def test_create_supervised_tbptt_trainer_callcounts(mock_detach_hidden):
# Mocking objects
model = mock.MagicMock()
# Necessary to unpack output
model.return_value = (1, 1)
optimizer = mock.MagicMock()
loss = mock.MagicMock()
trainer = create_supervised_tbptt_trainer(model, optimizer, loss, tbtt_step=2)
# Adding two mock handles to the trainer to monitor that TBPTT events are
# called correctly
handle_started = mock.MagicMock()
trainer.add_event_handler(Tbptt_Events.TIME_ITERATION_STARTED, handle_started)
handle_completed = mock.MagicMock()
trainer.add_event_handler(Tbptt_Events.TIME_ITERATION_COMPLETED, handle_completed)
# Fake data
X = torch.ones(6, 2, 1)
y = torch.ones(6, 2, 1)
data = [(X, y)]
# Running trainer
trainer.run(data)
# Verifications
assert handle_started.call_count == 3
assert handle_completed.call_count == 3
assert mock_detach_hidden.call_count == 2
assert model.call_count == 3
assert loss.call_count == 3
assert optimizer.zero_grad.call_count == 3
assert optimizer.step.call_count == 3
n_args_tuple = tuple(len(args) for args, kwargs in model.call_args_list)
assert n_args_tuple == (1, 2, 2)
def _test_create_supervised_tbptt_trainer(device):
# Defining dummy recurrent model with zero weights
model = nn.RNN(1, 1, bias=False)
for p in model.parameters():
p.data.zero_()
# Set some mock on forward to monitor
forward_mock = mock.MagicMock()
forward_mock.return_value = None
model.register_forward_hook(forward_mock)
# Defning optimizer and trainer
optimizer = optim.SGD(model.parameters(), 1)
trainer = create_supervised_tbptt_trainer(
model,
optimizer,
F.mse_loss,
tbtt_step=2,
device=device
)
# Fake data
X = torch.ones(6, 2, 1)
y = torch.ones(6, 2, 1)
data = [(X, y)]
# Running trainer
trainer.run(data)
# If tbptt is not use (one gradient update), the hidden to hidden weight
# should stay zero
assert not model.weight_hh_l0.item() == pytest.approx(0)
# Cheking forward calls
assert forward_mock.call_count == 3
for i in range(3):
inputs = forward_mock.call_args_list[i][0][1]
if i == 0:
assert len(inputs) == 1
else:
assert len(inputs) == 2
x, h = inputs
assert h.is_leaf
def _test_create_supervised_tbptt_trainer(device):
# Defining dummy recurrent model with zero weights
model = nn.RNN(1, 1, bias=False)
for p in model.parameters():
p.data.zero_()
# Defning optimizer and trainer
optimizer = optim.SGD(model.parameters(), 1)
trainer = create_supervised_tbptt_trainer(model,
optimizer,
F.mse_loss,
tbtt_step=2,
device=device)
# Adding two mock handles to the trainer to monitor that TBPTT events are
# called correctly
handle_started = MagicMock()
trainer.add_event_handler(Tbptt_Events.TIME_ITERATION_STARTED,
handle_started)
handle_completed = MagicMock()
trainer.add_event_handler(Tbptt_Events.TIME_ITERATION_COMPLETED,
handle_completed)
# Fake data
X = torch.ones(6, 2, 1)
y = X = torch.ones(6, 2, 1)
data = [(X, y)]
# Running trainer
trainer.run(data)
# Verifications
assert handle_started.call_count == 3
assert handle_completed.call_count == 3
# If tbptt is not use (one gradient update), the hidden to hidden weight
# should stay zero
assert not model.weight_hh_l0.item() == pytest.approx(0)
num_val_batches = len(valdl)
num_val_examples = num_val_batches * BATCH_SIZE
print("Num train examples: {} ({} batches)".format(num_train_examples,
num_train_batches))
print("Num validation examples: {} ({} batches)".format(
num_val_examples, num_val_batches))
print("\nStarting training for {} epochs...\n".format(NUM_EPOCHS))
print(type(traindl))
# create ignite trainer
trainer = create_supervised_tbptt_trainer(model,
optimizer,
loss_function,
tbtt_step=TBTT_STEP)
evaluator = create_supervised_tbptt_evaluator(
model,
metrics={
'accuracy': Accuracy(),
'nll': Loss(loss_function),
'precision':
Precision(output_transform=thresholded_output_transform),
'recall': Recall(output_transform=thresholded_output_transform)
})
# evaluator = create_supervised_evaluator(model, metrics=['accuracy'])
@trainer.on(Events.ITERATION_COMPLETED)