Add feature extractor for torch

[paulnovello]

Translate KerasFeatureExtractor into TorchFeatureExtractor. It is unnecessary to translate the methods gradient_pred, gradient_index, and gradient_full for now because they are likely to be deprecated soon.

The class should:

• Include the same remaining methods as KerasFeatureExtractor (an abstraction BaseFeatureExtractor will probably be created later but it is outside the scope of this MR)
• handle pytorch data loaders (equivalent of tf.data.Datasets).

[Matt3164]

For the interface, I have no use of

output_activation: str = None
flatten: bool = False
batch_size: int = 256

Can I remove them? Or should I keep them unused with default values?

[Matt3164]

For the output type of the predict method, is it better to output the default torch.Tensor or output a np.ndarray which could be more easy to use with methods implemented is Tensorflow?

[paulnovello]

I seems like these parameters are used:

• output_activation defines the output activation to apply to the output features (in use when you want to gather the output logits or probabilities)
• flatten toggles if the output features are flattened or not (e.g. when the output features are those of a pooling layer)
• batch_size is used in the predict method

For the predict method, it is indeed better to output a np.ndarray

[Matt3164]

Just a simple check, the output of feature extractor is the output of the given output layer (specified with output_layers_id) and not the input right.

I am not familiar with the default behavior of keras.

[paulnovello]

Yes. Notice that the output of the whole network is also returned self.output_layers.append(model.output) l.86 of feature_extractor.py.