def test_trainer_can_run_exponential_moving_average(self):
moving_average = ExponentialMovingAverage(self.model.named_parameters(), decay=0.9999)
trainer = Trainer(
model=self.model,
optimizer=self.optimizer,
data_loader=self.data_loader,
validation_data_loader=self.validation_data_loader,
num_epochs=2,
moving_average=moving_average,
)
trainer.train()
def test_trainer_can_run_exponential_moving_average(self):
moving_average = ExponentialMovingAverage(
self.model.named_parameters(), decay=0.9999)
trainer = Trainer(model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
train_dataset=self.instances,
validation_dataset=self.instances,
num_epochs=2,
moving_average=moving_average)
trainer.train()
def test_trainer_can_run_amp(self):
self.model.cuda()
trainer = Trainer(
self.model,
self.optimizer,
self.data_loader,
num_epochs=2,
cuda_device=0,
opt_level="O1",
)
_ = trainer.train()
def test_passing_trainer_multiple_gpus_raises_error(self):
self.model.cuda()
with pytest.raises(ConfigurationError):
Trainer(
self.model,
self.optimizer,
self.data_loader,
num_epochs=2,
cuda_device=[0, 1],
)
def test_trainer_can_run_cuda(self):
self.model.cuda()
trainer = Trainer(
self.model, self.optimizer, self.iterator, self.instances, num_epochs=2, cuda_device=0
)
metrics = trainer.train()
assert "peak_cpu_memory_MB" in metrics
assert isinstance(metrics["peak_cpu_memory_MB"], float)
assert metrics["peak_cpu_memory_MB"] > 0
assert "peak_gpu_0_memory_MB" in metrics
assert isinstance(metrics["peak_gpu_0_memory_MB"], int)
def test_trainer_can_log_learning_rates_tensorboard(self):
iterator = BasicIterator(batch_size=4)
iterator.index_with(self.vocab)
trainer = Trainer(self.model, self.optimizer,
iterator, self.instances, num_epochs=2,
serialization_dir=self.TEST_DIR,
should_log_learning_rate=True,
summary_interval=2)
trainer.train()
def test_trainer_saves_and_loads_best_validation_metrics_correctly_2(self):
# Use -loss and run 1 epoch of original-training, and one of restored-training
# Run 1 epoch of original training.
trainer = Trainer(
self.model,
self.optimizer,
self.iterator,
self.instances,
validation_dataset=self.instances,
validation_metric="+loss",
num_epochs=1,
serialization_dir=self.TEST_DIR,
)
trainer.train()
_ = trainer._restore_checkpoint()
best_epoch_1 = trainer._metric_tracker.best_epoch
best_validation_metrics_epoch_1 = trainer._metric_tracker.best_epoch_metrics
# best_validation_metrics_epoch_1: {'accuracy': 0.75, 'accuracy3': 1.0, 'loss': 0.6243013441562653}
assert isinstance(best_validation_metrics_epoch_1, dict)
assert "loss" in best_validation_metrics_epoch_1
# Run 1 more epoch of restored training.
restore_trainer = Trainer(
self.model,
self.optimizer,
self.iterator,
self.instances,
validation_dataset=self.instances,
validation_metric="+loss",
num_epochs=2,
serialization_dir=self.TEST_DIR,
)
restore_trainer.train()
_ = restore_trainer._restore_checkpoint()
best_epoch_2 = restore_trainer._metric_tracker.best_epoch
best_validation_metrics_epoch_2 = restore_trainer._metric_tracker.best_epoch_metrics
# Because of using +loss, 2nd epoch won't be better than 1st. So best val metrics should be same.
assert best_epoch_1 == best_epoch_2 == 0
assert best_validation_metrics_epoch_2 == best_validation_metrics_epoch_1
def test_trainer_can_run(self):
trainer = Trainer(
model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
train_dataset=self.instances,
validation_dataset=self.instances,
num_epochs=2,
)
metrics = trainer.train()
assert "best_validation_loss" in metrics
assert isinstance(metrics["best_validation_loss"], float)
assert "best_validation_accuracy" in metrics
assert isinstance(metrics["best_validation_accuracy"], float)
assert "best_validation_accuracy3" in metrics
assert isinstance(metrics["best_validation_accuracy3"], float)
assert "best_epoch" in metrics
assert isinstance(metrics["best_epoch"], int)
