Add Inverse Short Time Fourier Transform in ATen native #35569
Conversation
facebook-github-bot
added
the
oncall: jit
💊 Build failures summary and remediationsAs of commit de4d2e9 (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions on the GitHub issue tracker. This comment has been revised 101 times. |
mthrok
force-pushed
the
aten-isftf
branch
5 times, most recently
from
fed4286 to
3016752
Compare
mthrok
changed the title
mthrok
force-pushed
the
aten-isftf
branch
8 times, most recently
from
bb7c8b3 to
0c444a1
Compare
|
FYI: There is a ongoing discussion on ways to extend |
There was a problem hiding this comment.
Overall, LGTM :) see comments below
- Batch packing/unpacking is performed in Python. ATen implementation expects 4D input tensor.
I'm assuming you did so to mirror the stft implementation? This means that the libtorch implementation will behave differently from the python implementation, right?
- The way
hop_lengthis initialized in the same way asstftimplementation. The Torchaudio's version tried to mimic the same behavior but slightly different.
The initialization in ATen seems fine to me. Can you clarify if it changes something for the user in python?
Let's leave these changes for a follow-up PR. |
|
@mthrok -- can you import the PR so that the internal tests run too? |
There was a problem hiding this comment.
@mthrok has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
ISTFT got into PyTorch core. See pytorch/pytorch#35569
4 participants
Ported
torchaudio's implementation (test, and documentation as well) to ATen.Note
hop_lengthis initialized in the same way asstftimplementation. The Torchaudio's version tried to mimic the same behavior but slightly different.Closes #34827
Relates #3775