rcx986635
rcx986635
I found it should computes the L2 distance after Wl from the paper, but I didnt find in the lpips.py or the other area.
When I train the network, rpn_bbox_loss is too high, up to 0.4+. and others all ok less than 0.05 , my env is tf2.4.1 + python3.6, Anyone has the idea?
@Binbose, firstly,thank u for your sharing project. I try the GRU_LN, with layer_to_normalize=("input_gate", "input_recurrent", "recurrent_gate" , "recurrent_recurrent") and normalize_seperately=False, but I found that the training timecost takes twice as before...
if appling focal_loss to rpn_class ,mrcnn_class ,mrcnn_mask , the performance will be better ?
### Provide environment information python 3.7 labelme 4.5.6 ### What OS are you using? windows 10 ### Describe the Bug Cant edit Group Id with string instead of int: I...