tips for finetuning on private dataset based on the pretrained model

[alen-mask]

hi @sfzhang15 ,

I just test your uploaded model and better results is obtained on them compared to other models.
Do you have any plans to support finetuning on other datasets(different number of object classes) from the pretrained model-ATSS_dcnv2_X_101_64x4d_FPN_2x.pth?
Any tips or reference code here?

Many thanks!

[sfzhang15]

@alen-mask
You can refer to this. If you have any questions, please feel free to contact us.

[sfzhang15]

I close this issue because there was no response for a long time, please reopen it if there are still problems.

[alen-mask]

Hi @sfzhang15 ,
I just use the shapes dataset(coco format) for training and test since the ATSS_dcnv2_X_101_64x4d_FPN_2x.pth for coco you uploaded works fine.
However, no bounding box was returned by the new trained model, do i have to make extra modifications besides https://github.com/tianzhi0549/FCOS/issues/54 ?

test log:
1001.jpeg inference time: 0.19s
im_name
len(boxes): 0
1127.jpeg inference time: 0.19s
im_name
len(boxes): 0
1215.jpeg inference time: 0.18s
im_name
len(boxes): 0
1102.jpeg inference time: 0.18s
im_name
len(boxes): 0
1245.jpeg inference time: 0.18s
im_name
len(boxes): 0
1125.jpeg inference time: 0.19s
im_name
len(boxes): 0
1089.jpeg inference time: 0.19s
im_name
len(boxes): 0
1191.jpeg inference time: 0.20s

[alen-mask]

After some debug work, i found scors of the predictions from trained model are very low, and after resetting thresholds_for_classes values to 0.1 for my datasets(3 classes of target object+1 background)in atss_demo.py file, the result is not so good: all predicted bboxes before selecting are with wrong labels and wrong locations.
I guess much more modifications are required to train custom datasets.

[sfzhang15]

@alen-mask
Sorry for the late reply. After preparing your own dataset, you can debug the codes with one training iteration via PyCharm to ensure correctness. When all required changes are ensured, you can begin to train the model. The learning rate and its schedule may also need to be adjusted appropriately.