Joseph

I think libgtk2 is missing, but I installed libgtk2. What's the problem?

Is this a bug?

1

self.bn.append(nn.BatchNorm1d(N_filt,int((current_input-self.cnn_len_filt[i]+1)/self.cnn_max_pool_len[i]),momentum=0.05)) I pulled out the module and it looks like: BatchNorm1d(80, eps=2583, momentum=0.05, affine=True, track_running_stats=True) when I set batchnorm = [True, True, True] and layernorm = [False, False, False] in...

Shift vs overlap

2

Just to clarify, in the paper the overlap is said to be 10ms, but in the code the shift is said to be 10ms. Does that mean between 2 consecutive...

Hi, Could you publish the pre-trained models? Thanks, Best, Joseph

Pretraining code

1

Hi, How are the models pretrained? I notice that custom architectures like injector and extractors require rewriting the model, so I'm assuming you pretrained the model yourself? If that's correct,...

Input Normalization

1

I'm not sure if my mixing is exactly same as yours, but does your torchaudio read the wav files to int(value is typically around a couple hundred) or to float...

Can you use swin transformer as encoder?

1

As in title, can you replace the standard ViT encoder with a swin transformer + FPN? Would this be a reasonable thing to try out?

things are blurry

2

the rendered images are blurry. Any ideas on how to improve?

custom backward does not account for gradient scaling

7

### Description Hi, it seems like the backward() function in SpecifyGradient does not consider gradient scaling. I realized this because I tried to return torch.sum(gt_grad**2), and then skip the u-net...

HiFA and prolificdreamer+HiFA

6

Fixed z-variance loss by normalizing weights. This is crucial to ensure that the model isn't encourage to generate empty pixels(since empty pixels have zero z-variance) Added other HiFA(image space loss,...