Closed WoodieDudy closed 2 months ago
I am currently on last main commit
f2445da
I downloaded Voicebank Demand test set from here I evaluated DeepFilterNet versions 1-3 using the
test_voicebank_demand.py
script.Results of DeepFilterNet1 are the same as those reported in the DFN3 paper However, the results of DFN2 and 3 and are much lower.
How I can reproduce results similar to those in the paper?
python DeepFilterNet/df/scripts/test_voicebank_demand.py /datasets/testset2 --model-base-dir models/DeepFilterNet/ 2024-03-28 08:47:53 | INFO | DF | Running on torch 1.13.1+cpu 2024-03-28 08:47:53 | INFO | DF | Running on host sbx-4c5addaafaa7409c8eec87d5eb109b90-5968599-gstbl 2024-03-28 08:47:53 | INFO | DF | Git commit: 1097015, branch: torchDF-speedup-stft 2024-03-28 08:47:53 | INFO | DF | Loading model settings of DeepFilterNet 2024-03-28 08:47:53 | INFO | DF | Initializing model `deepfilternet` 2024-03-28 08:47:53 | WARNING | DF | Could not find `best` checkpoint. Checking for default... 2024-03-28 08:47:53 | INFO | DF | Found checkpoint models/DeepFilterNet/checkpoints/model_30.ckpt with epoch 30 2024-03-28 08:47:53 | INFO | DF | Running on device cpu 2024-03-28 08:47:53 | INFO | DF | Model loaded 2024-03-28 08:49:56 | INFO | DF | Progress: 25% 2024-03-28 08:52:02 | INFO | DF | Progress: 50% 2024-03-28 08:54:06 | INFO | DF | Progress: 75% 2024-03-28 08:56:09 | INFO | DF | Progress: 100% 2024-03-28 08:56:09 | INFO | DF | Waiting for metrics computation completion. This could take a few minutes. 2024-03-28 08:56:10 | INFO | DF | Enhanced STOI: 0.9420673982468806 2024-03-28 08:56:10 | INFO | DF | Enhanced PESQ: 2.809007406234741 2024-03-28 08:56:10 | INFO | DF | Enhanced CSIG: 4.124786853790283 2024-03-28 08:56:10 | INFO | DF | Enhanced CBAK: 3.3116347789764404 2024-03-28 08:56:10 | INFO | DF | Enhanced COVL: 3.472588062286377 2024-03-28 08:56:10 | INFO | DF | Enhanced SSNR: 7.947866439819336 2024-03-28 08:56:10 | INFO | DF | Enhanced SISDR: 16.64629923125494 0.9420673982468806,2.809007406234741,4.124786853790283,3.3116347789764404,3.472588062286377,16.64629923125494
python DeepFilterNet/df/scripts/test_voicebank_demand.py /datasets/testset2 --model-base-dir models/DeepFilterNet2/ 2024-03-28 08:58:36 | INFO | DF | Running on torch 1.13.1+cpu 2024-03-28 08:58:36 | INFO | DF | Running on host sbx-4c5addaafaa7409c8eec87d5eb109b90-5968599-gstbl 2024-03-28 08:58:36 | INFO | DF | Git commit: 1097015, branch: torchDF-speedup-stft 2024-03-28 08:58:36 | INFO | DF | Loading model settings of DeepFilterNet2 2024-03-28 08:58:36 | INFO | DF | Initializing model `deepfilternet2` 2024-03-28 08:58:36 | INFO | DF | Found checkpoint models/DeepFilterNet2/checkpoints/model_96.ckpt.best with epoch 96 2024-03-28 08:58:36 | INFO | DF | Running on device cpu 2024-03-28 08:58:36 | INFO | DF | Model loaded 2024-03-28 08:59:46 | INFO | DF | Progress: 25% 2024-03-28 09:00:55 | INFO | DF | Progress: 50% 2024-03-28 09:02:06 | INFO | DF | Progress: 75% 2024-03-28 09:03:19 | INFO | DF | Progress: 100% 2024-03-28 09:03:19 | INFO | DF | Waiting for metrics computation completion. This could take a few minutes. 2024-03-28 09:03:20 | INFO | DF | Enhanced STOI: 0.8134182890036015 2024-03-28 09:03:20 | INFO | DF | Enhanced PESQ: 2.714287281036377 2024-03-28 09:03:20 | INFO | DF | Enhanced CSIG: 3.8609461784362793 2024-03-28 09:03:20 | INFO | DF | Enhanced CBAK: 2.533511161804199 2024-03-28 09:03:20 | INFO | DF | Enhanced COVL: 3.253587484359741 2024-03-28 09:03:20 | INFO | DF | Enhanced SSNR: -2.2124998569488525 2024-03-28 09:03:20 | INFO | DF | Enhanced SISDR: -22.458605475868413
