recipe for Multi-Decoder DPRNN #463
Conversation
|
I just finished writing the eval script. Now only bash script is left. (I don't know how to write bash, btw, so I'd be nice if I could get some help) |
|
Also, I'm not entirely sure how to share models on huggingface. Can someone send me a tutorial somewhere? I currently have both .ckpt and .pth |
|
You just need to create a repo there, pull, add your model, change the readme and push. Thanks for your work on this recipe |
| train_dir: "../../../dataset/{}speakers/wav8k/min/tr" | ||
| valid_dir: "../../../dataset/{}speakers/wav8k/min/cv" |
There was a problem hiding this comment.
I think I said this before, but relative path in the config file doesn't seem right.
There was a problem hiding this comment.
ok. I'll change this to data/{}speakers/wav8k/min/tr. I just wrote relative paths because how it was on my system
|
|
||
| if [[ $stage -le 3 ]]; then | ||
| echo "Stage 3: Training" | ||
| echo "If you want to change n_srcs, please change the config file" |
There was a problem hiding this comment.
Printing this every time doesn't seem useful.
Also, why n_srcs is in the file if you can't change it.
There was a problem hiding this comment.
I don't think bash supports adding an integer list as argument, it just gets processed as strings automatically.
| @@ -0,0 +1,214 @@ | |||
| import os | |||
| self.log("avg_sdr", avg_sdr) | ||
|
|
||
|
|
||
| class WeightedPITLoss(nn.Module): |
There was a problem hiding this comment.
Can you add a docstring and some comments for understanding what this does on the high level?
There was a problem hiding this comment.
Yeah, you can see the docstring above the init line.
|
I fixed the stuff you mentioned, could we pass this PR now? |
|
Have you managed to share a model on HF Hub? |
| est_sources = model.separate(mix[None]) | ||
| p_si_snr = Penalized_PIT_Wrapper(pairwise_neg_sisdr_loose)(est_sources, sources) | ||
| utt_metrics = { | ||
| "P-Si-SNR": p_si_snr.item(), |
There was a problem hiding this comment.
It's defined in the paper.
| p_si_snr = Penalized_PIT_Wrapper(pairwise_neg_sisdr_loose)(est_sources, sources) | ||
| utt_metrics = { | ||
| "P-Si-SNR": p_si_snr.item(), | ||
| "Accuracy": float(sources.size(0) == est_sources.size(0)), |
There was a problem hiding this comment.
If this is a counting accuracy, then counting_accuracy is better.
| from scipy.optimize import linear_sum_assignment | ||
|
|
||
|
|
||
| class PairwiseNegSDR_Loose(_Loss): |
| self.take_log = take_log | ||
| self.EPS = EPS | ||
|
|
||
| def forward(self, est_targets, targets): |
There was a problem hiding this comment.
Is there a way not to duplicate all this code, but subclass it, with minor modif.
There was a problem hiding this comment.
The main problem is the line assert targets.size() == est_targets.size() in your original code, as it does not support variable number of speakers.
|
Do we have any further issue to fix, or should we pass this PR? |
|
We're close.
|
|
Hi Pariente, |
Don't have the bash script and separate function yet