Orca do not support preprocess inputs before training in forward()

[zhanglirong1999]

I rewrite yolox's code to run on orca as distributed, but errors were reported when forward(). Because orca directly use the input to forward() after the data loader, resulting in the inputs not being supported in yolox forward(). Yolox would preprocess input and target after the data loader before going to model forward.

Is it necessary for orca to support some special preprocess before training?

[zhanglirong1999]

This is the link that orca directly call the forward(). But the input exceeds the size of the forward() in yolox

https://github.com/intel-analytics/BigDL/blob/main/python/orca/src/bigdl/orca/learn/pytorch/training_operator.py#L282

[zhanglirong1999]

in yolox, there is some preprocess on input

https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/core/trainer.py#L100

https://github.com/analytics-zoo/YOLOX/blob/main/yolox/core/trainer.py#L95

[hkvision]

Probably we can add an argument for users to specify a preprocess function?

[zhanglirong1999]

The original yolox GPU version used a prefetcher, throwing away two inputs and only retaining two inputs. This is also a part of its data preprocessing.
https://github.com/Megvii-BaseDetection/YOLOX/blob/7ba9fd298413b79f2d07b6b7a3cdcbc34a9b244d/yolox/data/data_prefetcher.py#L25