Sharing training process

Aneureka / animate-me

🍭 A miniapp to convert selfie to anime.

https://zhuanlan.zhihu.com/p/79389141

222 stars 19 forks source link

Sharing training process #4

Open FantasyJXF opened 5 years ago

FantasyJXF commented 5 years ago

I use the selfie2anime datasets the Author provided, and replace like one thousand images that was not so good based on aesthetics.

It's so time consuming to train the model.

I use Tesla P40 and resize the image size to 128x128, it took me 7days to train the model.

OMG, the result is no better than the original.

Aneureka commented 5 years ago

Have you trained the model based on the pretrained one the author provided? I have tried to train it on my datasets (around 5000 pictures) also but it's very very life wasting...

FantasyJXF commented 5 years ago

Also tried to train base on the pretrained 100epoch model, the result was not so good. So I retrained the model with PYTORCH release.

And I don't know what to do to improve the model performance, the present model has too many badcase.

What's your opinion?