Transfer Learning on Small Dataset

[shen338]

Hi Xingyi:

Thank you for providing this implementation, Great work!

I was able to get the reported result on COCO dataset. But recently, I need to fine-tune CenterNet model on a small but unique dataset with only ~10000 images.

I plan to freeze the backbone network to train a new classifier and fine-tune whole model afterwards. Is this easy to implement on CenterNet by Torch? Could you provide some necessary information for this? For example, the layer name I need to freeze,

Thank you!

[xingyizhou]

Thanks for your interests. You can refer to DEVELOP.md if you want to try new datasets. 10k images are good enough, I won't freeze any layers. You can finetune from the COCO model by --load_model ../models/ctdet_coco_dla_2x.pth.

[shen338]

@xingyizhou Thank you for the quick reply! I will try to fine-tune the whole network with my dataset.

My dataset has some other categories other than COCO's 81 classes. I may need to build my own Dataset instance with src/lic/datasets.

I will report my procedure and result here after I am done. Thank you!

[PRYUS]

Hey @shen338 ! Have you fine-tuned the model? Can you share your trained model?

[Ai-is-light]

@shen338 have some finetuning results to sharing? I met some troubles with finetueing or retraining the network with new dataset?

[xingyizhou]

The model is related to your specific dataset and it is no need to look at other people's model. Feel free to open a separate issue for your case.

[sisrfeng]

Anyone finetunes on own dataset (custom dataset) and gets better results than model provided by the author (Xingyi)? I finetuned and get worset results.

[JSNaidu]

Any git hub path for centernet implementation