Detectron_FocalLoss

I add focal loss to rpn.But the effect is not good..

It defaults to close.

If you don't want to use focal loss to rpn, you can ignore it

I add focal loss to fast_rcnn(lib/modeling/fast_rcnn_heads_fl.py).But it's not work.

I'm confused about it.Welcome to communicate with me~

install caffe2

https://github.com/caffe2/caffe2

install Detectron_FocalLoss

git clone https://github.com/withyou1771/Detectron_FocalLoss.git

install Detectron_FocalLoss refer to https://github.com/facebookresearch/Detectron

train your data

(1)convert xml to json

python tools/xml_to_json.py

you should change

xml_path = ''

json_file = ''

(2)set catalog

add yourdata to

lib/datasets/dataset_catalog.py

(3)download model and save to experiments

(4)experiments/faster_rcnn_of_your_data.yaml

NUM_CLASSES:

STEPS:

WEIGHTS:

DATASETS:

(5)run

cd run_train

sh train_faster_fpn_fl.sh or sh train_faster_fpn.sh

(6)analysis loss

python tools/draw_loss_one.py

you should change

log_path =''

img_path = ''

(7)test one model

CUDA_VISIBLE_DEVICES=0 python2 tools/test_net.py --cfg experiments/faster_rcnn_R-50-FPN_Focalloss.yaml TEST.WEIGHTS

output/train/yourdata/generalized_rcnn/model_final.pkl NUM_GPUS 1

(8)Batch test model

Save once every iteration 4000 times

Batch test model and choose the best one

python tools/test_list.py --model_root /path/to/model --yaml_path experiments/faster_rcnn_R-50-FPN_Focalloss.yaml --res_path /path/to/result.txt