CNNLayers

Full name Microsoft.VisualBasic.MachineLearning.CNN.CNNLayers Assembly Microsoft.VisualBasic.DeepLearning Members 17

Factory functions for building the layer specifications of a CNN network.

00 Remarks

The function names correspond one to one with the MLkit of R# (studio\Rsharp_kit\MLkit\MachineLearning\CNN.vb), so a network definition written in an R# script can be translated line by line into VB. Taking tutorials\..\CNN_image\auto_encoder.R as an example:

 ' R#:
 let cnn = cnn()
     + input_layer([28, 28], 1)
     + conv_layer(5, 32, 1, 2)
     + pool_layer(2, 2, 0)
     + leaky_relu_layer()
     + softmax_layer();

 ' VB:
 Dim cnn = New LayerBuilder() +
     input_layer({28, 28}, 1) +
     conv_layer(5, 32, 1, 2) +
     pool_layer(2, 2, 0) +
     leaky_relu_layer() +
     softmax_layer()

The only difference is that VB requires the binary operator at the end of the previous line for an implicit line continuation, whereas R# allows it at the start of the next line.

These functions all return CNNLayerArguments objects (layers that have not been created yet). The LayerBuilder + operator instantiates them from left to right, so the resulting layer sequence is identical to calling buildXxxLayer for every layer individually.

01 Syntax

Microsoft.VisualBasic.MachineLearning.CNN.CNNLayers

02 Methods

Name	Overloads	Summary
spec	1	构造规格对象；name 只用于 CNNLayerArguments.ToString() 的可读输出
input_layer	2	Input layer variant that takes the image size directly as a Dimension.
conv_layer	1	Convolution layer: extracts local features (edges, textures and so on) with a bank of filters.
conv_transpose_layer	1	Transposed convolution layer: upsamples a feature map back to a larger spatial size (decoder side of an auto encoder, semantic segmentation, and similar scenarios).
pool_layer	1	Pooling layer: downsamples the feature map inside a sliding window.
full_connected_layer	1	Fully connected layer: every neuron is connected to the complete output of the previous layer.
relu_layer	1	ReLU activation: f(x) = max(0, x).
leaky_relu_layer	1	LeakyReLU activation: keeps a small slope on the negative half axis so neurons do not "die".
sigmoid_layer	1	Sigmoid activation: f(x) = 1 / (1 + exp(-x)), with output in (0, 1).
tanh_layer	1	Tanh activation: output in (-1, 1).
maxout_layer	1	Maxout activation: takes the maximum value inside each group.
gaussian_layer	1	Gaussian activation layer.
lrn_layer	1	Local response normalization (LRN): lateral inhibition around strongly responding neurons increases the contrast of high frequency features.
dropout_layer	1	Dropout layer: randomly drops a fraction of the activations during training to reduce overfitting.
softmax_layer	1	Softmax: turns the activations into a probability distribution in [0, 1] (multi-class output layer).
regression_layer	1	Regression loss layer: the loss used for continuous outputs (auto encoders, variational auto encoders).

03 Members

method spec #

spec(String, Func(Of LayerBuilder, LayerBuilder))

构造规格对象；name 只用于 CNNLayerArguments.ToString() 的可读输出

method input_layer overload 2 #

input_layer(Int32(), Int32)

Input layer: passes the data into the network and declares the image size and channel count.

Parameters

Name	Type	Description
`size`	`Int32()`	The image size, given as `{width, height}`.
`depth`	`Int32`	Number of channels; 1 for grayscale images.

Returns

The input layer specification.

method input_layer #

input_layer(Dimension, Int32)

Input layer variant that takes the image size directly as a Dimension.

Parameters

Name	Type	Description
`dims`	Dimension	The image size.
`depth`	`Int32`	Number of channels; 1 for grayscale images.

Returns

The input layer specification.

method conv_layer #

conv_layer(Int32, Int32, Int32, Int32)

Convolution layer: extracts local features (edges, textures and so on) with a bank of filters.

Parameters

Name	Type	Description
`sx`	`Int32`	Side length of the (square) filter window.
`filters`	`Int32`	Number of filters, i.e. the number of output channels.
`stride`	`Int32`	Sliding step of the filter window.
`padding`	`Int32`	Width of the zero padding applied to the input.

Returns

The convolution layer specification.

method conv_transpose_layer #

conv_transpose_layer(Int32(), Int32(), Int32, Int32)

Transposed convolution layer: upsamples a feature map back to a larger spatial size (decoder side of an auto encoder, semantic segmentation, and similar scenarios).

Parameters

Name	Type	Description
`dims`	`Int32()`	Target output size `{width, height, depth}`.
`filter`	`Int32()`	Filter window `{width, height}`.
`filters`	`Int32`	Number of filters.
`stride`	`Int32`	Stride.

Returns

The transposed convolution layer specification.

method pool_layer #

pool_layer(Int32, Int32, Int32)

Pooling layer: downsamples the feature map inside a sliding window.

Parameters

Name	Type	Description
`sx`	`Int32`	Side length of the (square) pooling window.
`stride`	`Int32`	Stride of the pooling window.
`padding`	`Int32`	Width of the zero padding applied to the input.

Returns

The pooling layer specification.

method full_connected_layer #

full_connected_layer(Int32)

Fully connected layer: every neuron is connected to the complete output of the previous layer.

Parameters

Name	Type	Description
`size`	`Int32`	Number of neurons.

Returns

The fully connected layer specification.

method relu_layer #

relu_layer

ReLU activation: f(x) = max(0, x).

Returns

The ReLU layer specification.

method leaky_relu_layer #

leaky_relu_layer

LeakyReLU activation: keeps a small slope on the negative half axis so neurons do not "die".

Returns

The LeakyReLU layer specification.

method sigmoid_layer #

sigmoid_layer

Sigmoid activation: f(x) = 1 / (1 + exp(-x)), with output in (0, 1).

Returns

The sigmoid layer specification.

method tanh_layer #

tanh_layer

Tanh activation: output in (-1, 1).

Returns

The tanh layer specification.

method maxout_layer #

maxout_layer

Maxout activation: takes the maximum value inside each group.

Returns

The maxout layer specification.

method gaussian_layer #

gaussian_layer

Gaussian activation layer.

Returns

The Gaussian layer specification.

method lrn_layer #

lrn_layer(Int32)

Local response normalization (LRN): lateral inhibition around strongly responding neurons increases the contrast of high frequency features.

Parameters

Name	Type	Description
`n`	`Int32`	Size of the neighborhood involved in the normalization.

Returns

The LRN layer specification.

method dropout_layer #

dropout_layer(Double)

Dropout layer: randomly drops a fraction of the activations during training to reduce overfitting.

Parameters

Name	Type	Description
`drop_prob`	`Double`	Probability of dropping an activation.

Returns

The dropout layer specification.

method softmax_layer #

softmax_layer

Softmax: turns the activations into a probability distribution in [0, 1] (multi-class output layer).

Returns

The softmax layer specification.

method regression_layer #

regression_layer

Regression loss layer: the loss used for continuous outputs (auto encoders, variational auto encoders).

Returns

The regression layer specification.