Use a Given PyTorch Data Loader

[RoyiAvital]

Is there a way to use a given PyTorch data loader instead of just a data set?

[BenjaminBossan]

You cannot pass a DataLoader instance to fit.

You can change the class of the DataLoader by passing it to the iterator_train and iterator_valid argument.

If you want to have control over the actual DataLoader instance, you'd have to subclass the net and override the get_iterator method. You can check the code, it should be quite straightforward to make it return something custom built.

[ottonemo]

Then all data in dsTrain will be used for training and all data in dsValid will be used for validation, right?

Correct!

Or should I set train_split to None as well?

That would deactivate the validation split completely (no validation run/metrics).

By the way, any way to set netCls.train_split.stratified = False in the NeuralNetClassifier() construction?

In a predefined split there is no stratification (splitting is already done).
The default value for the train_split parameter is skorch.dataset.ValidSplit(5) which, depending on the type of network you use, has stratification enabled or disabled (classification uses stratification, regression doesn't).

So, in essence, if you use the ValidSplit class for the train_split parameter, then you can deactivate stratification by passing stratified=False to ValidSplit but if you use a predefined split, there's nothing to deactivate because you already did the split.

[RoyiAvital]

@ottonemo , I see. So in general (Default split), to have the effect of netCls.train_split.stratified = False in the construction of the classifier I can do something like:

netCls = NeuralNetClassifier(
    modelCls,
    ValidSplit__stratified = False,
)

[githubnemo]

Almost :)

netCls = NeuralNetClassifier(
    modelCls,
    train_split__stratified=False,
)

[RoyiAvital]

Strange, I think I tried it:

I tried train_split__stratified = False, yet it didn't work.

But maybe I tried it in the case you wrote above, with the pre defined split.

[githubnemo]

Sorry, you are absolutely right. I thought we passed these arguments to the split as well but I remembered wrongly. It also doesn't really make sense since the VaildSplit is passed as an instance and not as a class (in contrast to, e .g., the module), so we can directly set the parameters we want.

from skorch.dataset import ValidSplit

netCls = NeuralNetClassifier(
    modelCls,
    train_split=ValidSplit(5, stratified=False),
)

[ottonemo]

To add to @BenjaminBossan's answer:

If you want to have control over the actual DataLoader instance, you'd have to subclass the net and override the get_iterator method. You can check the code, it should be quite straightforward to make it return something custom built.

While true you can also pass a generator function to the net instance that yields from an existing dataloader or constructs one on the fly.
For example:

def train_iterator(dataset, **kwargs):
    loader = DataLoader(dataset, **kwargs)

    for X, y in loader:
        yield X, y

net = NeuralNet(
    ...,
    iterator_train=train_iterator,
    ....,
)