length difference between input and output signal

sun-peach commented 3 years ago

I trained my own model and follow the quick start demo "Quickstart: How to use Pretrained Models". However I find the signal lengths of input and output are different. Is this right?

When I trained the model, I tuned the trim length, hop length and window length.

Thank you.

ws-choi commented 3 years ago

Hi @sun-peach,

TLDR: right, lengths of input and output might be different.

`separate`

at the low level, our model is designed to take input with the fixed length. The function separate takes an input with the fixed-length. https://github.com/ws-choi/Conditioned-Source-Separation-LaSAFT/blob/a3e60bfdc1d5b4d20f5d5df852241a0c8d80420a/lasaft/source_separation/conditioned/cunet/dcun_base.py#L191

If you use separate_track like in the QuickStart, you don't have to split the original track manually.

`separate_track`

https://github.com/ws-choi/Conditioned-Source-Separation-LaSAFT/blob/a3e60bfdc1d5b4d20f5d5df852241a0c8d80420a/lasaft/source_separation/conditioned/cunet/dcun_base.py#L252

separate_track automatically split the input to sub-chunks to fit separate. This operation is capsulated in the SingleTrackSet object.

https://github.com/ws-choi/Conditioned-Source-Separation-LaSAFT/blob/a3e60bfdc1d5b4d20f5d5df852241a0c8d80420a/lasaft/source_separation/conditioned/cunet/dcun_base.py#L258

`SingleTrackSet`

SingleTrackSet automatically split the input into chunks. For each chunk, it appends paddings if its length is not fit to the separate_track. https://github.com/ws-choi/Conditioned-Source-Separation-LaSAFT/blob/a3e60bfdc1d5b4d20f5d5df852241a0c8d80420a/lasaft/data/musdb_wrapper.py#L272

This is why lengths of input and output might be different. If you want to make them have the same length, then try

input_length = input.shape[1] # input => numpy array with shape: [channel, T]
output = output[:,:input_length]

sun-peach commented 3 years ago

Got it. Have done that. Thank you.

ws-choi / Conditioned-Source-Separation-LaSAFT