Layers API

keras4torch.layers provides KerasLayer for automatic shape inference as well as some other useful layers for quick experiment.

Built-in KerasLayer

The built-in KerasLayer is a replacement for some native torch module. The followings are supported.

• Conv1d
• Conv2d
• Conv3d
• Linear
• GRU
• LSTM
• BatchNorm1d
• BatchNorm2d
• BatchNorm3d
• LayerNorm

Compared with native torch modules, what you need to change is omitting the first shape parameter. For example, nn.Linear(128, 512) must be rewritten as k4t.layers.Linear(512); nn.Conv1d(32, 64, kernel_size=3) must be rewritten as k4t.layers.Conv1d(64, kernel_size=3).

If a model contains KerasLayer, you should build it after wrapping it by keras4torch.Model.

model = k4t.Model(model)        # the model contains at least one `KerasLayer`
model.build(input_shape=[128])

The argument input_shape should not include batch_size dimension.

Custom KerasLayer

In fact, what a KerasLayer do is to delay the module instantiation to the first .forward() call, thus it can get the output shape from its previous layer and decide how many channels should be created.

To write your own KerasLayer for automatic shape inference, you need to subclass KerasLayer and implement its abstract method build().

In KerasLayer.build(), you will get the current input shape to instantiate the actual module and return it. The arguments are stored by self.args and self.kwargs.

class Conv1d(KerasLayer):
    def build(self, in_shape):
        return nn.Conv1d(in_shape[1], *self.args, **self.kwargs)

Note that KerasLayer should be used as a wrapper only, which means your native torch module need to be available without KerasLayer. Never write logics in KerasLayer irrelevant with in_shape.

Others

• Lambda
• Add
• Reshape
• Concatenate