search

djiajunustc / TransVG

157 stars 26 forks source link

Questions about batch_size #22

Closed toke1220 closed 2 years ago

toke1220 commented 2 years ago

Hello! Thank you for your excellent work! I train the model from scratch on a GPU. I set bs to “4”. But the train_loss hardly coverages in the first few epochs(5 epoch). I want to know if this situation is normal. Is it because of the value of bs?

toke1220 commented 2 years ago

Hello! Thank you for your excellent work! I train the model from scratch on a GPU. I set bs to “4”. But the train_loss hardly coverages in the first few epochs(5 epoch). I want to know if this situation is normal. Is it because of the value of bs?

I changed the equipment and set bs to ‘8’(the same as train.sh), and the convergence situation was normal. Maybe bs has to be above 8.