This is a slightly tiny project espacailly for face recognition,
where we seperate several parts(networks,metrics & loss functions, data loaders) into different folders accordingly. moreover, in order to be extensible,
major training modules are written in modules.py which includes:
IOFactoryfor logging, reading and saving.OptimFactoryfor params optimization and lr schedulingHeaderfor pairwise metric learning or metrics with classification. In this project ,we use colPallelLinear and mpu vocab pallel cross entropy to build a pallel version Face train project. 1 To get training data,you can download from web link in dbfile.txt 2 we have make a scripy to convert images in folders to tfrecord files,which can help speed up data read in HDD disk.the scripy is data/create_tfrecord.py. Also if you have fast SSD disk, ignore this step. rebuild the dataloader in solver_ddp.py. 3 The overall training protocol has been written in file solver_ddp.py, while the configurations are set in file config.yaml, using shell run_face_main.sh to train a model.Make sure set right host IP for model pallel env. In config.yaml, set train_data for face_ms1s/images floder, set aldata_folder for valid data folder(lfw、agedb) download from link in dbfile.txt this project use full GPU in host,we have set defaults gpu to be 8. 4 After training ,you can use validation/vertification_owndata.py to get vertification performance of self-models on agedb_30 lfw and cfp_fp.(the validation has provided in dbfile.txt) 5 Experiment shows a consistent result with InsightFace, some more networks will be added!
The code has been tested using Pytorch 1.4.0 under Centos7.4 with python3.6. we reommand use docker image horovod/horovod:0.19.3-tf2.1.0-torch-mxnet1.6.0-py3.6-gpu . OS.txt give s information about OS and NVIDIA-driver. reqirements.txt give information about packages.