# Making sure that both increasing and decreasing validation metrics work.
trainer = Trainer(
model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
train_dataset=self.instances,
validation_dataset=self.instances,
validation_metric="+loss",
num_epochs=2,
)
metrics = trainer.train()
assert "best_validation_loss" in metrics
assert isinstance(metrics["best_validation_loss"], float)
assert "best_validation_accuracy" in metrics
assert isinstance(metrics["best_validation_accuracy"], float)
assert "best_validation_accuracy3" in metrics
assert isinstance(metrics["best_validation_accuracy3"], float)
assert "best_epoch" in metrics
assert isinstance(metrics["best_epoch"], int)
assert "peak_cpu_memory_MB" in metrics
assert isinstance(metrics["peak_cpu_memory_MB"], float)
assert metrics["peak_cpu_memory_MB"] > 0
def test_trainer_can_run_with_lr_scheduler(self):
lr_scheduler = ExponentialLearningRateScheduler(self.optimizer, gamma=0.5)
trainer = Trainer(
model=self.model,
optimizer=self.optimizer,
data_loader=self.data_loader,
learning_rate_scheduler=lr_scheduler,
validation_metric="-loss",
validation_data_loader=self.validation_data_loader,
num_epochs=2,
)
trainer.train()
def test_trainer_can_run_with_lr_scheduler(self):
lr_params = Params({"type": "reduce_on_plateau"})
lr_scheduler = LearningRateScheduler.from_params(self.optimizer, lr_params)
trainer = Trainer(model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
learning_rate_scheduler=lr_scheduler,
validation_metric="-loss",
train_dataset=self.instances,
validation_dataset=self.instances,
num_epochs=2)
trainer.train()
def test_restored_training_returns_best_epoch_metrics_even_if_no_better_epoch_is_found_after_restoring(
self, ):
# Instead of -loss, use +loss to assure 2nd epoch is considered worse.
# Run 1 epoch of original training.
original_trainer = Trainer(
self.model,
self.optimizer,
self.iterator,
self.instances,
validation_dataset=self.instances,
validation_metric="+loss",
num_epochs=1,
serialization_dir=self.TEST_DIR,
)
training_metrics = original_trainer.train()
# Run 1 epoch of restored training.
restored_trainer = Trainer(
self.model,
self.optimizer,
self.iterator,
self.instances,
validation_dataset=self.instances,
validation_metric="+loss",
num_epochs=2,
serialization_dir=self.TEST_DIR,
)
restored_metrics = restored_trainer.train()
assert "best_validation_loss" in restored_metrics
assert "best_validation_accuracy" in restored_metrics
assert "best_validation_accuracy3" in restored_metrics
assert "best_epoch" in restored_metrics
# Epoch 2 validation loss should be lesser than that of Epoch 1
assert training_metrics["best_validation_loss"] == restored_metrics[
"best_validation_loss"]
assert training_metrics["best_epoch"] == 0
assert training_metrics["validation_loss"] > restored_metrics[
"validation_loss"]
def test_trainer_can_log_histograms(self):
# enable activation logging
for module in self.model.modules():
module.should_log_activations = True
trainer = Trainer(self.model,
self.optimizer,
self.iterator,
self.instances,
num_epochs=3,
serialization_dir=self.TEST_DIR,
histogram_interval=2)
trainer.train()
def test_should_stop_early_with_flat_lining_metric(self):
# pylint: disable=protected-access
flatline = [.2] * 6
tracker = Trainer(self.model, self.optimizer,
self.iterator, self.instances,
validation_dataset=self.instances,
num_epochs=3,
serialization_dir=self.TEST_DIR,
patience=5,
validation_metric="+test")._metric_tracker
tracker.add_metrics(flatline)
assert tracker.should_stop_early
tracker = Trainer(self.model, self.optimizer,
self.iterator, self.instances,
validation_dataset=self.instances,
num_epochs=3,
serialization_dir=self.TEST_DIR,
patience=5,
validation_metric="-test")._metric_tracker
tracker.add_metrics(flatline)
assert tracker.should_stop_early
