Error of pretrained model

[Vegetebird]

Hello~ Thanks for your great work, I have some questions:

1. When I run the "python trainval.py -k cpn_ft_h36m_dbb -arc 3,3,3,3 -c checkpoint/gastnet --evaluate 81_frame_model.bin", I got the error of "size mismatch for layers_graph_conv.3.cat_bn.running_var: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([2048])."

2. What's the performance of single frame as input, and could yout provide the pretrained model of single frame?

[fabro66]

Hi~

1. To test the 81 receptive field model, we need to set the channels parameter to 64.

   python trainval.py -k cpn_ft_h36m_dbb -arc 3,3,3,3 -c checkpoint/gastnet --evaluate 81_frame_model.bin -ch 64

2. Single-frame model is worse than video-based models. The performance comparison can be seen in the video we provide. We do not consider providing the pretrained model of single frame.

[Vegetebird]

Hi~Thank you for your reply.

So when I train the 81 receptive field model, should I set "-arc 3,3,3,3 -ch 64" ? Train the 243 model, set "-arc 3,3,3,3,3 -ch 64" ? Train the 9 model, set "-arc 3,3 -ch 128" ? Train the 3 model, set "-arc 3 -ch 128" ?

Could you provide the performance value of single frame model? And how can I train the single frame model, just set the "-arc 1 -ch 128"?

[fabro66]

Hi~ You can set the following configurations to train different receptive field models. 1 frame model: -arc 1,1,1 -ch 128 -drop 0.10 3 frame model: -arc 3,1,1 -ch 128 -drop 0.10 9 frame model: -arc 3,3 -ch 128 (results in paper) or -arc 3,3,1 -ch 128 (higher precision but increasing computation) 27 frame model: -arc 3,3,3 -ch 128 81 frame model: -arc 3,3,3,3 -ch 64 243 frame model: -arc 3,3,3,3,3 -ch 32 --downsample 1