def test_concat_scheduler_3_schedulers():
tensor = torch.zeros([1], requires_grad=True)
optimizer = torch.optim.SGD([tensor], lr=0)
scheduler_1 = LinearCyclicalScheduler(optimizer, "lr", start_value=1.0, end_value=0.5, cycle_size=20)
scheduler_2 = LinearCyclicalScheduler(optimizer, "lr", start_value=0.5, end_value=0.45, cycle_size=10)
scheduler_3 = LinearCyclicalScheduler(optimizer, "lr", start_value=0.5, end_value=0.0, cycle_size=20)
durations = [10, 5]
concat_scheduler = ConcatScheduler(
schedulers=[scheduler_1, scheduler_2, scheduler_3], durations=durations, save_history=True
)
state_dict = concat_scheduler.state_dict()
data = [0] * 10
max_epochs = 2
simulated_values = ConcatScheduler.simulate_values(
num_events=len(data) * max_epochs, schedulers=[scheduler_1, scheduler_2, scheduler_3], durations=durations
)
def save_lr(engine):
lrs.append(optimizer.param_groups[0]["lr"])
trainer = Engine(lambda engine, batch: None)
trainer.add_event_handler(Events.ITERATION_STARTED, concat_scheduler)
trainer.add_event_handler(Events.ITERATION_COMPLETED, save_lr)
for _ in range(2):
lrs = []
trainer.run(data, max_epochs=max_epochs)
assert lrs == list(
map(
pytest.approx,
[
# Cycle 1 of the first LinearCyclicalScheduler
1.0,
0.95,
0.9,
0.85,
0.8,
0.75,
0.7,
0.65,
0.6,
0.55,
# Cycle 1 of the second LinearCyclicalScheduler
0.5,
0.49,
0.48,
0.47,
0.46,
# Cycle 1 of the third LinearCyclicalScheduler
0.5,
0.45,
0.4,
0.35,
0.3,
],
)
)
state_lrs = trainer.state.param_history["lr"]
assert len(state_lrs) == len(lrs)
# Unpack singleton lists
assert [group[0] for group in state_lrs] == lrs
assert lrs == pytest.approx([v for i, v in simulated_values])
concat_scheduler.load_state_dict(state_dict)
trainer.state.param_history = None
def test_concat_scheduler_two_schedulers(duration_vals_as_np_int):
tensor = torch.zeros([1], requires_grad=True)
optimizer = torch.optim.SGD([tensor], lr=0)
scheduler_1 = LinearCyclicalScheduler(optimizer, "lr", start_value=1.0, end_value=0.0, cycle_size=10)
scheduler_2 = CosineAnnealingScheduler(optimizer, "lr", start_value=0.0, end_value=1.0, cycle_size=10)
durations = [10]
if duration_vals_as_np_int:
durations = [np.int64(t) for t in durations]
concat_scheduler = ConcatScheduler(schedulers=[scheduler_1, scheduler_2], durations=durations, save_history=True)
state_dict = concat_scheduler.state_dict()
data = [0] * 10
max_epochs = 2
simulated_values = ConcatScheduler.simulate_values(
num_events=len(data) * max_epochs, schedulers=[scheduler_1, scheduler_2], durations=durations
)
def save_lr(engine):
lrs.append(optimizer.param_groups[0]["lr"])
trainer = Engine(lambda engine, batch: None)
trainer.add_event_handler(Events.ITERATION_STARTED, concat_scheduler)
trainer.add_event_handler(Events.ITERATION_COMPLETED, save_lr)
for _ in range(2):
lrs = []
trainer.run(data, max_epochs=max_epochs)
assert lrs == list(
map(
pytest.approx,
[
# Cycle 1 of the LinearCyclicalScheduler
1.0,
0.8,
0.6,
0.4,
0.2,
0.0,
0.2,
0.4,
0.6,
0.8,
# Cycle 1 of the CosineAnnealingScheduler
0.0,
0.02447174185242318,
0.09549150281252627,
0.20610737385376332,
0.3454915028125263,
0.5,
0.6545084971874737,
0.7938926261462365,
0.9045084971874737,
0.9755282581475768,
],
)
)
state_lrs = trainer.state.param_history["lr"]
assert len(state_lrs) == len(lrs)
# Unpack singleton lists
assert [group[0] for group in state_lrs] == lrs
assert lrs == pytest.approx([v for i, v in simulated_values])
concat_scheduler.load_state_dict(state_dict)
trainer.state.param_history = None
def test_concat_scheduler_asserts():
tensor = torch.zeros([1], requires_grad=True)
optimizer = torch.optim.SGD([tensor], lr=0)
scheduler_1 = LinearCyclicalScheduler(optimizer, "lr", start_value=1.0, end_value=0.0, cycle_size=10)
scheduler_2 = CosineAnnealingScheduler(optimizer, "lr", start_value=0.0, end_value=1.0, cycle_size=10)
with pytest.raises(TypeError, match=r"Argument schedulers should be a sequence"):
ConcatScheduler(schedulers=None, durations=[])
with pytest.raises(ValueError, match=r"Argument schedulers should be of more than one parameter schedulers"):
ConcatScheduler(schedulers=[], durations=[])
with pytest.raises(ValueError, match=r"Argument schedulers should be of more than one parameter schedulers"):
ConcatScheduler(schedulers=[scheduler_1], durations=[10])
with pytest.raises(TypeError, match=r"Value at index 1 of schedulers should be a parameter scheduler"):
ConcatScheduler(schedulers=[scheduler_1, 12], durations=[10])
with pytest.raises(ValueError, match=r"Incorrect number schedulers or duration values"):
ConcatScheduler(schedulers=[scheduler_1, scheduler_2], durations=[10, 5])
