def test_optimizer_state_change(self, data_batch, model,
optimizer: OptimizerAdapter):
optimizer.register_model_params(
model_params=dict(model.named_parameters()))
optimizer_state_0 = deepcopy(optimizer.state_dict())
x, y = data_batch
loss_fn = torch.nn.MSELoss(reduction='sum')
optimizer.zero_grad()
output = model(x)
loss = loss_fn(output, y)
loss.backward()
optimizer.step()
optimizer_state_1 = deepcopy(optimizer.state_dict())
optimizer.zero_grad()
output = model(x)
loss = loss_fn(output, y)
loss.backward()
optimizer.step()
optimizer_state_2 = deepcopy(optimizer.state_dict())
assert optimizer_state_0 != optimizer_state_1
assert (optimizer_state_1["state"][0]["momentum_buffer"] !=
optimizer_state_2["state"][0]["momentum_buffer"]).all()
def test_register_model_params_without_restore_state(
self, optimizer: OptimizerAdapter, model, false_restore_state):
# If not restore_state,
# reinitialize the optimizer._optimizer with given model_params,
# no matter the optimizer._optimizer is None or not
optimizer.register_model_params(model_params=dict(
model.named_parameters()),
restore_state=false_restore_state)
assert len(optimizer._optimizer.param_groups) > 0
def optimizer_state_dict(self, model, data_batch) -> Dict[str, Any]:
# To have a state dict
optimizer = OptimizerAdapter(SGD, {"lr": 1.0, "momentum": 0.9})
optimizer.register_model_params(
model_params=dict(model.named_parameters()))
x, y = data_batch
loss_fn = torch.nn.MSELoss(reduction='sum')
optimizer.zero_grad()
output = model(x)
loss = loss_fn(output, y)
loss.backward()
optimizer.step()
optimizer_state_1 = deepcopy(optimizer.state_dict())
return optimizer_state_1
def test_load_state_dict(self, optimizer: OptimizerAdapter, model,
optimizer_state_dict):
# test when the optimizer._optimizer is None and self._state_dict is not None
# Firstly load state dict for optimizer, to make optimizer._state_dict not None
assert optimizer._optimizer is None
optimizer.load_state_dict(state_dict=optimizer_state_dict)
assert optimizer.state_dict() == optimizer_state_dict
# initialize a optimizer._optimizer,
# it would load the state dict from optimizer._state_dict for optimizer._optimizer
assert optimizer._optimizer is None and optimizer._state_dict is not None
optimizer.register_model_params(
model_params=dict(model.named_parameters()))
# check if optimizer._optimizer.state_dict() same with the optimizer_state_dict
assert (
optimizer._optimizer.state_dict()["state"][0]["momentum_buffer"] ==
optimizer.state_dict()["state"][0]["momentum_buffer"]).all()
assert (optimizer.state_dict()["state"][0]["momentum_buffer"] ==
optimizer_state_dict["state"][0]["momentum_buffer"]).all()
assert optimizer._state_dict is None
def test_optimizer_state_recovery(self, data_batch, model,
optimizer: OptimizerAdapter):
optimizer.register_model_params(
model_params=dict(model.named_parameters()))
x, y = data_batch
loss_fn = torch.nn.MSELoss(reduction='sum')
optimizer.zero_grad()
output = model(x)
loss = loss_fn(output, y)
loss.backward()
optimizer.step()
optimizer_state = deepcopy(optimizer.state_dict())
optimizer.register_model_params(model_params=dict(
model.named_parameters()),
restore_state=False)
assert optimizer.state_dict() != optimizer_state
optimizer.load_state_dict(optimizer_state)
assert (optimizer.state_dict()["state"][0]["momentum_buffer"] ==
optimizer_state["state"][0]["momentum_buffer"]).all()
def test_register_model_params(self, optimizer: OptimizerAdapter, model):
assert optimizer._optimizer is None
optimizer.register_model_params(
model_params=dict(model.named_parameters()))
assert len(optimizer._optimizer.param_groups) > 0