python 卷积_15行python代码手写卷积操作

import numpy as np


def conv_naive(x, out_c, ksize, padding=0, stride=1):
    # x = [b, h, w, in_c]
    b, h, w, in_c = x.shape
    kernel = np.random.rand(ksize, ksize, in_c, out_c)
    if padding > 0:
        pad_x = np.zeros((b, h+2*padding, w+2*padding, in_c))
        pad_x[:,padding:-padding,padding:-padding,:] = x

    out_h = (h+2*padding-ksize)//stride+1
    out_w = (w+2*padding-ksize)//stride+1
    out = np.zeros((b, out_h, out_w, out_c))

    for i in range(out_h):
        for j in range(out_w):
            roi_x = pad_x[:,i*stride:i*stride+ksize,j*stride:j*stride+ksize,:]
            # roi_x = [b, ksize, ksize, in_c] -> [b, ksize, ksize, in_c, out_c]
            # kernel = [ksize, ksize, in_c, out_c]
            # conv = [b, ksize, ksize, in_c, out_c] -> [b, 1, 1, out_c]
            conv = np.tile(np.expand_dims(roi_x, -1), (1,1,1,1,out_c))*kernel
            out[:,i,j,:] = np.squeeze(np.sum(conv, axis=(1,2,3), keepdims=True), axis=3)
    return out

if __name__ == '__main__':
    x = np.random.rand(1,10,10,3)
    out = conv_naive(x, 15, ksize=3, padding=1, stride=2)
    print(out.shape)