search

luoyetx / mini-caffe-example

detect facial landmark with mini-caffe
https://youtu.be/oNiAtu0erEk
18 stars 24 forks source link

how to train with the five split patches end to end #1

Open YIFanH opened 8 years ago

YIFanH commented 8 years ago

Hi,luoyetx:

I watched your works on face five landmarks which help me a lot,now i want to try this work myself,but i don't know how to train this with split patches,could you send me an training log of this work. Thank you and wish your reply. 

                                                                                                                                                          YiFan
luoyetx commented 8 years ago

@YIFanH glad to hear this project helps you. The model is trained by deep-landmark

YIFanH commented 8 years ago

@luoyetx Thanks for your reply, I will try to make it,If i have some problems can not be solved , I will write letter to you, Thanks !!! YiFan

YIFanH commented 8 years ago

@luoyetx Could you send me a log file to me,my email is bingshan_shuilei@163.com,my train loss and test loss is not convergence, I'm sure about the face image dataset and image points labels(range is 0 - 38,not be normalized to 0 - 1). thank you and best wishes for you. YIFan

luoyetx commented 8 years ago

@YIFanH I have found the log file and uploaded in this issue.

log.tar.gz

YIFanH commented 8 years ago

@luoyetx My train_loss and test_loss not drop, the range of image points labels is 0 ~ 1。

luoyetx commented 8 years ago

@YIFanH You should carefully tune the parameters. I recommend you using caffe command not the bootstrap.sh. The loss should converge quickly in the most situation.

YIFanH commented 8 years ago

@luoyetx Could I send you my train and test log to you,the log you give me is the training log, I see the loss is 0.0044 below 20000 iteration,could you give me your QQ or email address to me ?

luoyetx commented 8 years ago

luoyetx@gmail.com

YIFanH commented 8 years ago

@luoyetx
The model could give the landmarks have 3-5 pixels offset, which may have many iterations to train it. I send the log files to your emai. Thank you and best wishes!!!