Export ONNX Dropout for opset 10 #20710
Conversation
|
|
||
| # we should only export the onnx Dropout op in training mode; test both modes | ||
|
|
||
| # test training mode |
There was a problem hiding this comment.
I think the training proposal is still in progress. So in ONNX, the dropout operator still only support inference mode. We should error out and print out warning if it's in training mode. Tell user, you should call .eval before exporting to onnx.
There was a problem hiding this comment.
So for both opset 9 and 10, error out instead of exporting Dropout() ?
There was a problem hiding this comment.
could we maybe still export the model with Dropout in training mode, but add a warning to inform the user?
There was a problem hiding this comment.
the only difference between opset 7 and opset 10's dropout is that their mask type (2nd output) is different. So they both only support inference, no training... So please don't add support for training yet. (ONNX's training proposal is still under development)
houseroad
left a comment
houseroad
left a comment
There was a problem hiding this comment.
Looks good, could you rebase again to let the CI pass.
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@houseroad has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
…aidar/remove_onnx_dropout_blacklist
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@houseroad has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
|
@houseroadcan we merge this? |
|
@houseroad merged this pull request in 27d1daa. |
6 participants
Remove Dropout from the opset 10 blacklist.
ONNX Dropout was modified in opset 10, but only the output "mask" was modified, which is not exported in pytorch opset 9. So we can still fallback on the opset 9 op.