def test_train_step_epoch_end(tmpdir):
"""
Checks train_step + training_epoch_end (NO training_step_end)
"""
model = DeterministicModel()
model.training_step = model.training_step_dict_return
model.training_step_end = None
model.training_epoch_end = model.training_epoch_end_dict
model.val_dataloader = None
trainer = Trainer(max_epochs=1, weights_summary=None)
trainer.fit(model)
# make sure correct steps were called
assert model.training_step_called
assert not model.training_step_end_called
assert model.training_epoch_end_called
# assert epoch end metrics were added
assert trainer.callback_metrics['epoch_end_log_1'] == 178
assert trainer.progress_bar_metrics['epoch_end_pbar_1'] == 234
# make sure training outputs what is expected
for batch_idx, batch in enumerate(model.train_dataloader()):
break
out = trainer.run_training_batch(batch, batch_idx)
assert out.signal == 0
assert out.batch_log_metrics['log_acc1'] == 12.0
assert out.batch_log_metrics['log_acc2'] == 7.0
train_step_end_out = out.training_step_output_for_epoch_end
pbar_metrics = train_step_end_out['progress_bar']
assert pbar_metrics['pbar_acc1'] == 17.0
assert pbar_metrics['pbar_acc2'] == 19.0
def test_no_callbacks_with_train_loop_only(tmpdir):
"""
Make sure early stop + checkpoint work with only a train loop
"""
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_no_callbacks_result_obj
model.training_epoch_end = None
model.val_dataloader = None
batches = 3
epochs = 3
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
log_every_n_steps=1,
limit_train_batches=batches,
weights_summary=None,
)
trainer.fit(model)
all_losses = trainer.dev_debugger.saved_train_losses
assert len(all_losses) == batches * epochs
assert trainer.early_stop_callback is None
assert len(trainer.dev_debugger.checkpoint_callback_history) == 3
assert len(trainer.dev_debugger.early_stopping_history) == 0
def test_full_training_loop_dict(tmpdir):
"""
Checks train_step + training_step_end + training_epoch_end
"""
model = DeterministicModel()
model.training_step = model.training_step_for_step_end_dict
model.training_step_end = model.training_step_end_dict
model.training_epoch_end = model.training_epoch_end_dict
model.val_dataloader = None
trainer = Trainer(max_epochs=1, weights_summary=None)
trainer.fit(model)
# make sure correct steps were called
assert model.training_step_called
assert model.training_step_end_called
assert model.training_epoch_end_called
# assert epoch end metrics were added
assert trainer.callback_metrics['epoch_end_log_1'] == 178
assert trainer.progress_bar_metrics['epoch_end_pbar_1'] == 234
# make sure training outputs what is expected
for batch_idx, batch in enumerate(model.train_dataloader()):
break
out = trainer.run_training_batch(batch, batch_idx)
signal, grad_norm_dic, all_log_metrics, training_step_output_for_epoch_end = out
assert signal == 0
assert all_log_metrics['log_acc1'] == 12.0
assert all_log_metrics['log_acc2'] == 7.0
pbar_metrics = training_step_output_for_epoch_end['pbar_on_batch_end']
assert pbar_metrics['pbar_acc1'] == 17.0
assert pbar_metrics['pbar_acc2'] == 19.0
def test_val_step_only_epoch_metrics(tmpdir):
"""
Make sure the logged + pbar metrics are allocated accordingly when auto-reduced at epoch end
"""
# enable internal debugging actions
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_and_step_for_callbacks
model.training_step_end = None
model.training_epoch_end = None
model.validation_step = model.validation_step_result_only_epoch_metrics
model.validation_step_end = None
model.validation_epoch_end = None
batches = 3
epochs = 3
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
row_log_interval=1,
limit_train_batches=batches,
weights_summary=None,
)
trainer.fit(model)
# make sure correct steps were called
assert model.validation_step_called
assert not model.validation_step_end_called
assert not model.validation_epoch_end_called
# no early stopping
assert len(trainer.dev_debugger.early_stopping_history) == 0
# make sure we logged the exact number of metrics
assert len(trainer.dev_debugger.logged_metrics) == epochs
assert len(trainer.dev_debugger.pbar_added_metrics) == epochs
# make sure we logged the correct epoch metrics
for metric in trainer.dev_debugger.logged_metrics:
assert 'no_val_no_pbar' not in metric
assert 'val_step_pbar_acc' not in metric
assert metric['val_step_log_acc'] == (12 + 13) / 2
assert metric['val_step_log_pbar_acc'] == (13 + 14) / 2
# make sure we logged the correct epoch pbar metrics
for metric in trainer.dev_debugger.pbar_added_metrics:
