TensorFlow r1.4 or above.

NCCL 1.3.2

1. Basic train & feature extracting pipeline for deep face recognition.
2. Stable and efficient Multi-GPU training support.
3. TensorFlow Dataset API support for efficient I/O from Caffe style lists.
4. Random mirror/rotation/brightness/contrast/hue(RGB only)/saturation(RGB only) data augmentation.
5. Network support: ResNeXt, MobileNet, ShuffleNet, SENet, SphereFaceNet, LightCNN(coming soon).
6. Loss support: center loss, triplet loss, A-softmax loss*, Coco Loss(coming soon).

Results on mainstream face recognition benchmarks are coming soon.

* As far as we know, our code is the first A-softmax loss implementation in TensorFlow. @shangwenxiang claims to reproduce the LFW accuracy 99.4% on SphereFaceNet-20 with our implementation. It exceed the results for original implementation.

For training:

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python train.py --num_gpus=4 \
--model_name='Your model name.' \
--net_name='Your net name' \
--data_list_path='Your caffe-style list path.' \
--batch_size=512 \
...
...

For feature extraction:

CUDA_VISIBLE_DEVICES=0 \
python evaluate.py \
--model_name='Your model name. \
--net_name='Your net name' \
--fea_name='Your feature name' \
--data_list_path='Your caffe-style list path.' \
--batch_size=200 \
...
...

Comming soon...