[BUG] NeuMF - PyTorch backend ignores num_factors, causing assertion failure

[Gabriel-Kissin]

Description

When using the PyTorch backend to run the NeuMF model, Cornac ignores the num_factors argument provided during model initialisation.
As a result, the PyTorch NeuMF model always receives the default num_factors=8, even when the user explicitly sets a different value.

This leads to an assertion failure in backend_pt.py#L140 on assert layers[-1] == num_factors. As layers[-1] is the user-provided layers, but num_factors is from the default 8, not the value given by the user.

How do we replicate the issue?

model = cornac.models.NeuMF(
    num_factors=11,
    layers=(64, 16, 8, 11),
    backend="pytorch",
)
model._build_model_pt()    # AssertionError 

Cause of the problem

Looks to me that this is in (1) recom_neumf.py#L263

        model = NeuMF(
            num_users=self.num_users,
            num_items=self.num_items,
            layers=self.layers,
            act_fn=self.act_fn,
        )

which instantiates the class in (2) backend_pt.py#L123

class NeuMF(nn.Module):
    def __init__(
        self,
        num_users: int,
        num_items: int,
        num_factors: int = 8,
        layers=(64, 32, 16, 8),
        act_fn="relu",
    ):

Note that the class in (2) accepts param num_factors with default of 8; however this param is not passed at initialisation in (1).

Proposed fix

        model = NeuMF(
            num_users=self.num_users,
            num_items=self.num_items,
            layers=self.layers,
            num_factors=self.num_factors,  # add this!
            act_fn=self.act_fn,
        )

My other suggestion would be to have a more helpful error message which displays what layers and num_factors are

Workaround

Use tensorflow backend - this correctly uses num_factors here.

model = cornac.models.NeuMF(
    num_factors=11,
    layers=(64, 16, 8, 11),
    backend="tensorflow",
)
model._build_model_tf()

should work though I've not tested it.

Note: setting num_factors=None doesn't help with pytorch backend, as the None doesn't get passed either.

Thanks for the great package!