Self-Correction-Human-Parsing
Self-Correction-Human-Parsing copied to clipboard
GoGoDuck912
trained on custom dataset
Forgive my english skill :)
This is a good work. I tried to train on a custom data set for human parsing. It work good on my dataset. The performance rise about 4%.
But current model is too big for development. @PeikeLi Do you have any idea to speed up the model?
@conan2333 There are several things you can have a try to boost the inference speed.
- replace a lightweight model as the backbone, e.g. MobileNetV3
- network pruning
- deploy on TensorRT or TVM
However, have to notice, all these approaches may slightly decrease the model performance.
I search the TensorRT. But it need a onnx model. Can you please provide the onnx model?
Please refer to #6 when you want to export the model to onnx. And also welcome to pull request if you have finished it.
Please refer to #6 when you want to export the model to onnx. And also welcome to pull request if you have finished it.
@conan2333 There are several things you can have a try to boost the inference speed.
- replace a lightweight model as the backbone, e.g. MobileNetV3
- network pruning
- deploy on TensorRT or TVM
However, have to notice, all these approaches may slightly decrease the model performance.
Please refer to #6 when you want to export the model to onnx. And also welcome to pull request if you have finished it.
Hi， I find you have offered the lightweight model of mobilenetv2, but only do a classification job. refer to some comment "invert residual network layer2,layer3,layer4,layer5" in backbone/mobilenet. So i think i should modify this script according AugumentCE2P.py . Am I right? Can you offer the mIoU score of MobileNetV2 ? Thank you very much!
