state(self.model)
train.train(self.hparams.training_hparams,
self.model,
self.train_loader,
self.root,
callbacks=self.callbacks,
start_step=Step.from_epoch(0, 0, len(self.train_loader)),
end_step=Step.from_epoch(3, 0, len(self.train_loader)))
end_state = TestStandardCallbacks.get_state(self.model)
# Check that final state has been saved.
self.assertTrue(os.path.exists(init_loc))
self.assertTrue(os.path.exists(end_loc))
# Check that the checkpoint file still exists.
self.assertTrue(os.path.exists(paths.checkpoint(self.root)))
# Check that the initial and final states match those that were saved.
self.model.load_state_dict(torch.load(init_loc))
saved_state = TestStandardCallbacks.get_state(self.model)
self.assertStateEqual(init_state, saved_state)
self.model.load_state_dict(torch.load(end_loc))
saved_state = TestStandardCallbacks.get_state(self.model)
self.assertStateEqual(end_state, saved_state)
# Check that the logger has the right number of entries.
self.assertTrue(os.path.exists(paths.logger(self.root)))
logger = MetricLogger.create_from_file(self.root)
self.assertEqual(len(logger.get_data('train_loss')), 4)
self.assertEqual(len(logger.get_data('test_loss')), 4)
self.assertEqual(len(logger.get_data('train_accuracy')), 4)
self.assertEqual(len(logger.get_data('test_accuracy')), 4)
def train(training_hparams: hparams.TrainingHparams,
model: Model,
train_loader: DataLoader,
output_location: str,
callbacks: typing.List[typing.Callable] = [],
start_step: Step = None,
end_step: Step = None):
"""The main training loop for this framework.
Args:
* training_hparams: The training hyperparameters whose schema is specified in hparams.py.
* model: The model to train. Must be a models.base.Model
* train_loader: The training data. Must be a datasets.base.DataLoader
* output_location: The string path where all outputs should be stored.
* callbacks: A list of functions that are called before each training step and once more
after the last training step. Each function takes five arguments: the current step,
the output location, the model, the optimizer, and the logger.
Callbacks are used for running the test set, saving the logger, saving the state of the
model, etc. The provide hooks into the training loop for customization so that the
training loop itself can remain simple.
* start_step: The step at which the training data and learning rate schedule should begin.
Defaults to step 0.
* end_step: The step at which training should cease. Otherwise, training will go for the
full `training_hparams.training_steps` steps.
"""
# Create the output location if it doesn't already exist.
if not get_platform().exists(output_location) and get_platform(
).is_primary_process:
get_platform().makedirs(output_location)
# Get the optimizer and learning rate schedule.
model.to(get_platform().torch_device)
optimizer = optimizers.get_optimizer(training_hparams, model)
step_optimizer = optimizer
lr_schedule = optimizers.get_lr_schedule(training_hparams, optimizer,
train_loader.iterations_per_epoch)
# Adapt for FP16.
if training_hparams.apex_fp16:
if NO_APEX:
raise ImportError('Must install nvidia apex to use this model.')
model, step_optimizer = apex.amp.initialize(model,
optimizer,
loss_scale='dynamic',
verbosity=0)
# Handle parallelism if applicable.
if get_platform().is_distributed:
model = DistributedDataParallel(model,
device_ids=[get_platform().rank])
elif get_platform().is_parallel:
model = DataParallel(model)
# Get the random seed for the data order.
data_order_seed = training_hparams.data_order_seed
# Restore the model from a saved checkpoint if the checkpoint exists.
cp_step, cp_logger = restore_checkpoint(output_location, model, optimizer,
train_loader.iterations_per_epoch)
start_step = cp_step or start_step or Step.zero(
train_loader.iterations_per_epoch)
logger = cp_logger or MetricLogger()
with warnings.catch_warnings(): # Filter unnecessary warning.
warnings.filterwarnings("ignore", category=UserWarning)
for _ in range(start_step.iteration):
lr_schedule.step()
# Determine when to end training.
end_step = end_step or Step.from_str(training_hparams.training_steps,
train_loader.iterations_per_epoch)
if end_step <= start_step: return
# The training loop.
for ep in range(start_step.ep, end_step.ep + 1):
# Ensure the data order is different for each epoch.
train_loader.shuffle(None if data_order_seed is None else (
data_order_seed + ep))
for it, (examples, labels) in enumerate(train_loader):
# Advance the data loader until the start epoch and iteration.
if ep == start_step.ep and it < start_step.it: continue
# Run the callbacks.
step = Step.from_epoch(ep, it, train_loader.iterations_per_epoch)
for callback in callbacks:
callback(output_location, step, model, optimizer, logger)
# Exit at the end step.
if ep == end_step.ep and it == end_step.it: return
# Otherwise, train.
examples = examples.to(device=get_platform().torch_device)
labels = labels.to(device=get_platform().torch_device)
step_optimizer.zero_grad()
model.train()
loss = model.loss_criterion(model(examples), labels)
if training_hparams.apex_fp16:
with apex.amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# Step forward. Ignore extraneous warnings that the lr_schedule generates.
step_optimizer.step()
with warnings.catch_warnings(): # Filter unnecessary warning.
warnings.filterwarnings("ignore", category=UserWarning)
lr_schedule.step()
get_platform().barrier()
def distill(
training_hparams: hparams.TrainingHparams,
distill_hparams: hparams.DistillHparams,
student: Model,
teacher: Model,
train_loader: DataLoader,
output_location: str,
callbacks: typing.List[typing.Callable] = [],
start_step: Step = None,
end_step: Step = None
):
"""The main training loop for this framework.
