Add Tracks dataset - Githubissues

feldberlin / wavenet

An unconditioned Wavenet implementation with fast generation.

3 stars 0 forks source link

Add Tracks dataset #6

Closed purzelrakete closed 3 years ago

purzelrakete commented 3 years ago

What

Tracks dataset is able to read over a corpus of many tracks on disk, given metadata about the location of each track on disk.

Why

Needed to create custom audio corpus, e.g. anything not included in torchaudio already.

Acceptance Criteria

[x] Loads multiple tracks off disk.
[x] Produces appropriately framed non-overlapping frames
[x] Manage the speed of resampling by making the method configurable
[x] Cache the reads by moving data to the SSD and writing out resampled 8 bit versions
[x] Speed up reads by seeking into the files instead of reading the whole audio file for every sample
[x] Make sure that the compression step is also cached
[x] Inspect the results in the datasets notebook

purzelrakete commented 3 years ago

Questions

how long in seconds should each example be
should examples overlap a bit, or not at all
how should we deal with the tail frame if it's not full – should we pad?

purzelrakete commented 3 years ago

Answers

There's a longer discussion about batching in #7

each sample should be sample_length long. Make it around 2 seconds? It doesn't really matter for now, as long as we can fill the GPU with each batch.
make them overlap by receptive_field
drop the tail frame