assert 'no_val_no_pbar' not in metric
assert 'val_step_log_acc' not in metric
assert metric['val_step_log_pbar_acc'] == (13 + 14) / 2
assert metric['val_step_pbar_acc'] == (14 + 15) / 2
# only 1 checkpoint expected since values didn't change after that
assert len(trainer.dev_debugger.checkpoint_callback_history) == 1
# make sure the last known metric is correct
assert trainer.logger_connector.callback_metrics[
'val_checkpoint_on'] == 171
def test_val_step_epoch_end_result(tmpdir):
"""
Make sure val step + val epoch end works with EvalResult
"""
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_and_step_for_callbacks
model.training_step_end = None
model.training_epoch_end = None
model.validation_step = model.validation_step_for_epoch_end_result
model.validation_step_end = None
model.validation_epoch_end = model.validation_epoch_end_result
batches = 3
epochs = 3
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
log_every_n_steps=1,
limit_train_batches=batches,
limit_val_batches=batches,
weights_summary=None,
)
trainer.fit(model)
assert len(trainer.logger_connector.callback_metrics) == 6
# make sure correct steps were called
assert model.validation_step_called
assert not model.validation_step_end_called
assert model.validation_epoch_end_called
# no early stopping
assert len(trainer.dev_debugger.early_stopping_history) == 0
# make sure we logged the exact number of metrics
assert len(trainer.dev_debugger.logged_metrics) == epochs
assert len(trainer.dev_debugger.pbar_added_metrics) == epochs
# make sure we logged the correct metrics
for metric in trainer.dev_debugger.logged_metrics:
assert metric['val_epoch_end_metric'] == 189
assert 'val_step_metric' in metric
# make sure we pbar logged the correct metrics
for metric in trainer.dev_debugger.pbar_added_metrics:
assert metric['val_epoch_end_metric'] == 189
assert 'val_step_metric' in metric
# only 1 checkpoint expected since values didn't change after that
assert len(trainer.dev_debugger.checkpoint_callback_history) == 1
# make sure the last known metric is correct
assert trainer.logger_connector.callback_metrics['checkpoint_on'] == 189
def test_val_step_result_callbacks(tmpdir):
"""
Tests that val step can be used:
- val step
- no other val_xxx
- train loop
- callbacks coming from val loop (not train loop)
"""
# enable internal debugging actions
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_and_step_for_callbacks
model.training_step_end = None
model.training_epoch_end = None
model.validation_step = model.validation_step_result_callbacks
model.validation_step_end = None
model.validation_epoch_end = None
batches = 3
epochs = 300
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
early_stop_callback=True,
log_every_n_steps=1,
limit_train_batches=batches,
weights_summary=None,
)
trainer.fit(model)
# make sure correct steps were called
assert model.validation_step_called
assert not model.validation_step_end_called
assert not model.validation_epoch_end_called
# assert that early stopping happened after the requested num of steps
# if it used the train step for ES then it wouldn't be 5
assert len(trainer.dev_debugger.early_stopping_history) == 5
# only 2 checkpoints expected
assert len(trainer.dev_debugger.checkpoint_callback_history) == 2
# make sure the last known metric is correct
assert trainer.logger_connector.callback_metrics[
'checkpoint_on'] == 171 + 15
# did not request any metrics to log (except the metrics saying which epoch we are on)
assert len(trainer.logger_connector.progress_bar_metrics) == 0
assert len(trainer.dev_debugger.logged_metrics) == 0
def test_full_training_loop_scalar(tmpdir):
"""
Checks train_step + training_step_end + training_epoch_end
(all with scalar return from train_step)
"""
model = DeterministicModel()
model.training_step = model.training_step_scalar_return
model.training_step_end = model.training_step_end_scalar
model.training_epoch_end = model.training_epoch_end_scalar
model.val_dataloader = None
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
weights_summary=None,
)
trainer.fit(model)
# make sure correct steps were called
assert model.training_step_called
assert model.training_step_end_called
assert model.training_epoch_end_called
# assert epoch end metrics were added
assert 'epoch' in trainer.callback_metrics and len(
trainer.callback_metrics) == 1
assert len(trainer.progress_bar_metrics) == 0
# make sure training outputs what is expected
for batch_idx, batch in enumerate(model.train_dataloader()):
break
