Training losses remain unchanged

I have built an image enhancement network based on the paper , which requires convolution of the generated filter kernel and input image. However, after writing the convolution process, the model gradient will not change during training

my code is :

def extract_patche_onechannel(img, k_size, stride=1):

    kernel_size = torch.zeros(k_size, k_size).cuda()

    kernel_list = []
    for i in range(k_size):
        for j in range(k_size):
            k = kernel_size.clone()
            k[i][j] = 1

    kernel = torch.stack(kernel_list)
    kernel = kernel.unsqueeze(1)
    weight = nn.Parameter(data=kernel, requires_grad=True)
    out = F.conv2d(img[:, :1, ...], weight, padding=int((k_size - 1) / 2), stride=stride)
    return out

def forward(self, input_img):
    dynamic_filters = self.general_model(input_img)

    dynamic_filters = self.softmax(dynamic_filters)
    input_transformed = self.extract_patche_onechannel(input_img, k_size=1) 
    output_dynconv = torch.sum(dynamic_filters * input_transformed.cuda(), dim=1, keepdim=True)
    return output_dynconv

Based on your code snippet you are initializing the trainable weight parameter from scratch in each new forward pass which would break the training process. Instead, initialize the weight once and reuse it in the forward method.

Thank you for your answer. I set the weight parameter attribute to False for retraining, and the training loss will not decrease. I suspect that my code may have caused gradient loss. Another possibility is that the code I used to use convolutional kernels to convolution images is an error code

Sorry, I’m unsure how you are setting the parameter to False, but did you check my previous recommendation in initializing the parameter once?