Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Implementation of D3Net #57

Open
tky823 opened this issue Mar 11, 2021 · 11 comments
Open

Implementation of D3Net #57

tky823 opened this issue Mar 11, 2021 · 11 comments

Comments

@tky823
Copy link
Owner

tky823 commented Mar 11, 2021

Reference: "D3Net: Densely connected multidilated DenseNet for music source separation"

@tky823
Copy link
Owner Author

tky823 commented Mar 13, 2021

I'm not sure of # channels before frequency concatenation.
The # of channels depends on the growth rate and # of D2 blocks.
I added bottleneck convolution so that both frequency bands have the same channels.

self.encoder = nn.Sequential(*encoder)
self.decoder = nn.Sequential(*decoder)
self.bottleneck_conv2d = nn.Conv2d(num_features, bottleneck_channels, kernel_size=(1,1), stride=(1,1))

@tky823
Copy link
Owner Author

tky823 commented Jun 5, 2021

What needs to be fixed

  • multi dilated convolution
    • timing of batch normalization
  • # of output channels of D2 block
  • order of D3 block and downsampling layer in Down D3 block
  • upsampling layer

@tky823
Copy link
Owner Author

tky823 commented Jun 6, 2021

Now, I updated D3Net architecture.

@lyghter
Copy link

lyghter commented Jun 27, 2021

Hello, @tky823. I participate in the Music Demixing Challenge (4th place on leaderboard A). I suggest you write a training script for D3Net and join a team with me.

@tky823
Copy link
Owner Author

tky823 commented Jul 4, 2021

Hi, @lyghter. I'm now writing the training code. I am not sure if it will be available soon, but I plan to add it.

@lyghter
Copy link

lyghter commented Jul 4, 2021

The challenge will end on July 31st. If you write the training code this month, I will try to train the model and use it in my solution. Sony's nnabla implementation has too slow inference on CPU. It cannot be used in the challenge.

@lyghter
Copy link

lyghter commented Jul 4, 2021

I invite you to join my team and suggest you keep the new code private until the end of the challenge.

@lyghter
Copy link

lyghter commented Jul 4, 2021

I am currently 4th on Leaderboard A and 5th on Leaderboard B. Top-3 from A and top-3 from B will receive prizes.

@tky823
Copy link
Owner Author

tky823 commented Jul 6, 2021

I'm not sure how my implementation of D3Net will work, so I don't know if I'll be able to participate anytime soon. If I can help, I will join your team. I work on other tasks for about a week. Maybe I will be able to join after that.

@lyghter
Copy link

lyghter commented Jul 12, 2021

Hello @tky823
Take a look at this. It looks like this repo contains pytorch implementation of D3Net and training code. I just found it and haven't tried it yet.

@tky823
Copy link
Owner Author

tky823 commented Sep 3, 2021

@lyghter
I'm sorry I couldn't help you. Now, I'm sharing the scripts and results in egs/musdb18/d3net.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants