def corr2d_multi_in(X, K):
"""多个输入通道的互相关运算
"""
res = d2l.corr2d(X[0, :, :], K[0, :, :])
for i in range(1, X.shape[0]):
res += d2l.corr2d(X[i, :, :], K[i, :, :])
return res
def main():
# 5.1.1 二维互相关运算
X = torch.tensor([[0, 1, 2], [3, 4, 5], [6, 7, 8]])
K = torch.tensor([[0, 1], [2, 3]])
print(d2l.corr2d(X, K))
# 5.1.2 二维卷积层
layer = Conv2D(kernel_size=(2, 2))
output = layer(X)
print(output.shape)
# 5.1.3 图像中物体边缘检测
X = torch.ones(6, 8)
X[:, 2:6] = 0
K = torch.tensor([[1, -1]])
Y = d2l.corr2d(X, K)
print(Y)
# 5.1.4 通过数据学习核数组
conv2d = Conv2D(kernel_size=(1, 2))
step = 20
lr = 0.01
for i in range(step):
Y_hat = conv2d(X)
loss = ((Y_hat - Y)**2).sum()
loss.backward()
conv2d.weight.data -= lr * conv2d.weight.grad
conv2d.bias.data -= lr * conv2d.bias.grad
conv2d.weight.grad.fill_(0)
conv2d.bias.grad.fill_(0)
if (i + 1) % 5 == 0:
print('Step %d, loss %.3f' % (i + 1, loss.item()))
print(conv2d.weight.data, conv2d.bias.data)
def corr2d_multi_in(X, K):
# 沿着X和K的第0维（通道维）分别计算再相加
res = d2l.corr2d(X[0, :, :], K[0, :, :])
for i in range(1, X.shape[0]):
res += d2l.corr2d(X[i, :, :], K[i, :, :])
return res
def forward(self, x):
return d2l.corr2d(x, self.weight) + self.bias
# 5.1.2 ⼆维卷积层
class Conv2D(nn.Module):
def __init__(self, kernel_size):
super(Conv2D, self).__init__()
self.weight = nn.Parameter(torch.randn(kernel_size))
self.bias = nn.Parameter(torch.randn(1))
def forward(self, x):
return d2l.corr2d(x, self.weight) + self.bias
# 5.1.3
X = torch.ones(6, 8)
X[:, 2:6] = 0
K = torch.tensor([[1, -1]])
Y = d2l.corr2d(X, K)
# print(Y)
# 5.1.4
# 构造⼀个核数组形状是(1, 2)的⼆维卷积层
conv2d = Conv2D(kernel_size=(1, 2))
step = 20
lr = 0.01
for i in range(step):
Y_hat = conv2d(X)
l = ((Y_hat - Y)**2).sum() # 平方误差
l.backward()
# 梯度下降
conv2d.weight.data -= lr * conv2d.weight.grad
