ZheC / Realtime_Multi-Person_Pose_Estimation

Code repo for realtime multi-person pose estimation in CVPR'17 (Oral)
Other
5.1k stars 1.37k forks source link

"meta.write_number" is always 0 #218

Open allan030213 opened 5 years ago

allan030213 commented 5 years ago

Hi! Thanks for your brilliant work!

I encounter some problems during training with my own dataset. I've successfully build up an annotation file and LMDB, but during training, "meta.write_numbe"r is always 0, even total loss is down to about 20. Does it mean I didn't train successfully even 1 image? Since there is no error message, how can I debug it?

the following is my training log: I0212 15:30:31.223089 1433 caffe.cpp:217] Using GPUs 1 I0212 15:30:31.272265 1433 caffe.cpp:222] GPU 1: GeForce GTX 750 Ti I0212 15:30:31.502509 1433 solver.cpp:48] Initializing solver from parameters: base_lr: 4e-05 display: 5 max_iter: 600000 lr_policy: "step" gamma: 0.333 momentum: 0.9 weight_decay: 0.0005 stepsize: 136106 snapshot: 2000 snapshot_prefix: "/usr/local/RMPPE/training/model_allan/" solver_mode: GPU device_id: 1 net: "pose_train_test.prototxt" train_state { level: 0 stage: "" } I0212 15:30:31.502662 1433 solver.cpp:91] Creating training net from net file: pose_train_test.prototxt I0212 15:30:31.505437 1433 net.cpp:58] Initializing net from parameters: . . . I0213 13:00:19.271257 1442 cpm_data_transformer.cpp:73] dataset: COCO; img_size: [720 x 960]; meta.annolist_index: 7; meta.write_number: 0; meta.total_write_number: 12; meta.epoch: 7883 I0213 13:00:29.130000 1442 cpm_data_transformer.cpp:73] dataset: COCO; img_size: [720 x 960]; meta.annolist_index: 7; meta.write_number: 0; meta.total_write_number: 12; meta.epoch: 7884 I0213 13:00:33.296797 1433 solver.cpp:228] Iteration 23650, loss = 22.878

Thanks for your help!