[Bug] To Fix the Hang Problem in Training PSPNet

[barry-jin]

Error when training PSPNet on Cityscapes dataset using GluonCV #17439

Problem Description

The problem is when I train a PSPNet using GluonCV semantic segmentation library on the Cityscapes dataset, the training will stuck (hang) right after it started.

Debugging

After bisect the date of failure, I find the first bad commit is PR 13896, which introduced this problem.

Proposed solutions

Need more efforts.

References

• list reference and related literature
  Issue #17439, PR #13896
• list known implementations

[leezu]

Why not fix the hang instead of disabling the feature?

[sxjscience]

This does not sound like a solution. Problems related to CUDNN Dropout has a very long history and we should try to

• Fix cudnn dropout
• Consider to drop CuDNN Dropout if we can accelerate our native dropout

In fact, we haven't used cuda calls like curand4 (curandStatePhilox4_32_10_t *state) when implementing the random operators.

[sxjscience]

In addition, I guess is that the root cause is related to multiprocessing + cudnn dropout. Thus, we will need a minimal reproducible code snippet first.

[zhreshold]

+1 to @sxjscience , the segmentation model training adopts the DataParallel pipeline(https://github.com/dmlc/gluon-cv/blob/master/gluoncv/utils/parallel.py#L138), but it's using multithreading instead of mp