with pytest.raises(ValueError, match=r"Argument durations should be list/tuple of integers"):
ConcatScheduler(schedulers=[scheduler_1, scheduler_2, scheduler_2], durations=[15, 12.0])
with pytest.raises(TypeError, match=r"Argument durations should be list/tuple"):
ConcatScheduler(schedulers=[scheduler_1, scheduler_2], durations="abc")
with pytest.raises(TypeError, match=r"Argument param_names should be list or tuple"):
ConcatScheduler.simulate_values(
num_events=123, schedulers=[scheduler_1, scheduler_2], durations=[15], param_names="abc"
)
with pytest.raises(ValueError, match=r"Argument param_names should be list or tuple of strings"):
ConcatScheduler.simulate_values(
num_events=123, schedulers=[scheduler_1, scheduler_2], durations=[15], param_names=[1]
)
optimizer_2 = torch.optim.SGD([tensor], lr=0)
scheduler_3 = CosineAnnealingScheduler(optimizer_2, "lr", start_value=0.0, end_value=1.0, cycle_size=10)
with pytest.raises(ValueError, match=r"schedulers should be related to same optimizer"):
ConcatScheduler([scheduler_1, scheduler_3], durations=[30])
scheduler_4 = CosineAnnealingScheduler(optimizer, "lr2", start_value=0.0, end_value=1.0, cycle_size=10)
with pytest.raises(ValueError, match=r"schedulers should be related to same param_name"):
ConcatScheduler([scheduler_1, scheduler_4], durations=[30])
with pytest.raises(ValueError, match=r"schedulers should be related to same optimizer"):
ConcatScheduler.simulate_values(3, [scheduler_1, scheduler_3], durations=[30])
def test_concat_scheduler_state_dict():
tensor = torch.zeros([1], requires_grad=True)
optimizer = torch.optim.SGD([tensor], lr=0)
scheduler_1 = LinearCyclicalScheduler(optimizer, "lr", start_value=1.0, end_value=0.0, cycle_size=10)
scheduler_2 = CosineAnnealingScheduler(optimizer, "lr", start_value=0.0, end_value=1.0, cycle_size=10)
durations = [10]
concat_scheduler = ConcatScheduler(schedulers=[scheduler_1, scheduler_2], durations=durations, save_history=False)
state_dict = concat_scheduler.state_dict()
assert state_dict["durations"] == durations
assert state_dict["_current_duration"] == durations[0]
assert state_dict["_scheduler_index"] == 0
for _ in range(20):
concat_scheduler(None, None)
concat_scheduler.load_state_dict(state_dict)
assert concat_scheduler.durations == durations
assert concat_scheduler._current_duration == durations[0]
assert id(concat_scheduler._current_scheduler) == id(scheduler_1)
with pytest.raises(ValueError, match=r"Required state attribute 'schedulers' is absent in provided state_dict"):
concat_scheduler.load_state_dict({"a": 1})
with pytest.raises(ValueError, match=r"Input state_dict contains 0 state_dicts of concatenated schedulers"):
concat_scheduler.load_state_dict({"schedulers": []})
with pytest.raises(TypeError, match=r"Argument state_dict should be a dictionary, but given"):
concat_scheduler.load_state_dict(None)
def test_create_lr_scheduler_with_warmup_on_combined_scheduler(save_history):
# Test with a complex scheduler
tensor = torch.ones([1], requires_grad=True)
optimizer = torch.optim.SGD([tensor], lr=0.001)
max_epochs = 25
lr_max_value = 0.4
num_iterations_per_epoch = 128
num_iterations = max_epochs * num_iterations_per_epoch
warmup_duration = 5 * num_iterations_per_epoch
cooldown_duration = 5 * num_iterations_per_epoch
scheduler_1 = LinearCyclicalScheduler(
optimizer,
"lr",
start_value=lr_max_value,
end_value=lr_max_value * 0.9,
cycle_size=(num_iterations - warmup_duration - cooldown_duration) * 2,
)
scheduler_2 = LinearCyclicalScheduler(
optimizer, "lr", start_value=lr_max_value, end_value=0.0, cycle_size=cooldown_duration * 2
)
lr_scheduler = ConcatScheduler(
schedulers=[scheduler_1, scheduler_2],
durations=[num_iterations - warmup_duration - cooldown_duration],
save_history=False,
)
lr_values = [None] * num_iterations
scheduler = create_lr_scheduler_with_warmup(
lr_scheduler,
warmup_start_value=0.0,
warmup_end_value=lr_max_value,
warmup_duration=warmup_duration,
save_history=save_history,
output_simulated_values=lr_values,
)
state_dict = scheduler.state_dict()
trainer = Engine(lambda engine, batch: None)
@trainer.on(Events.ITERATION_COMPLETED)
def save_lr(engine):
lrs.append(optimizer.param_groups[0]["lr"])
trainer.add_event_handler(Events.ITERATION_STARTED, scheduler)
data = [0] * num_iterations_per_epoch
for _ in range(2):
lrs = []
trainer.run(data, max_epochs=max_epochs)
assert lrs == pytest.approx([v for i, v in lr_values])
if save_history:
param_history = trainer.state.param_history["lr"]
assert lrs == pytest.approx([v[0] for v in param_history])
trainer.state.param_history = None
scheduler.load_state_dict(state_dict)