Closed GreyZzzzzzXh closed 5 years ago
@huningxin ,yes SRGAN is in my consideration and i'll try it in the future. This project is a good reference, thanks!
I trained the SRGAN model on DIV2K dataset using tensorlayer. Reference: tensorlayer/srgan
The generator of SRGAN is called SRResNet, being like this. We can achieve a 4x upscaling. According to the origin paper, set B=16.
SRResNet uses operation PixelShuffler(DepthToSpace) which is not supported in tflite yet, see TfLite buildin ops. So i use transpose + batchToSpace + transpose instead. New ops we need: transpose, batchToSpace, tanh
A simple image demo with WebGL backend. input size [96, 96, 3] => output size [384, 384, 3] It takes several seconds to infer now. Visit https://greyzzzzzzxh.github.io/webml-examples/index.html to test.
Very nice! We need to investigate the platform API support for the three new ops. BTW, the link doesn't work for me. Would you mind to double check it?
the link doesn't work
updated it.
The link works, however it reports WebGL backend failure to me. Did I miss any thing?
Is there any error or warning? I can get results on linux and android via this link.
The failure disappears now. It's super (for super resolution). 👍
OpenVINO pre-trained models also contains a SR model (https://software.intel.com/en-us/openvino-toolkit/documentation/pretrained-models). Please take a look.
I trained FSRCNN, a small network for super resolution.
New ops we need: sigmoid, maximum, tanh, transpose_conv
Simple image demo with WebGL backend:
Reference: Accelerating the Super-Resolution Convolutional Neural Network
Visit https://greyzzzzzzxh.github.io/webml-examples/index.html to test.
Move to Feb 2019 milestone for this issue.
@GreyZzzzzzXh , SRGAN might be interested by you.