This is a TensorFlow implementation of the face recognizer described in the paper "FaceNet: A Unified Embedding for Face Recognition and Clustering". The project also uses ideas from the paper "Deep Face Recognition" from the Visual Geometry Group at Oxford.
The code is tested using Tensorflow r1.7 under Ubuntu 14.04 with Python 2.7 and Python 3.5. The test cases can be found here and the results can be found here.
| Date | Update |
|---|---|
| 2018-04-10 | Added new models trained on Casia-WebFace and VGGFace2 (see below). Note that the models uses fixed image standardization (see wiki). |
| 2018-03-31 | Added a new, more flexible input pipeline as well as a bunch of minor updates. |
| 2017-05-13 | Removed a bunch of older non-slim models. Moved the last bottleneck layer into the respective models. Corrected normalization of Center Loss. |
| 2017-05-06 | Added code to train a classifier on your own images. Renamed facenet_train.py to train_tripletloss.py and facenet_train_classifier.py to train_softmax.py. |
| 2017-03-02 | Added pretrained models that generate 128-dimensional embeddings. |
| 2017-02-22 | Updated to Tensorflow r1.0. Added Continuous Integration using Travis-CI. |
| 2017-02-03 | Added models where only trainable variables has been stored in the checkpoint. These are therefore significantly smaller. |
| 2017-01-27 | Added a model trained on a subset of the MS-Celeb-1M dataset. The LFW accuracy of this model is around 0.994. |
| 2017‑01‑02 | Updated to run with Tensorflow r0.12. Not sure if it runs with older versions of Tensorflow though. |
| Model name | LFW accuracy | Training dataset | Architecture |
|---|---|---|---|
| 20180408-102900 | 0.9905 | CASIA-WebFace | Inception ResNet v1 |
| 20180402-114759 | 0.9965 | VGGFace2 | Inception ResNet v1 |
NOTE: If you use any of the models, please do not forget to give proper credit to those providing the training dataset as well.
The code is heavily inspired by the OpenFace implementation.
The CASIA-WebFace dataset has been used for training. This training set consists of total of 453 453 images over 10 575 identities after face detection. Some performance improvement has been seen if the dataset has been filtered before training. Some more information about how this was done will come later. The best performing model has been trained on the VGGFace2 dataset consisting of ~3.3M faces and ~9000 classes.
One problem with the above approach seems to be that the Dlib face detector misses some of the hard examples (partial occlusion, silhouettes, etc). This makes the training set too "easy" which causes the model to perform worse on other benchmarks. To solve this, other face landmark detectors has been tested. One face landmark detector that has proven to work very well in this setting is the Multi-task CNN. A Matlab/Caffe implementation can be found here and this has been used for face alignment with very good results. A Python/Tensorflow implementation of MTCNN can be found here. This implementation does not give identical results to the Matlab/Caffe implementation but the performance is very similar.
Currently, the best results are achieved by training the model using softmax loss. Details on how to train a model using softmax loss on the CASIA-WebFace dataset can be found on the page Classifier training of Inception-ResNet-v1 and .
A couple of pretrained models are provided. They are trained using softmax loss with the Inception-Resnet-v1 model. The datasets has been aligned using MTCNN.
The accuracy on LFW for the model 20180402-114759 is 0.99650+-0.00252. A description of how to run the test can be found on the page Validate on LFW. Note that the input images to the model need to be standardized using fixed image standardization (use the option --use_fixed_image_standardization when running e.g. validate_on_lfw.py).
