用numpy实现二维卷积

用numpy实现二维卷积

import numpy as np

# img = np.random.rand(2, 5, 5, 3)
img = np.ones([2, 5, 5, 4])
print("输入图像的尺寸", img.shape)
ker = np.ones([3, 3, 4, 6])  # 6 表示输出通道


def conv(img, ker, stride=1, padding=0):
    size = list(img.shape)
    pad_img = np.zeros([size[0], size[1] + 2 * padding, size[2] + 2 * padding, size[3]])
    pad_img[:, padding:-padding, padding:-padding, :] = img
    img = pad_img
    out_size = (img.shape[1] - ker.shape[0]) // stride + 1
    res = np.zeros([img.shape[0], out_size, out_size, ker.shape[3]])
    for i in range(img.shape[0]):  # minibatch的维度
        for out_c in range(ker.shape[3]):  # 卷积后的输出的通道数
            for hi in range(0, out_size * stride, stride):  # hi表示padding后原图的坐标 每一次都是移动stride大小
                for wi in range(0, out_size * stride, stride):
                    region = img[i, hi:hi + ker.shape[0], wi:wi + ker.shape[0], :]
                    res[i, hi // stride, wi // stride, out_c] = np.sum(region * ker[:, :, :, out_c])
    return res


stride = 2
padding = 1
fm = conv(img, ker, stride=stride, padding=padding)
print("stride = {}  padding = {}".format(stride, padding))
print("输出的feature map 尺寸", fm.shape)
print(fm)