The BCE loss is small at the beginning of train

KichangKim / DeepDanbooru

AI based multi-label girl image classification system, implemented by using TensorFlow.

MIT License

2.58k stars 258 forks source link

The BCE loss is small at the beginning of train #59

Closed gqy1025 closed 2 years ago

gqy1025 commented 2 years ago

Hi,

i tried to rerun your code step by step, and i randomly picked 10,000 pictures as my dataset from Danbooru. when i started training, i found the loss is very small (around 0.02) even at the very beginning.

Is this normal?

KichangKim commented 2 years ago

Loss may be varying by your training parameter, but 0.02 seems to be normal. DeepDanbooru web's final loss is about 0.006.