When running to the sixth step of finetune, there will be inconsistencies in the tensor dimension. How can I solve this problem? I look forward to your reply.

RuntimeError: The expanded size of the tensor (64) must match the existing size (58) at non

Would be good to show Lasso/sparsity converging too.

I replaced the resnet with the VGG and still had the same problem. In the sixth step, finetune's new model went wrong.The number of channels after pruning does not match the number of channels before pruning, is it because the pruning channel is not saved, and the original channel does not match?How do you solve this problem? Looking forward to your reply.Thank you.

RuntimeError: Given groups=1, weight of size [209, 128, 3, 3], expected input[64, 115, 12, 12] to have 128 channels, but got 115 channels instead

Will also have to update the training function in utils

hi, I run main.py, but an erro occured， what's wrong?thanks very much;

I found that the function "expand_model" may cause "out of memory" for a network with huge parameters.

thank for your codes.
but I can not find any code that implements the process “absorb constant channels into next layer”,
does it implement or not ?
if done, can you tell me in where,thank you.

Output a model with channels removed where the batchnorm weights are zero.

in paper, the ista algorithm described as rl+1 = max{(rl-u▽l),0}*sgn(rl-u▽l) , but in sgd.py, the ista algorithm described as:
x = p.data.add(-group['lr'],d_p)
x = torch.clamp((torch.abs(x) - ista), min=0.)
p.data = x * torch.sign(x)
I think the true code should be:
x = p.data.add(-group['lr'],d_p)
y = torch.clamp((torch.abs(x) - ista), min=0.)
p.data = y * torch.sign(x)