About Keras Layers

Overview

Keras layers are the fundamental building block of keras models. Layers are created using a wide variety of layer_ functions and are typically composed together by stacking calls to them using the pipe %>% operator. For example:

model <- keras_model_sequential() 
model %>% 
  layer_dense(units = 32, input_shape = c(784)) %>% 
  layer_activation('relu') %>% 
  layer_dense(units = 10) %>% 
  layer_activation('softmax')

A wide variety of layers are available, including:

• Core Layers

• Convolutional Layers

• Pooling Layers

• Activation Layers

• Dropout Layers

• Locally-connected Layers

• Recurrent Layers

• Embedding Layers

• Normalization Layers

• Noise Layers

• Merge Layers

• Layer Wrappers

Properties

All layers share the following properties:

• layer$name — String, must be unique within a model.

• layer$input_spec — List of input specifications. Each entry describes one required input: (ndim, dtype). A layer with n input tensors must have an input_spec of length n.

• layer$trainable — Boolean, whether the layer weights will be updated during training.

• layer$uses_learning_phase – Whether any operation of the layer uses K.in_training_phase() or K.in_test_phase().

• layer$input_shape — Input shape. Provided for convenience, but note that there may be cases in which this attribute is ill-defined (e.g. a shared layer with multiple input shapes), in which case requesting input_shape will result in an error. Prefer using get_input_shape_at(layer, node_index).

• layer$output_shape — Output shape. See above.

• layer$inbound_nodes — List of nodes.

• layer$outbound_nodes — List of nodes.

• layer$input, layer$output — Input/output tensor(s). Note that if the layer is used more than once (shared layer), this is ill-defined and will result in an error. In such cases, use get_input_at(layer, node_index).

• layer$input_mask, layer$output_mask — Same as above, for masks.

• layer$trainable_weights — List of variables.

• layer$non_trainable_weights — List of variables.

• layer$weights — The concatenation of the lists trainable_weights and non_trainable_weights (in this order).

• layer$constraints — Mapping of weights to constraints.

Functions

The following functions are available for interacting with layers: