请问resnet50 kl loss 256*192配置

[zonghaofan]

我用resnet50 kl loss 256*192 sigma=1训练出来ap是0.685，没有文章0.71高。

[leeyegy]

Could you provide more details you are using? The recommended setting is in experiments/coco/resnet/simdr/original/nmt_res50_256x192_d256x3_adam_lr1e-3_deconv3_split2.yaml. The parameter sigma should do nothing with the proposed method.

[zonghaofan]

Could you provide more details you are using? The recommended setting is in experiments/coco/resnet/simdr/original/nmt_res50_256x192_d256x3_adam_lr1e-3_deconv3_split2.yaml. The parameter sigma should do nothing with the proposed method.

AUTO_RESUME: true CUDNN: BENCHMARK: true DETERMINISTIC: false ENABLED: true DATA_DIR: '' GPUS: (0,1,2,3) OUTPUT_DIR: 'output' LOG_DIR: 'log' WORKERS: 24 PRINT_FREQ: 20

DATASET: COLOR_RGB: false DATASET: 'coco' ROOT: '/share/public/coco-icp' TEST_SET: 'val2017' TRAIN_SET: 'train2017' FLIP: true ROT_FACTOR: 40 SCALE_FACTOR: 0.3 MODEL: NAME: 'pose_resnet_upfree' SIMDR_SPLIT_RATIO: 2.0 HEAD_INPUT: 1344 PRETRAINED: '/data/pretrained/imagenet/resnet50-19c8e357.pth' IMAGE_SIZE:

• 192
• 256 HEATMAP_SIZE:
• 192
• 256 SIGMA: 1 NUM_JOINTS: 17 COORD_REPRESENTATION: 'sa-simdr'

  COORD_REPRESENTATION: 'simdr'

  EXTRA: FINAL_CONV_KERNEL: 1 CHANNEL_PER_JOINT: 28 NUM_LAYERS: 50 LOSS: USE_TARGET_WEIGHT: true TYPE: 'KLDiscretLoss'

  TYPE: 'NMTCritierion'

  TRAIN: BATCH_SIZE_PER_GPU: 32 SHUFFLE: true BEGIN_EPOCH: 0 END_EPOCH: 140 OPTIMIZER: 'adam' LR: 0.001 LR_FACTOR: 0.1 LR_STEP:

• 90
• 120 WD: 0.0001 GAMMA1: 0.99 GAMMA2: 0.0 MOMENTUM: 0.9 NESTEROV: false TEST: BATCH_SIZE_PER_GPU: 32 COCO_BBOX_FILE: '/share/public/coco-icp/annotations/COCO_val2017_detections_AP_H_56_person.json' BBOX_THRE: 1.0 IMAGE_THRE: 0.0 IN_VIS_THRE: 0.01 MODEL_FILE: '/home/fanzonghao/AI/papercode/SimDR/pretrained/pose_upfree_resnet_50_256x192_cpj28.pth' NMS_THRE: 1.0 OKS_THRE: 0.9 FLIP_TEST: false POST_PROCESS: false SHIFT_HEATMAP: true USE_GT_BBOX: true DEBUG: DEBUG: true SAVE_BATCH_IMAGES_GT: true SAVE_BATCH_IMAGES_PRED: true SAVE_HEATMAPS_GT: true SAVE_HEATMAPS_PRED: true

[leeyegy]

The paper result 71.4 AP can be reproduced by using the recommended setting(experiments/coco/resnet/simdr/original/nmt_res50_256x192_d256x3_adam_lr1e-3_deconv3_split2.yaml), which keeps the three-layers deconvolutional modules.

[leeyegy]

It seems that the yaml file you are using has been changed a lot. For example, the FLIP_TEST should be set as TRUE, which may affect the performance. AND, if you want to reproduce the 70.8 AP(the version without any deconvolutional module), we recommend you to follow the setting in experiments/coco/resnet/simdr/upsample_free/nmt_res50_256x192_d256x3_adam_lr1e-3_split_2_cpj28.yaml

[zonghaofan]

It seems that the yaml file you are using has been changed a lot. For example, the FLIP_TEST should be set as TRUE, which may affect the performance. AND, if you want to reproduce the 70.8 AP(the version without any deconvolutional module), we recommend you to follow the setting in experiments/coco/resnet/simdr/upsample_free/nmt_res50_256x192_d256x3_adam_lr1e-3_split_2_cpj28.yaml

ok thanks

[leeyegy]

Feel free to reopen this issue, if there is any problems in reproducing.