search

dessa-oss / fake-voice-detection

Using temporal convolution to detect Audio Deepfakes
http://www.atlas.dessa.com
Apache License 2.0
342 stars 86 forks source link

Training my own dataset #16

Open BenjaminChua opened 3 years ago

BenjaminChua commented 3 years ago

I wish to train on my own dataset which consists of real and fake wav files. May I know how I can do so in terms of preprocessing and tuning of the hyperparameters?

ranasac19878 commented 3 years ago

Hi Benjamin, you can take a look at unlabeled_inference.py file in which I have called the preprocessing function. You can use that function in order to preprocess your wav files. For hyperparam tuning, you can follow the readme to do this using foundations software or you can use any other software you like. You should fix a validation metric such as accuracy, f-1 score or roc-auc and whichever combination of hyperparams maximizes these metrics should the optimum parameters for your architecture.

BenjaminChua commented 3 years ago

It seems like the training is running OOM on a single GPU. May I know what are the specs of the GPUs used in this project?

yzslry commented 2 years ago

The link to download the data in this project seems to be invalid. Can you provide the data or link in the project?