def test_trainer_raises_on_model_with_no_loss_key(self):
class FakeModel(Model):
def forward(self, **kwargs): # pylint: disable=arguments-differ,unused-argument
return {}
with pytest.raises(RuntimeError):
trainer = Trainer(FakeModel(None),
self.optimizer,
self.iterator,
self.instances,
num_epochs=2,
serialization_dir=self.TEST_DIR)
trainer.train()
def test_trainer_can_run_with_lr_scheduler(self):
lr_scheduler = ExponentialLearningRateScheduler(self.optimizer, gamma=0.5)
trainer = Trainer(
model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
learning_rate_scheduler=lr_scheduler,
validation_metric="-loss",
train_dataset=self.instances,
validation_dataset=self.instances,
num_epochs=2,
)
trainer.train()
def test_trainer_can_log_learning_rates_tensorboard(self):
data_loader = DataLoader(self.instances, batch_size=4, collate_fn=allennlp_collate)
trainer = Trainer(
self.model,
self.optimizer,
data_loader,
num_epochs=2,
serialization_dir=self.TEST_DIR,
should_log_learning_rate=True,
summary_interval=2,
)
trainer.train()
def test_trainer_can_run_and_resume_with_momentum_scheduler(self):
scheduler = MomentumScheduler.from_params(
self.optimizer,
Params({
"type": "inverted_triangular",
"cool_down": 2,
"warm_up": 2
}))
trainer = Trainer(model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
momentum_scheduler=scheduler,
validation_metric="-loss",
train_dataset=self.instances,
validation_dataset=self.instances,
num_epochs=4,
serialization_dir=self.TEST_DIR)
trainer.train()
new_scheduler = MomentumScheduler.from_params(
self.optimizer,
Params({
"type": "inverted_triangular",
"cool_down": 2,
"warm_up": 2
}))
new_trainer = Trainer(model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
momentum_scheduler=new_scheduler,
validation_metric="-loss",
train_dataset=self.instances,
validation_dataset=self.instances,
num_epochs=6,
serialization_dir=self.TEST_DIR)
epoch = new_trainer._restore_checkpoint()
assert epoch == 4
assert new_trainer._momentum_scheduler.last_epoch == 3
new_trainer.train()
def test_trainer_respects_num_serialized_models_to_keep(self):
trainer = Trainer(self.model, self.optimizer,
self.iterator, self.instances, num_epochs=5,
serialization_dir=self.TEST_DIR,
num_serialized_models_to_keep=3)
trainer.train()
# Now check the serialized files
for prefix in ['model_state_epoch_*', 'training_state_epoch_*']:
file_names = glob.glob(os.path.join(self.TEST_DIR, prefix))
epochs = [int(re.search(r"_([0-9])\.th", fname).group(1))
for fname in file_names]
assert sorted(epochs) == [2, 3, 4]
def test_production_rule_field_with_multiple_gpus(self):
wikitables_dir = 'allennlp/tests/fixtures/data/wikitables/'
wikitables_reader = WikiTablesDatasetReader(tables_directory=wikitables_dir,
dpd_output_directory=wikitables_dir + 'dpd_output/')
instances = wikitables_reader.read(wikitables_dir + 'sample_data.examples')
archive_path = self.FIXTURES_ROOT / 'semantic_parsing' / 'wikitables' / 'serialization' / 'model.tar.gz'
model = load_archive(archive_path).model
model.cuda()
multigpu_iterator = BasicIterator(batch_size=4)
multigpu_iterator.index_with(model.vocab)
trainer = Trainer(model, self.optimizer, multigpu_iterator, instances, num_epochs=2, cuda_device=[0, 1])
trainer.train()
def test_trainer_saves_models_at_specified_interval(self):
iterator = BasicIterator(batch_size=4)
iterator.index_with(self.vocab)
trainer = Trainer(self.model, self.optimizer,
iterator, self.instances, num_epochs=2,
serialization_dir=self.TEST_DIR,
model_save_interval=0.0001)
trainer.train()
# Now check the serialized files for models saved during the epoch.
prefix = 'model_state_epoch_*'
file_names = sorted(glob.glob(os.path.join(self.TEST_DIR, prefix)))
epochs = [re.search(r"_([0-9\.\-]+)\.th", fname).group(1)
for fname in file_names]
