FAQ

How do I apply L2 regularization?

To apply L2 regularization (aka weight decay), PyTorch supplies the weight_decay parameter, which must be supplied to the optimizer. To pass this variable in skorch, use the double-underscore notation for the optimizer:

net = NeuralNet(
    ...,
    optimizer__weight_decay=0.01,
)

How can I continue training my model?

By default, when you call fit() more than once, the training starts from zero instead of from where it was left. This is in line with sklearn’s behavior but not always desired. If you would like to continue training, use partial_fit() instead of fit(). Alternatively, there is the warm_start argument, which is False by default. Set it to True instead and you should be fine.

How do I shuffle my train batches?

skorch uses DataLoader from PyTorch under the hood. This class takes a couple of arguments, for instance shuffle. We therefore need to pass the shuffle argument to DataLoader, which we achieve by using the double-underscore notation (as known from sklearn):

net = NeuralNet(
    ...,
    iterator_train__shuffle=True,
)

Note that we have an iterator_train for the training data and an iterator_valid for validation and test data. In general, you only want to shuffle the train data, which is what the code above does.

How do I use sklearn GridSeachCV when my data is in a dictionary?

skorch supports dicts as input but sklearn doesn’t. To get around that, try to wrap your dictionary into a SliceDict. This is a data container that partly behaves like a dict, partly like an ndarray. For more details on how to do this, have a look at the coresponding data section in the notebook.

How do I use sklearn GridSeachCV when my data is in a dataset?

skorch supports datasets as input but sklearn doesn’t. If it’s possible, you should provide your data in a non-dataset format, e.g. as a numpy array or torch tensor, extracted from your original dataset.

Sometimes, this is not possible, e.g. when your data doesn’t fit into memory. To get around that, try to wrap your dataset into a SliceDataset. This is a data container that partly behaves like a dataset, partly like an ndarray. Further information can be found here: SliceDataset.

I want to use sample_weight, how can I do this?

Some scikit-learn models support to pass a sample_weight argument to fit calls as part of the fit_params. This allows you to give different samples different weights in the final loss calculation.

In general, skorch supports fit_params, but unfortunately just calling net.fit(X, y, sample_weight=sample_weight) is not enough, because the fit_params are not split into train and valid, and are not batched, resulting in a mismatch with the training batches.

Fortunately, skorch supports passing dictionaries as arguments, which are actually split into train and valid and then batched. Therefore, the best solution is to pass the sample_weight with X as a dictionary. Below, there is example code on how to achieve this:

X, y = get_data()
# put your X into a dict if not already a dict
X = {'data': X}
# add sample_weight to the X dict
X['sample_weight'] = sample_weight

class MyModule(nn.Module):
    ...
    def forward(self, data, sample_weight):
        # when X is a dict, its keys are passed as kwargs to forward, thus
        # our forward has to have the arguments 'data' and 'sample_weight';
        # usually, sample_weight can be ignored here
        ...

class MyNet(NeuralNet):
    def __init__(self, *args, criterion__reduce=False, **kwargs):
        # make sure to set reduce=False in your criterion, since we need the loss
        # for each sample so that it can be weighted
        super().__init__(*args, criterion__reduce=criterion__reduce, **kwargs)

    def get_loss(self, y_pred, y_true, X, *args, **kwargs):
        # override get_loss to use the sample_weight from X
        loss_unreduced = super().get_loss(y_pred, y_true, X, *args, **kwargs)
        sample_weight = skorch.utils.to_tensor(X['sample_weight'], device=self.device)
        loss_reduced = (sample_weight * loss_unreduced).mean()
        return loss_reduced

net = MyNet(MyModule, ...)
net.fit(X, y)

I already split my data into training and validation sets, how can I use them?

If you have predefined training and validation datasets that are subclasses of PyTorch Dataset, you can use predefined_split() to wrap your validation dataset and pass it to NeuralNet’s train_split parameter:

from skorch.helper import predefined_split

net = NeuralNet(
    ...,
    train_split=predefined_split(valid_ds)
)
net.fit(train_ds)

If you split your data by using train_test_split(), you can create your own skorch Dataset, and then pass it to predefined_split():

from sklearn.model_selection import train_test_split
from skorch.helper import predefined_split
from skorch.dataset import Dataset

X_train, X_test, y_train, y_test = train_test_split(X, y)

valid_ds = Dataset(X_test, y_test)

net = NeuralNet(
    ...,
    train_split=predefined_split(valid_ds)
)

net.fit(X_train, y_train)

What happens when NeuralNet is passed an initialized Pytorch module?

