--use_additional_encoders

[dsdsdddd]

I found that when adding the --use_additional_encoders parameter, the training speed is ten times slower than without this parameter. I think there might be some issues with your image feature extraction code.