search

JasonSWFu / VQscore

35 stars 3 forks source link

Do I need to fill in the path under data in the config/SE. yaml file to train this model? #4

Closed qiushanjun closed 5 months ago

qiushanjun commented 5 months ago

data: path: "" subset: clean_train: "./VCTK_clean_train.csv" clean_valid: "./VCTK_clean_test.csv" # not actually used

JasonSWFu commented 5 months ago

It depends on where you put your 'clean_train' .csv files files. (As defined in https://github.com/JasonSWFu/VQscore/blob/main/trainVQVAE.py#L88)

Maybe I should write it like this to avoid confusion: data: path: "./" subset: clean_train: "VCTK_clean_train.csv"

qiushanjun commented 5 months ago

Thank you for your reply, which enabled me to successfully run the project. Due to limitations in graphics memory, I adjusted the batch size to 32, resulting in almost half of the evaluation metrics being lowered. Have you ever encountered this situation?

JasonSWFu commented 5 months ago

No, I don't think the batch size will matter so much. Could you share some training results?

qiushanjun commented 5 months ago

No, I don't think the batch size will matter so much. Could you share some training results?

Of course, I followed the Readme steps to train locally, but the results were not very ideal, as it seems that the model did not converge. ![Uploading 56d9d087-256b-48d9-abe5-b357c8a257c8.png…]()

qiushanjun commented 5 months ago

[train]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 500000/500000 [24:43:17<00:00, 3.00it/s] 2024-06-14 19:25:29,265 (trainerAE:246) INFO: (Steps: 500000) train/noisy_cross_entropy_loss = 7.6729. 2024-06-14 19:25:29,265 (trainerAE:246) INFO: (Steps: 500000) train/clean_cross_entropy_loss = 7.5350. 2024-06-14 19:25:29,265 (trainerAE:246) INFO: (Steps: 500000) train/noisy_SP_loss = 17.6327. 2024-06-14 19:25:29,266 (trainerAE:246) INFO: (Steps: 500000) train/clean_SP_loss = 15.5765. 2024-06-14 19:25:29,266 (trainerAE:246) INFO: (Steps: 500000) train/VQVAE_loss = 48.4172. 2024-06-14 19:25:35,306 (trainerAE:127) INFO: (Steps: 500000) Start evaluation. [eval]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████| 26/26 [00:02<00:00, 10.25it/s] _eval_vctk_ValidSet.........██████████████████████████████████████████████████████████████████████████████████████████▊ | 25/26 [00:02<00:00, 21.35it/s] _eval_vctk_TestSet......... _eval_DNS1_test......... Evaluation takes 5.32m 2024-06-14 19:30:59,683 (trainerAE:158) INFO: (Steps: 500000) Finished evaluation (26 steps per epoch). 2024-06-14 19:30:59,683 (trainerAE:167) INFO: (Steps: 500000) eval/vqloss = 0.0258. 2024-06-14 19:30:59,683 (trainerAE:167) INFO: (Steps: 500000) eval/SP_loss = 15.3382. 2024-06-14 19:30:59,683 (trainerAE:167) INFO: (Steps: 500000) eval/waveform_loss = 0.0126. 2024-06-14 19:30:59,683 (trainerAE:167) INFO: (Steps: 500000) eval/VQVAE_loss = 15.3382. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk_valid/enhanced_dnsmos_sig = 1.3587. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk_valid/enhanced_dnsmos_bak = 1.4254. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk_valid/enhanced_dnsmos_ovr = 1.2162. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk_valid/enhanced_pesq = 1.6643. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/clean_recovery_dnsmos_sig = 2.2858. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/clean_recovery_dnsmos_bak = 2.6473. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/clean_recovery_dnsmos_ovr = 1.8547. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/clean_recovery_pesq = 2.4566. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/enhanced_dnsmos_sig = 2.1190. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/enhanced_dnsmos_bak = 2.2957. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/enhanced_dnsmos_ovr = 1.6867. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/enhanced_pesq = 1.9137. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/Clean_spectral_convergence_loss = 0.3409. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/Noisy_spectral_convergence_loss = 0.4947. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/Noisy_to_clean_spectral_convergence_loss = 0.4306. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/index_accuracy = 0.2694. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/Noisy_index_entropy = 10.6187. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) vctk/Clean_index_entropy = 10.4819. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) dns1/Real_enhanced_dnsmos_sig = 2.4086. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) dns1/Real_enhanced_dnsmos_bak = 2.6141. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) dns1/Real_enhanced_dnsmos_ovr = 1.9597. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) dns1/Noreverb_enhanced_dnsmos_sig = 2.9696. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) dns1/Noreverb_enhanced_dnsmos_bak = 3.2448. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) dns1/Noreverb_enhanced_dnsmos_ovr = 2.4561. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) dns1/Reverb_enhanced_dnsmos_sig = 2.2393. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) dns1/Reverb_enhanced_dnsmos_bak = 2.3688. 2024-06-14 19:30:59,684 (trainerAE:167) INFO: (Steps: 500000) dns1/Reverb_enhanced_dnsmos_ovr = 1.7672. 2024-06-14 19:30:59,731 (trainerAE:234) INFO: Successfully saved checkpoint @ 500000 steps. [train]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 500000/500000 [24:48:49<00:00, 5.60it/s]

JasonSWFu commented 5 months ago

Are there any changes in config/SE_cbook_4096_1_128_lr_1m5_1m5_github.yaml other than batch size?