PyTorch-RetinaNet

Train RetinaNet with Focal Loss in PyTorch.

Reference:
[1] Focal Loss for Dense Object Detection

Code Structure

1. datagen.py for ListDataset, based on torch.dataset, it uses the other files:

   • transform.py: data transforms [eg. random_flip, random_crop, resize, etc.]
   • utils.py: calculate box_iou, box_nms & some utils for displaying
   • encoder.py: encode data to 9 anchor boxes.

2. train.py:

   • fpn.py: Feature Pyramid Network definition & architecture. TODO: Review the code & refer to the paper
   • loss.py: Focal loss definition
   • retinanet.py: RetinaNet definition, also based on FPN50 inside fpn.py

3. test.py: test using the trained model & decode the results.

4. script/get_state_dict.py: use the pre-trained weights & generate FPN50 feature_extractor weights

conclusion

the code is good, TODO: refer to the paper & review the code, understanding all the code snippets, and train you own data

Train-You-Own-Data

20180808: train with Metro data and it seems ok. train with 1e-3