out = trainer.run_training_batch(batch, batch_idx)
assert out.signal == 0
assert len(out.batch_log_metrics) == 0 and isinstance(
out.batch_log_metrics, dict)
assert len(out.grad_norm_dic) == 0 and isinstance(out.grad_norm_dic, dict)
train_step_out = out.training_step_output_for_epoch_end
assert len(train_step_out) == 1
train_step_out = train_step_out[0][0]
assert isinstance(train_step_out, torch.Tensor)
assert train_step_out.item() == 171
# make sure the optimizer closure returns the correct things
opt_closure_result = trainer.optimizer_closure(batch, batch_idx, 0,
trainer.optimizers[0],
trainer.hiddens)
assert opt_closure_result['loss'].item() == 171
def test_val_step_using_train_callbacks(tmpdir):
"""
ES conditioned in train
CKPT conditioned in val
"""
# enable internal debugging actions
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_and_step_for_callbacks
model.training_step_end = None
model.training_epoch_end = None
model.validation_step = model.validation_step_result_no_callbacks
model.validation_step_end = None
model.validation_epoch_end = None
batches = 3
epochs = 300
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
early_stop_callback=True,
log_every_n_steps=1,
limit_train_batches=batches,
weights_summary=None,
)
trainer.fit(model)
expected_epochs = 10
# make sure correct steps were called
assert model.validation_step_called
assert not model.validation_step_end_called
assert not model.validation_epoch_end_called
# early stopping was not conditioned in val loop, but instead in train loop
assert len(trainer.dev_debugger.early_stopping_history) == expected_epochs
# only 2 checkpoints expected
assert len(trainer.dev_debugger.checkpoint_callback_history) == 2
# make sure the last known metric is correct
assert trainer.logger_connector.callback_metrics[
'checkpoint_on'] == 171 + 20
# did not request any metrics to log (except the metrics saying which epoch we are on)
assert len(trainer.logger_connector.progress_bar_metrics) == 0
assert len(trainer.dev_debugger.logged_metrics) == 0
def test_use_callbacks_with_train_loop_only(tmpdir):
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_and_step_for_callbacks
model.training_epoch_end = None
model.val_dataloader = None
batches = 3
epochs = 300
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
early_stop_callback=True,
log_every_n_steps=1,
limit_train_batches=batches,
weights_summary=None,
)
trainer.fit(model)
num_expected_epochs = 10
# ----------------------------------
# VERIFY EARLY STOPPING BEHAVIOR
# ----------------------------------
# with train loop only it happens on every epoch
early_stop_vals = trainer.dev_debugger.early_stopping_history
assert len(early_stop_vals) == num_expected_epochs
min_val = min([x['best'] for x in early_stop_vals])
assert min_val == 171 + 9
all_losses = trainer.dev_debugger.saved_train_losses
from collections import Counter
batch_idxs = Counter([x['batch_idx'] for x in all_losses])
for i, val in batch_idxs.items():
assert val == num_expected_epochs
assert i in [0, 1, 2]
# ----------------------------------
# VERIFY CHECKPOINTING BEHAVIOR
# ----------------------------------
ckpt_vals = trainer.dev_debugger.checkpoint_callback_history
assert len(ckpt_vals) == 5, '5 ckpts should have been saved'
for ckpt_val, expected_epoch in zip(ckpt_vals, [0, 1, 2, 3, 6]):
assert ckpt_val['epoch'] == expected_epoch
assert ckpt_val['monitor'] == 'checkpoint_on'
def test_val_step_only_step_metrics(tmpdir):
"""
Make sure the logged + pbar metrics are allocated accordingly at every step when requested
"""
# enable internal debugging actions
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_and_step_for_callbacks
model.training_step_end = None
model.training_epoch_end = None
model.validation_step = model.validation_step_result_only_step_metrics
model.validation_step_end = None
model.validation_epoch_end = None
batches = 3
epochs = 3
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
log_every_n_steps=1,
limit_train_batches=batches,
limit_val_batches=batches,
weights_summary=None,
)
trainer.fit(model)
# make sure correct steps were called
assert model.validation_step_called
assert not model.validation_step_end_called
assert not model.validation_epoch_end_called
# no early stopping
assert len(trainer.dev_debugger.early_stopping_history) == 0
# make sure we logged the exact number of metrics
assert len(trainer.dev_debugger.logged_metrics) == epochs * batches
assert len(
trainer.dev_debugger.pbar_added_metrics) == epochs * batches + (epochs)
