为什么我这边训练出来的结果很差呢？比Yolact和Solov2差远了

[caolianxue]

这个是配置文件：
BASE: Fast-COCO-InstanceSegmentation.yaml
MODEL:
BACKBONE:
NAME: "build_resnet_vd_backbone"
WEIGHTS: "checkpoints/model_0004999.pth"
RESNETS:
DEFORM_ON_PER_STAGE: [ False, False, True, True ] # dcn on res4, res5
FASTINST:
DEC_LAYERS: 3
DATASETS:
TRAIN: ("gangban_train",)
TEST: ("gangban_val",)
OUTPUT_DIR: "output/fastinst_r50-vd-dcn_ppm-fpn_bs16_50ep_x3_640_gangban"

SOLVER:
IMS_PER_BATCH: 4
MAX_ITER: 10000
WARMUP_ITERS: 500

[caolianxue]

这个是solov2的结果：