Error in Training about lip_ssl

[Engineering-Course]

Oh, train.txt should contain both image and label split by blank, which is different from test.txt.

[Engineering-Course]

E0516 17:22:28.599076 34580 io.cpp:81] Could not open or find file human/data/

Please check your directory of the image data.

[CPFLAME]

Thank you for your reply.
my train.txt is the same as test.txt,it is just like this:

And my files in human/data:

I'm so confused that I can run TEST but I failed when training.
I run the code in the root path : LIP_SSL

[CPFLAME]

Thanks for your reply. It works fine now!

[staywithme23]

@CPFLAME May I know if you uncomment USE_CUDA when make caffe folder?
I ran into #2 problem.
Best,

[CPFLAME]

@adwin5 I used cuda when I make caffe folder.
It seems that the memory of your GPU is not big enough to run test, I think you can choose a GPU of larger memory to run this code.
By the way, I used the memory of 2579M to run the test code. You can check your momery size.

[staywithme23]

@CPFLAME Thank you!
I will try uncomment USE_CUDA
I have 8G, should be OK for testing!

[Engineering-Course]

@adwin5 You may use cuDNN or a larger memory GPU.

[staywithme23]

@Engineering-Course
Thanks it works.

[staywithme23]

@Engineering-Course @CPFLAME
I can run training, and I reference init.model from deeplap attention model.
I want to ask how to set path to use pose ground truth for training.
The run_human.sh code seems just reading ground truth labels not including pose, right?

My second question is what does strong and weak means in the training section in the run_human.sh file
Thank you.

[Engineering-Course]

@adwin5
Sorry, these scripts are only used for parsing tasks.
Strong and weak mean nothing which are used by deeplab not this project.

[staywithme23]

@Engineering-Course
Can I confirm with you that the code is just using parsing loss (pixel-wise softmax loss) instead of structure loss (structure-sensitive loss) which mentioned in paper.
Thank you.

[Engineering-Course]

@adwin5 NONONO, it does use a structure loss. You can see the configuration in train.prototxt.