[multimodal] fix num_gpus #3070
Conversation
|
Job PR-3070-68fa0f3 is done. |
| @@ -482,7 +482,7 @@ def predict( | |||
| if predictor._problem_type == OBJECT_DETECTION: | |||
| strategy = "ddp" | |||
|
|
|||
| if strategy == "ddp" and predictor._fit_called: | |||
| if strategy == "ddp" and predictor._fit_called and predictor._problem_type == OBJECT_DETECTION: | |||
There was a problem hiding this comment.
Have you tested if it works for other problem type? While working on object detection task, the blocker is a cuda barrier which I think may not only affect the detection problem type
There was a problem hiding this comment.
multi-gpu works for other problem types, since we are using DP. I think @zhiqiangdon would have more detail.
There was a problem hiding this comment.
multi-gpu works for other problem types, since we are using DP. I think @zhiqiangdon would have more detail.
Here we try to allow using DDP for other problem types. Right?
There was a problem hiding this comment.
this single-line change isn't quite related to DDP yet, actually it would be DP for other problem types.
There was a problem hiding this comment.
This if condition is about ddp. Wondering how it affects dp.
There was a problem hiding this comment.
Yes, you're right, this if condition is about ddp, but it should only affect OBJECT_DETECTION problem type, not other problem types.
it shouldn't affect dp for other problem types.
Issue #, if available:
num_gpus always become 1 when calling predictor.evaluate after fit, therefore prevent us from supporting multi-gpu.
Description of changes:
By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.