Args:
* training_hparams: The training hyperparameters whose schema is specified in hparams.py.
* distll_hparams: The knowledge distillation hyperparameters whose schema is specified in hparams.py.
* student: The student model to train. Must be a models.base.Model
* teacher: The teacher model to distill the knowledge. Must be a models.base.Model
* train_loader: The training data. Must be a datasets.base.DataLoader
* output_location: The string path where all outputs should be stored.
* callbacks: A list of functions that are called before each training step and once more
after the last training step. Each function takes five arguments: the current step,
the output location, the model, the optimizer, and the logger.
Callbacks are used for running the test set, saving the logger, saving the state of the
model, etc. The provide hooks into the training loop for customization so that the
training loop itself can remain simple.
* start_step: The step at which the training data and learning rate schedule should begin.
Defaults to step 0.
* end_step: The step at which training should cease. Otherwise, training will go for the
full `training_hparams.training_steps` steps.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
# Create the output location if it doesn't already exist.
if not get_platform().exists(output_location) and get_platform().is_primary_process:
get_platform().makedirs(output_location)
# Get the optimizer and learning rate schedule.
student.to(get_platform().torch_device)
teacher.to(get_platform().torch_device)
optimizer = optimizers.get_optimizer(training_hparams, student)
step_optimizer = optimizer
lr_schedule = optimizers.get_lr_schedule(training_hparams, optimizer, train_loader.iterations_per_epoch)
ce_loss_fct = nn.KLDivLoss(reduction="batchmean")
if distill_hparams.alpha_mse > 0.0:
mse_loss_fct = nn.MSELoss(reduction='sum')
if distill_hparams.alpha_cos > 0.0:
cos_loss_fct = nn.CosineEmbeddingLoss(reduction='mean')
# Adapt for FP16.
if training_hparams.apex_fp16:
if NO_APEX: raise ImportError('Must install nvidia apex to use this model.')
(student, teacher), step_optimizer = apex.amp.initialize(
[student, teacher], optimizer, loss_scale='dynamic', verbosity=0
)
# Handle parallelism if applicable.
if get_platform().is_distributed:
student = DistributedDataParallel(student, device_ids=[get_platform().rank])
teacher = DistributedDataParallel(teacher, device_ids=[get_platform().rank])
elif get_platform().is_parallel:
student = DataParallel(student)
teacher = DataParallel(teacher)
# Get the random seed for the data order.
data_order_seed = training_hparams.data_order_seed
# Restore the model from a saved checkpoint if the checkpoint exists.
cp_step, cp_logger = restore_checkpoint(output_location, student, optimizer, train_loader.iterations_per_epoch)
start_step = cp_step or start_step or Step.zero(train_loader.iterations_per_epoch)
logger = cp_logger or MetricLogger()
with warnings.catch_warnings(): # Filter unnecessary warning.
warnings.filterwarnings("ignore", category=UserWarning)
for _ in range(start_step.iteration): lr_schedule.step()
# Determine when to end training.
end_step = end_step or Step.from_str(training_hparams.training_steps, train_loader.iterations_per_epoch)
if end_step <= start_step: return
# The training loop.
for ep in range(start_step.ep, end_step.ep + 1):
# Ensure the data order is different for each epoch.
train_loader.shuffle(None if data_order_seed is None else (data_order_seed + ep))
for it, (examples, labels) in enumerate(train_loader):
# Advance the data loader until the start epoch and iteration.
if ep == start_step.ep and it < start_step.it: continue
# Run the callbacks.
step = Step.from_epoch(ep, it, train_loader.iterations_per_epoch)
for callback in callbacks: callback(output_location, step, student, optimizer, logger)
# Exit at the end step.
if ep == end_step.ep and it == end_step.it: return
# Otherwise, train.
examples = examples.to(device=get_platform().torch_device)
labels = labels.to(device=get_platform().torch_device)
loss = 0.0
step_optimizer.zero_grad()
student.train()
teacher.eval()
student_outputs = student(examples)
with torch.no_grad():
teacher_outputs = teacher(examples)
s_logits = student_outputs
t_logits = teacher_outputs
# KL Divergence loss for the knowledge distillation
loss_ce = ce_loss_fct(
F.log_softmax(s_logits / distill_hparams.temperature, dim=-1),
F.softmax(t_logits / distill_hparams.temperature, dim=-1),
) * distill_hparams.temperature**2
loss += distill_hparams.alpha_ce * loss_ce
if distill_hparams.alpha_cls > 0.0:
loss_cls = student.loss_criterion(student_outputs, labels)
loss += distill_hparams.alpha_cls * loss_cls
if distill_hparams.alpha_mse > 0.0:
loss_mse = mse_loss_fct(s_logits, t_logits) / s_logits.size(0)
loss += distill_hparams.alpha_mse * loss_mse
if training_hparams.apex_fp16:
with apex.amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# Step forward. Ignore extraneous warnings that the lr_schedule generates.
step_optimizer.step()
with warnings.catch_warnings(): # Filter unnecessary warning.
warnings.filterwarnings("ignore", category=UserWarning)
lr_schedule.step()
get_platform().barrier()
def test_zero(self):
self.assertStepEquals(Step.zero(100), 0, 0, 0)