# only 1 checkpoint expected since values didn't change after that
assert len(trainer.dev_debugger.checkpoint_callback_history) == 1
# make sure the last known metric is correct
assert trainer.logger_connector.callback_metrics['checkpoint_on'] == 189
def test_result_obj_lr_scheduler_step(tmpdir):
"""
test that the LR scheduler was called at the correct time with the correct metrics
"""
model = DeterministicModel()
model.training_step = model.training_step_for_step_end_dict
model.training_step_end = model.training_step_end_dict
model.training_epoch_end = model.training_epoch_end_dict
model.val_dataloader = None
model.configure_optimizers = model.configure_optimizers__lr_on_plateau_step
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=2,
weights_summary=None,
)
trainer.fit(model)
assert len(trainer.dev_debugger.saved_lr_scheduler_updates) == 8
def test_train_step_epoch_end(tmpdir):
"""
Checks train_step + training_epoch_end (NO training_step_end)
"""
model = DeterministicModel()
model.training_step = model.training_step_dict_return
model.training_step_end = None
model.training_epoch_end = model.training_epoch_end_dict
model.val_dataloader = None
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
weights_summary=None,
)
trainer.fit(model)
# make sure correct steps were called
assert model.training_step_called
assert not model.training_step_end_called
assert model.training_epoch_end_called
# assert epoch end metrics were added
assert trainer.logger_connector.callback_metrics['epoch_end_log_1'] == 178
assert trainer.logger_connector.progress_bar_metrics[
'epoch_end_pbar_1'] == 234
# make sure training outputs what is expected
batch_idx, batch = 0, next(iter(model.train_dataloader()))
out = trainer.train_loop.run_training_batch(batch, batch_idx, 0)
assert out.signal == 0
assert trainer.logger_connector.logged_metrics['log_acc1'] == 12.0
assert trainer.logger_connector.logged_metrics['log_acc2'] == 7.0
# outputs are for 1 optimizer and no tbptt
train_step_end_out = out.training_step_output_for_epoch_end
assert len(train_step_end_out) == 1
train_step_end_out = train_step_end_out[0][0]
pbar_metrics = train_step_end_out['progress_bar']
assert pbar_metrics['pbar_acc1'] == 17.0
assert pbar_metrics['pbar_acc2'] == 19.0
def test_train_step_epoch_end_scalar(tmpdir):
"""
Checks train_step + training_epoch_end (NO training_step_end)
(with scalar return)
"""
os.environ['PL_DEV_DEBUG'] = '0'
model = DeterministicModel()
model.training_step = model.training_step_scalar_return
model.training_step_end = None
model.training_epoch_end = model.training_epoch_end_scalar
model.val_dataloader = None
trainer = Trainer(max_epochs=1, weights_summary=None)
trainer.fit(model)
# make sure correct steps were called
assert model.training_step_called
assert not model.training_step_end_called
assert model.training_epoch_end_called
# assert epoch end metrics were added
assert len(trainer.logger_connector.callback_metrics) == 0
assert len(trainer.logger_connector.progress_bar_metrics) == 0
# make sure training outputs what is expected
for batch_idx, batch in enumerate(model.train_dataloader()):
break
out = trainer.train_loop.run_training_batch(batch, batch_idx, 0)
assert out.signal == 0
assert len(out.grad_norm_dic) == 0 and isinstance(out.grad_norm_dic, dict)
train_step_out = out.training_step_output_for_epoch_end
assert len(train_step_out) == 1
train_step_out = train_step_out[0][0]
assert isinstance(train_step_out['minimize'], torch.Tensor)
assert train_step_out['minimize'].item() == 171
# make sure the optimizer closure returns the correct things
opt_closure_result = trainer.train_loop.training_step_and_backward(
batch, batch_idx, 0, trainer.optimizers[0], trainer.hiddens)
assert opt_closure_result['loss'].item() == 171
def test_no_auto_callbacks_with_train_loop_only(tmpdir):
"""
Make sure early stop + checkpoint work with only a train loop
"""
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_no_default_callbacks_for_train_loop
model.training_epoch_end = None
model.val_dataloader = None
batches = 3
epochs = 3
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
row_log_interval=1,
limit_train_batches=batches,
weights_summary=None,
)
trainer.fit(model)
assert len(trainer.logger_connector.callback_metrics) == 1
all_losses = trainer.dev_debugger.saved_train_losses
assert len(all_losses) == batches * epochs
assert trainer.checkpoint_callback.monitor == 'checkpoint_on'
assert trainer.early_stop_callback is None
trainer = Trainer(
default_root_dir=tmpdir,