python DeepFilterNet/df/scripts/test_voicebank_demand.py /datasets/testset2 --model-base-dir models/DeepFilterNet3/ 2024-03-28 08:40:25 | INFO | DF | Running on torch 1.13.1+cpu 2024-03-28 08:40:25 | INFO | DF | Running on host sbx-4c5addaafaa7409c8eec87d5eb109b90-5968599-gstbl 2024-03-28 08:40:25 | INFO | DF | Git commit: 1097015, branch: torchDF-speedup-stft 2024-03-28 08:40:25 | INFO | DF | Loading model settings of DeepFilterNet3 2024-03-28 08:40:25 | INFO | DF | Initializing model `deepfilternet3` 2024-03-28 08:40:25 | INFO | DF | Found checkpoint models/DeepFilterNet3/checkpoints/model_120.ckpt.best with epoch 120 2024-03-28 08:40:25 | INFO | DF | Running on device cpu 2024-03-28 08:40:25 | INFO | DF | Model loaded 2024-03-28 08:41:33 | INFO | DF | Progress: 25% 2024-03-28 08:42:42 | INFO | DF | Progress: 50% 2024-03-28 08:43:52 | INFO | DF | Progress: 75% 2024-03-28 08:45:04 | INFO | DF | Progress: 100% 2024-03-28 08:45:04 | INFO | DF | Waiting for metrics computation completion. This could take a few minutes. 2024-03-28 08:45:05 | INFO | DF | Enhanced STOI: 0.6192451893910225 2024-03-28 08:45:05 | INFO | DF | Enhanced PESQ: 3.1094107627868652 2024-03-28 08:45:05 | INFO | DF | Enhanced CSIG: 3.632720708847046 2024-03-28 08:45:05 | INFO | DF | Enhanced CBAK: 2.6275570392608643 2024-03-28 08:45:05 | INFO | DF | Enhanced COVL: 3.30583119392395 2024-03-28 08:45:05 | INFO | DF | Enhanced SSNR: -2.231890916824341 2024-03-28 08:45:05 | INFO | DF | Enhanced SISDR: -24.651835330047653 0.6192451893910225,3.1094107627868652,3.632720708847046,2.6275570392608643,3.30583119392395,-24.651835330047653
That's what I'm getting. Did you manage to fix it?
I have updated torch to 2.2 and that fixed problem. There's is my new requirements reqs.txt But I haven't figured out what exactly was causing the problem yet. I'll text when I find out.
我已将火炬更新到 2.2 并解决了这个问题。这是我的新要求reqs.txt但我还没有弄清楚到底是什么导致了这个问题。当我发现时,我会发短信。
I have updated torch to 2.2 and that fixed problem. There's is my new requirements reqs.txt But I haven't figured out what exactly was causing the problem yet. I'll text when I find out.
I appreciate you writing back. I'll try it.
我已将火炬更新到 2.2 并解决了这个问题。这是我的新要求reqs.txt但我还没有弄清楚到底是什么导致了这个问题。当我发现时,我会发短信。
I have updated torch to 2.2 and that fixed problem. There's is my new requirements reqs.txt But I haven't figured out what exactly was causing the problem yet. I'll text when I find out. Hi, I upgraded torch=2.2.2. the last result of Deepfilter2 is as follows: 2024-04-04 16:45:36 | INFO | DF | Running on torch 2.2.2+cu118 2024-04-04 16:45:36 | INFO | DF | Running on host DESKTOP-J0UR71S 2024-04-04 16:45:36 | INFO | DF | Loading model settings of outputs 2024-04-04 16:45:36 | INFO | DF | Initializing model
deepfilternet2
2024-04-04 16:45:38 | INFO | DF | Found checkpoint D:\DeepFilterNet-main\DeepFilterNet\outputs\checkpoints\model_60.ckpt.best with epoch 60 2024-04-04 16:45:38 | INFO | DF | Running on device cuda:0 2024-04-04 16:45:38 | INFO | DF | Model loaded 2024-04-04 16:45:53 | INFO | DF | Progress: 25% 2024-04-04 16:46:08 | INFO | DF | Progress: 50% 2024-04-04 16:46:22 | INFO | DF | Progress: 75% 2024-04-04 16:46:36 | INFO | DF | Progress: 100% 2024-04-04 16:46:36 | INFO | DF | Waiting for metrics computation completion. This could take a few minutes. 2024-04-04 16:48:02 | INFO | DF | Enhanced STOI: 0.925880108458049 2024-04-04 16:48:02 | INFO | DF | Enhanced PESQ: 2.823953866958618 2024-04-04 16:48:02 | INFO | DF | Enhanced CSIG: 4.110456466674805 2024-04-04 16:48:02 | INFO | DF | Enhanced CBAK: 3.1988513469696045 2024-04-04 16:48:02 | INFO | DF | Enhanced COVL: 3.4739575386047363 2024-04-04 16:48:02 | INFO | DF | Enhanced SSNR: 6.198051452636719 2024-04-04 16:48:02 | INFO | DF | Enhanced SISDR: 13.75959416670562 It still doesn't achieve the results in the original article, so may I ask if you succeeded? I found your best epoch at 96 while I'm at 60, can you share your config.ini? I'm very much looking forward to your reply
I took weights and config.ini from there
@WoodieDudy upgrading torch to 2.2 surprisingly worked!! I also installed pystoi and pesq package versions from your requirements file. Thank you.
This issue is stale because it has been open 90 days with no activity. Remove stale label or comment or this will be closed in 7 days.
I am currently on last main commit
f2445da
I downloaded Voicebank Demand test set from here
I evaluated DeepFilterNet versions 1-3 using the
test_voicebank_demand.py
script.Results of DeepFilterNet1 are the same as those reported in the DFN3 paper However, the results of DFN2 and 3 and are much lower.
How I can reproduce results similar to those in the paper?
my_requirements.txt