When NeuralNet is passed an initialized Pytorch module, skorch will usually leave the module alone. In the following example, the resulting module will be trained for 20 epochs:

class MyModule(nn.Module):
    def __init__(self, hidden=10):
        ...

module = MyModule()
net1 = NeuralNet(module, max_epochs=10, ...)
net1.fit(X, y)

net2 = NeuralNet(module, max_epochs=10, ...)
net2.fit(X, y)

When the module is passed to the second NeuralNet, it will not be re-initialized and will keep its parameters from the first 10 epochs.

When the module parameters are set through keywords arguments, NeuralNet will re-initialized the module:

net = NeuralNet(module, module__hidden=10, ...)
net.fit(X, y)

Although it is possible to pass an initialized Pytorch module to NeuralNet, it is recommended to pass the module class instead:

net = NeuralNet(MyModule, ...)
net.fit(X, y)

In this case, fit() will always re-initialize the model and partial_fit() won’t after the network is initialized once.

How do I use a PyTorch Dataset with skorch?

skorch supports PyTorch’s Dataset as arguments to fit() or partial_fit(). We create a dataset by subclassing PyTorch’s Dataset:

import torch.utils.data

class RandomDataset(torch.utils.data.Dataset):
    def __init__(self):
        self.X = torch.randn(128, 10)
        self.Y = torch.randn(128, 10)

    def __getitem__(self, idx):
        return self.X[idx], self.Y[idx]

    def __len__(self):
        return 128

skorch expects the output of __getitem__ to be a tuple of two values. The RandomDataset can be passed directly to fit():

from skorch import NeuralNet
import torch.nn as nn

train_ds = RandomDataset()

class MyModule(nn.Module):
    def __init__(self):
        super().__init__()
        self.layer = torch.nn.Linear(10, 10)

    def forward(self, X):
        return self.layer(X)

net = NeuralNet(MyModule, criterion=torch.nn.MSELoss)
net.fit(train_ds)

How can I deal with multiple return values from forward?

skorch supports modules that return multiple values. To do this, simply return a tuple of all values that you want to return from the forward method. However, this tuple will also be passed to the criterion. If the criterion cannot deal with multiple values, this will result in an error.

To remedy this, you need to either implement your own criterion that can deal with the output or you need to override get_loss() and handle the unpacking of the tuple.

To inspect all output values, you can use either the forward() method (eager) or the forward_iter() method (lazy).

For an example of how this works, have a look at this notebook.

How can I perform gradient accumulation with skorch?

There is no direct option to turn on gradient accumulation (at least for now). However, with a few modifications, you can implement gradient accumulation yourself:

ACC_STEPS = 2  # number of steps to accumulate before updating weights

class GradAccNet(NeuralNetClassifier):
    """Net that accumulates gradients"""
    def __init__(self, *args, acc_steps=ACC_STEPS, **kwargs):
        super().__init__(*args, **kwargs)
        self.acc_steps = acc_steps

    def get_loss(self, *args, **kwargs):
        loss = super().get_loss(*args, **kwargs)
        return loss / self.acc_steps  # normalize loss

    def train_step(self, batch, **fit_params):
        """Perform gradient accumulation

        Only optimize every nth batch.

        """
        # note that n_train_batches starts at 1 for each epoch
        n_train_batches = len(self.history[-1, 'batches'])
        step = self.train_step_single(batch, **fit_params)

        if n_train_batches % self.acc_steps == 0:
            self.optimizer_.step()
            self.optimizer_.zero_grad()
        return step

This is not a complete recipe. For example, if you optimize every 2nd step, and the number of training batches is uneven, you should make sure that there is an optimization step after the last batch of each epoch. However, this example can serve as a starting point to implement your own version gradient accumulation.

Alternatively, make use of skorch’s accelerate integration provided by AccelerateMixin and use the gradient accumulation feature from that library.

How can I dynamically set the input size of the PyTorch module based on the data?

Typically, it’s up to the user to determine the shape of the input data when defining the PyTorch module. This can sometimes be inconvenient, e.g. when the shape is only known at runtime. E.g., when using sklearn.feature_selection.VarianceThreshold, you cannot know the number of features in advance. The best solution would be to set the input size dynamically.

In most circumstances, this can be achieved with a few lines of code in skorch. Here is an example:

class InputShapeSetter(skorch.callbacks.Callback):
    def on_train_begin(self, net, X, y):
        net.set_params(module__input_dim=X.shape[1])


net = skorch.NeuralNetClassifier(
    ClassifierModule,
    callbacks=[InputShapeSetter()],
)

This assumes that your module accepts an argument called input_units, which determines the number of units of the input layer, and that the number of features can be determined by X.shape[1]. If those assumptions are not true for your case, adjust the code accordingly. A fully working example can be found on stackoverflow.