early_stop_callback=True,
max_epochs=epochs,
row_log_interval=1,
limit_train_batches=batches,
weights_summary=None,
)
trainer.fit(model)
assert trainer.early_stop_callback.monitor == 'early_stop_on'
def test_full_training_loop_dict(tmpdir):
"""
Checks train_step + training_step_end + training_epoch_end
"""
model = DeterministicModel()
model.training_step = model.training_step_for_step_end_dict
model.training_step_end = model.training_step_end_dict
model.training_epoch_end = model.training_epoch_end_dict
model.val_dataloader = None
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
weights_summary=None,
)
trainer.fit(model)
# make sure correct steps were called
assert model.training_step_called
assert model.training_step_end_called
assert model.training_epoch_end_called
# assert epoch end metrics were added
assert trainer.callback_metrics['epoch_end_log_1'] == 178
assert trainer.progress_bar_metrics['epoch_end_pbar_1'] == 234
# make sure training outputs what is expected
batch_idx, batch = 0, next(iter(model.train_dataloader()))
out = trainer.run_training_batch(batch, batch_idx)
assert out.signal == 0
assert out.batch_log_metrics['log_acc1'] == 14.0
assert out.batch_log_metrics['log_acc2'] == 9.0
# get the output of the first optimizer
train_step_end_out = out.training_step_output_for_epoch_end
assert len(train_step_end_out) == 1
train_step_end_out = train_step_end_out[0][0]
pbar_metrics = train_step_end_out['progress_bar']
assert pbar_metrics['pbar_acc1'] == 19.0
assert pbar_metrics['pbar_acc2'] == 21.0
def test_training_step_result_log_epoch_only(tmpdir):
"""
Tests that only training_step can be used with TrainResult
Makes sure that things are routed to pbar, loggers and loss accordingly
Makes sure pbar and logs happen on epoch only when requested
"""
# enable internal debugging actions
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_only
model.training_step_end = None
model.training_epoch_end = None
model.val_dataloader = None
epochs = 3
batches = 2
trainer = Trainer(
default_root_dir=tmpdir,
limit_train_batches=batches,
limit_val_batches=batches,
log_every_n_steps=1,
max_epochs=epochs,
weights_summary=None,
)
trainer.fit(model)
# make sure correct steps were called
assert model.training_step_called
assert not model.training_step_end_called
assert not model.training_epoch_end_called
assert len(trainer.logger_connector.callback_metrics) == 11
# make sure correct metrics are logged (one per batch step as requested)
assert len(trainer.dev_debugger.logged_metrics) == epochs
epoch_metrics = trainer.dev_debugger.logged_metrics
assert len(epoch_metrics) == epochs
for batch_idx, logged_metrics in enumerate(epoch_metrics):
assert logged_metrics[f'epoch_log_and_pbar_acc1_e{batch_idx}'] == 14.0
assert logged_metrics[f'epoch_log_acc2_e{batch_idx}'] == 15.0
assert f'epoch_pbar_acc3_e{batch_idx}' not in logged_metrics
assert len(logged_metrics) == 4
# make sure we are using the correct metrics for callbacks
assert trainer.logger_connector.callback_metrics['checkpoint_on'] == 171
# make sure pbar metrics are correct ang log metrics did not leak
for epoch_idx in range(epochs):
assert trainer.logger_connector.progress_bar_metrics[
f'epoch_log_and_pbar_acc1_e{epoch_idx}'] == 14
assert trainer.logger_connector.progress_bar_metrics[
f'epoch_pbar_acc3_e{epoch_idx}'] == 16
assert f'epoch_log_acc2_e{epoch_idx}' not in trainer.logger_connector.progress_bar_metrics
# make sure training outputs what is expected
for batch_idx, batch in enumerate(model.train_dataloader()):
break
out = trainer.train_loop.run_training_batch(batch, batch_idx, 0)
assert out.signal == 0
assert len(out.batch_log_metrics) == 0
train_step_out = out.training_step_output_for_epoch_end
assert len(train_step_out) == 1
train_step_out = train_step_out[0][0]
assert isinstance(train_step_out, TrainResult)
assert 'minimize' in train_step_out
assert f'epoch_log_and_pbar_acc1_e{trainer.current_epoch}' in train_step_out
assert f'epoch_log_acc2_e{trainer.current_epoch}' in train_step_out
# make sure the optimizer closure returns the correct things
opt_closure_result = trainer.train_loop.training_step_and_backward(
batch, batch_idx, 0, trainer.optimizers[0], trainer.hiddens)
assert opt_closure_result['loss'] == (42.0 * 3) + (15.0 * 3)
def test_val_step_only_step_metrics(tmpdir):
"""
Make sure the logged + pbar metrics are allocated accordingly at every step when requested
"""
# enable internal debugging actions
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_and_step_for_callbacks
