[Bug] Model's state_dict not loading properly when loading a ckpt

[inafergra]

Right now, when loading model checkpoints (QuantumModel or QNN) the state_dict of the model is not loaded properly. The issue is in the load_model() function, which treats the QuantumModel.from_dict() classmethod as an in-place operation

qadence/qadence/ml_tools/saveload.py

Lines 114 to 116 in 70e12d4

	iteration, model_dict = torch.load(folder / model_ckpt_name, *args, **kwargs)
	if isinstance(model, (QuantumModel, QNN)):
	model._from_dict(model_dict, as_torch=True)

when in reality .from_dict() is a class method and does not act as an in-place operation, as it returns a new class instance with the parameters taken from the input model state_dict:

qadence/qadence/model.py

Lines 276 to 308 in 70e12d4

	@classmethod
	def _from_dict(cls, d: dict, as_torch: bool = False) -> QuantumModel:
	from qadence.serialization import deserialize
	qm: QuantumModel
	try:
	qm_dict = d[cls.__name__]
	qm = cls(
	circuit=QuantumCircuit._from_dict(qm_dict["circuit"]),
	observable=(
	None
	if not isinstance(qm_dict["observable"], list)
	else [deserialize(q_obs) for q_obs in qm_dict["observable"]] # type: ignore[misc]
	),
	backend=qm_dict["backend"],
	diff_mode=qm_dict["diff_mode"],
	measurement=Measurements._from_dict(qm_dict["measurement"]),
	noise=Noise._from_dict(qm_dict["noise"]),
	configuration=config_factory(qm_dict["backend"], qm_dict["backend_configuration"]),
	)
	if as_torch:
	conv_pd = torch.nn.ParameterDict()
	param_dict = d["param_dict"]
	for n, param in param_dict.items():
	conv_pd[n] = torch.nn.Parameter(param)
	qm._params = conv_pd
	logger.debug(f"Initialized {cls.__name__} from {d}.")
	except Exception as e:
	logger.warning(f"Unable to deserialize object {d} to {cls.__name__} due to {e}.")
	return qm

The most straightforward solution would be to do model = model.from_dict() when loading the state_dict. However, this doesn't work since then, when using the train_grad() loop, the parameters passed to the optimizer before the training loop do not correspond to the newly created instance of the model, and therefore the parameters won't be updated in optimizer.step().

One solution would be to make .from_dict() an instance method instead of a class method so that it acts as an in-place operation, by mutating the instance attributes instead of creating a new instance. This might cause some additional issues (@dominikandreasseitz mentioned the uuid assignments might cause problems?)

[inafergra]

Tagging @dominikandreasseitz @jpmoutinho @Roland-djee

[Roland-djee]

Tagging @dominikandreasseitz @jpmoutinho @Roland-djee

Thanks for this @inafergra. I guess @dominikandreasseitz or @chMoussa are the best placed to have a look ?

[dominikandreasseitz]

Tagging @dominikandreasseitz @jpmoutinho @Roland-djee

Thanks for this @inafergra. I guess @dominikandreasseitz or @chMoussa are the best placed to have a look ?

thanks @inafergra. @Roland-djee , since @inafergra already knows how to solve it he could coordinate with @chMoussa on fixing this. i can ofc also do it. Let me know

[Roland-djee]

Tagging @dominikandreasseitz @jpmoutinho @Roland-djee

Thanks for this @inafergra. I guess @dominikandreasseitz or @chMoussa are the best placed to have a look ?

thanks @inafergra. @Roland-djee , since @inafergra already knows how to solve it he could coordinate with @chMoussa on fixing this. i can ofc also do it. Let me know

If that is fine with them sure.

[Roland-djee]

@inafergra @chMoussa Would you be able to coordinate on this ?