why is the recognition loss does not reduce?

[hoainv99]

i trained model but i recognize that the recognition loss does not reduce? it is due to box coordinates? i see this code does not scale to 128 correspond to feature_map

[xansar]

i trained model but i recognize that the recognition loss does not reduce? it is due to box coordinates? i see this code does not scale to 128 correspond to feature_map

I have the same problem with yours. When training, the output of recognition branch contain different chars. However, from the second epoch, this branch just ouput blank string "", its loss can be around 4 and never decreased anymore.