Any t-PIT loss calculation included?

[shincling]

Seems the loss in current code is based on uPIT.
In your papers, the tPIT seems get obvious better performance.
Will you add this part?
另外，Yuzhou不知道可不可以加个微信认识一下？ 我是自动化所，现在在JHU访学。

[shincling]

Update:
Seems I made some misunderstanding, after reading the code, I fould current version is actually based on tPIT. You treated the fist size of Y1R ( Y1I, Y2R.....) as batch_size*Time. And take each frame in each mixed sample as an independent element to calculate the loss and permutation. So it is the tPIT I guess. Is it right ?

deep-casa/nn/simul_group.py

Line 335 in 90c6e8f

def PIT_cost(Y1R,Y1I,Y2R,Y2I,EST1,EST2,SIG,_seq_len):

[yuzhou-git]

Yes. The current version uses tPIT loss to reorganize frames, and calculate SNR on an utterance level.