train my own dataset , but AP is very low

[ing-NT]

Recently I run HRnet keypoint detection code, I didn't use any PRETRAINED MODEL (including ImageNet and COCO datasets), then train and test（use groundtruth bbox） my own datasets(My own dataset is already in coco format) ，AP only around 1%. However, using pretrained (coco trained in the paper), AP can reach about 88%. I don't know why, and do you know the reason?@leoxiaobin My configuration file is as follows： AUTO_RESUME: true CUDNN: BENCHMARK: true DETERMINISTIC: false ENABLED: true DATA_DIR: '' GPUS: (0,) OUTPUT_DIR: 'output' LOG_DIR: 'log' WORKERS: 24 PRINT_FREQ: 100 DATASET: COLOR_RGB: true DATASET: 'coco' DATA_FORMAT: jpg FLIP: true NUM_JOINTS_HALF_BODY: 8 PROB_HALF_BODY: 0.3 ROOT: 'data/coco/' ROT_FACTOR: 45 SCALE_FACTOR: 0.35 TEST_SET: 'val2017' TRAIN_SET: 'train2017' MODEL: INIT_WEIGHTS: true NAME: pose_hrnet NUM_JOINTS: 17 PRETRAINED: ' ' TARGET_TYPE: gaussian IMAGE_SIZE:

• 192
• 256 HEATMAP_SIZE:
• 48
• 64 SIGMA: 2 EXTRA: PRETRAINED_LAYERS:
  • 'conv1'
  • 'bn1'
  • 'conv2'
  • 'bn2'
  • 'layer1'
  • 'transition1'
  • 'stage2'
  • 'transition2'
  • 'stage3'
  • 'transition3'
  • 'stage4' FINAL_CONV_KERNEL: 1 STAGE2: NUM_MODULES: 1
    NUM_BRANCHES: 2 BLOCK: BASIC NUM_BLOCKS:
    • 4
    • 4 NUM_CHANNELS:
    • 32
    • 64 FUSE_METHOD: SUM STAGE3: NUM_MODULES: 4 NUM_BRANCHES: 3 BLOCK: BASIC NUM_BLOCKS:
    • 4
    • 4
    • 4 NUM_CHANNELS:
    • 32
    • 64
    • 128 FUSE_METHOD: SUM STAGE4: NUM_MODULES: 3 NUM_BRANCHES: 4 BLOCK: BASIC NUM_BLOCKS:
    • 4
    • 4
    • 4
    • 4 NUM_CHANNELS:
    • 32
    • 64
    • 128
    • 256 FUSE_METHOD: SUM LOSS: USE_TARGET_WEIGHT: true TRAIN: BATCH_SIZE_PER_GPU: 32 SHUFFLE: true BEGIN_EPOCH: 0 END_EPOCH: 210 OPTIMIZER: adam LR: 0.001 LR_FACTOR: 0.1 LR_STEP:
• 170
• 200 WD: 0.0001 GAMMA1: 0.99 GAMMA2: 0.0 MOMENTUM: 0.9 NESTEROV: false TEST: BATCH_SIZE_PER_GPU: 32 COCO_BBOX_FILE: 'data/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: '' NMS_THRE: 1.0 OKS_THRE: 0.9 USE_GT_BBOX: true FLIP_TEST: true POST_PROCESS: true SHIFT_HEATMAP: true DEBUG: DEBUG: true SAVE_BATCH_IMAGES_GT: true SAVE_BATCH_IMAGES_PRED: true SAVE_HEATMAPS_GT: true SAVE_HEATMAPS_PRED: true

[khandriod]

hi, how you used this architecture on a custom dataset?

Thanks

[ing-NT]

Make the same data set format as COCO2017 and put it in the same place as the code

------------------ 原始邮件 ------------------ 发件人: "leoxiaobin/deep-high-resolution-net.pytorch" @.>; 发送时间: 2021年12月23日(星期四) 上午10:41 @.>; @.@.>; 主题: Re: [leoxiaobin/deep-high-resolution-net.pytorch] train my own dataset , but AP is very low (#264)

hi, how you used this architecture on a custom dataset?

Thanks

— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: @.***>

[khandriod]

Thanks for your reply. I do not have labeled data. So, should I use pre-trained model weights for my dataset? and where are the results stored (images with pose skeleton)?

[ing-NT]

How can you train your own data without  the label?? train result stored:output/coco/pose_hrnet/w32_256x192_adam_lr1e-3/results

------------------ 原始邮件 ------------------ 发件人: "leoxiaobin/deep-high-resolution-net.pytorch" @.>; 发送时间: 2021年12月23日(星期四) 下午2:30 @.>; @.@.>; 主题: Re: [leoxiaobin/deep-high-resolution-net.pytorch] train my own dataset , but AP is very low (#264)

Thanks for your reply. I do not have labeled data. So, should I use pre-trained model weights for my dataset? and where are the results stored (images with pose skeleton)?

— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: @.***>

[khandriod]

I don't want to train from scratch. That's why I want to use a pre-trained model (transfer learning, if I am not wrong). Can I use pre-trained weight on a custom dataset?

Thanks

[ing-NT]

Of course you can

------------------ 原始邮件 ------------------ 发件人: "leoxiaobin/deep-high-resolution-net.pytorch" @.>; 发送时间: 2021年12月23日(星期四) 下午3:07 @.>; @.@.>; 主题: Re: [leoxiaobin/deep-high-resolution-net.pytorch] train my own dataset , but AP is very low (#264)

I don't want to train from scratch. That's why I want to use a pre-trained model (transfer learning, if I am not wrong). Can I use pre-trained weight on a custom dataset?

Thanks

— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: @.***>

[khandriod]

could you please write down the steps that I should follow in order to achieve the goal? my goal is to use HRNET to estimate the human pose on my dataset using pre-trained weights.

Thanks

[Aliiiu]

@ing-NT Thanks for the feedbacks, please i will like to know if i can train this model, using a pretrained model, on a images with lesser joint label. If not how can i go about it without it causing conflict with the pretrained model.

[inever789]

Make the same data set format as COCO2017 and put it in the same place as the code

Hello, may I ask, if the custom dataset is used, does each picture need to be marked manually and then converted into coco format? Thank you very much.

[inever789]

could you please write down the steps that I should follow in order to achieve the goal? my goal is to use HRNET to estimate the human pose on my dataset using pre-trained weights.

Thanks

Hello, have you implemented training on custom datasets now? Could you tell me the operation procedure? I'm having the same problem right now. Thank you very much.