zma-c-137 / VarGFaceNet

312 stars 84 forks source link

无法复现Paper里面的结果,请教一下超参 #15

Open zys1994 opened 4 years ago

zys1994 commented 4 years ago

你好,我训了多版vargfacenet,但最好的一版也与论文的准确率有很大的出入,想请教一下我的超参有什么大问题吗。

Namespace(batch_size=512, ckpt=3, ctx_num=8, dataset='retina', frequent=20, image_channel=3, kvstore='device', loss='combined', lr=0.01, lr_steps='100000,160000,220000', models_root='./models', mom=0.9, network='vargfacenet', per_batch_size=64, pretrained='', pretrained_epoch=1, rescale_threshold=0, verbose=4000, wd=0.0005) {'bn_mom': 0.9, 'workspace': 256, 'emb_size': 512, 'ckpt_embedding': True, 'net_se': 0, 'net_act': 'prelu', 'net_unit': 3, 'net_input': 1, 'net_blocks': [1, 4, 6, 2], 'net_output': 'J', 'net_multiplier': 1.25, 'val_targets': ['lfw', 'cfp_fp', 'agedb_30'], 'ce_loss': True, 'fc7_lr_mult': 1.0, 'fc7_wd_mult': 1.0, 'fc7_no_bias': False, 'max_steps': 0, 'data_rand_mirror': True, 'data_cutoff': False, 'data_color': 0, 'data_images_filter': 0, 'count_flops': True, 'memonger': False, 'loss_name': 'margin_softmax', 'loss_s': 64.0, 'loss_m1': 1.0, 'loss_m2': 0.3, 'loss_m3': 0.2, 'net_name': 'vargfacenet', 'dataset': 'retina', 'dataset_path': '/data/cv/faceRec_data/ms1m-retinaface-t1', 'num_classes': 93431, 'image_shape': [112, 112, 3], 'loss': 'combined', 'network': 'vargfacenet', 'num_workers': 1, 'batch_size': 512, 'per_batch_size': 64}

我训到后面的准确率

[lfw][308000]XNorm: 9.591558
[lfw][308000]Accuracy-Flip: 0.99483+-0.00353
testing verification..
(14000, 512)
infer time 10.176672000000002
[cfp_fp][308000]XNorm: 2856.294986
[cfp_fp][308000]Accuracy-Flip: 0.83286+-0.03168
testing verification..
(12000, 512)
infer time 8.675411
[agedb_30][308000]XNorm: 5.247274
[agedb_30][308000]Accuracy-Flip: 0.96350+-0.00758