Skip to content

jacke121/arcface_retinaface_mxnet2onnx

master
Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?
Code
This branch is up to date with zheshipinyinMc/arcface_retinaface_mxnet2onnx:master.

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 

Update2 models

arcface 提取码:jks7
retinaface 提取码:otx7

Update1 fix_gamma

retinaface_mnet025_v1
retinaface_mnet025_v2
In mxnet symbol, BN has fix_gamma, if fix_gamma is true, then set gamma to 1 and its gradient to 0, you can find this in mxnet API.
In retinaface_mnet025_v1, fix_gamma in 'conv_3_dw_batchnorm' is true,but its value is 0.000007107922556315316(you can see weight by Netron).However, forward mxnet model, the gamma of 'conv_3_dw_batchnorm' is 1.This bug may cause mxnet output is different from onnx model.

fix bn gamma model have upload(/Retinaface/retinaface_mnet025_v1, /Retinaface/retinaface_mnet025_v2).

Update

Retinaface fixed softmax bug.
Upsample is implemented using Resize.
Upsample is implemented using ConvTranspose.

arcface_retinaface_mxnet2onnx

arcface and retinaface model convert mxnet to onnx

environment

Ubuntu 18.04
MxNet 1.5.0
onnx 1.7.0 (protobuf 3.0.0)
onnxruntime 1.3.0
Python 3.6.9
cv2 3.3.1

tested models

arcface:from ZQCNN mobilefacenet-res2-6-10-2-dim512
retinaface:mnet.25
arcface/model-r34-amf-slim and retinaface-R50 also can convert successfully, but model file is too big to upload.

arcface

Insightface中ArcFace MxNet2ONNX踩坑

retinaface

Insightface中Retinaface MxNet2ONNX踩坑

Results

arcface onnx_mxnet_output
retinaface onnx_mxnet_output

reference

insightface

About

arcface and retinaface model convert mxnet to onnx.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%