Fine-tune your trained pose detection model?

[YinghaoHuang91]

Hi Eldear,

Thanks for sharing with us such a great library!

I wonder whether it's possible to train the model directly from your trained model instead of the pretrained ImageNet model.

I tried to modify your demo code, but it always give me very weird results. Seems it fails to restore the trained model from Tensorflow saved model files.

Any clue?

Thanks!

[eldar]

Hey,

Sure, you can fine-tune from pre-trained model. I just modified pose_cfg.yaml with this line:

init_weights: path/to/snapshot-1000000 

and started training - it worked. What exactly is wrong for you?

[YinghaoHuang91]

Yeah, it runs smoothly. But it always yields more or less the same output, that's very close to 0. Like 1e-6.

I'm trying to adjust your trained model to some other tasks related to human body. And now the extra layers are built on the score_map produced by your trained MPII model. I found it's very close to 0. Maybe that's the reason.

Thank you for your response! I will try to do some normalization about the score_map.

[eldar]

I understand now. I think you shouldn't be using the prediction layers of this model, but replace them with your own ones. They should feed from the last layer of the original ResNet. You will obviously have to modify the code.