yule-li / CosFace

Tensorflow implementation for paper CosFace: Large Margin Cosine Loss for Deep Face Recognition
285 stars 99 forks source link

Where is the list_file of CASIA-WebFace-112X96 dataset #10

Closed Paleve closed 5 years ago

Paleve commented 5 years ago

I download the preprocessed dataset of CASIA-WebFace-112X96, however I didn't get the list_file for training. Could you provide me the it, or give some methods?

chenyyx commented 5 years ago

I recreated a list of data myself. Put it here --------- Links : https://pan.baidu.com/s/1Vq-hwmEDI4eULYH6TmiApg password: buyc

Paleve commented 5 years ago

I recreated a list of data myself. Put it here --------- Links : https://pan.baidu.com/s/1Vq-hwmEDI4eULYH6TmiApg password: buyc

非常感谢,想知道你使用CASIA-WebFace-112X96训练得到的结果怎么样,我训练60000 epoch得到的精度只有0.959,嵌入纬度1024,m设置为0.35,参数设置为: 4 pretrained_model: None 5 loss_type: cosface 6 network: sphere_network 7 data_dir: dataset/CASIA-maxpy-clean 8 list_file: dataset/lst 9 model_def: models.inception_resnet_v1 10 max_nrof_epochs: 100 11 batch_size: 90 12 image_size: 160 13 image_src_size: 256 14 image_height: 112 15 image_width: 96 16 people_per_batch: 30 17 num_gpus: 1 18 images_per_person: 5 19 epoch_size: 600 20 alpha: 0.35 21 scale: 64.0 22 weight: 2.0 23 embedding_size: 1024 24 random_crop: False 25 random_flip: True 26 fc_bn: False 27 keep_probability: 1.0 28 weight_decay: 0.0001 29 optimizer: MOM 30 center_loss_factor: 0.0 31 center_loss_alfa: 0.95 32 learning_rate: -1.0 33 learning_rate_decay_epochs: 100 34 learning_rate_decay_factor: 1.0 35 moving_average_decay: 0.9999 36 seed: 666 37 learning_rate_schedule_file: lr_coco.txt

Paleve commented 5 years ago

@chenyyx 难道使用112x96的图最好设置嵌入纬度512吗?

chenyyx commented 5 years ago

@Paleve 你好哈,我训练的 embedding_size 为 128,1024,唯独 512 的没有训练过,23333。128 的好一些,忘记 epoch 大小了,但是 1080ti 训练了 3 天左右,得到的精度是 99.3%,还不错。如果你实在没有提升的地方了,可以换换图片呀,换换 image_size ,我们这儿有人就是这么做的。效果也挺好的。

Paleve commented 5 years ago

@chenyyx 我用1080ti几个小时就跑了60000步,你这么说可能是我训的模型没完全收敛。非常感谢,这几天刚要入门人脸识别,方便留个邮箱吗有些问题希望能请教一下