arcface_retinaface_mxnet2onnx
arcface_retinaface_mxnet2onnx copied to clipboard
arcface and retinaface model convert mxnet to onnx.
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