Number of classes in training set: 10575 Number of examples in training set: 462406 Number of classes in validation set: 10575 Number of examples in validation set: 29136
160 + 22
160 + 64 python src/align/align_dataset_mtcnn.py ./dataset/CASIA-WebFace ./dataset/casia_mtcnnpy_182/ --image_size 182 --margin 44
python src/train_softmax.py
--logs_base_dir ./log/facenet/
--models_base_dir ./model/
--pretrained_model ./model/20191029-134852/model-20191029-134852.ckpt-90
--data_dir ./dataset/casia_mtcnnpy_182/
--image_size 160
--model_def models.inception_resnet_v1
--lfw_dir /home/rqy/Project/data/lfw_mtcnnpy_160/
--optimizer ADAM
--learning_rate -1
--max_nrof_epochs 150
--keep_probability 0.8
--random_crop
--random_flip
--use_fixed_image_standardization
--learning_rate_schedule_file data/learning_rate_schedule_classifier_casia.txt
--weight_decay 5e-4
--embedding_size 512
--lfw_distance_metric 1
--lfw_use_flipped_images
--lfw_subtract_mean
--validation_set_split_ratio 0.05
--validate_every_n_epochs 5
--prelogits_norm_loss_factor 5e-4
sh
python src/train_softmax_mobilenet.py
--logs_base_dir ./log/facenet/
--models_base_dir ./model/
--data_dir ./dataset/casia_mtcnnpy_182/
--batch_size 90
--image_size 160
--lfw_dir /home/rqy/Project/data/lfw_mtcnnpy_160/
--pretrained_model ./model/20191101-100725/model-20191101-100725.ckpt-150
--optimizer ADAM
--learning_rate -1
--max_nrof_epochs 150
--keep_probability 0.8
--random_crop
--random_flip
--use_fixed_image_standardization
--learning_rate_schedule_file data/learning_rate_schedule_classifier_casia_third.txt
--weight_decay 5e-4
--embedding_size 512
--lfw_distance_metric 1
--lfw_use_flipped_images
--lfw_subtract_mean
--validation_set_split_ratio 0.05
--validate_every_n_epochs 5
--prelogits_norm_loss_factor 5e-4
sh quant
python src/train_softmax_mobilenet.py
--logs_base_dir ./log/facenet/
--models_base_dir ./model/
--data_dir ./dataset/casia_mtcnnpy_182/
--batch_size 90
--image_size 160
--lfw_dir /home/rqy/Project/data/lfw_mtcnnpy_160/
--pretrained_model ./model/20191101-100725/model-20191101-100725.ckpt-150
--optimizer ADAM
--learning_rate -1
--max_nrof_epochs 150
--keep_probability 0.8
--random_crop
--random_flip
--use_fixed_image_standardization
--learning_rate_schedule_file data/learning_rate_schedule_classifier_casia_third.txt
--weight_decay 5e-4
--embedding_size 512
--lfw_distance_metric 1
--lfw_use_flipped_images
--lfw_subtract_mean
--validation_set_split_ratio 0.05
--validate_every_n_epochs 5
--prelogits_norm_loss_factor 5e-4
Saving variables Variables saved in 0.34 seconds Runnning forward pass on LFW images ........................ Accuracy: 0.91967+-0.01655 Validation rate: 0.46067+-0.04341 @ FAR=0.00100 Saving statistics Epoch: [90][1/1000] Time 0.155 Loss 3.817 Xent 2.415 RegLoss 1.403 Accuracy 0.594 Lr 0.00050 Cl 0.279 Epoch: [90][101/1000] Time 0.200 Loss 4.485 Xent 3.081 RegLoss 1.404 Accuracy 0.375 Lr 0.00050 Cl 0.299 Epoch: [90][201/1000] Time 0.195 Loss 5.035 Xent 3.633 RegLoss 1.402 Accuracy 0.375 Lr 0.00050 Cl 0.318 Epoch: [90][301/1000] Time 0.194 Loss 5.070 Xent 3.669 RegLoss 1.401 Accuracy 0.469 Lr 0.00050 Cl 0.320 Epoch: [90][401/1000] Time 0.199 Loss 4.102 Xent 2.699 RegLoss 1.403 Accuracy 0.625 Lr 0.00050 Cl 0.270 Epoch: [90][501/1000] Time 0.169 Loss 4.702 Xent 3.300 RegLoss 1.402 Accuracy 0.469 Lr 0.00050 Cl 0.323 Epoch: [90][601/1000] Time 0.194 Loss 3.858 Xent 2.456 RegLoss 1.402 Accuracy 0.500 Lr 0.00050 Cl 0.265 Epoch: [90][701/1000] Time 0.187 Loss 4.404 Xent 3.001 RegLoss 1.403 Accuracy 0.375 Lr 0.00050 Cl 0.280 Epoch: [90][801/1000] Time 0.194 Loss 4.031 Xent 2.631 RegLoss 1.400 Accuracy 0.594 Lr 0.00050 Cl 0.276 Epoch: [90][901/1000] Time 0.129 Loss 5.756 Xent 4.355 RegLoss 1.401 Accuracy 0.281 Lr 0.00050 Cl 0.335 Running forward pass on validation set ............................. Validation Epoch: 90 Time 37.555 Loss 5.772 Xent 4.373 Accuracy 0.287 Saving variables Variables saved in 0.29 seconds Runnning forward pass on LFW images ........................ Accuracy: 0.92067+-0.01679
lf
python3 src/train_softmax_mobilenet.py
--logs_base_dir ./log/facenet/
--models_base_dir ./model/
--data_dir ./dataset/casia_mtcnnpy_224/
--batch_size 64
--image_size 160
--lfw_dir ./dataset/lfw_mtcnnpy_160/
--pretrained_model ./model/20191031-180131/model-20191031-180131.ckpt-90
--optimizer ADAM
--learning_rate -1
--max_nrof_epochs 150
--keep_probability 0.8
--random_crop
--random_flip
--use_fixed_image_standardization
--learning_rate_schedule_file data/learning_rate_schedule_classifier_casia_second.txt
--weight_decay 5e-4
--embedding_size 512
--lfw_distance_metric 1
--lfw_use_flipped_images
--lfw_subtract_mean
--validation_set_split_ratio 0.05
--validate_every_n_epochs 5
--prelogits_norm_loss_factor 5e-4
python3 src/train_softmax_mobilenet.py
--logs_base_dir ./log/facenet/
--models_base_dir ./model/
--data_dir ./dataset/casia_mtcnnpy_224/
--batch_size 90
--image_size 224
--lfw_dir ./dataset/lfw_mtcnnpy_160/
--pretrained_model ./model/20191101-060848/model-20191101-060848.ckpt-249
--optimizer ADAM
--learning_rate -1
--max_nrof_epochs 250
--keep_probability 0.8
--random_flip
--use_fixed_image_standardization
--learning_rate_schedule_file data/learning_rate_schedule_classifier_casia_third.txt
--weight_decay 5e-4
--embedding_size 512
--lfw_distance_metric 1
--lfw_use_flipped_images
--lfw_subtract_mean
--validation_set_split_ratio 0.05
--validate_every_n_epochs 5
--prelogits_norm_loss_factor 5e-4
lf quant
python3 src/train_softmax_mobilenet.py
--logs_base_dir ./log/facenet/
--models_base_dir ./model/
--data_dir ./dataset/casia_mtcnnpy_224/
--batch_size 32
--image_size 224
--lfw_dir ./dataset/lfw_mtcnnpy_160/
--pretrained_model ./model/20191104-031643/model-20191104-031643.ckpt-233
--optimizer ADAM
--learning_rate -1
--max_nrof_epochs 250
--keep_probability 0.8
--random_flip
--use_fixed_image_standardization
--learning_rate_schedule_file data/learning_rate_schedule_classifier_casia_third.txt
--
A3D6
weight_decay 5e-4
--embedding_size 512
--lfw_distance_metric 1
--lfw_use_flipped_images
--lfw_subtract_mean
--validation_set_split_ratio 0.05
--validate_every_n_epochs 5
--prelogits_norm_loss_factor 5e-4
Saving statistics Epoch: [149][1/1000] Time 0.362 Loss 2.659 Xent 1.334 RegLoss 1.325 Accuracy 0.733 Lr 0.00005 Cl 0.816 Epoch: [149][101/1000] Time 0.426 Loss 2.439 Xent 1.114 RegLoss 1.325 Accuracy 0.822 Lr 0.00005 Cl 0.785 Epoch: [149][201/1000] Time 0.416 Loss 2.697 Xent 1.371 RegLoss 1.326 Accuracy 0.800 Lr 0.00005 Cl 0.818 Epoch: [149][301/1000] Time 0.436 Loss 2.218 Xent 0.893 RegLoss 1.326 Accuracy 0.856 Lr 0.00005 Cl 0.755 Epoch: [149][401/1000] Time 0.411 Loss 2.187 Xent 0.863 RegLoss 1.324 Accuracy 0.878 Lr 0.00005 Cl 0.734 Epoch: [149][501/1000] Time 0.421 Loss 2.721 Xent 1.396 RegLoss 1.326 Accuracy 0.833 Lr 0.00005 Cl 0.792 Epoch: [149][601/1000] Time 0.333 Loss 2.289 Xent 0.964 RegLoss 1.325 Accuracy 0.889 Lr 0.00005 Cl 0.772 Epoch: [149][701/1000] Time 0.357 Loss 2.344 Xent 1.020 RegLoss 1.325 Accuracy 0.844 Lr 0.00005 Cl 0.769 Epoch: [149][801/1000] Time 0.334 Loss 2.389 Xent 1.066 RegLoss 1.323 Accuracy 0.856 Lr 0.00005 Cl 0.779 Epoch: [149][901/1000] Time 0.450 Loss 2.374 Xent 1.049 RegLoss 1.325 Accuracy 0.844 Lr 0.00005 Cl 0.793 Saving variables Variables saved in 0.35 seconds Runnning forward pass on LFW images ........................ Accuracy: 0.94300+-0.01513 Validation rate: 0.66067+-0.03803 @ FAR=0.00100 Saving statistics Epoch: [150][1/1000] Time 0.341 Loss 2.164 Xent 0.840 RegLoss 1.324 Accuracy 0.867 Lr 0.00005 Cl 0.683 Epoch: [150][101/1000] Time 0.328 Loss 2.229 Xent 0.905 RegLoss 1.324 Accuracy 0.878 Lr 0.00005 Cl 0.717 Epoch: [150][201/1000] Time 0.437 Loss 2.451 Xent 1.126 RegLoss 1.325 Accuracy 0.833 Lr 0.00005 Cl 0.756 Epoch: [150][301/1000] Time 0.445 Loss 2.413 Xent 1.089 RegLoss 1.324 Accuracy 0.856 Lr 0.00005 Cl 0.779 Epoch: [150][401/1000] Time 0.438 Loss 2.591 Xent 1.268 RegLoss 1.323 Accuracy 0.800 Lr 0.00005 Cl 0.803 Epoch: [150][501/1000] Time 0.440 Loss 2.178 Xent 0.854 RegLoss 1.324 Accuracy 0.844 Lr 0.00005 Cl 0.712 Epoch: [150][601/1000] Time 0.411 Loss 2.380 Xent 1.055 RegLoss 1.325 Accuracy 0.822 Lr 0.00005 Cl 0.747 Epoch: [150][701/1000] Time 0.437 Loss 2.896 Xent 1.572 RegLoss 1.324 Accuracy 0.722 Lr 0.00005 Cl 0.803 Epoch: [150][801/1000] Time 0.430 Loss 2.830 Xent 1.506 RegLoss 1.324 Accuracy 0.756 Lr 0.00005 Cl 0.803 Epoch: [150][901/1000] Time 0.398 Loss 2.649 Xent 1.324 RegLoss 1.325 Accuracy 0.822 Lr 0.00005 Cl 0.797 Running forward pass on validation set ............................. Validation Epoch: 150 Time 37.541 Loss 3.856 Xent 2.529 Accuracy 0.601 Saving variables Variables saved in 0.28 seconds Runnning forward pass on LFW images ........................ Accuracy: 0.94267+-0.01566 Validation rate: 0.66900+-0.04222 @ FAR=0.00100 Saving statistics
－－－－－－ 20191101-060848/model-20191101-060848.ckpt-249 Epoch: [245][1/1000] Time 0.588 Loss 3.117 Xent 1.282 RegLoss 1.835 Accuracy 0.767 Lr 0.00000 Cl 0.673 Epoch: [245][101/1000] Time 0.606 Loss 2.829 Xent 0.994 RegLoss 1.835 Accuracy 0.867 Lr 0.00000 Cl 0.632 Epoch: [245][201/1000] Time 0.588 Loss 3.666 Xent 1.830 RegLoss 1.836 Accuracy 0.733 Lr 0.00000 Cl 0.741 Epoch: [245][301/1000] Time 0.599 Loss 2.662 Xent 0.825 RegLoss 1.837 Accuracy 0.889 Lr 0.00000 Cl 0.636 Epoch: [245][401/1000] Time 0.565 Loss 3.271 Xent 1.436 RegLoss 1.835 Accuracy 0.789 Lr 0.00000 Cl 0.712 Epoch: [245][501/1000] Time 0.582 Loss 3.035 Xent 1.199 RegLoss 1.836 Accuracy 0.844 Lr 0.00000 Cl 0.675 Epoch: [245][601/1000] Time 0.611 Loss 2.904 Xent 1.068 RegLoss 1.835 Accuracy 0.856 Lr 0.00000 Cl 0.652 Epoch: [245][701/1000] Time 0.607 Loss 3.057 Xent 1.222 RegLoss 1.835 Accuracy 0.844 Lr 0.00000 Cl 0.663 Epoch: [245][801/1000] Time 0.593 Loss 2.963 Xent 1.128 RegLoss 1.835 Accuracy 0.833 Lr 0.00000 Cl 0.699 Epoch: [245][901/1000] Time 0.636 Loss 2.629 Xent 0.795 RegLoss 1.834 Accuracy 0.900 Lr 0.00000 Cl 0.629 Running forward pass on validation set ............................. Validation Epoch: 245 Time 77.314 Loss 4.532 Xent 2.694 Accuracy 0.572 Saving variables Variables saved in 0.44 seconds Runnning forward pass on LFW images ........................ Accuracy: 0.86583+-0.01948 Validation rate: 0.24467+-0.05596 @ FAR=0.00100 Saving statistics Epoch: [246][1/1000] Time 0.662 Loss 3.007 Xent 1.172 RegLoss 1.835 Accuracy 0.822 Lr 0.00000 Cl 0.681 Epoch: [246][101/1000] Time 0.599 Loss 3.183 Xent 1.347 RegLoss 1.835 Accuracy 0.800 Lr 0.00000 Cl 0.722 Epoch: [246][201/1000] Time 0.630 Loss 2.674 Xent 0.838 RegLoss 1.836 Accuracy 0.878 Lr 0.00000 Cl 0.624 Epoch: [246][301/1000] Time 0.575 Loss 2.897 Xent 1.063 RegLoss 1.835 Accuracy 0.811 Lr 0.00000 Cl 0.676 Epoch: [246][401/1000] Time 0.573 Loss 3.189 Xent 1.354 RegLoss 1.836 Accuracy 0.822 Lr 0.00000 Cl 0.684 Epoch: [246][501/1000] Time 0.608 Loss 2.878 Xent 1.042 RegLoss 1.836 Accuracy 0.867 Lr 0.00000 Cl 0.655 Epoch: [246][601/1000] Time 0.613 Loss 2.997 Xent 1.162 RegLoss 1.835 Accuracy 0.822 Lr 0.00000 Cl 0.680 Epoch: [246][701/1000] Time 0.677 Loss 3.072 Xent 1.238 RegLoss 1.835 Accuracy 0.811 Lr 0.00000 Cl 0.661 Epoch: [246][801/1000] Time 0.636 Loss 3.617 Xent 1.782 RegLoss 1.835 Accuracy 0.733 Lr 0.00000 Cl 0.748 Epoch: [246][901/1000] Time 0.593 Loss 3.046 Xent 1.211 RegLoss 1.835 Accuracy 0.811 Lr 0.00000 Cl 0.653 Saving variables Variables saved in 8.19 seconds Runnning forward pass on LFW images ........................ Accuracy: 0.86250+-0.01781 Validation rate: 0.24600+-0.05438 @ FAR=0.00100 Saving statistics Epoch: [247][1/1000] Time 0.624 Loss 3.165 Xent 1.331 RegLoss 1.834 Accuracy 0.811 Lr 0.00000 Cl 0.692 Epoch: [247][101/1000] Time 0.602 Loss 2.928 Xent 1.094 RegLoss 1.834 Accuracy 0.856 Lr 0.00000 Cl 0.653 Epoch: [247][201/1000] Time 0.596 Loss 3.028 Xent 1.194 RegLoss 1.834 Accuracy 0.833 Lr 0.00000 Cl 0.691 Epoch: [247][301/1000] Time 0.596 Loss 3.228 Xent 1.393 RegLoss 1.835 Accuracy 0.767 Lr 0.00000 Cl 0.699 Epoch: [247][401/1000] Time 0.601 Loss 3.192 Xent 1.357 RegLoss 1.835 Accuracy 0.789 Lr 0.00000 Cl 0.703 Epoch: [247][501/1000] Time 0.641 Loss 3.147 Xent 1.311 RegLoss 1.836 Accuracy 0.778 Lr 0.00000 Cl 0.686 Epoch: [247][601/1000] Time 0.610 Loss 2.848 Xent 1.014 RegLoss 1.834 Accuracy 0.844 Lr 0.00000 Cl 0.662 Epoch: [247][701/1000] Time 0.593 Loss 3.070 Xent 1.236 RegLoss 1.834 Accuracy 0.844 Lr 0.00000 Cl 0.642 Epoch: [247][801/1000] Time 0.602 Loss 3.000 Xent 1.166 RegLoss 1.834 Accuracy 0.822 Lr 0.00000 Cl 0.661 Epoch: [247][901/1000] Time 0.635 Loss 2.846 Xent 1.011 RegLoss 1.836 Accuracy 0.833 Lr 0.00000 Cl 0.647 Saving variables Variables saved in 0.59 seconds Runnning forward pass on LFW images ........................ Accuracy: 0.86050+-0.01773 Validation rate: 0.24667+-0.05395 @ FAR=0.00100 Saving statistics Epoch: [248][1/1000] Time 0.635 Loss 2.595 Xent 0.762 RegLoss 1.834 Accuracy 0.878 Lr 0.00000 Cl 0.589 Epoch: [248][101/1000] Time 0.606 Loss 2.905 Xent 1.071 RegLoss 1.835 Accuracy 0.856 Lr 0.00000 Cl 0.620 Epoch: [248][201/1000] Time 0.605 Loss 2.806 Xent 0.971 RegLoss 1.836 Accuracy 0.833 Lr 0.00000 Cl 0.635 Epoch: [248][301/1000] Time 0.602 Loss 2.697 Xent 0.861 RegLoss 1.836 Accuracy 0.878 Lr 0.00000 Cl 0.603 Epoch: [248][401/1000] Time 0.603 Loss 2.868 Xent 1.032 RegLoss 1.836 Accuracy 0.889 Lr 0.00000 Cl 0.635 Epoch: [248][501/1000] Time 0.600 Loss 3.367 Xent 1.533 RegLoss 1.834 Accuracy 0.778 Lr 0.00000 Cl 0.679 Epoch: [248][601/1000] Time 0.602 Loss 2.886 Xent 1.050 RegLoss 1.836 Accuracy 0.856 Lr 0.00000 Cl 0.666 Epoch: [248][701/1000] Time 0.589 Loss 2.989 Xent 1.155 RegLoss 1.834 Accuracy 0.789 Lr 0.00000 Cl 0.645 Epoch: [248][801/1000] Time 0.606 Loss 3.121 Xent 1.286 RegLoss 1.835 Accuracy 0.811 Lr 0.00000 Cl 0.679 Epoch: [248][901/1000] Time 0.616 Loss 2.724 Xent 0.889 RegLoss 1.835 Accuracy 0.844 Lr 0.00000 Cl 0.619 Saving variables Variables saved in 0.56 seconds Runnning forward pass on LFW images ........................ Accuracy: 0.86467+-0.01810 Validation rate: 0.24733+-0.05341 @ FAR=0.00100 Saving statistics Epoch: [249][1/1000] Time 0.604 Loss 2.781 Xent 0.948 RegLoss 1.834 Accuracy 0.878 Lr 0.00000 Cl 0.601 Epoch: [249][101/1000] Time 0.600 Loss 2.783 Xent 0.948 RegLoss 1.835 Accuracy 0.856 Lr 0.00000 Cl 0.633 Epoch: [249][201/1000] Time 0.584 Loss 2.837 Xent 1.001 RegLoss 1.836 Accuracy 0.800 Lr 0.00000 Cl 0.688 Epoch: [249][301/1000] Time 0.612 Loss 2.887 Xent 1.052 RegLoss 1.836 Accuracy 0.856 Lr 0.00000 Cl 0.671 Epoch: [249][401/1000] Time 0.595 Loss 2.940 Xent 1.105 RegLoss 1.835 Accuracy 0.811 Lr 0.00000 Cl 0.700 Epoch: [249][501/1000] Time 0.600 Loss 3.024 Xent 1.189 RegLoss 1.835 Accuracy 0.811 Lr 0.00000 Cl 0.672 Epoch: [249][601/1000] Time 0.607 Loss 2.984 Xent 1.148 RegLoss 1.836 Accuracy 0.844 Lr 0.00000 Cl 0.701 Epoch: [249][701/1000] Time 0.584 Loss 2.890 Xent 1.054 RegLoss 1.836 Accuracy 0.800 Lr 0.00000 Cl 0.631 Epoch: [249][801/1000] Time 0.602 Loss 2.894 Xent 1.057 RegLoss 1.836 Accuracy 0.822 Lr 0.00000 Cl 0.647 Epoch: [249][901/1000] Time 0.568 Loss 2.749 Xent 0.914 RegLoss 1.836 Accuracy 0.844 Lr 0.00000 Cl 0.604 Saving variables Variables saved in 0.48 seconds Runnning forward pass on LFW images
sh 20191106-110538 Accuracy: 0.94283+-0.01461 Validation rate: 0.66533+-0.04003 @ FAR=0.00100 Saving statistics Epoch: [150][1/1000] Time 0.361 Loss 2.126 Xent 0.819 RegLoss 1.307 Accuracy 0.911 Lr 0.00000 Cl 0.677 Epoch: [150][101/1000] Time 0.415 Loss 2.267 Xent 0.959 RegLoss 1.308 Accuracy 0.844 Lr 0.00000 Cl 0.734 Epoch: [150][201/1000] Time 0.447 Loss 2.417 Xent 1.109 RegLoss 1.308 Accuracy 0.833 Lr 0.00000 Cl 0.760 Epoch: [150][301/1000] Time 0.421 Loss 2.347 Xent 1.039 RegLoss 1.308 Accuracy 0.867 Lr 0.00000 Cl 0.782 Epoch: [150][401/1000] Time 0.410 Loss 2.562 Xent 1.256 RegLoss 1.306 Accuracy 0.789 Lr 0.00000 Cl 0.807 Epoch: [150][501/1000] Time 0.337 Loss 2.178 Xent 0.871 RegLoss 1.307 Accuracy 0.900 Lr 0.00000 Cl 0.732 Epoch: [150][601/1000] Time 0.427 Loss 2.158 Xent 0.851 RegLoss 1.307 Accuracy 0.867 Lr 0.00000 Cl 0.739 Epoch: [150][701/1000] Time 0.333 Loss 2.824 Xent 1.516 RegLoss 1.308 Accuracy 0.733 Lr 0.00000 Cl 0.800 Epoch: [150][801/1000] Time 0.426 Loss 2.817 Xent 1.510 RegLoss 1.308 Accuracy 0.733 Lr 0.00000 Cl 0.809 Epoch: [150][901/1000] Time 0.408 Loss 2.683 Xent 1.374 RegLoss 1.308 Accuracy 0.800 Lr 0.00000 Cl 0.800 Running forward pass on validation set ............................. Validation Epoch: 150 Time 37.689 Loss 3.825 Xent 2.515 Accuracy 0.606 Saving variables Variables saved in 0.29 seconds
lf 20191104-031643/
Epoch: [232][1/1000] Time 0.690 Loss 2.877 Xent 1.076 RegLoss 1.801 Accuracy 0.856 Lr 0.00000 Cl 0.702 Epoch: [232][101/1000] Time 0.610 Loss 2.818 Xent 1.020 RegLoss 1.798 Accuracy 0.833 Lr 0.00000 Cl 0.658 Epoch: [232][201/1000] Time 0.606 Loss 2.737 Xent 0.939 RegLoss 1.798 Accuracy 0.822 Lr 0.00000 Cl 0.656 Epoch: [232][301/1000] Time 0.603 Loss 2.977 Xent 1.178 RegLoss 1.799 Accuracy 0.833 Lr 0.00000 Cl 0.680 Epoch: [232][401/1000] Time 0.605 Loss 2.958 Xent 1.160 RegLoss 1.798 Accuracy 0.833 Lr 0.00000 Cl 0.693 Epoch: [232][501/1000] Time 0.599 Loss 3.283 Xent 1.485 RegLoss 1.798 Accuracy 0.778 Lr 0.00000 Cl 0.705 Epoch: [232][601/1000] Time 0.645 Loss 2.898 Xent 1.101 RegLoss 1.797 Accuracy 0.856 Lr 0.00000 Cl 0.673
Epoch: [232][701/1000] Time 0.578 Loss 3.046 Xent 1.247 RegLoss 1.799 Accuracy 0.856 Lr 0.00000 Cl 0.665 Epoch: [232][801/1000] Time 0.577 Loss 2.592 Xent 0.794 RegLoss 1.798 Accuracy 0.844 Lr 0.00000 Cl 0.598 Epoch: [232][901/1000] Time 0.620 Loss 2.892 Xent 1.092 RegLoss 1.799 Accuracy 0.856 Lr 0.00000 Cl 0.670 Saving variables Variables saved in 0.47 seconds Runnning forward pass on LFW images ........................ Accuracy: 0.86150+-0.01843 Validation rate: 0.24367+-0.04202 @ FAR=0.00100 Saving statistics Epoch: [233][1/1000] Time 0.610 Loss 2.658 Xent 0.860 RegLoss 1.798 Accuracy 0.878 Lr 0.00000 Cl 0.604 Epoch: [233][101/1000] Time 0.577 Loss 2.844 Xent 1.044 RegLoss 1.800 Accuracy 0.878 Lr 0.00000 Cl 0.637 Epoch: [233][201/1000] Time 0.662 Loss 2.984 Xent 1.186 RegLoss 1.798 Accuracy 0.856 Lr 0.00000 Cl 0.659 Epoch: [233][301/1000] Time 0.636 Loss 2.776 Xent 0.978 RegLoss 1.798 Accuracy 0.844 Lr 0.00000 Cl 0.646 Epoch: [233][401/1000] Time 0.691 Loss 2.707 Xent 0.907 RegLoss 1.799 Accuracy 0.889 Lr 0.00000 Cl 0.660 Epoch: [233][501/1000] Time 0.597 Loss 2.542 Xent 0.745 RegLoss 1.798 Accuracy 0.878 Lr 0.00000 Cl 0.623 Epoch: [233][601/1000] Time 0.608 Loss 2.950 Xent 1.150 RegLoss 1.800 Accuracy 0.867 Lr 0.00000 Cl 0.677 Epoch: [233][701/1000] Time 0.607 Loss 2.834 Xent 1.036 RegLoss 1.798 Accuracy 0.867 Lr 0.00000 Cl 0.664 Epoch: [233][801/1000] Time 0.612 Loss 3.015 Xent 1.217 RegLoss 1.798 Accuracy 0.811 Lr 0.00000 Cl 0.684 Epoch: [233][901/1000] Time 0.609 Loss 2.882 Xent 1.082 RegLoss 1.800 Accuracy 0.856 Lr 0.00000 Cl 0.679 Saving variables Variables saved in 0.94 seconds Runnning forward pass on LFW images ........................ Accuracy: 0.86217+-0.01821 Validation rate: 0.22567+-0.04867 @ FAR=0.00100 Saving statistics Epoch: [234][1/1000] Time 0.644 Loss 2.838 Xent 1.039 RegLoss 1.799 Accuracy 0.867 Lr 0.00000 Cl 0.691 Epoch: [234][101/1000] Time 0.615 Loss 3.006 Xent 1.206 RegLoss 1.800 Accuracy 0.833 Lr 0.00000 Cl 0.722 Epoch: [234][201/1000] Time 0.604 Loss 2.626 Xent 0.827 RegLoss 1.799 Accuracy 0.867 Lr 0.00000 Cl 0.638 Epoch: [234][301/1000] Time 0.565 Loss 2.833 Xent 1.035 RegLoss 1.798 Accuracy 0.822 Lr 0.00000 Cl 0.680 Epoch: [234][401/1000] Time 0.658 Loss 3.097 Xent 1.298 RegLoss 1.799 Accuracy 0.822 Lr 0.00000 Cl 0.707
triplet loss first training python3 src/train_tripletloss_mobilenet.py --logs_base_dir ./log/facenet/ --models_base_dir ./model/ --data_dir ./dataset/casia_mtcnnpy_224/ --batch_size 90 --image_size 224 --lfw_dir ./dataset/lfw_mtcnnpy_160/ --optimizer ADAM --learning_rate -1 --max_nrof_epochs 250 --keep_probability 0.8 --random_flip --use_fixed_image_standardization --learning_rate_schedule_file data/learning_rate_schedule_classifier_casia.txt--weight_decay 5e-4 --embedding_size 512 --lfw_distance_metric 1 --lfw_use_flipped_images --lfw_subtract_mean 20191113-065711/model-20191113-065711
python3 src/train_tripletloss_mobilenet.py --pretrained_model ./model/20191113-065711/model-20191113-065711.ckpt-79033 --logs_base_dir ./log/facenet/ --models_base_dir ./model/ --data_dir ./dataset/casia_mtcnnpy_224/ --batch_size 90 --image_size 224 --lfw_dir ./dataset/lfw_mtcnnpy_160/ --optimizer ADAM --learning_rate -1 --max_nrof_epochs 250 --keep_probability 0.8 --random_flip --use_fixed_image_standardization --learning_rate_schedule_file data/learning_rate_schedule_classifier_casia_second.txt --weight_decay 5e-4 --embedding_size 512 --lfw_distance_metric 1 --lfw_use_flipped_images --lfw_subtract_mean
20191114-031120/ Epoch: [248][1201/1000] Time 0.535 Loss 0.131 Epoch: [248][1401/1000] Time 0.607 Loss 0.139 Saving variables Variables saved in 0.09 seconds Running forward pass on LFW images: 82.690 Accuracy: 0.803+-0.015 Validation rate: 0.09233+-0.01687 @ FAR=0.00100
python3 src/train_tripletloss_mobilenet.py --pretrained_model ./model/20191114-031120/model-20191114-031120.ckpt-246603 --logs_base_dir ./log/facenet/ --models_base_dir ./model/ --data_dir ./dataset/casia_mtcnnpy_224/ --batch_size 90 --image_size 224 --lfw_dir ./dataset/lfw_mtcnnpy_160/ --optimizer ADAM --learning_rate -1 --max_nrof_epochs 250 --keep_probability 0.8 --random_flip --use_fixed_image_standardization --learning_rate_schedule_file data/learning_rate_schedule_classifier_casia_second.txt --weight_decay 5e-4 --embedding_size 512 --lfw_distance_metric 1 --lfw_use_flipped_images --lfw_subtract_mean
python3 src/train_tripletloss_mobilenet.py --pretrained_model ./model/20191113-065711/model-20191113-065711.ckpt-79033 --logs_base_dir ./log/facenet/ --models_base_dir ./model/ --data_dir ./dataset/casia_mtcnnpy_224/ --batch_size 90 --image_size 224 --lfw_dir ./dataset/lfw_mtcnnpy_160/ --optimizer ADAM --learning_rate -1 --max_nrof_epochs 250 --keep_probability 0.8 --random_flip --use_fixed_image_standardization --learning_rate_schedule_file data/learning_rate_schedule_classifier_casia_third.txt --weight_decay 5e-4 --embedding_size 512 --lfw_distance_metric 1 --lfw_use_flipped_images --lfw_subtract_mean