DBNet customizing

[simanast]

I'm trying to customize pretrained model DB_r50. From what i know, DBNet should be able to process images of any shape. But in model's config file i found EastRandomCrop to (640, 640). Whenever i try to delete it and replace with Resize, the pipeline won't work, i get a warning like below for every file in my dataset :

prepare index 343 with error 
Warning: skip broken file {...}

But EastRandomCrop to different shape works fine.
So i wonder how can i make DBNet train on resized images, not cropped.

[simanast]

Or is EastRandomCrop an augmentation in this case? Since it's not wrapped in such dict (like ImgAug) i'm not sure

[gaotongxiao]

In MMOCR 0.x, the augmentations do not necessarily follow the same protocol and thus may not be interchangeable. Hence, simply replacing one transform with another could result in errors.

In MMOCR 1.0, EastRandomCrop has been replaced by three atomic transforms, and you can safely modify the order or change the parameters of these transforms by your need. (See DBNet's config and Migration Guide for details]. I'd suggest you upgrade to MMOCR 1.0 to carry out more solid investigations.