model.training_step_end = None
model.training_epoch_end = None
model.validation_step = model.validation_step_result_only_step_metrics
model.validation_step_end = None
model.validation_epoch_end = None
batches = 3
epochs = 3
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
row_log_interval=1,
limit_train_batches=batches,
limit_val_batches=batches,
weights_summary=None,
)
trainer.fit(model)
# make sure correct steps were called
assert model.validation_step_called
assert not model.validation_step_end_called
assert not model.validation_epoch_end_called
# no early stopping
assert len(trainer.dev_debugger.early_stopping_history) == 0
# make sure we logged the exact number of metrics
assert len(
trainer.dev_debugger.logged_metrics) == epochs * batches + (epochs)
assert len(
trainer.dev_debugger.pbar_added_metrics) == epochs * batches + (epochs)
# make sure we logged the correct epoch metrics
total_empty_epoch_metrics = 0
epoch = 0
for metric in trainer.dev_debugger.logged_metrics:
if 'epoch' in metric:
epoch += 1
if len(metric) > 2:
assert 'no_val_no_pbar' not in metric
assert 'val_step_pbar_acc' not in metric
assert metric[f'val_step_log_acc/epoch_{epoch}']
assert metric[f'val_step_log_pbar_acc/epoch_{epoch}']
else:
total_empty_epoch_metrics += 1
assert total_empty_epoch_metrics == 3
# make sure we logged the correct epoch pbar metrics
total_empty_epoch_metrics = 0
for metric in trainer.dev_debugger.pbar_added_metrics:
if 'epoch' in metric:
epoch += 1
if len(metric) > 2:
assert 'no_val_no_pbar' not in metric
assert 'val_step_log_acc' not in metric
assert metric['val_step_log_pbar_acc']
assert metric['val_step_pbar_acc']
else:
total_empty_epoch_metrics += 1
assert total_empty_epoch_metrics == 3
# only 1 checkpoint expected since values didn't change after that
assert len(trainer.dev_debugger.checkpoint_callback_history) == 1
# make sure the last known metric is correct
assert trainer.logger_connector.callback_metrics[
'val_checkpoint_on'] == 171
def test_training_step_epoch_end_result(tmpdir):
"""
Makes sure training_step and epoch_end can be used with Results (without batch_end)
"""
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_and_step
model.training_epoch_end = model.training_epoch_end_return_for_log_epoch_and_step
model.val_dataloader = None
batches = 3
epochs = 1
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
log_every_n_steps=1,
limit_train_batches=batches,
weights_summary=None,
)
trainer.fit(model)
assert len(trainer.logger_connector.callback_metrics) == 17
# make sure correct steps were called
assert model.training_step_called
assert not model.training_step_end_called
assert model.training_epoch_end_called
# make sure correct metrics were logged
logged_metrics = trainer.dev_debugger.logged_metrics
assert len(logged_metrics) == (epochs * batches) + epochs
last_logged = logged_metrics[-1]
assert last_logged['step_epoch_log_and_pbar_acc1_epoch'] == 210.0
assert last_logged['step_epoch_log_acc2_epoch'] == 336.0
assert last_logged['epoch_end_log_acc_epoch'] == 1212.0
assert last_logged['epoch_end_log_pbar_acc_epoch'] == 1214.0
assert 'epoch_end_pbar_acc' not in last_logged
# make sure pbar metrics are correct
logged_pbar = trainer.dev_debugger.pbar_added_metrics
assert len(logged_pbar) == (epochs * batches) + epochs
assert trainer.logger_connector.progress_bar_metrics[
'step_epoch_log_and_pbar_acc1_epoch'] == 210.0
assert trainer.logger_connector.progress_bar_metrics[
'step_epoch_log_and_pbar_acc1_step'] == 7.0
assert trainer.logger_connector.progress_bar_metrics[
'step_epoch_pbar_acc3_epoch'] == 504.0
assert trainer.logger_connector.progress_bar_metrics[
'epoch_end_pbar_acc_epoch'] == 1213.0
assert trainer.logger_connector.progress_bar_metrics[
'epoch_end_log_pbar_acc_epoch'] == 1214.0
assert 'epoch_end_log_acc' not in trainer.logger_connector.progress_bar_metrics
assert 'log_acc2' not in trainer.logger_connector.progress_bar_metrics
# make sure callback metrics didn't change
assert trainer.logger_connector.callback_metrics['checkpoint_on'] == 171
# -----------------------------------------
# make sure training outputs what is expected
# -----------------------------------------
for batch_idx, batch in enumerate(model.train_dataloader()):
break
out = trainer.train_loop.run_training_batch(batch, batch_idx, 0)
assert out.signal == 0
assert len(out.batch_log_metrics) == 4
train_step_out = out.training_step_output_for_epoch_end
assert len(train_step_out) == 1
train_step_out = train_step_out[0][0]