# We should have checkpoints at the end of each epoch and during each, e.g.
# [0.timestamp, 0, 1.timestamp, 1]
assert len(epochs) == 4
assert epochs[3] == '1'
assert '.' in epochs[0]
# Now make certain we can restore from timestamped checkpoint.
# To do so, remove the checkpoint from the end of epoch 1&2, so
# that we are forced to restore from the timestamped checkpoints.
for k in range(2):
os.remove(os.path.join(self.TEST_DIR, 'model_state_epoch_{}.th'.format(k)))
os.remove(os.path.join(self.TEST_DIR, 'training_state_epoch_{}.th'.format(k)))
os.remove(os.path.join(self.TEST_DIR, 'best.th'))
restore_trainer = Trainer(self.model, self.optimizer,
self.iterator, self.instances, num_epochs=2,
serialization_dir=self.TEST_DIR,
model_save_interval=0.0001)
epoch = restore_trainer._restore_checkpoint() # pylint: disable=protected-access
assert epoch == 2
# One batch per epoch.
assert restore_trainer._batch_num_total == 2 # pylint: disable=protected-access
def test_trainer_raises_on_model_with_no_loss_key(self):
class FakeModel(Model):
def forward(self, **kwargs):
return {}
with pytest.raises(RuntimeError):
trainer = Trainer(
FakeModel(None),
self.optimizer,
self.data_loader,
num_epochs=2,
serialization_dir=self.TEST_DIR,
)
trainer.train()
def test_should_stop_early_with_invalid_patience(self):
for patience in [0, -1, -2, 1.5, 'None']:
with pytest.raises(ConfigurationError,
match='.* is an invalid value for "patience": '
'it must be a positive integer or None '
'\\(if you want to disable early stopping\\)'):
Trainer(self.model,
self.optimizer,
self.iterator,
self.instances,
validation_dataset=self.instances,
num_epochs=100,
patience=patience,
validation_metric="+test")
def test_should_stop_early_with_invalid_patience(self):
for patience in [0, -1, -2, 1.5, 'None']:
with pytest.raises(
ConfigurationError,
message='No ConfigurationError for patience={}'.format(
patience)):
Trainer(self.model,
self.optimizer,
self.iterator,
self.instances,
validation_dataset=self.instances,
num_epochs=100,
patience=patience,
validation_metric="+test")
def test_passing_trainer_multiple_gpus_raises_error(self):
self.model.cuda()
multigpu_iterator = BasicIterator(batch_size=4)
multigpu_iterator.index_with(self.vocab)
with pytest.raises(ConfigurationError):
Trainer(
self.model,
self.optimizer,
multigpu_iterator,
self.instances,
num_epochs=2,
cuda_device=[0, 1],
)
def test_trainer_can_run(self):
trainer = Trainer(model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
train_dataset=self.instances,
validation_dataset=self.instances,
num_epochs=2)
metrics = trainer.train()
assert 'best_validation_loss' in metrics
assert isinstance(metrics['best_validation_loss'], float)
assert 'best_validation_accuracy' in metrics
assert isinstance(metrics['best_validation_accuracy'], float)
assert 'best_validation_accuracy3' in metrics
assert isinstance(metrics['best_validation_accuracy3'], float)
assert 'best_epoch' in metrics
assert isinstance(metrics['best_epoch'], int)
