a caffe implementation of mobilefacenet,with the record of trainnig log
update the deploy.prototxt & training log of Amsoftmax on Webface @ 24/4/2018
update the solver.prototxt & training log of 6W iteration in ./proto_revise/ @ 30/4/2018
update @ 7/5/2018:
since there is a re-implementation of mobilefacenet by insightface: https://github.com/deepinsight/insightface
you can train on public dataset using insightface then transform the mxnet model to caffe model using MxNet2Caffe
pay attention that:
1.there is something to be fixed in insightface/src/symbols/fmobilefacenet.py
you can just download the fmobilefacent.py in this project and replace it in insightface.
then type:
CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train_softmax.py --network y1 --loss-type 4 --margin-m 0.5 --data-dir ../datasets/faces_ms1m_112x112 --prefix ../model-y1 --emb-size = 128
to train the mobilefacenet with insightface.
2.if you are using cudnn,the group convolution with cause a memory error while using Mxnet2Caffe so you have to add engine:CAFFE in each group convolution layer
or you can download the prototxt_basic.py in this project and replace it in MXNet2Caffe
Links:
paper:MobileFaceNets: Efficient CNNs for Accurate Real-time Face Verification on Mobile Devices
amsoftmax:happynear/AMSoftmax
depthwise convolution:yonghenglh6/DepthwiseConvolution
mobilenetv2-caffe:shicai/MobileNet-Caffe