assert isinstance(train_step_out, TrainResult)
assert 'minimize' in train_step_out
assert 'step_epoch_log_and_pbar_acc1_step' in train_step_out
assert 'step_epoch_log_and_pbar_acc1_epoch' in train_step_out
assert 'step_epoch_log_acc2_step' in train_step_out
assert 'step_epoch_log_acc2_epoch' in train_step_out
# make sure the optimizer closure returns the correct things
opt_closure_result = trainer.train_loop.training_step_and_backward(
batch, batch_idx, 0, trainer.optimizers[0], trainer.hiddens)
assert opt_closure_result['loss'] == (42.0 * 3) + (15.0 * 3)
def test_training_step_result_log_step_and_epoch(tmpdir):
"""
Tests that only training_step can be used with TrainResult
Makes sure that things are routed to pbar, loggers and loss accordingly
Makes sure pbar and logs happen on epoch only when requested
"""
# enable internal debugging actions
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_and_step
model.training_step_end = None
model.training_epoch_end = None
model.val_dataloader = None
epochs = 3
batches = 2
trainer = Trainer(
default_root_dir=tmpdir,
limit_train_batches=batches,
limit_val_batches=batches,
log_every_n_steps=1,
max_epochs=epochs,
weights_summary=None,
)
trainer.fit(model)
# make sure correct steps were called
assert model.training_step_called
assert not model.training_step_end_called
assert not model.training_epoch_end_called
assert len(trainer.logger_connector.callback_metrics) == 11
# make sure correct metrics are logged (one per batch step as requested)
assert len(
trainer.dev_debugger.logged_metrics) == (epochs * batches) + epochs
epoch_metrics = trainer.dev_debugger.logged_metrics
epoch_idx = -1
for i_start in range(0, len(epoch_metrics), batches + 1):
epoch_idx += 1
epoch_outputs = epoch_metrics[i_start:i_start + batches + 1]
mean_vals = {
'epoch_step_epoch_log_and_pbar_acc1': [],
'epoch_step_epoch_log_acc2': []
}
# make sure each batch logged the expected value
for batch_idx in range(len(epoch_outputs) - 1):
logged_metrics = epoch_outputs[batch_idx]
expected_val_1 = (5 + batch_idx) * (epoch_idx + 1)
expected_val_2 = (6 + batch_idx) * (epoch_idx + 1)
mean_vals['epoch_step_epoch_log_and_pbar_acc1'].append(
torch.tensor(expected_val_1).float())
mean_vals['epoch_step_epoch_log_acc2'].append(
torch.tensor(expected_val_2).float())
assert logged_metrics[
'step_epoch_log_and_pbar_acc1_step'] == expected_val_1
assert logged_metrics['step_epoch_log_acc2_step'] == expected_val_2
assert 'step_epoch_pbar_acc3' not in logged_metrics
assert len(logged_metrics) == 6
# make sure the metrics for the epoch end are actual means (the default reduce fx) or all the batches
epoch_end_metrics = epoch_outputs[-1]
eval_1 = torch.stack(
mean_vals['epoch_step_epoch_log_and_pbar_acc1']).mean()
eval_2 = torch.stack(mean_vals['epoch_step_epoch_log_acc2']).mean()
assert epoch_end_metrics[
'step_epoch_log_and_pbar_acc1_epoch'] == eval_1
assert epoch_end_metrics['step_epoch_log_acc2_epoch'] == eval_2
assert 'step_epoch_pbar_acc3' not in epoch_end_metrics
assert len(logged_metrics) == 6
# make sure we are using the correct metrics for callbacks
assert trainer.logger_connector.callback_metrics['checkpoint_on'] == 171
# -------------------------------
# VERIFY PBAR METRICS
# -------------------------------
# make sure pbar metrics are correct ang log metrics did not leak
all_pbar_metrics = trainer.dev_debugger.pbar_added_metrics
assert len(all_pbar_metrics) == (epochs * batches) + epochs
epoch_idx = -1
for i_start in range(0, len(all_pbar_metrics), batches + 1):
epoch_idx += 1
epoch_outputs = all_pbar_metrics[i_start:i_start + batches + 1]
mean_vals = {
'epoch_step_epoch_log_and_pbar_acc1': [],
'epoch_step_epoch_pbar_acc3': []
}
# make sure each batch logged the expected value
for batch_idx in range(len(epoch_outputs) - 1):
logged_metrics = epoch_outputs[batch_idx]
expected_val_1 = (5 + batch_idx) * (epoch_idx + 1)
expected_val_2 = (7 + batch_idx) * (epoch_idx + 1)
mean_vals['epoch_step_epoch_log_and_pbar_acc1'].append(
torch.tensor(expected_val_1).float())
mean_vals['epoch_step_epoch_pbar_acc3'].append(
torch.tensor(expected_val_2).float())
assert logged_metrics[
'step_epoch_log_and_pbar_acc1_step'] == expected_val_1
assert logged_metrics[
'step_epoch_pbar_acc3_step'] == expected_val_2
assert 'epoch_log_acc2_step' not in logged_metrics
assert len(logged_metrics) == 5
# make sure the metrics for the epoch end are actual means (the default reduce fx) or all the batches
epoch_end_metrics = epoch_outputs[-1]
eval_1 = torch.stack(
