[arcface_mxnet] Can't convert arcface fp16 model to onnx format.

[nhmnhat1997]

• I trained my arcface model using arcface mxnet with fp16 precision. Training code: CUDA_VISIBLE_DEVICES='0,1,2,3,4,5' python -u train_parall.py --network r100 --loss arcface --dataset webface --fp16-scale 1.0

• After finish training, I tried to convert it to onnx format using mxnet's onnx converter. Code:

  
  import onnx
  import mxnet as mx
  from onnx import helper
  from mxnet import onnx as onnx_mxnet

sym, arg_params, aux_params = mx.model.load_checkpoint('./r100-arcface-webface/model', 20) input_shape = (1,) + tuple([int(x) for x in '3,112,112'.split(',')] ) all_args = {} all_args.update(arg_params) all_args.update(aux_params) onnx_mxnet.export_model(sym, all_args, [input_shape], np.float32, './r100_arcface.onnx')

- It returns this error:
![image](https://user-images.githubusercontent.com/40227850/159461353-0a282634-f2f1-4159-a2c9-37139d95aa51.png)

- Can you help me ? Thanks for your help!

- Dependencies:

mxnet==1.9.0 onnx==1.8.0

[anxiangsir]

create a new fp32 symbol and cast params to fp32.