lower mAP

[JWSunny]

hello，I modified the config xxx.py file into a yaml file and used the hrnet or higherhrnet framework code for training. I found that the mAP on the coco validation dataset was only about 0.51.

2022-08-22 15:25:59,519 Epoch: [179][0/2341] Time 3.316s (3.316s) Speed 19.3 samples/s Data 2.263s (2.263s) Loss 0.00042 (0.00042) Accuracy 0.751 (0.751)
2022-08-22 15:30:13,151 Epoch: [179][300/2341] Time 0.813s (0.854s) Speed 78.8 samples/s Data 0.000s (0.019s) Loss 0.00032 (0.00038) Accuracy 0.803 (0.748)
2022-08-22 15:34:33,721 Epoch: [179][600/2341] Time 0.813s (0.861s) Speed 78.7 samples/s Data 0.000s (0.014s) Loss 0.00039 (0.00038) Accuracy 0.725 (0.747)
2022-08-22 15:42:29,909 Epoch: [179][900/2341] Time 1.648s (1.103s) Speed 38.8 samples/s Data 0.000s (0.012s) Loss 0.00035 (0.00038) Accuracy 0.737 (0.746)
2022-08-22 15:50:49,289 Epoch: [179][1200/2341] Time 1.665s (1.243s) Speed 38.4 samples/s Data 0.000s (0.013s) Loss 0.00035 (0.00038) Accuracy 0.756 (0.747)
2022-08-22 15:59:08,989 Epoch: [179][1500/2341] Time 1.639s (1.328s) Speed 39.1 samples/s Data 0.000s (0.013s) Loss 0.00035 (0.00038) Accuracy 0.775 (0.747)
2022-08-22 16:07:28,549 Epoch: [179][1800/2341] Time 1.668s (1.384s) Speed 38.4 samples/s Data 0.000s (0.013s) Loss 0.00041 (0.00038) Accuracy 0.752 (0.748)
2022-08-22 16:15:47,927 Epoch: [179][2100/2341] Time 1.674s (1.424s) Speed 38.2 samples/s Data 0.000s (0.012s) Loss 0.00033 (0.00038) Accuracy 0.785 (0.748)
2022-08-22 16:22:31,716 Test: [0/199] Time 1.750 (1.750) Loss 0.0004 (0.0004) Accuracy 0.816 (0.816)
2022-08-22 16:24:33,818 => writing results json to LiteHRNet_w18_output/coco/HigherLiteHRNet/LiteHRNet_w18_256x256_coco_correct_lr1e-3/results/keypoints_val2017_results_0.json
2022-08-22 16:24:44,456 | Arch | AP | Ap .5 | AP .75 | AP (M) | AP (L) | AR | AR .5 | AR .75 | AR (M) | AR (L) |
2022-08-22 16:24:44,457 |---|---|---|---|---|---|---|---|---|---|---|
2022-08-22 16:24:44,457 | HigherLiteHRNet | 0.511 | 0.807 | 0.544 | 0.501 | 0.530 | 0.557 | 0.830 | 0.598 | 0.539 | 0.583 |

[JWSunny]

config.yaml 内容如下：

AUTO_RESUME: true
CUDNN:
BENCHMARK: true
DETERMINISTIC: false
ENABLED: true
DATA_DIR: ''
GPUS: (0,1)
OUTPUT_DIR: 'LiteHRNet_w18_output'
LOG_DIR: 'LiteHRNet_w18_log'
WORKERS: 8
PRINT_FREQ: 300

DATASET:
COLOR_RGB: false
DATASET: 'coco'
ROOT: '/mnt/share/COCO/'
TEST_SET: 'val2017'
TRAIN_SET: 'train2017'
NUM_JOINTS_HALF_BODY: 8
PROB_HALF_BODY: 0.3
FLIP: true
ROT_FACTOR: 45
SCALE_FACTOR: 0.35
MODEL:
NAME: 'LiteHRNet'
MODEL_FILE: ''
INIT_WEIGHTS: true
IMAGE_SIZE:

• 256
• 256
  HEATMAP_SIZE:
• 64
• 64
  SIGMA: 2
  NUM_JOINTS: 17
  BASE_CHANNEL: 40
  TARGET_TYPE: 'gaussian'
  RATIO: 0.5
  NUM_STAGES: 3
  STAGE_REPEATS:
• 2
• 4
• 2
  STAGE_BRANCHES:
• 2
• 3
• 4
  STAGE_BLOCKS:
• 2
• 2
• 2
  MODULE_TYPE:
• 'LITE'
• 'LITE'
• 'LITE'
  WITH_FUSE:
• True
• True
• True
  REDUCE_RATIOS:
• 8
• 8
• 8
  WITH_HEAD: True

LOSS:
USE_TARGET_WEIGHT: true
TRAIN:
BATCH_SIZE_PER_GPU: 32
SHUFFLE: true
BEGIN_EPOCH: 0
END_EPOCH: 210
OPTIMIZER: 'adam'
LR: 0.002
LR_FACTOR: 0.1
LR_STEP:

• 160
• 190
  WD: 0.0001
  GAMMA1: 0.99
  GAMMA2: 0.0
  MOMENTUM: 0.9
  NESTEROV: false
  TEST:
  BATCH_SIZE_PER_GPU: 32
  COCO_BBOX_FILE: '/mnt/share/COCO/person_detection_results/COCO_val2017_detections_AP_H_56_person.json'
  BBOX_THRE: 1.0
  IMAGE_THRE: 0.0
  IN_VIS_THRE: 0.2
  MODEL_FILE: 'LiteHRNet_w18_output/coco/LiteHRNet/LiteHRNet_w18_256x256_coco_better_lr1e-3/model_best.pth'
  NMS_THRE: 1.0
  OKS_THRE: 0.9
  FLIP_TEST: true
  POST_PROCESS: true
  BLUR_KERNEL: 11
  USE_GT_BBOX: true
  DEBUG:
  DEBUG: true
  SAVE_BATCH_IMAGES_GT: true
  SAVE_BATCH_IMAGES_PRED: true
  SAVE_HEATMAPS_GT: false
  SAVE_HEATMAPS_PRED: false