mean_vals['epoch_step_epoch_log_and_pbar_acc1']).mean()
eval_2 = torch.stack(mean_vals['epoch_step_epoch_pbar_acc3']).mean()
assert epoch_end_metrics[
'step_epoch_log_and_pbar_acc1_epoch'] == eval_1
assert epoch_end_metrics['step_epoch_pbar_acc3_epoch'] == eval_2
assert 'epoch_log_acc2_step' not in epoch_end_metrics
assert len(logged_metrics) == 5
# -----------------------------------------
# make sure training outputs what is expected
# -----------------------------------------
for batch_idx, batch in enumerate(model.train_dataloader()):
break
out = trainer.train_loop.run_training_batch(batch, batch_idx, 0)
assert out.signal == 0
assert len(out.batch_log_metrics) == 4
train_step_out = out.training_step_output_for_epoch_end
assert len(train_step_out) == 1
train_step_out = train_step_out[0][0]
assert isinstance(train_step_out, TrainResult)
assert 'minimize' in train_step_out
assert 'step_epoch_log_and_pbar_acc1_step' in train_step_out
assert 'step_epoch_log_acc2_step' in train_step_out
assert 'step_epoch_log_and_pbar_acc1_epoch' in train_step_out
assert 'step_epoch_log_acc2_epoch' in train_step_out
# make sure the optimizer closure returns the correct things
opt_closure_result = trainer.train_loop.training_step_and_backward(
batch, batch_idx, 0, trainer.optimizers[0], trainer.hiddens)
assert opt_closure_result['loss'] == (42.0 * 3) + (15.0 * 3)
def test_val_step_epoch_step_metrics(tmpdir):
"""
Make sure the logged + pbar metrics are allocated accordingly at every step when requested
"""
# enable internal debugging actions
os.environ['PL_DEV_DEBUG'] = '1'
model = DeterministicModel()
model.training_step = model.training_step_result_log_epoch_and_step_for_callbacks
model.training_step_end = None
model.training_epoch_end = None
model.validation_step = model.validation_step_result_epoch_step_metrics
model.validation_step_end = None
model.validation_epoch_end = None
batches = 3
epochs = 3
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=epochs,
log_every_n_steps=1,
limit_train_batches=batches,
limit_val_batches=batches,
weights_summary=None,
)
trainer.fit(model)
assert len(trainer.logger_connector.callback_metrics) == 11
expected_metrics = {
'early_stop_on', 'checkpoint_on', 'val_step_pbar_acc',
'val_step_pbar_acc_epoch', 'val_step_log_acc',
'val_step_log_acc_epoch', 'val_step_log_pbar_acc',
'val_step_log_pbar_acc_epoch', 'val_step_batch_idx',
'val_step_batch_idx_epoch'
}
expected_metrics.add('debug_epoch')
seen_metrics = set(trainer.logger_connector.callback_metrics)
assert expected_metrics == seen_metrics
# make sure correct steps were called
assert model.validation_step_called
assert not model.validation_step_end_called
assert not model.validation_epoch_end_called
# no early stopping
assert len(trainer.dev_debugger.early_stopping_history) == 0
# make sure we logged the exact number of metrics
assert len(
trainer.dev_debugger.logged_metrics) == epochs * batches + (epochs)
assert len(
trainer.dev_debugger.pbar_added_metrics) == epochs * batches + (epochs)
# make sure we logged the correct epoch metrics
for metric_idx in range(0, len(trainer.dev_debugger.logged_metrics),
batches + 1):
batch_metrics = trainer.dev_debugger.logged_metrics[
metric_idx:metric_idx + batches]
epoch_metric = trainer.dev_debugger.logged_metrics[metric_idx +
batches]
epoch = epoch_metric['epoch']
# make sure the metric was split
for batch_metric in batch_metrics:
assert f'val_step_log_acc_step/epoch_{epoch}' in batch_metric
assert f'val_step_log_pbar_acc_step/epoch_{epoch}' in batch_metric
# make sure the epoch split was correct
assert 'val_step_log_acc_epoch' in epoch_metric
assert 'val_step_log_pbar_acc_epoch' in epoch_metric
# make sure we logged the correct pbar metrics
for metric_idx in range(0, len(trainer.dev_debugger.pbar_added_metrics),
batches + 1):
batch_metrics = trainer.dev_debugger.pbar_added_metrics[
metric_idx:metric_idx + batches]
epoch_metric = trainer.dev_debugger.pbar_added_metrics[metric_idx +
batches]
# make sure the metric was split
for batch_metric in batch_metrics:
assert 'val_step_pbar_acc_step' in batch_metric
assert 'val_step_log_pbar_acc_step' in batch_metric
# make sure the epoch split was correct
assert 'val_step_pbar_acc_epoch' in epoch_metric
assert 'val_step_log_pbar_acc_epoch' in epoch_metric
# only 1 checkpoint expected since values didn't change after that
assert len(trainer.dev_debugger.checkpoint_callback_history) == 1
# make sure the last known metric is correct
assert trainer.logger_connector.callback_metrics['checkpoint_on'] == 189
