search

yaojieliu / ICCVW2017-DenseFaceAlignment

http://cvlab.cse.msu.edu/project-pifa.html
MIT License
26 stars 4 forks source link

how can I compile the MatConvNet #3

Open wqz960 opened 5 years ago

wqz960 commented 5 years ago

`Error using vl_nnconv An input is not a numeric array (or GPU support not compiled).

Error in vl_test_simplenn (line 297) output = vl_nnconv(res(i).x, l.weights{1}, l.weights{2}, ...

Error in runDeFA (line 18) res = vl_test_simplenn(net, im4d, [], [],...

Error in test (line 37) [A3D,landmark68] = runDeFA(im,DeFA);`

how to solve it?

Guyume commented 5 years ago

do you use the gpu matconvnet? you can type the command of --vl_testnn('gpu', true) in your matlib if you have compiled but dont know whether it is successful. And if you haven't compiled it yet, you can follow the instructions of compiling, it works for me http://www.vlfeat.org/matconvnet/install/#compiling

wqz960 commented 5 years ago

@Guyume `Failure Summary:

 Name                                                                    Failed  Incomplete  Reason(s)
=======================================================================================================
 nnmnist[dataType=single,device=cpu]/valErrorRate(networkType=dagnn)       X         X       Errored.
-------------------------------------------------------------------------------------------------------
 nnmnist[dataType=single,device=cpu]/valErrorRate(networkType=simplenn)    X         X       Errored.
-------------------------------------------------------------------------------------------------------
 nnmnist[dataType=single,device=gpu]/valErrorRate(networkType=dagnn)       X         X       Errored.
-------------------------------------------------------------------------------------------------------
 nnmnist[dataType=single,device=gpu]/valErrorRate(networkType=simplenn)    X         X       Errored.

result =

1x2996 TestResult array with properties:

Name
Passed
Failed
Incomplete
Duration
Details

Totals: 2992 Passed, 4 Failed, 4 Incomplete. 368.085 seconds testing time.` It is my test log with vl_testsimplenn('gpu',true), I am new in Matlab, i do not know whether it is right? and the test.m also failed after compiling again.

Guyume commented 5 years ago

have you replced the original matlab file by the compiled one? I raplaced it and it works, and make some change with the test.m if you use cpu not gpu. these all are what i have done.

wqz960 commented 5 years ago

@Guyume really thank you for your kind respond! the codes finally run, but it seems that the performance is not so good in AFLW2000 dataset which was not cropped, the NMS reaches 0.17!!! do you know whether the author provide the cropped image of AFLW2000 on his project network? thank you!

Guyume commented 5 years ago

@wqz960 you can search the homepage of the author, he has done some awsome work

yaojieliu commented 5 years ago

@wqz960 please run vl_compilenn.m to setup gpu support and mex the c code. You may need to change the setting to fit your need.

DeFA needs to process a pre-cropped face. Popular MTCNN, Viola-Jones, and Dlib face detector can work properly with DeFA. More specifically, the training faces are cropped with 15% larger of the tightest bounding box (from 68 landmarks).

I provide the annotations of AFLW2000 in /data folder, but you need to download the data from AFLW since I'm not allowed to redistribute the original data.

lxzkeyan commented 4 years ago

@yaojieliu Thanks for sharing the code.I would like to ask about how to train and test the AFLW_2000 dataset.Can you give me some detailed training steps?