How to improve performance?

[ccl-private]

First, thank you for sharing. I tested your pre-trained model on my Linux(GTX 1080), but my fps was about 5. Do you have some ideas to may it real-time? Really appreciate.

[YangZeyu95]

you can resize the input image to a small one at line 103

[ccl-private]

Thanks :)
It worked! FPS reached 10 when I changed the size to 320*320.
I also notice that there are two smaller network models in vgg.py file, named vgg_a and vgg_16. But when I made some adjustments in train.py and try to train on these models, it always tells me the size is not matched. How can I train on these smaller models?

[ccl-private]

Here is where I adjusted in train.py, including two scope names to args.name.

[YangZeyu95]

That because VGGs, with different layers, downsample the image to the different scales. You need to know which downsample scale is used in the backbone network you chose and edit the 'scale' in line 78.

[ccl-private]

I have some trouble with vgg_a and vgg_16 in vgg.py. How 'scale' inline 78 related to VGG downsample. First I think it was the max pool layer that related to the 'scale', so I edited the 'scale' to 10 for vgg_16. However, it reported the same error:

I also set the 'scale' to 5,6, 8, and so on, the error report did not change even the dimension value.

[YangZeyu95]

make sure you have changed the backbone network.

[ccl-private]

I have set it to None in line: 19.

And here is where that report the error above:

[YangZeyu95]

you need to make sure that the tensor shape of ground truth and mode output have the same shape.

[ccl-private]

I am still suffering from vgg_16 and vgg_a. It seemed that only the vgg_19 can get the same output shape of the ground truth. Maybe I am in the wrong direction.

I really wish a smaller network, such as mobile-net. Will it be possible a mobile-net version in the future?

[YangZeyu95]

yes, the shape of each vggs' output is fixed, while you can change the heatmap and paf shape when generating ground truth to match vgg's output shape. you can change that use 'scale' at line 78 in 'train.py'.