# Making sure that both increasing and decreasing validation metrics work.
trainer = Trainer(model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
train_dataset=self.instances,
validation_dataset=self.instances,
validation_metric='+loss',
num_epochs=2)
metrics = trainer.train()
assert 'best_validation_loss' in metrics
assert isinstance(metrics['best_validation_loss'], float)
assert 'best_validation_accuracy' in metrics
assert isinstance(metrics['best_validation_accuracy'], float)
assert 'best_validation_accuracy3' in metrics
assert isinstance(metrics['best_validation_accuracy3'], float)
assert 'best_epoch' in metrics
assert isinstance(metrics['best_epoch'], int)
assert 'peak_cpu_memory_MB' in metrics
assert isinstance(metrics['peak_cpu_memory_MB'], float)
assert metrics['peak_cpu_memory_MB'] > 0
def test_trainer_can_resume_with_lr_scheduler(self):
lr_scheduler = LearningRateScheduler.from_params(
self.optimizer, Params({
"type": "exponential",
"gamma": 0.5
}))
trainer = Trainer(
model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
learning_rate_scheduler=lr_scheduler,
train_dataset=self.instances,
validation_dataset=self.instances,
num_epochs=2,
serialization_dir=self.TEST_DIR,
)
trainer.train()
new_lr_scheduler = LearningRateScheduler.from_params(
self.optimizer, Params({
"type": "exponential",
"gamma": 0.5
}))
new_trainer = Trainer(
model=self.model,
optimizer=self.optimizer,
iterator=self.iterator,
learning_rate_scheduler=new_lr_scheduler,
train_dataset=self.instances,
validation_dataset=self.instances,
num_epochs=4,
serialization_dir=self.TEST_DIR,
)
epoch = new_trainer._restore_checkpoint()
assert epoch == 2
assert new_trainer._learning_rate_scheduler.lr_scheduler.last_epoch == 1
new_trainer.train()
def test_trainer_can_resume_training_for_exponential_moving_average(self):
moving_average = ExponentialMovingAverage(
self.model.named_parameters())
trainer = Trainer(
self.model,
self.optimizer,
self.iterator,
self.instances,
validation_dataset=self.instances,
num_epochs=1,
serialization_dir=self.TEST_DIR,
moving_average=moving_average,
)
trainer.train()
new_moving_average = ExponentialMovingAverage(
self.model.named_parameters())
new_trainer = Trainer(
self.model,
self.optimizer,
self.iterator,
self.instances,
validation_dataset=self.instances,
num_epochs=3,
serialization_dir=self.TEST_DIR,
moving_average=new_moving_average,
)
epoch = new_trainer._restore_checkpoint()
assert epoch == 1
tracker = trainer._metric_tracker
assert tracker.is_best_so_far()
assert tracker._best_so_far is not None
new_trainer.train()
def train():
model.cuda(device)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
trainer = Trainer(model=model,
optimizer=optimizer,
iterator=train_iterator,
validation_iterator=validation_iterator,
train_dataset=train_dataset,
validation_dataset=validation_dataset,
num_epochs=num_epochs,
serialization_dir=serialization_dir,
grad_clipping=grad_clipping,
cuda_device=device)
trainer.train()
def test_multigpu_qanet(self):
params = Params.from_file(self.param_file)
vocab = Vocabulary.from_instances(self.instances)
model = Model.from_params(vocab=vocab, params=params["model"]).cuda()
optimizer = torch.optim.SGD(self.model.parameters(),
0.01,
momentum=0.9)
self.instances.index_with(model.vocab)
loader = DataLoader(self.instances, batch_size=4)
trainer = Trainer(model,
optimizer,
loader,
num_epochs=2,
cuda_device=[0, 1])
trainer.train()
def test_multigpu_qanet(self):
params = Params.from_file(self.param_file)
vocab = Vocabulary.from_instances(self.instances)
model = Model.from_params(vocab=vocab, params=params['model']).cuda()
optimizer = torch.optim.SGD(self.model.parameters(),
0.01,
momentum=0.9)
multigpu_iterator = BasicIterator(batch_size=4)
multigpu_iterator.index_with(model.vocab)
trainer = Trainer(model,
optimizer,
multigpu_iterator,
self.instances,
num_epochs=2,
cuda_device=[0, 1])
trainer.train()
