Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

how to check the hd5 data ? #6

Open
wuqiangch opened this issue Mar 11, 2016 · 18 comments
Open

how to check the hd5 data ? #6

wuqiangch opened this issue Mar 11, 2016 · 18 comments

Comments

@wuqiangch
Copy link

After training by using bootstrap.sh, I get the wrong result .
4febec6cd3fa9e75d2a28023c0b0bc76
b656dd97cb9670160ad455223cd61a6c

I use HDFView the 1_F/train.hd5 file,what are the meaning of data and landmark?
(1)landmark,first six rows:
0.27142859 0.27142859 0.6571429 0.27142859 0.3 0.6142857 0.31428573 0.78571427 0.62857145 0.78571427
0.32422328 0.1958884 0.8004285 0.24294785 0.49501193 0.4644839 0.2713394 0.6769809 0.68675554 0.74030954
0.21656966 0.22970417 0.7015257 0.23771922 0.58244133 0.60628456 0.28283 0.8140927 0.68875426 0.7836232
0.2470238 0.22058824 0.6875 0.24411765 0.39583334 0.5029412 0.24107143 0.7617647 0.6279762 0.7852941
0.26837537 0.24787492 0.765999 0.2968669 0.5151268 0.6077065 0.19847062 0.6727469 0.6902065 0.7266793
0.4111675 0.30964467 0.7817259 0.20812182 0.75126904 0.6040609 0.5329949 0.8071066 0.78680205 0.7360406
(2)data's six rows
1.5018264
-1.6148947
-1.7687876
-1.742761
-2.0221322
-1.746762
“data” only has a column and "landmark" has 10 columns . Is it right?

@luoyetx
Copy link
Owner

luoyetx commented Mar 11, 2016

@wuqiangch the result here is 3 level ? and the level-1 includes EN and NM ?

what's more, did you monitor the loss during network training? Maybe some cnns are not trained well.

@luoyetx
Copy link
Owner

luoyetx commented Mar 11, 2016

the training data you have put out is ok. 5 landmark has 10 values and image data is a multi-dimension variable so it may be hard to view all values.

you may refer to this pdf, in latter half part, I put some more detailed data about the training status.

@wuqiangch
Copy link
Author

@luoyetx yes,I used my trained 23 models.But two models(1_EN,1_NM) was not be used. I used my 21 models to repace your models under webapp/v0.0.6 and test it. I used my 8 k80 gpus to train together. Its error is very low.

@wuqiangch
Copy link
Author

@luoyetx I can't open your pdf. So if it is possible,please send it to my email ([email protected]).Thanks!

@luoyetx
Copy link
Owner

luoyetx commented Mar 11, 2016

the model I provided under webapp/v0.0.6, did it have the same problem?

@wuqiangch
Copy link
Author

@luoyetx no, it is ok.

@wuqiangch
Copy link
Author

@luoyetx I have another question. Caffe with different version will get different train.hd5.
Why?

@luoyetx
Copy link
Owner

luoyetx commented Mar 11, 2016

the pdf is on github, the loading may cost some time but should be able to view it, if not, you can just download directly fro github

the training data is processed by Python code which has nothing to do with Caffe

@wuqiangch
Copy link
Author

This wrong result may be caused by caffe with diffrent version. You can upload your caffe to github.

@luoyetx
Copy link
Owner

luoyetx commented Mar 11, 2016

I don't think Caffe is the problem, the Caffe version I use is about 10 months ago. I already test the model with recently Caffe, it's OK.

What you mean here may be the training part is not compact with different Caffe version? I think it's impossible.

Maybe you should pay more attention to the training status of 21 networks to see if the loss is small enough and stable.

@luoyetx
Copy link
Owner

luoyetx commented Mar 11, 2016

By the way, you can test your model by

python test/test.py 0 # to test 1_F
python test/test.py 2 # to test 1_F and level2
python test/test.py 3 # to test 1_F and level2 and level3

you can also modify the webapp code to view different result of different level

@wuqiangch
Copy link
Author

I used bootstrap.sh to train it and i got 23 models

@luoyetx
Copy link
Owner

luoyetx commented Mar 11, 2016

@wuqiangch I recommend you to train the models manually using Caffe command which you can control the training status. The models I provide are well tuned which may not use the default parameters in prototxt files.

@wuqiangch
Copy link
Author

There are my results:
LEVEL - 0
Mean Error:
Left Eye = 0.022836
Right Eye = 0.253351
Nose = 0.029818
Left Mouth = 0.028378
Right Mouth = 0.028626
Failure:
Left Eye = 0.061743
Right Eye = 1.000000
Nose = 0.138488
Left Mouth = 0.108194
Right Mouth = 0.118003
f_test

LEVEL - 2
Mean Error:
Left Eye = 0.015997
Right Eye = 0.336471
Nose = 0.018802
Left Mouth = 0.128772
Right Mouth = 0.020509
Failure:
Left Eye = 0.033468
Right Eye = 1.000000
Nose = 0.038950
Left Mouth = 0.997403
Right Mouth = 0.058857

LEVEL - 3
Mean Error:
Left Eye = 0.062053
Right Eye = 0.337876
Nose = 0.018445
Left Mouth = 0.175577
Right Mouth = 0.019996
Failure:
Left Eye = 0.905366
Right Eye = 1.000000
Nose = 0.037507
Left Mouth = 0.999711
Right Mouth = 0.056261
Uploading level3_test.png…

@luoyetx
Copy link
Owner

luoyetx commented Mar 11, 2016

It seems 1_F's right eye is wrong, the error is too large, what's more in level-2, left mouth's error is too large, you should retrain related cnns.

@wuqiangch
Copy link
Author

How to modify my prototxts?

@luoyetx
Copy link
Owner

luoyetx commented Mar 11, 2016

I think re-train 1_F and 2_LM1, 2_LM2 will give you a good result if the loss of these network is pretty low and stable. There's no need to modify the prototxts only if you want a better result.

@wuqiangch
Copy link
Author

Thanks,I will try it.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants