Python 如何提高FaceNet的准确性
我目前正在写论文，我将比较面部识别技术Eigenface、Fisherface和CNN在准确性和速度方面的差异。我目前正试图用电脑为CNN训练我自己的模特 我的问题是，我似乎无法将模型的准确度提高到71%以上，而我为分类器管理的最大准确度是80%。我使用的是LFW数据集的一小部分，它包含10个类，每个类有40个图像用于训练，每个类有4个图像用于测试 我尝试了以下参数，使用分类器获得77.5%的准确率，在模型向前传递时获得约71%的准确率：Python 如何提高FaceNet的准确性,python,tensorflow,deep-learning,conv-neural-network,tensorboard,Python,Tensorflow,Deep Learning,Conv Neural Network,Tensorboard,我目前正在写论文，我将比较面部识别技术Eigenface、Fisherface和CNN在准确性和速度方面的差异。我目前正试图用电脑为CNN训练我自己的模特 我的问题是，我似乎无法将模型的准确度提高到71%以上，而我为分类器管理的最大准确度是80%。我使用的是LFW数据集的一小部分，它包含10个类，每个类有40个图像用于训练，每个类有4个图像用于测试 我尝试了以下参数，使用分类器获得77.5%的准确率，在模型向前传递时获得约71%的准确率： /* * Alignment */ python3
/*
* Alignment
*/
python3 src/align/align_dataset_mtcnn.py ~/datasets/custom2/raw/training/ ~/datasets/custom2/aligned_182/training --image_size 182 --margin 44
python3 src/align/align_dataset_mtcnn.py ~/datasets/custom2/raw/testing/ ~/datasets/custom2/aligned_182/testing --image_size 182 --margin 44
/*
* Training
*/
python3 src/train_softmax.py \
--logs_base_dir ~/logs/facenet/ \
--models_base_dir ~/models/facenet/ \
--data_dir ~/datasets/custom2/aligned_182/training/ \
--image_size 160 \
--model_def models.inception_resnet_v1 \
--lfw_dir ~/datasets/lfw/lfw_aligned_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
python3 src/freeze_graph.py ~/models/facenet/20190402-225816/ ~/models/facenet/20190402-225816/model.pb
python3 src/classifier.py TRAIN ~/datasets/custom2/aligned_182/training/ ~/models/facenet/20190402-225816/model.pb ~/models/facenet/20190402-225816/classifier.pkl --batch_size 1000
/*
* Test verification
*/
python3 src/classifier.py CLASSIFY ~/datasets/custom2/aligned_182/testing/ ~/models/facenet/20190402-225816/model.pb ~/models/facenet/20190402-225816/classifier.pkl --batch_size 1000
/*
* Alignment
*/
for N in {1..4}; do python3 src/align/align_dataset_mtcnn.py ~/datasets/custom2/raw/training ~/datasets/custom2/aligned_160/training --image_size 160 --margin 32 --random_order --gpu_memory_fraction 0.2 & done
/*
* Training
*/
python3 src/train_softmax.py \
--logs_base_dir ~/logs/facenet/ \
--models_base_dir ~/models/facenet/ \
--data_dir ~/datasets/custom2/aligned_160/training \
--image_size 160 \
--model_def models.inception_resnet_v1 \
--lfw_dir ~/datasets/lfw/lfw_aligned_160 \
--optimizer RMSPROP \
--learning_rate 0.01 \
--weight_decay 1e-4 \
--max_nrof_epochs 100 \
--gpu_memory_fraction 0.95
python3 src/freeze_graph.py ~/models/facenet/20190401-114728/ ~/models/facenet/20190401-114728/model.pb
python3 src/classifier.py TRAIN ~/datasets/custom2/aligned_160/training/ ~/models/facenet/20190401-114728/model.pb ~/models/facenet/20190401-114728/classifier.pkl --batch_size 1000
/*
* Verification
*/
python3 src/classifier.py CLASSIFY ~/datasets/custom2/aligned_160/testing/ ~/models/facenet/20190401-114728/model.pb ~/models/facenet/20190401-114728/classifier.pkl --batch_size 1000
/*
* Alignment
*/
python3 src/align/align_dataset_mtcnn.py ~/datasets/custom2/raw/training/ ~/datasets/custom2/aligned_182/training --image_size 182 --margin 44
python3 src/align/align_dataset_mtcnn.py ~/datasets/custom2/raw/testing/ ~/datasets/custom2/aligned_182/testing --image_size 182 --margin 44
/*
* Training
*/
python3 src/train_softmax.py \
--logs_base_dir ~/logs/facenet/ \
--models_base_dir ~/models/facenet/ \
--data_dir ~/datasets/custom2/aligned_182/training/ \
--image_size 160 \
--model_def models.inception_resnet_v1 \
--lfw_dir ~/datasets/lfw/lfw_aligned_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
python3 src/freeze_graph.py ~/models/facenet/20190402-225816/ ~/models/facenet/20190402-225816/model.pb
python3 src/classifier.py TRAIN ~/datasets/custom2/aligned_182/training/ ~/models/facenet/20190402-225816/model.pb ~/models/facenet/20190402-225816/classifier.pkl --batch_size 1000
/*
* Test verification
*/
python3 src/classifier.py CLASSIFY ~/datasets/custom2/aligned_182/testing/ ~/models/facenet/20190402-225816/model.pb ~/models/facenet/20190402-225816/classifier.pkl --batch_size 1000
/*
* Alignment
*/
for N in {1..4}; do python3 src/align/align_dataset_mtcnn.py ~/datasets/custom2/raw/training ~/datasets/custom2/aligned_160/training --image_size 160 --margin 32 --random_order --gpu_memory_fraction 0.2 & done
/*
* Training
*/
python3 src/train_softmax.py \
--logs_base_dir ~/logs/facenet/ \
--models_base_dir ~/models/facenet/ \
--data_dir ~/datasets/custom2/aligned_160/training \
--image_size 160 \
--model_def models.inception_resnet_v1 \
--lfw_dir ~/datasets/lfw/lfw_aligned_160 \
--optimizer RMSPROP \
--learning_rate 0.01 \
--weight_decay 1e-4 \
--max_nrof_epochs 100 \
--gpu_memory_fraction 0.95
python3 src/freeze_graph.py ~/models/facenet/20190401-114728/ ~/models/facenet/20190401-114728/model.pb
python3 src/classifier.py TRAIN ~/datasets/custom2/aligned_160/training/ ~/models/facenet/20190401-114728/model.pb ~/models/facenet/20190401-114728/classifier.pkl --batch_size 1000
/*
* Verification
*/
python3 src/classifier.py CLASSIFY ~/datasets/custom2/aligned_160/testing/ ~/models/facenet/20190401-114728/model.pb ~/models/facenet/20190401-114728/classifier.pkl --batch_size 1000