YuQianzi
commented
2 years ago @JosephKJ :Hello! My configs is shown below:
BASE: "../Base-RCNN-C4.yaml" MODEL: WEIGHTS: "/sharefs/wwen/incremental/iOD/IOD/configs/MSRA-R-50.pkl" MASK_ON: False RESNETS: DEPTH: 50 ROI_HEADS: NUM_CLASSES: 80 LEARN_INCREMENTALLY: True TRAIN_ON_BASE_CLASSES: True NUM_BASE_CLASSES: 40 NUM_NOVEL_CLASSES: 40 NMS_THRESH_TEST: 0.4 DATASETS: TRAIN: ("coco_2014_train",) TEST: ("coco_2014_val",)
OUTPUT_DIR: ./output/coco_base_40 SEED: 3074309 VERSION: 2
SOLVER: IMS_PER_BATCH: 16 BASE_LR: 0.02 STEPS: (140000, 160000) MAX_ITER: 180000
WG: ENABLE: False TRAIN_WARP_AT_ITR_NO: 20 WARP_LAYERS: ("module.roi_heads.res5.2.conv3.weight",) NUM_FEATURES_PER_CLASS: 100 NUM_IMAGES_PER_CLASS: 10 BATCH_SIZE: 2 USE_FEATURE_STORE: True IMAGE_STORE_LOC: '/40_image_store.pth'
DISTILL: ENABLE: False BACKBONE: True RPN: False ROI_HEADS: True ONLY_FG_ROIS: False LOSS_WEIGHT: 0.2
the result of AP50 of the base 40 classes is 30.4642, which made me confused.
I cannnot reopen issue #12. Actually, when I only train the first 40 base classes using the configs in the file "warp_faster_rcnn_R_50_C4_1x.yaml", the AP50 was about 30%. I think it is not reasonable.