Add CrossEntropyLoss to classifier

[amine759]

I had to use CrossEntropyLoss for my use case, I thought of creating this PR since NLLLoss is the only loss function supported :).

[BenjaminBossan]

Thanks for proposing the PR. A few points:

• We don't have type annotations in PEFT. Even if we did, we would support more than those two criteria.
• CrossEntropyLoss should already work with NeuralNetClassifier.
• The actual addition of your PR seems to be converting one-hot encoded targets to label-encoded targets. I don't think that this is specifically related to CrossEntropyLoss, is it?
• Also, AFAIK, sklearn does not generally accept one-hot encoded targets for classification, so I don't think skorch should either.

[amine759]

@BenjaminBossan Hi, Thanks for your reply,
As far as I understood it seems we're restricted to using only the NLLLoss loss function with skorch.NeuralNetClassifier. When attempting to use any other loss function, like CrossEntropyLoss, we encounter an error indicating that the criterion parameter only accepts NLLLoss.
Whereas the optimizer we can use any other than the default SGD. I assumed this is because get_loss does not handle the instance of a torch.nn.CrossEntropyLoss and since skorch don't use type annotation, I see now it wasn't necessary in the first place. I could have just changed get_loss and that's it, right?

[BenjaminBossan]

When attempting to use any other loss function, like CrossEntropyLoss, we encounter an error indicating that the criterion parameter only accepts NLLLoss.

Could you please provide an example to reproduce this error? CE should absolutely work in skorch as is, e.g.:

import numpy as np
from sklearn.datasets import make_classification
from torch import nn

from skorch import NeuralNetClassifier


X, y = make_classification(1000, 20, n_informative=10, random_state=0)
X = X.astype(np.float32)
y = y.astype(np.int64)

class MyModule(nn.Module):
    def __init__(self, num_units=10, nonlin=nn.ReLU()):
        super().__init__()

        self.dense0 = nn.Linear(20, num_units)
        self.nonlin = nonlin
        self.dropout = nn.Dropout(0.5)
        self.dense1 = nn.Linear(num_units, num_units)
        self.output = nn.Linear(num_units, 2)

    def forward(self, X, **kwargs):
        X = self.nonlin(self.dense0(X))
        X = self.dropout(X)
        X = self.nonlin(self.dense1(X))
        X = self.output(X)
        return X


net = NeuralNetClassifier(
    MyModule,
    max_epochs=10,
    criterion=nn.CrossEntropyLoss(),
    lr=0.1,
    # Shuffle training data on each epoch
    iterator_train__shuffle=True,
)

net.fit(X, y)
y_proba = net.predict_proba(X)