Using a custom dataset with 512x512 shape produces an error about hila

[1234gary]

Hey Denis,

Looking at your config I can see two things that might cause issues.

1. Around line 34 you should use 'whole' testing instead of sliding window testing
   test_cfg=dict(mode='whole'))

2. Around line 115 and 139 your resize should have keep_ratio=False, otherwise the dimensions will keep the original image ratio.
   dict(type='Resize', keep_ratio=False),

I've ran your config and wasn't able to get the error you got. My guess is some of these things are happening:

1. The validation inputs might not be loaded with the correct size
2. Ht and Wt aren't matching the actual dimensions of the input (perhaps the input isn't being downsampled properly?)

Could you try checking those values?

Those dimension numbers look quite strange to me, as I would expect 512x512 to be downsampled maybe 16x in layer 3 to be 32x32 = 1024. Ht * Wt = 920 in your case isn't a square number as you would expect, and 1802240 corresponds to a [1, 128, 16, 880] dimension.

[DenisN03]

Thanks for your help Gary.

I changed the keep_ratio setting to False and that did the trick.

Now I am facing another problem. After testing is complete, an error occurs:

/usr/lib/python3.6/multiprocessing/semaphore_tracker.py:143: UserWarning: semaphore_tracker: There appear to be 20 leaked semaphores to clean up at shutdown
  len(cache))

During the testing phase, RAM consumption increases very much.
At the end of the test swap overflow and the script stops.

[1234gary]

Hey Denis,

Are you able to get the test results? What is the exact line that you are running for evaluation? It's hard for me to diagnose what the issue might be without those.

I know that F-Score can be very heavy on the RAM and for that reason, if you are running F-Score you might want to run it in batches and average the scores together. This will result in the same results as if you ran the F-Score on all images at once.