How do I implement a score method on the net that returns the loss?

Sometimes, it is useful to be able to compute the loss of a net from within skorch (e.g. when a net is part of an sklearn pipeline). The function skorch.scoring.loss_scoring() achieves this. Two examples are provided below. The first demonstrates how to use skorch.scoring.loss_scoring() as a function on a trained net object.

from skorch.scoring import loss_scoring

X = np.random.randn(250, 25).astype('float32')
y = (X.dot(np.ones(25)) > 0).astype(int)

module = nn.Sequential(
    nn.Linear(25, 25),
    nn.ReLU(),
    nn.Linear(25, 2),
    nn.Softmax(dim=1)
)
net = skorch.NeuralNetClassifier(module).fit(X, y)
print(loss_scoring(net, X, y))

The second example shows how to sub-class skorch.classifier.NeuralNetClassifier to implement a score method. In this example, the score method returns the negative of the loss value, because we want sklearn.model_selection.GridSearchCV to return the run with least loss and sklearn.model_selection.GridSearchCV searches for the run with the greatest score.

class ScoredNet(skorch.NeuralNetClassifier):
    def score(self, X, y=None):
        loss_value = loss_scoring(self, X, y)
        return -loss_value

net = ScoredNet(module)
grid_searcher = GridSearchCV(
    net, {'lr': [1e-2, 1e-3], 'batch_size': [8, 16]},
)
grid_searcher.fit(X, y)
best_net = grid_searcher.best_estimator_
print(best_net.score(X, y))

How can I set the random seed of my model?

skorch does not provide an unified way for setting the seed of your model. You can utilize the numpy and torch seeding interfaces to set random seeds before model initialization and will get consistent results if you do not employ other libraries that introduce randomness.

Here’s an example:

seed = 42
numpy.random.seed(seed)
torch.manual_seed(seed)

net = NeuralNet(
    module=...
)
net.fit(X, y)

Note: torch.manual_seed calls torch.cuda.manual_seed if the current process employs CUDA, so you don’t have to worry about seeding CUDA RNG explicitly.

There are cases where you want a fixed train/validation split (by default the split is not seeded). This can be done by passing the random_state parameter to ValidSplit:

from skorch.dataset import ValidSplit

seed = 42
net = NeuralNet(
    module=...,
    train_split=ValidSplit(random_state=seed),
)
net.fit(X, y)

Note that there are other places where randomness is introduced that skorch does not control, such as the torch DataLoader when setting shuffle=True. See the corresponding documentation on how to fix the random seed in this case.

Migration guide

Migration from 0.10 to 0.11

With skorch 0.11, we pushed the tuple unpacking of values returned by the iterator to methods lower down the call chain. This way, it is much easier to work with iterators that don’t return exactly two values, as per the convention.

A consequence of this is a change in signature of these methods:

Instead of receiving the unpacked tuple of X and y, they just receive a batch, which is whatever is returned by the iterator. The tuple unpacking needs to be performed inside these methods.

If you have customized any of these methods, it is easy to retrieve the previous behavior. E.g. if you wrote your own on_batch_begin, this is how to make the transition:

# before
def on_batch_begin(self, net, X, y, ...):
    ...

# after
def on_batch_begin(self, net, batch, ...):
    X, y = batch
    ...

The same goes for the other three methods.

Migration from 0.11 to 0.12

In skorch 0.12, we made a change regarding the training step. Now, we initialize the torch.utils.data.DataLoader only once per fit call instead of once per epoch. This is accomplished by calling skorch.net.NeuralNet.get_iterator() only once at the beginning of the training process. For the majority of the users, this should make no difference in practice.

However, you might be affected if you wrote a custom skorch.net.NeuralNet.run_single_epoch(). The first argument to this method is now the initialized DataLoader instead of a Dataset. Therefore, this method should no longer call skorch.net.NeuralNet.get_iterator(). You only need to change a few lines of code to accomplish this, as shown below:

# before
def run_single_epoch(self, dataset, ...):
    ...
    for batch in self.get_iterator(dataset, training=training):
        ...

# after
def run_single_epoch(self, iterator, ...):
    ...
    for batch in iterator:
        ...

Your old code should still work for the time being but will give a DeprecationWarning. Starting from skorch v0.13, old code will raise an error instead.

If it is necessary to have access to the Dataset inside of run_single_epoch, you can access it on the DataLoader object using iterator.dataset.