Inference on pre-trained model

dessa-oss / fake-voice-detection

Using temporal convolution to detect Audio Deepfakes

http://www.atlas.dessa.com

Apache License 2.0

349 stars 86 forks source link

Inference on pre-trained model #11

Closed Parag0506 closed 4 years ago

Parag0506 commented 4 years ago

Hello, Can anyone help me understand the inference logic used in the code? I used the pre-trained model and I can't seem to get correct results. No video is crossing the probability of 10% while classifying any audio file. What does the model prediction actually output? what does [[0.00045425]] mean?

ranasac19878 commented 4 years ago

Hi Parag, thanks for your inquiry. The pretrained model works good for the training and validation data. It is currently not working good for arbitrary audio files. We are working to improve the model currently and will release the new pretrained model soon.

As of now, the model seems to predict low probabilities since the original training data had class imbalance (fake data was around 90% of total data).

Thanks, Sachin