search

davidsandberg / facenet

Face recognition using Tensorflow
MIT License
13.85k stars 4.81k forks source link

Improving accuracy #889

Closed Zumbalamambo closed 5 years ago

Zumbalamambo commented 6 years ago

I used the following hyperparams to train on casia.

python src/train_tripletloss.py --logs_base_dir ./logs --models_base_dir ./models --data_dir facenet/data/casia_maxpy_mtcnnpy_182 --image_size 160 --model_def models.inception_resnet_v1 --optimizer RMSPROP --learning_rate 0.01 --weight_decay 1e-4 --max_nrof_epochs 600 --gpu_memory_fraction 0.1

The Embedding is 512 dimensions. people_per_batch is 6 and batch_size is 9 and images_per_person is also 6.

However after training the model , i tried to validate on lfw.

This is the output on validation.

Accuracy: 0.76133+-0.01545 Validation rate: 0.07233+-0.01984 @ FAR=0.00100 Area Under Curve (AUC): 0.845 Equal Error Rate (EER): 0.238

why is my accuracy just 76 %? How do I improve it?

test4fest commented 6 years ago

Did you aligned the LFW dasaset before running validation? would you provide more detail about (code) for validaion part?

yuwgle commented 5 years ago

try to increase batch_size