Random output after several early epoch then start training

aimagelab / meshed-memory-transformer

Meshed-Memory Transformer for Image Captioning. CVPR 2020

BSD 3-Clause "New" or "Revised" License

517 stars 136 forks source link

Random output after several early epoch then start training #13

Open TrungThanhTran opened 4 years ago

TrungThanhTran commented 4 years ago

Hi @marcellacornia,

When I started my train, I got random outputs for about the first five epochs, I mean it generated words. Then, it produced nothing, and I had to train for several epochs to get good results. Do you have any idea? Because of initialization?