torch.nn.functional.interpolate is unsupport!

[Wwupup]

I meet error when run: python demo.py -m model.onnx -i imgfile.jpg It seems that the latest version don`t support upsample operation. The onnx model in Netron:

[fengyuentau]

The problems are:

1. OpenCV DNN does not support ONNX model with dynamic input shape. However, we have a clear plan to support this in OpenCV 5.x.
2. For fixed input shape, the resize operators have fixed target size as input. So for OpenCV DNN, when it takes a input of different size, the ouputs of resize operators will have a wrong shape.

Current solution is:

1. Export this model with resize operators with scale 2 as input instead of fixed size.
2. Resize the input so that its width and height can be divisible by 64.
3. Get the lastest OpenCV: pip install opencv-python-rolling.

Code needs to be changed in libfacedetection.train:

-                size=[feats[i].size(2), feats[i].size(3)], 
+                scale = 2,

We will have a discussion on whether we should replace the one currently in the zoo with this model today.