Layer - DepthwiseConv2d

DepthwiseConv2d

DepthwiseConv2d performs the first step in a depthwise spatial convolution, apply convolution operation on each input channel separately.

Constructor

Based on whether TensorSpace Model load a pre-trained model before initialization, configure Layer in different ways. Checkout Layer Configuration documentation for more information about the basic configuration rules.

〔Case 1〕If TensorSpace Model has loaded a pre-trained model before initialization, there is no need to configure network model related parameters.

TSP.layers.DepthwiseConv2d();

〔Case 2〕If there is no pre-trained model before initialization, it is required to configure network model related parameters.

Two methods to create a new layer. Arguments are required.

[Method 1] Use kernelSize and strides

TSP.layers.DepthwiseConv2d( { kernelSize: Int, strides: Int } );

[Method 2] Use shape

TSP.layers.DepthwiseConv2d( { shape : [ Int, Int, Int ] } );

Fig. 1 - DepthwiseConv2d layer close and open

Arguments

Name Tag	Type	Instruction	Usage Notes and Examples
kernelSize	Int \| Int[]	The dimension of the convolution window, Network Model Related	The 2d convolutional window. For example, Square kernel, configre kernelSize to be 3 Rectangle kernel, configure kernelSize to [1, 7].
strides	Int \| Int[]	The strides of the convolution, Network Model Related	Strides in both dimensions, [default] Automatic inference strides: [1, 1]. For example, Strides in width and height is same, configure strides to be 2 Strides in width and height is different, configure strides to be [1, 2]
shape	Int[]	Output shape, Network Model Related	For example, shape: [ 28, 28, 6 ] means the output is 3-dimensional, 6 feature maps and each one is 28 by 28. Data format is channel last.
name	String	Name of the layer	For example, name: "layerName" In Sequential Model: Highly recommend to add a name attribute to make it easier to get Layer object from model. In Functional Model: It is required to configure name attribute for TensorSpace Layer, and the name should be the same as the name of corresponding Layer in pre-trained model.
padding	String	Padding mode, Network Model Related	valid[default]: Without padding. It drops the right-most columns (or bottom-most rows). same: With padding. Output size is the same as input size.
depthMultiplier	Int	The number of depthwise convolution output channels for each input channel.	Default to 1. For example, if input channel is 32, the default output channel would be 32, if depthMultiplier = 2, the output channel would be 64.
color	Color Format	Color of layer	DepthwiseConv2d's default color is #FBBF1F
closeButton	Dict	Close button appearance control dict. More about close button	display: Bool. true[default] Show button, false Hide button ratio: Int. Times to close button's normal size, default is 1, for example, set ratio to be 2, close button will become twice the normal size
initStatus	String	Layer initial status. Open or Close. More about Layer initial Status	close[default]: Closed at beginning, open: Open at beginning
animeTime	Int	The speed of open and close animation	For example, animeTime: 2000 means the animation time will last 2 seconds. Note: Configure animeTime in a specific layer will override model's animeTime configuration.

Properties

.inputShape : Int[]

The shape of input tensor. For example, inputShape = [ 28, 28, 3 ] represents 3 feature maps and each one is 28 by 28.

After model.init() data is available, otherwise is undefined.

.outputShape : Int[]

The shape of the output tensor is 3-dimensional. 3️⃣

The dataFormat is "channel last". For example, outputShape = [ 32, 32, 4 ] represents the output through this layer has 4 feature maps and each one is 32 by 32.

After model.init() data is available, otherwise is undefined.

.neuralValue : Float[]

The intermediate raw data after this layer.

After load and model.predict(), the data are available. Otherwise is undefined.

.name : String

The customized name of this layer.

Once created, you can get it.

.layerType : String

Type of this layer, return a constant: string DepthwiseConv2d.

Once created, you can get it.

Methods

.apply( previous_layer ) : void

Link this layer to previous layer.

This method can be use to construct topology in Functional Model.

See Construct Topology for more details.

.openLayer() : void

Open Layer, if layer is already in "open" status, the layer will keep open.

See Layer Status for more details.

.closeLayer() : void

Close Layer, if layer is already in "close" status, the layer will keep close.

See Layer Status for more details.

Examples

If TensorSpace Model load a pre-trained model before initialization, there is no need to configure network model related parameters.

let dwConv2d = new TSP.layers.DepthwiseConv2d( {

    // Recommend Configuration. Required for TensorSpace Functional Model.
    name: "depthwiseConv2d1",
    // Optional Configuration.
    animeTime: 4000,
    initStatus: "open"

} );

If there is no pre-trained model before initialization, it is required to configure network model related parameters.

let dwConv2d = new TSP.layers.DepthwiseConv2d( {

    // Required network model related Configuration.
    kernelSize: 3,
    strides: 1,
    depthMultiplier: 2,
    // Recommend Configuration. Required for TensorSpace Functional Model.
    name: "depthwiseConv2d2",
    // Optional Configuration.
    animeTime: 4000,
    initStatus: "open"

} );

Use Case

When you add convolutional layer with Keras | TensorFlow | tfjs in your model the corresponding API is DepthwiseConv2d in TensorSpace.

Framework	Documentation
Keras	keras.layers.SeparableConv2D(filters, kernel_size, strides=(1, 1))
TensorFlow	tf.nn.depthwise_conv2d( input, filter, strides )
TensorFlow.js	tf.layers.depthwiseConv2d (config)

Source Code

tensorspace/src/layer/intermediate/DepthwiseConv2d.js