Memory Leak: PANet

[seekingdeep]

• When trying to predict a dense image using PANet, the prediction gets stuck, and the memory usage constantly increase.

  sh tools/test_imgs.sh configs/textdet/panet/panet_r18_fpem_ffm_600e_ctw1500.py ./panet_r18_fpem_ffm_sbn_600e_ctw1500_20210219-3b3a9aa3.pth ./input ./input/list.txt ./output

• But, when modifying Line72 in the config, from 3000 into 940, then it works. Also anything below 940 works.

• Anything above 940 will get stuck.

• There is something that is causing a Memory Leak with PANet.

Input image used:

[jeffreykuang]

1 did upgrade your mmcv to 1.3.4? 2 when predicting this dense text img, I suggest to use pannet icdar2015 config.

I guess it is not due to memory leakage. It might be the memory usage complexity of the postprocessing of panet is positively correlated to the pixel number. I have optimized its memory usage already.

I will try to recover your bug to check what's happen.

[seekingdeep]

i reinstalled the python environment, by using this conda environment now PANet works works, thanks