Layers

Core レイヤー

レイヤー	備考
Dense	全結合層
Flatten	入力の平滑化 (1次元にする)
Dropout	過学習防止のために前のレイヤーの出力の一部をランダムに無視する

CNN (Convolutional neural network)

レイヤー	備考
Conv1D, Convolution1D
Conv2D, Convolution2D
Conv3D, Convolution3D
AveragePooling1D, AvgPool1D
AveragePooling2D, AvgPool2D
AveragePooling3D, AvgPool3D
GlobalAvgPool1D, GlobalAveragePooling1D
GlobalAvgPool2D, GlobalAveragePooling2D
GlobalAvgPool3D, GlobalAveragePooling3D
MaxPool1D, MaxPooling1D
MaxPool2D, MaxPooling2D
MaxPool3D, MaxPooling3D
GlobalMaxPool1D, GlobalMaxPooling1D
GlobalMaxPool2D, GlobalMaxPooling2D
GlobalMaxPool3D, GlobalMaxPooling3D

RNN (Recurrent neural network)

レイヤー	備考
SimpleRNN	Fully-connected RNN where the output is to be fed back to input.
GRU	Gated Recurrent Unit - Cho et al.
LSTM	Long-Short Term Memory unit - Hochreiter 1997.

活性化 レイヤー

レイヤー	備考
LeakyReLU	Leaky version of a Rectified Linear Unit.
PReLU	Parametric Rectified Linear Unit.
ELU	Exponential Linear Unit.
ThresholdedReLU	Thresholded Rectified Linear Unit.

Merge レイヤー

• class Add: Layer that adds a list of inputs.
• class Multiply: Layer that multiplies (element-wise) a list of inputs.
• class Average: Layer that averages a list of inputs.
• class Maximum: Layer that computes the maximum (element-wise) a list of inputs.
• class Concatenate: Layer that concatenates a list of inputs.
• class Dot: Layer that computes a dot product between samples in two tensors.
• add(...): Functional interface to the Add layer.
• multiply(...): Functional interface to the Multiply layer.
• average(...): Functional interface to the Average layer.
• maximum(...): Functional interface to the Maximum layer.
• concatenate(...): Functional interface to the Concatenate layer.
• dot(...): Functional interface to the Dot layer.

参考資料

• Module: tf.contrib.keras.layers  |  TensorFlow
• レイヤーについて - Keras Documentation
• Guide to the Sequential model - Keras Documentation
• How to concatenate two layers in keras? - Stack Overflow

未分類

レイヤー	備考
Conv2DTranspose, Convolution2DTranspose
ConvLSTM2D

Classes

class Activation: Applies an activation function to an output.

class ActivityRegularization: Layer that applies an update to the cost function based input activity.






class BatchNormalization: Batch normalization layer (Ioffe and Szegedy, 2014).



class Cropping1D: Cropping layer for 1D input (e.g. temporal sequence).

class Cropping2D: Cropping layer for 2D input (e.g. picture).

class Cropping3D: Cropping layer for 3D data (e.g.





class Embedding: Turns positive integers (indexes) into dense vectors of fixed size.



class GaussianDropout: Apply multiplicative 1-centered Gaussian noise.

class GaussianNoise: Apply additive zero-centered Gaussian noise.


class InputLayer: Layer to be used as an entry point into a graph.

class InputSpec: Specifies the ndim, dtype and shape of every input to a layer.


class Lambda: Wraps arbitrary expression as a Layer object.

class Layer: Abstract base layer class.


class LocallyConnected1D: Locally-connected layer for 1D inputs.

class LocallyConnected2D: Locally-connected layer for 2D inputs.

class Masking: Masks a sequence by using a mask value to skip timesteps.





class Permute: Permutes the dimensions of the input according to a given pattern.

class RepeatVector: Repeats the input n times.

class Reshape: Reshapes an output to a certain shape.

class SeparableConv2D: Depthwise separable 2D convolution.

class SeparableConvolution2D: Depthwise separable 2D convolution.


class SpatialDropout1D: Spatial 1D version of Dropout.

class SpatialDropout2D: Spatial 2D version of Dropout.

class SpatialDropout3D: Spatial 3D version of Dropout.


class UpSampling1D: Upsampling layer for 1D inputs.

class UpSampling2D: Upsampling layer for 2D inputs.

class UpSampling3D: Upsampling layer for 3D inputs.

class ZeroPadding1D: Zero-padding layer for 1D input (e.g. temporal sequence).

class ZeroPadding2D: Zero-padding layer for 2D input (e.g. picture).

class ZeroPadding3D: Zero-padding layer for 3D data (spatial or spatio-temporal).

Functions


Input(...): Input() is used to instantiate a Keras tensor.