search

breizhn / DTLN

Tensorflow 2.x implementation of the DTLN real time speech denoising model. With TF-lite, ONNX and real-time audio processing support.
MIT License
567 stars 160 forks source link

validation loss -16.83, but performance is not better than model.h5. #33

Open LeeGyuHa opened 3 years ago

LeeGyuHa commented 3 years ago

model.h5 : /pretrained_model/model.h5 (not normalized)

As a result of performance comparison between my_model.h5 and model.h5, model.h5 is the best. I looked at the reason. The size of the weights(weight,bias) values of "my_model.h5" is small compared to the weights values of "model.h5"

figure1. weights result image

figure2. signal result image

tensorflow-gpu == 2.4.0 tensorflow==2.4.0 GPU : GeForce GTX 1050 The parameters are the same as the code. The dataset was DNS-Challenge2020. I used clean,noise. snr_lowr : -5 snr_upper : 25 total_hours : 40 norm_stft : False

Training data was created with the code provided by DNS-Challenge2020. I set it to epoch 200, but because it is set to patient 10, it stopped at 96. validation loss : -16.83

Is there a way to increase the weight value? Is it a dataset issue? Any advice would be appreciated. Thank you.

yemifeng commented 3 years ago

Have you solved this problem? The same phenomena also happens in my experiment.

ghost commented 3 years ago

hello, the parameters of your training model is 4003624？ 3990352？ Thanks！