2021 08 17

Participants: Kayode and Dan

Some points from our discussion on refactoring:

• Have a single train.py script parameterized by the model type. I will start a branch called refactor, in which I will upload the suggested changes. I assume a similar refactoring can be achieved for the testing scripts, although some care has to be taken for the slightly different detection mechanisms employed by each of the models; maybe we can have a detect method (with type (audio, word_id) → (time, score)) in each of the models.
• Fix small discrepance between training and testing data processing. In particular, the standardization step is applied only at training and not at testing. Maybe it would have been better to use the __getitem__ method from Flickr8kDataset in the prediction step as well (although in this case we should make sure not to apply SpecAugment at test time).
• Clean up unused information, such as command line arguments, variables or imports.