Open SeeseeYep opened 6 months ago
""" PyTorch的backward参数, 默认是标量参数。因此如果计算非张量的梯度输出，需要使用 一个张量作为梯度参数, 用于存储各个元素梯度参数 """
X = torch.randn(size=(8 ,8)) conv2d_layer = Conv2D(kernel_size=(2, 2))
X.requires_grad = True output = conv2d_layer(X)
print(output)
grad_output = torch.ones_like(output)
output.backward(grad_output)
print("the gradients of weights :", conv2d_layer.weight.grad) print("the gradients of bias :", conv2d_layer.bias.grad)
RuntimeError: grad can be implicitly created only for scalar outputs
""" PyTorch的backward参数, 默认是标量参数。因此如果计算非张量的梯度输出，需要使用 一个张量作为梯度参数, 用于存储各个元素梯度参数 """
torch.conv2d 默认输入为 4维 code as:
X = torch.randn(size=(8 ,8)) conv2d_layer = Conv2D(kernel_size=(2, 2))
get grtadients
set auto gradients for x
X.requires_grad = True output = conv2d_layer(X)
compute gradients
print(output)
运行报错
output.backward()
assess gradients
print("the gradients of weights :", conv2d_layer.weight.grad)
解决方法
假设output是卷积操作的输出，它是一个矩阵
创建一个与output形状相同的张量，并将其所有元素设置为1
grad_output = torch.ones_like(output)
调用backward并传递grad_output作为参数
output.backward(grad_output)
现在，可以正确地评估权重和偏差的梯度
print("the gradients of weights :", conv2d_layer.weight.grad) print("the gradients of bias :", conv2d_layer.bias.grad)