The model is copied from https://github.com/HRNet/HRNet-Image-Classification with the following two modifications:
1. The Upsample module is replaced by BilinearResize2D because there is no native support of nearest resizing in mxnet.
2. All item of HybridSequential must be a sub-class of HybridBlock, so it is not possible to put None into a
HybridSequential. I workaround this by a new class named NoneHybridBlock.
Please see models/backbones/hrnet/cls_hrnet_mx.py.The keys of the original pytorch model and the converted mxnet model are almost same. So it is possible to load the original
pretrained model with a little extra codes. And because of that, I do not provide converted params. You can get the converted params
by model.save_parameters() if you like.
I strongly suggest validating the classification accuracy before using the converted models. In order to achieve this goal,
you need to :
1. follow https://github.com/pytorch/examples/tree/master/imagenet and https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh
to prepare the imagenet validation dataset with 50k images.
2. Download pytorch params from https://github.com/HRNet/HRNet-Image-Classification.
3. Change the dataset path, config path, and params path in script models/backbones/hrnet/cls_hrnet_mx.py.
After having done these things, you can run python models/backbones/hrnet/cls_hrnet_mx.py directly, and you will see the Top1 acc result.
The following is my testing results:
| ModelName | Top1 Acc | Dataset | config path |
|---|---|---|---|
| HRNetv2-W32 | 0.78246 | ImageNet val 50k | cls_hrnet_w32_sgd_lr5e-2_wd1e-4_bs32_x100.yaml |