- export PYTHONPATH=face/facenet/src
编译opencv: cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local -D WITH_TBB=ON -D WITH_EIGEN=ON -D WITH_OPENCL=ON -D WITH_CUDA=ON -D BUILD_opencv_gpu=ON -D BUILD_opencv_gpuarithm=ON -D BUILD_opencv_gpubgsegm=ON -D BUILD_opencv_gpucodec=ON -D BUILD_opencv_gpufeatures2d=ON -D BUILD_opencv_gpufilters=ON -D BUILD_opencv_gpuimgproc=ON -D BUILD_opencv_gpulegacy=ON -D BUILD_opencv_gpuoptflow=ON -D BUILD_opencv_gpustereo=ON -D BUILD_opencv_gpuwarping=ON ..
图片集：
| Tables | du -sh | classes | files |
|---|---|---|---|
| megaface | 489G | 672057 | |
| megdatasets | 139G | 546412 | 3440176 |
| casia | 5.4G | 10591 | 500452 |
| casdatasets | 21G | 10591 | 497549 |
1、训练 export CUDA_HOME=/usr/local/cuda export PATH=$PATH:$CUDA_HOME/bin export LD_LIBRARY_PATH=/usr/local/cuda-8.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}} export PYTHONPATH=/home/taohui/code/facenet/src
对齐
for N in {1..4}; do
python src/align/align_dataset_mtcnn.py /data/faceimg/CASIA-WebFace
/data/datasets/CASIA-WebFace_182 --image_size 182 --margin 44
--random_order --gpu_memory_fraction 0.22
& done
python src/align/align_dataset.py /data/thface/CASIA-WebFace /data/datasets/casia/casia_maxpy_dlib_182 --image_size 182
python src/my_train_softmax.py --logs_base_dir /home/taohui/face/logs/
--models_base_dir /home/taohui/face/models/facenet/
--data_dir /data/datasets/casia_maxpy_mtcnnpy_182/
--image_size 160 --model_def models.inception_resnet_v1 --optimizer RMSPROP
--learning_rate -1 --max_nrof_epochs 200 --batch_size 100 --epoch_size 1000
--keep_probability 0.8 --random_crop --random_flip
--learning_rate_schedule_file data/learning_rate_schedule_classifier_casia.txt
--weight_decay 5e-5 --center_loss_factor 1e-2 --center_loss_alfa 0.9
--lfw_dir /home/taohui/face/lfw_mtcnnpy_160
--pretrained_model /home/taohui/face/models/facenet/20170912-101532/model-20170912-101532.ckpt-200000
tensorboard --logdir=/home/taohui/face/logs/ --port 6006
megaface:
python src/align/my_align_mtcnn.py
/data/faceimg/images
/data/datasets/images_182
--image_size 182 --margin 44 --random_order --gpu_memory_fraction 0.25
2.lfw验证
for N in {1..4}; do
python src/align/align_dataset_mtcnn.py /data/faceimg/lfw /data/datasets/lfw/lfw_mtcnnpy_160
--image_size 160 --margin 32 --random_order --gpu_memory_fraction 0.25
& done
3、图片集注册并验证
纯脸（非160*160）
python test/testcompare.py /home/taohui/face/models/facenet/20170912-101532
--logfolder /home/taohui/face/logs/mycomparetest/ --resize
/home/taohui/face/datasets/images_182
找脸并对齐
python test/testcompare.py /home/taohui/face/models/facenet/20170911-190042
/data/faceimg/images/ --detect_face --resize
--logfolder /home/taohui/face/logs/mycomparetest/
4、分类器
python src/classifier.py TRAIN /home/taohui/face/datasets/images_182
/home/taohui/face/models/facenet/20170911-190042
/home/taohui/face/models/images_classifier.pkl
--batch_size 1000 --min_nrof_images_per_class 7 --nrof_train_images_per_class 5
--use_split_dataset;
python src/classifier.py CLASSIFY /home/taohui/face/datasets/images_182
/home/taohui/face/models/facenet/20170911-190042
/home/taohui/face/models/images_classifier.pkl
--batch_size 1000 --min_nrof_images_per_class 7 --nrof_train_images_per_class 5
--use_split_dataset > aaa