search

JusperLee / SonicSim

http://cslikai.cn/SonicSim/
Creative Commons Attribution Share Alike 4.0 International
202 stars 25 forks source link

could you share the train code for those model? thanks #5

Open haha010508 opened 3 weeks ago

haha010508 commented 3 weeks ago

could you share the train code for those model? thanks

JusperLee commented 3 weeks ago

https://github.com/JusperLee/SonicSim/commit/2eae1f07b09124af5632572e7886811e98294f54

haha010508 commented 3 weeks ago

some error in separation/audio_train.py 20# import look2hear.system 21# import look2hear.datas and the train code do not choose train dataset

JusperLee commented 3 weeks ago

I have updated the codebase. Please check the latest version.

Feel free to let me know if you need any further assistance!

haha010508 commented 3 weeks ago

could you share the train config file? thanks! include cfg.datas.target, cfg.model.target,cfg.optimizer.target,cfg.scheduler.target,cfg.loss.target,cfg.metrics.target,cfg.system.target, in audio_train.py file, thanks a lot!

JusperLee commented 3 weeks ago

I'll recheck the code later.

JusperLee commented 2 weeks ago

https://github.com/JusperLee/SonicSim/commit/4387f55f8d457480cc23151400f12a50abba96c3

haha010508 commented 1 week ago

i found a bug at SonicSim/separation/train.py line: 104 to_save = system.audio_model.serialize() do not define.

JusperLee commented 1 week ago

https://github.com/JusperLee/SonicSim/blob/c29219acf73893151ff8e6b03be85b9a5195f3ff/separation/look2hear/models/base_model.py#L66

haha010508 commented 1 week ago

i run the train code, and finish the train. but i got an error: No such file or directory: "" is mean the checkpoint.best_model_path is not set in yaml file, so what the code do? i am very confuse that. by the way. i rewrite the best_k_models.json file, and i forget the content, if i retrain, i need a lot of time, so, could you give me an example for the best_k_models.json, thanks a lot. trainer.fit(system, datamodule=datamodule) print_only("Training finished!") best_k = {k: v.item() for k, v in checkpoint.best_k_models.items()} with open(os.path.join(cfg.exp.dir, cfg.exp.name, "best_k_models.json"), "w") as f: json.dump(best_k, f, indent=0) state_dict = torch.load(checkpoint.best_model_path) system.load_state_dict(state_dict=state_dict["state_dict"]) system.cpu()

this is the yaml file for checkpoints

checkpoint: _target_: pytorch_lightning.callbacks.ModelCheckpoint dirpath: ${exp.dir}/${exp.name}/checkpoints monitor: val_loss mode: min verbose: true save_top_k: 5 save_last: true filename: '{epoch}-{val_loss:.4f}'

JusperLee commented 6 days ago 
{
"Exps/bsrnn-2spk-noise-tencent/checkpoints/epoch=8-val_loss=-8.5521.ckpt": -8.552091598510742,
"Exps/bsrnn-2spk-noise-tencent/checkpoints/epoch=11-val_loss=-8.7302.ckpt": -8.730206489562988,
"Exps/bsrnn-2spk-noise-tencent/checkpoints/epoch=12-val_loss=-8.6079.ckpt": -8.6078519821167,
"Exps/bsrnn-2spk-noise-tencent/checkpoints/epoch=13-val_loss=-8.6066.ckpt": -8.606582641601562,
"Exps/bsrnn-2spk-noise-tencent/checkpoints/epoch=15-val_loss=-8.6824.ckpt": -8.682445526123047
}