Preprocessing to save some files for conditional inference

[mustass]

@sandorfoldi

When running preprocessing before training, we need to make sure that some portion of files is saved for conditional inference. I.e. we need to be sure to keep track of these files and make sure they won't be used for training or validation or testing steps.

Here we think about actually saving like 15 audiofiles and their spectrograms for evaluating the model. We would generate audio based on these 15 audiofiles and listen-compare the results. Maybe also do metrics on this is easier @panosapos ?

[panosapos]

That's what I meant yesterday evening as reconstruction (couldn't find a better term back then). Indeed, this will be very convenient for calculating multiple evaluation metrics.

@mustass, just to make sure: During conditional inference, we expect to generate an audio that will be very close to the original one, by providing the model the corresponding spectrogram, right?

[mustass]

@mustass, just to make sure: During conditional inference, we expect to generate an audio that will be very close to the original one, by providing the model the corresponding spectrogram, right?