Think of label classification as Multi-Task Learning
OysterQAQ opened this issue · comments
Have you tried to split the system label and text label to form a multi-output model with different tasks? How many epochs did you end up training before the model converged?
No, I didn't try to split model. It always has combined model. All of released models are trained with 30 epochs. 1 epoch has about 1.7m images.