About the gap between Training loss and Inference loss

gaopengcuhk / Pretrained-Pix2Seq

Replication of Pix2Seq with Pretrained Model

60 stars 5 forks source link

About the gap between Training loss and Inference loss #9

Open T-FOUNTAIN opened 2 years ago

T-FOUNTAIN commented 2 years ago

I found that the loss gap between training and inference stage is very large. The loss in inference stage is 10 times that in training. Even making inference on the training set ,the situation is the same. Can you give some advice?