In train() (in utils/train.py) evaluation on the test set is run on the best model (not the last one) according to the accuracy on dev set
load_audio() in classe SpeechDataset modified to always use the cached audio for test and dev sets. Otherwise results change when using the loader multiple time, even if using the same model. shuffle set to False for test and dev loaders.
daemon
commented
4 years ago
In train() (in utils/train.py) evaluation on the test set is run on the best model (not the last one) according to the accuracy on dev set
load_audio() in classe SpeechDataset modified to always use the cached audio for test and dev sets. Otherwise results change when using the loader multiple time, even if using the same model. shuffle set to False for test and dev loaders.