deepid2_caffe icon indicating copy to clipboard operation
deepid2_caffe copied to clipboard

Published 20 hours ago • verified publisher icon wuqianliang

Training MS-Celeb-1M

Open xhzcyc opened this issue 7 years ago • 10 comments

Hi, are you training ms-celeb-1m dataset now? How long did you extract image from ms-celeb-1m.tsv file?

xhzcyc avatar Jan 04 '18 01:01 xhzcyc

I use the SSD disk; Extracting image from ms-celeb-1m.tsv may take a long while , about several hours. I only extract 20000 persons' images for training. @xhzcyc

wuqianliang avatar Jan 08 '18 14:01 wuqianliang

How about your training result? @wuqianliang

xhzcyc avatar Jan 16 '18 01:01 xhzcyc

What's the size of your lmdb file that generated from 20000 person's image?

xhzcyc avatar Jan 16 '18 02:01 xhzcyc

deepid1 , test softmax accuracy 70% deepid2 , test softmax accuracy 37%

DeepID1_train_lmdb 19G DeepID1_test_lmdb 4.5G

This week , I will try multipatch preprocess on train dataset. @xhzcyc

wuqianliang avatar Jan 16 '18 05:01 wuqianliang

Why are your test accuracy so low? I have trained 10000 people subset and get a good result!

xhzcyc avatar Jan 18 '18 09:01 xhzcyc

Could you share the sovler txt for deepid1 and deepid2 ? @xhzcyc my emailbox : [email protected]

wuqianliang avatar Jan 18 '18 10:01 wuqianliang

Do you have a wechat ? @xhzcyc

My wechat is 15715164075

wuqianliang avatar Jan 19 '18 07:01 wuqianliang

用的deepid2跑了四万次 准确率63是不是有点低了 I0305 19:09:06.048393 22020 solver.cpp:337] Iteration 40000, Testing net (#0) I0305 19:09:47.081173 22020 solver.cpp:404] Test net output #0: accuracy1 = 0.634328 I0305 19:09:47.081173 22020 solver.cpp:404] Test net output #1: sim_loss = 0.0132883 (* 0.3 = 0.00398649 loss) I0305 19:09:47.081173 22020 solver.cpp:404] Test net output #2: softmax_loss = 2.3179 (* 0.7 = 1.62253 loss)

Yodamt avatar Mar 05 '19 11:03 Yodamt

用的deepid2跑了四万次 准确率63是不是有点低了 I0305 19:09:06.048393 22020 solver.cpp:337] Iteration 40000, Testing net (#0) I0305 19:09:47.081173 22020 solver.cpp:404] Test net output #0: accuracy1 = 0.634328 I0305 19:09:47.081173 22020 solver.cpp:404] Test net output #1: sim_loss = 0.0132883 (* 0.3 = 0.00398649 loss) I0305 19:09:47.081173 22020 solver.cpp:404] Test net output #2: softmax_loss = 2.3179 (* 0.7 = 1.62253 loss)

用的是lfw-aligned的train.txt 一共三千左右样本

Yodamt avatar Mar 05 '19 11:03 Yodamt

用的deepid2跑了四万次 准确率63是不是有点低了 I0305 19:09:06.048393 22020 solver.cpp:337] Iteration 40000, Testing net (#0) I0305 19:09:47.081173 22020 solver.cpp:404] Test net output #0: accuracy1 = 0.634328 I0305 19:09:47.081173 22020 solver.cpp:404] Test net output #1: sim_loss = 0.0132883 (* 0.3 = 0.00398649 loss) I0305 19:09:47.081173 22020 solver.cpp:404] Test net output #2: softmax_loss = 2.3179 (* 0.7 = 1.62253 loss)

用的是lfw-aligned的train.txt 一共三千左右样本

建议你用sphereface,效果比较好。 LFW数据集太小，效果相对会比较差一些。

wuqianliang avatar Mar 06 '19 07:03 wuqianliang