How to use TrainWeight.py?

[qqsh0214]

Hi @dragonfly90 I finish my own data.json for a new dataset according to annotation.ipynb and change GenerateLabelCPM.py to fit the keypoints. I want to use TrainWeight.py to fine tune. But I get the following error: RuntimeError: simple_bind error. Arguments: data: (1, 3, 368, 368) heatmaplabel: (1, 19, 46, 46) heatweight: (1, 19, 46, 46) partaffinityglabel: (1, 38, 46, 46) vecweight: (1, 38, 46, 46) This is caused by File "TrainWeight.py", line 88, in <module> cmodel.fit(cocodata, num_epoch = config.TRAIN.num_epoch, arg_params=arg_params, aux_params=aux_params) File "TrainWeight.py", line 19, in fit ('vecweight',(1,38,46,46))]) I think it may be the shapes are not matched with pre-trained model. So I change my Mydata.json to the previous data.json for COCO. But it still has the same error. Do you know the reason?

Thank you very much!

[dragonfly90]

@qqsh0214, I fix one bug in generating mask(but this should not cause your infer shape error). Could you use the following code to check if you are generating the right heatmap?

https://github.com/dragonfly90/mxnet_Realtime_Multi-Person_Pose_Estimation/blob/master/test_generateLabel.ipynb

Check if heat map, part affinity graph map, mask are generated correctly in training: test_generateLabel.ipynb

[qqsh0214]

@dragonfly90 Thanks for your reply! I have tried your test_generateLabel.ipynb and visualize the heat map and mask. In fact, they are correct. So maybe there is something else causing the infer shape error.

[dragonfly90]

@qqsh0214 I don't know the problem then, I think you should crop the image size to (368, 368).In test_generateLabel.ipynb I checked heatmap, mask but only pafmap is not checked.

[dragonfly90]

@qqsh0214 I update the test_generateLabel.ipynb to check paf map, could you have a look at it?

[qqsh0214]

@dragonfly90 OK, I see. I will try later. Thank you very much!

[dragonfly90]

@qqsh0214 could you clone the new code and change two parts of the code Except that

#COCO_to_ours_1 = [14, 6, 8, 10, 5, 7, 9, 12, 13, 11, 2, 1, 4, 3]
#COCO_to_ours_2 = [14, 6, 8, 10, 5, 7, 9, 12, 13, 11, 2, 1, 4, 3]

You may also need to change

mid_1 = [2, 9, 10, 2, 12, 13, 2, 3, 4, 3, 2, 6, 7, 6, 2, 1, 1, 15, 16]
mid_2 = [9, 10, 11, 12, 13, 14, 3, 4, 5, 17, 6, 7, 8, 18, 1, 15, 16, 17, 18]

[qqsh0214]

@dragonfly90 Thanks. Your new code works without previous error on COCO dataset. I will try to change other parts of the code to fit my own dataset.

[dragonfly90]

@qqsh0214 I fix the error. I will clean the code and talk to you later.