hi, does the movenet.pth is trained or ported from tf?

[lucasjinreal]

I tried to run, but seems default config didn't fit pretrained weights under models:

Creating model...
Skip loading parameter backbone.body.6.conv.2.weight, required shapetorch.Size([56, 336, 1, 1]), loaded shapetorch.Size([32, 192, 1, 1]). If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Skip loading parameter backbone.body.16.conv.2.weight, required shapetorch.Size([280, 1680, 1, 1]), loaded shapetorch.Size([160, 960, 1, 1]). If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Skip loading parameter backbone.body.11.conv.1.0.bias, required shapetorch.Size([672]), loaded shapetorch.Size([384]). If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Skip loading parameter backbone.body.5.conv.2.bias, required shapetorch.Size([56]), loaded shapetorch.Size([32]). If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Skip loading parameter backbone.body.16.conv.2.bias, required shapetorch.Size([280]), loaded shapetorch.Size([160]). If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Skip loading parameter backbone.body.13.conv.2.weight, required shapetorch.Size([168, 1008, 1, 1]), loaded shapetorch.Size([96, 576, 1, 1]). If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Skip loading parameter backbone.body.3.conv.2.weight, required shapetorch.Size([40, 240, 1, 1]), loaded shapetorch.Size([24, 144, 1, 1]). If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.

any idea?

[lee-man]

It's ported from tf directly. You may check the model definition. I think there's some inconsistency there.

[lucasjinreal]

@lee-man I just running instructions in your readme. Does it work out of box? On my side it's not.

[lee-man]

@jinfagang Actually I am not very sure cus this repo is taken over by other guys later. I suggest that you may check the previous few commits. I am pretty sure that at some point this repo work out correctly.

[lucasjinreal]

Thanks , for users later on comes into this repo but not able to run it, please using my fork:

https://github.com/jinfagang/movenet

original op indeed have some errors in code. I have fixed all of them. Meanwhile, I will port the model to ncnn as well.

[lee-man]

Great! Thanks for your work. I will add the link in README later.

[lucaskyle]

@lee-man

i checked ur setting.
if u guys dont change anything, it will build thunder movenet model structure and load lightning pre-train model