Open alexisdrakopoulos opened 4 years ago
It seems that the problem arises from the key. You need to carefully deal with the keys.
Hi, I have met the same problems. Actually I think it is an error caused by vesion mismatched. In ../codes/data/REDS_dataset.py line 52, It seems that opt['cache_keys'] stores a list. However, in ../codes/data_scripts/create_lmdb_mp.py, meta_info.pkl stores a dict like {name:' REDS ', resolution:' 3_180_320 ', keys:[000_00000000, .....]} We can detect the different between 'REDS_trainval_keys.pkl' and self 'meta_info.pkl' files by simple test function So, change either of them @xinntao , @alexisdrakopoulos
Thanks @yushizhiyao
The cache_keys
indeed mismatches with that generated in lmdb. We have updated it to read the cache_keys from the lmdb files.
I've created my own dataset with exact same structure as REDS other than different resolution. I successfully created the LMDB dataset and it passes all the tests I throw at it.
However when I attempt to train, it states there are only 3 images and then crashes.
I also notice that it seems to create the model/load things multiple times when compared to the example log provided in the repo. I can't figure out what causes this behavior.
I have over 30,000 images.
Full traceback: