Bowen Cheng

I got the same issue, with single GPU (I tried both 1080TI and 2080TI). Not sure if this is related to PyTorch version or not. I used Python 3.7, PyTorch...

@gngdb Did you use the linear learning rate decay as indicated in the paper? PyTorch does not have an implementation of that learning rate decay. BTW, you also need to...

Please follow https://github.com/HRNet/Higher-HRNet-Human-Pose-Estimation#installation to install CrowdPose API. If you plan to train on CrowdPose, you will need to fix a bug in the CrowdPose API to get the correct number...

You also need to change [this line](https://github.com/Jeff-sjtu/CrowdPose/blob/785e70d269a554b2ba29daf137354103221f479e/crowdpose-api/PythonAPI/crowdposetools/cocoeval.py#L620). Add `int` before `np.round`.

Hi, if you are still having this problem, please make sure you install the crowdpose api and it is in your PYTHONPATH.

Unfortunately, we do not have this model. WITH_CENTER is experimental code and we did not use it in our final model.

You can follow CrowdPose to avoid segmentation mask https://github.com/HRNet/HigherHRNet-Human-Pose-Estimation/blob/b4610aecaa5cf3de3cd69bfb13c7c79c8d514c7c/lib/dataset/CrowdPoseKeypoints.py#L123-L129

Grouping is done on CPU. Not available means FLOPs is not calculated for this layer (which is ReLU).

Yes, it is already included in the json file: https://github.com/HRNet/HigherHRNet-Human-Pose-Estimation/blob/master/lib/dataset/COCODataset.py#L290

It is tested with the torchvision accompanied with PyTorch v1.4.0. You can install it with `conda install pytorch==1.4.0 torchvision==0.5.0 cudatoolkit=10.1 -c pytorch`