bwang514 / performancenet Goto Github PK

Hi, your work was amazing! I am reproducing it but I got some problems when using my own cello solo midi file to run the synthesize_audio.sh.

At first, the ERROR was :
TypeError: conv1d(): argument 'input' (position 1) must be Tensor, not numpy.ndarray
Then I turned score and onoff to Tensor, but not fixed, the error was:
RuntimeError: input has less dimensions than expected
I found that my midi file's shape is torch.Size([22277, 128]), while your TEST_DATA is torch.Size([5, 128, 860])
I tried other midi files but also got same problems. So maybe the process_custom_midi function not transform the midi file with suitable shape?

Again, I appreciate your time reading this issue. Have a great day!

Dear authors,

Thank you for the wonderful paper and code. I just to confirm the implementation of MBRBlock.

I think these two lines intend to save the output from the conv-instancenorm block into bands and add bands to x.

However, if I correctly understand torch.0.4.0 doc (https://pytorch.org/docs/0.4.0/torch.html?highlight=torch%20add#torch.add), I think the code "torch.add(bands[i],1,t)" will not save t into bands.

Since bands originally stores the value of x, this line "x = torch.add(x,1,torch.cat(bands, dim = 1))" will do nothing but simply add x with itself (well, not the same object but the same value). As the result, what MBRBlock returns becomes x+x.

To testify that, I debugged and stopped at line 151 before executing it, then I compared the values in x and bands

(Pdb) p torch.std(x - torch.cat(bands, dim = 1))
tensor(0., device='cuda:0')
(Pdb) p torch.mean(x - torch.cat(bands, dim = 1))
tensor(0., device='cuda:0')

This indicates that bands stores the same value as x

With the two lines above, these four lines only multiplies x by 16, without using any conv in the MBRBlock. I can confirm this
through debugging .

So, could you check the implementation of MBRBlock?

FYI
I installed dependency according to requirement.txt except scikt-learn and pretty_midi.
But I think these two packages are not related to the above issue.

Hi guys, nice work. In the paper you tried the cello, violin and flute data and I plan to train it on piano midi data, have you guys tried that before? If so, what are the differences compared to other instruments? More specifically, 1) How many minutes of training data does piano need? (In paper you use different duration for different instruments). 2) What are the overlaps between chunks used in piano data? 3) Have you guys tried on shorter clips (like 2-second instead of 5-second) and would it easier to train? 4) Is high sampling rate (44.1k) necessary to get reasonable audio output?

Look forward to your replies. Thank you.

I found that the musicnet.npz file is not available now. Only the original data can be downloaded from https://zenodo.org/record/5120004#.YlP5K9a-s5l . How can I convert the original musicnet files to musicnet.npz which can be used in this project? Thanks.