Code architectures of Denoising and Super-resolution are different from Inpainting

[Ares960826]

Dear Dr. Li.

Sorry for bothering you and hope you are doing well!

May I ask why Code architectures of Denoising and Super-resolution are different from Inpainting?

For example, there is no codebase like this in train_denoising.py:

# @train_ip.py
... ... ... ...
        optimizer_bn.step()
        epoch_loss.append(loss.item())

        ### after one update, we would like to change our network weight back to its initializaiton
        old_weight_dict_after_opt = net.state_dict()
        new_weight_dict_after_opt = {}
        for param_key in net.state_dict():
            # custom initialization in new_weight_dict,
            # You can initialize partially i.e only some of the variables and let others stay as it is
            #print(param_key)
            # we only want to add noise to the CNN layers; we do not want to change the BN paprameters
            if param_key in CNN_weight_name:  # it's a CNN layer, so we want to copy back the init weights
                old_weight = old_weight_dict[param_key]
                new_weight_dict_after_opt[param_key] = old_weight
            else: # it's a BN layer, so we want to update its weights
                old_weight = old_weight_dict_after_opt[param_key]
                new_weight_dict_after_opt[param_key] = old_weight
        net.load_state_dict(new_weight_dict_after_opt)
        # let's fix the weights of networks
        for name, param in net.named_parameters():
            ### here we do not want to fix the BN
            if name in CNN_weight_name:
                param.requires_grad = False

        ### let's update our learning rate
        #scheduler_net.step()
        #scheduler_noise.step()
        #scheduler_bn.step()

        # ------ after one epoch: end to measure the time in this iteration
        if epoch != 0:
            t_end = time.time()
            per_used_time = t_end - t_start
            all_used_time += per_used_time
... ... ... ...

Best regards,
Mingze
PhD Candidate
University of Strathclyde