ResNet.ts

ResNet.ts

import * as tf from "@tensorflow/tfjs-node";

// In my case I have 256x256 pictures
const IMG_SIZE = 256;
const NUMBER_OF_CHANNELS = 3;

// Batch normalisation and ReLU always go together, let's add them to the separate function
const batchNormRelu = (input: tf.Tensor4D) => {
    const batch = tf.layers.batchNormalization().apply(input);
    return tf.layers.reLU().apply(batch);
};

// Residual block
const residualBlock = (input: tf.Tensor4D, filters: number, noDownSample: boolean = false) => {
    const filter1 = tf.layers
        .separableConv2d({
            kernelSize: 3,
            filters,
            activation: "relu",
            padding: "same",
            strides: noDownSample ? 1 : 2,
            depthwiseInitializer: "glorotNormal",
            pointwiseInitializer: "glorotNormal",
        })
        .apply(input) as tf.Tensor4D;
    const filter1norm = batchNormRelu(filter1);
    const filter2 = tf.layers
        .separableConv2d({
            kernelSize: 3,
            filters,
            activation: "relu",
            padding: "same",
            depthwiseInitializer: "glorotNormal",
            pointwiseInitializer: "glorotNormal",
        })
        .apply(filter1norm) as tf.Tensor4D;
    const dropout = tf.layers.dropout({ rate: 0.3 }).apply(filter2) as tf.Tensor4D;
    const batchNorm = batchNormRelu(dropout);
    
    // Residual connection - here we sum up first matrix and the result of 2 convolutions
    const residual = tf.layers.add().apply([filter1 as any, batchNorm as any]);
    return residual as tf.Tensor4D;
};


// ResNet - put all together
const modelResNet = () => {
    const input = tf.input({ shape: [IMG_SIZE, IMG_SIZE, NUMBER_OF_CHANNELS] });
    const conv1_filter = tf.layers
        .conv2d({
            kernelSize: 5,
            filters: 16,
            strides: 2,
            activation: "relu",
            padding: "same",
            kernelInitializer: "glorotNormal",
        })
        .apply(input) as tf.Tensor4D;
    const conv1 = tf.layers
        .maxPooling2d({
            poolSize: [3, 3],
            strides: [2, 2],
            padding: "same",
        })
        .apply(batchNormRelu(conv1_filter)) as tf.Tensor4D;

    // conv 2
    const residual2 = residualBlock(conv1, 16, true);

    // conv3
    const residual3 = residualBlock(residual2, 32);

    // conv4
    const residual4 = residualBlock(residual3, 64);

    // conv5
    const residual5 = residualBlock(residual4, 128);
    const conv5 = tf.layers
        .avgPool2d({
            poolSize: [8, 8],
            strides: [1, 1],
        })
        .apply(residual5) as tf.Tensor4D;

    const flatten = tf.layers.flatten().apply(conv5);
    const dropout = tf.layers.dropout({ rate: 0.5 }).apply(flatten);
    const dense = tf.layers
        .dense({
            units: 2,
            kernelInitializer: "glorotNormal",
            activation: "softmax", // softmax for categorical / relu
        })
        .apply(dropout);
    return tf.model({
        inputs: input as any,
        outputs: dense as any,
    });
};

Hi, Is there any real case that can demonstrate (or run) the above resnet network?

Network above is an example of the resnet that could be done in tensorflow JS. I demonstrate how to implement residual blocks and implement smaller version of the resent 18, you can read more about it and its applicable areas here https://arxiv.org/abs/1512.03385 . To learn more why this peace of code exists, you can check my article https://urakozz.medium.com/residual-model-from-scratch-with-tensorflow-js-part-1-dfee9ab8cbfe

@urakozz Your article I have already read, It's so great！
Maybe just I don't quite understand the output result after the resnet, I hope I can see some real case, for example, Yolo have the output picture after we run the code. But I'm not quite sure that the resent have the similar output or not.

I see, good question. First I need to correct you that for yolo Image is an input, output is category of the image (truck, bike) and frame (x1,y1,x2,y2,width, height). Image could be an output for the generative networks.
This particular network in example consumes 256x256 image and results with 2 outputs with softmax activation, so answer is 0,1 or 1,0, meaning 2 categories or yes/no answer. I used really similar lightweight network to determine if barcode exists on the image, yes/no.
But you can use residual blocks in different areas, such as nlp, image/sound recognition.