Results are not as good as the pretrained network after a train with the pretrained option.

[Leonard-GEHIN]

I got an FOTS checkpoint, trained from scratch, which has pretty good results on my database.
I want to finetune this checkpoint: when I try to use the pretrained option, everything goes smoothly. After the training is complete, I test the checkpoint with the main_test.py script, and the detection is really bad.

For example, on an image with 10 texts, correctly detected by the correctly trained FOTS, there is 2000 detection before nms on the finetuned FOTS.

I tried to finetune with only one step with a low learning rate to not change weights to much for the purpose of testing, but results were the same: 2000 detection when testing.

Did you ever got in a situation like this?