ranasac19878
commented
5 years ago Hello,
Thanks for your note:
-
Now you should see requirements.txt inside code/ folder. It was due to
.gitignorethat all the text files were ignored automatically in git push -
We are still trying to improve our model as we speak. If you observe, the pre-trained model performs very well on the training and validation set (i.e. above 0.9 f1-score) but not so much good on test set or any other arbitrary audio file. This shows that the pre-trained model is overfitting on the algorithms that were used to produce fake voices in train and validation set.
We would like to encourage other users to improve upon this model by performing hyperparam search to get a more generalized model. A couple of things to try are:
a) Try to use regularization techniques
b) Try to add more audio data to train and possibly validation set
c) Try different model architectures (i.e. using conv1d on raw audio instead of spectrograms, or using LSTMs etc.)
d) Analyze the pre-trained model on different audio clips to find patterns in the spectrograms where the model seems to overfit.
c) The classes are highly imbalanced in the original dataset: 90% fake and 10% real audio clips. So, the model generally has more tendency to predict audio signals to be fake. I took some steps to overcome these by augmenting the data by using nlpaug library (see in utils.py) but it can be further improved by adding more dataset.
You should not get the bash: scp_spectrogram.sh: No such file or directory error anymore since I removed that line. It was only there for some remote debugging processes
You can try pre-trained model without Atlas and that doesn't change anything.
Please see my suggestions in order to get a better model and let me know your thoughts.
Thanks!
Kinjal05
Harrshit
Hey, I was trying to use the model and some issues came up.
Do you have any suggestion what do I miss/should install, modify to get more reasonable probabilities? Thanks!