Tensorflow，conv2d和过滤器

Question

I'm beginner in deep learning, and trying to understand how algorithms works, writing them using JavaScript. Now I'm working on JavaScript implementation of conv2d like Tensorflow does, and misunderstand how to handle different count of filters, I have succeeded for one output filter and multiple output, but I'm confused how to produce operations with multiple filters input eg 32 -> 64

Here is example of code using ndarray :

const outCount = 32 // count of inputs filters
const inCount = 1 // count of output features
const filterSize = 3
const stride = 1
const inShape = [1, 10, 10, outCount]
const outShape = [
  1,
  Math.ceil((inShape[1] - filterSize + 1) / stride),
  Math.ceil((inShape[2] - filterSize + 1) / stride),
  outCount
];
const filters = ndarray([], [filterSize, filterSize, inCount, outCount])

const conv2d = (input) => {
  const result = ndarray(outShape)
   // for each output feature

  for (let fo = 0; fo < outCount; fo += 1) { 
    for (let x = 0; x < outShape[1]; x += 1) {
      for (let y = 0; y < outShape[2]; y += 1) {
      const fragment = ndarray([], [filterSize, filterSize]);
      const filter = ndarray([], [filterSize, filterSize]);

      // agregate fragment of image and filter
      for (let fx = 0; fx < filterSize; fx += 1) {
        for (let fy = 0; fy < filterSize; fy += 1) {
          const dx = (x * stride) + fx;
          const dy = (y * stride) + fy;

          fragment.data.push(input.get(0, dx, dy, 0));
          filter.data.push(filters.get(fx, fy, 0, fo));
        }
      }

      // calc dot product of filter and image fragment
      result.set(0, x, y, fo, dot(filter, fragment));
      }
    }
  }

  return result
}

For test I'm using a Tenforflow as a source of true and it algorithm works correct but with 1 -> N . But my question how to add a support of multiple filters in input value like N -> M .

Could someone explain how to modify this algorithm to make it more compatible with Tensorflow tf.nn.conv2d A lot of thanks.

Answer 1

You would need to add another for loop. You didn't specify all of your input shapes and dimensions so it's actually kind of hard to write it exactly but it would look like this.

  // agregate fragment of image and filter
  for (let fx = 0; fx < filterSize; fx += 1) {
    for (let fy = 0; fy < filterSize; fy += 1) {
      //addition
      for (let ch = 0; ch < input.get_channels) {
        const dx = (x * stride) + fx;
        const dy = (y * stride) + fy;

        fragment.data.push(input.get(0, dx, dy, ch));
        filter.data.push(filters.get(fx, fy, ch, fo));
      }
    }
  }