This project is forked from davidsandberg/facenet, which is a TensorFlow implementation of the face recognizer described in the papers "FaceNet: A Unified Embedding for Face Recognition and Clustering", "A Discriminative Feature Learning Approach for Deep Face Recognition", "Deep Face Recognition".
In comparison to the original project:
- more args are added to the train_softmax script to improve the training
- a new network implementation: SqueezeNet v1.1, which has 2.4 times less computation thant the original one.
- new pre-trained models for tensorflow and tensorflow.js
- new contributed scripts
Most of these changes are done in train_softmax.py:
- added "--continue_ckpt_dir" to allow continue training from one checkpoint
- added "--random_brightness" for random brightness data augmentation
- added "--snapshot_at_step" to make a checkpoint at the specific step
- added "--lfw_epoch_interval" to control how often to test against the lfw dataset
- added "--summary_iteration_interval" to control how often the summary is recorded
- made "--data_dir argument" support more train datasets (separated by comma)
- fixed a bug about "--nrof_preprocess_threads", it was not used and hardcoded to 4 before
The learning rate can be adjusted manually either by changing the schedule file on the fly (the file content is read in each epoch, train will be stopped if a 0 learning rate is found) or by changing the "--learning_rate" and continue the training with "--continue_ckpt_dir".
The authors of the original paper improved the SqueezeNet, based on their web page, SqueezeNet_v1.1 should be much faster than the original one.
The implementation is src/models/squeezenet_v1_1.py which is based on src/models/squeezenet.py with new network structure. It also uses He weights initialization instead of Xavier.
| Model | LFW accuracy | Training dataset | Architecture |
|---|---|---|---|
| squeezenet_10_vggface2 (tensorflow.js) | VGGFace2 | SqueezeNet | |
| squeezenet_v1_1_vggface2 (tensorflow.js) | VGGFace2 | SqueezeNet v1.1 | |
| inception_resnet_v2_vggface2 | 0.992 | VGGFace2 | Inception Resnet v2 |
| squeezenet_v1_1_vggface2 | 0.980 | VGGFace2 | SqueezeNet v1.1 |
- contributed/tensorflow_js_demo.zip: demo project for face recognition with tensorflow.js on browser (requires webcam, nodejs, pre-trained tensorflow.js models)
- contributed/photo_face_recognition.py: draw face recognition boxes on a single photo
- contributed/vggface2_gender.py: use vggface2 embedding feature vectors to train a gender classifier
- src/validate_on_lfw.py: add Precision and Recall measure in lfw validation