zymk9 / Yet-Another-Anime-Segmenter

Instance segmentation for anime characters based on CondInst and SOLOv2
MIT License
189 stars 13 forks source link

how to "transfer learn"? #4

Open taibii opened 3 years ago

taibii commented 3 years ago

Hi, your work is great and very interesting. I'm wondering if I could segment only head and hairs of the anime character. I want to train with my own dataset, but can it be realized by doing the way this link says? https://github.com/aim-uofa/AdelaiDet/tree/master/configs/SOLOv2#installation--quick-start Or did you build your own model and trained with pretrained models? If so, can you teach me further details how to do these things? I just started studying deeplearnings, so I apolgize if this is a dumb question.

zymk9 commented 3 years ago

The models I released were trained based on the pretrained models from AdelaiDet. In my earlier experiments transfer learning and training from scratch achieved similar performance on my dataset.

As for your case - yes, you can train either from scratch or with the pretrained models, depending on the size of your dataset. Sometimes it's a good idea to try both.

taibii commented 3 years ago

Thanks for you reply. I'll make a try starting from scratch.