这是一个用python实现的卷积。
1、可以选择加padding或者不加,如果选择加padding,在input周围加上“1”
2、为了提高运行速度,尽量减少for循环。在卷积部分,将input和卷积核均转换成矩阵,使用矩阵相乘完成卷积,仅对batch做循环
代码如下:
import numpy as np
import math
class Conv2D(object):
def __init__(self, shape, output_channels, ksize=3, stride=1, method='VALID'):
self.input_shape = shape
self.output_channels = output_channels
self.input_channels = shape[-1]
self.batchsize = shape[0]
self.stride = stride
self.ksize = ksize
self.method = method
weights_scale = math.sqrt(ksize*ksize*self.input_channels/2)
#卷积核初始化,标准正态分布
self.weights = np.random.standard_normal((ksize, ksize, self.input_channels, self.output_channels)) // weights_scale
self.bias = np.random.standard_normal(self.output_channels) // weights_scale#bias初始化
if method == 'VALID':
self.eta = np.zeros((shape[0], (shape[1] - ksize ) // self.stride + 1, (shape[2] - ksize ) // self.stride + 1,self.output_channels))
if method == 'SAME':
self.eta = np.zeros((shape[0], shape[1]//self.stride, shape[2]//self.stride,self.output_channels))
self.w_gradient = np.zeros(self.weights.shape)
self.b_gradient = np.zeros(self.bias.shape)
self.output_shape = self.eta.shape
def forward(self,x):
col_weights = self.weights.reshape([-1,self.output_channels])
##如果保持输出feature map的shape保持不变,那么对边缘进行填充
if self.method == 'SAME':
x = np.pad(x, ((0, 0), (self.ksize // 2, self.ksize // 2), (self.ksize // 2, self.ksize // 2), (0, 0)),'constant', constant_values=0)
self.col_image = []
conv_out = np.zeros(self.eta.shape)
for i in range(self.batchsize):
img_i = x[i][np.newaxis,...]
self.col_image_i = self.im2col(img_i,self.ksize,self.stride)
#使用矩阵相乘得到卷积后的结果
conv_out[i] = np.reshape(np.dot(self.col_image_i,col_weights)+self.bias, self.eta[0].shape)
return conv_out
#将图像取patch,patch的大小为k_size*k_size*3,将patch reshape一行为(k_size*k_size*3,1),若有col个patch,则整个图像转换为(k_size*k_size*3,col
def im2col(self,image,k_size,stride):
image_col = []
for i in range(0,image.shape[1] - k_size+1,stride):
for j in range(0,image.shape[2]-k_size+1,stride):
col = image[:,i:i+k_size,j:j+k_size,:].reshape([-1]) #image2col
image_col.append(col)
image_col = np.array(image_col)
return image_col
if __name__ == '__main__':
conv2d = Conv2D([4,3,3,3],32,3,1,'VALID')
input_data = np.random.standard_normal((4,3,3,3)) #input_data=np.ones((4,3,3,3))
print("input:",input_data.shape)
conv_out = conv2d.forward(input_data)
print(conv_out.shape)
由于卷积那块有点繁琐,记录一下思考过程。
下图是一个例子,左边是一个
H
×
W
×
3
H \times W \times 3
H×W×3的input,右图是一个
3
×
3
×
3
3 \times 3 \times 3
3×3×3。浅色部分代表空域,深色部分代表通道,数字代表坐标。将input和卷积核reshape成两个矩阵,卷积过程就变成了两个矩阵相乘。
下图是input和卷积核reshape后的两个矩阵。input经过im2col函数变成矩阵,col表示input取patch的个数,patch大小为
3
×
3
×
3
3 \times 3\times3
3×3×3。卷积核矩阵的每一列均为一个卷积核,如果输出的feature map的通道为output_channel个,则需要output_channel个卷积核。