Add conv-tasnet training script #896
Conversation
mthrok
force-pushed
the
conv-tasnet-train-loop
branch
from
25403d0 to
1f7df3c
Compare
fmassa
left a comment
fmassa
left a comment
There was a problem hiding this comment.
I had a quick looks and overall the structure looks pretty good!
I've made some comments about a few things I've done differently in my past scripts, but they are by no means meant to be the "ground-truth" on how to do things, and I'd be happy to hear your thoughts as well.
mthrok
force-pushed
the
conv-tasnet-train-loop
branch
from
152db73 to
b970f1c
Compare
mthrok
force-pushed
the
conv-tasnet-train-loop
branch
2 times, most recently
from
c3bd902 to
058620a
Compare
|
|
||
| </details> | ||
|
|
||
| ### SLURM |
There was a problem hiding this comment.
submitit is great for user experience: https://github.com/facebookresearch/detr/blob/master/run_with_submitit.py
There was a problem hiding this comment.
E.g. can do something like
python run_with_submitit.py --nodes 4 --timeout 3000 --coco_path /datasets01/COCO/022719
There was a problem hiding this comment.
See https://github.com/facebookincubator/submitit for submitit repo
There was a problem hiding this comment.
I looked at it, but it feels like it adds the more learning curve to go through this example.
The code is right now completely agnostic to SLURM yet runnable with SLURM.
I think that leaving the decision of the adaptation of submitit to users will keep our code simpler.
As long as the bare minimum of how to use in SLURM environment is explained in README, then it should do.
|
|
||
| The following is the evaluation result of training the model on WSJ0-2mix and WSJ0-3mix datasets. | ||
|
|
||
| ### wsj0-mix 2speakers |
There was a problem hiding this comment.
Do you want to upload a model to make it easier to reproduce these numbers?
There was a problem hiding this comment.
Yeah, I think that will be great.
mthrok
force-pushed
the
conv-tasnet-train-loop
branch
from
c10bf1b to
c97b349
Compare
mthrok
force-pushed
the
conv-tasnet-train-loop
branch
2 times, most recently
from
203cc7c to
2a9714d
Compare
examples/source_separation/README.md
Outdated
|
|
||
| ### Overview | ||
|
|
||
| To traing a model, you can use [`train.py`](./train.py). This script takes the form of |
mthrok
force-pushed
the
conv-tasnet-train-loop
branch
from
2a9714d to
1f9fbfb
Compare
|
Thanks! |
remove experimental from rpc tutorial
fix the error of generated code in subgraph rewrite
3 participants
#897