Hey,
I'm trying to re-implemented your method using tensorflow, while I'm not very proficient in torch. I'm quite confused about *** line 24 *** in AttentionPartsCRF.lua, does that mean just copy the same tensor ( *** Q[itersize] ) on its channel for many( numIn ***) times?
I'm quite confused about that and hope to get an reply. Thanks. @bearpaw

Thanks to your README, I've accomplished test on MPII dataset and got a '.mat' file, which is a 11731163 matrix. But MPII README confuses me. They didn't explicitly indicate the input for their 'evaluatePCKh.m’.
So how can I convert '.mat' to calculate the metric, just as shown on the leaderboard?

Hi,
When I was training, the accuracy of the training set was very high, but the accuracy of the validation set was very low, about 30%. I trained a total of 100 epochs, increased the number of iterations in the source code, and correspondingly reduced the batchsize, but I can ensure that every picture has been trained, so who can tell me why? thx.

Thanks for you code! could you give me some advice? thanks a lot!

sun@sunwin:~/0newcodedown/pose-attention-master/test$ qlua main.lua demo
qlua: /home/sun/torch/install/share/lua/5.1/hdf5/ffi.lua:42: Error: unable to locate HDF5 header file at /usr/local/HDF_Group/HDF5/1.8.18/include;/hdf5.h
stack traceback:
[C]: at 0x7f6031a5bf50
[C]: in function 'error'
/home/sun/torch/install/share/lua/5.1/hdf5/ffi.lua:42: in function 'loadHDF5Header'
/home/sun/torch/install/share/lua/5.1/hdf5/ffi.lua:60: in main chunk
[C]: in function 'dofile'
/home/sun/torch/install/share/lua/5.1/torch/init.lua:54: in function 'include'
/home/sun/torch/install/share/lua/5.1/hdf5/init.lua:29: in main chunk
[C]: in function 'require'
/home/sun/0newcodedown/pose-attention-master/test/util.lua:7: in main chunk
[C]: in function 'dofile'
main.lua:9: in main chunk

As title mentioned.

Hello , I download your pretrain model(model.t7).But when I run 'qlua main.lua demo ',the terminal give me an error :
{{{ THCudaCheck FAIL file=/home/zhangjiqing/torch/extra/cutorch/lib/THC/generic/THCStorage.cu line=66 error=2 : out of memory
qlua: ...zhangjiqing/torch/install/share/lua/5.1/nn/Container.lua:67:
In 1 module of nn.Sequential:
In 1 module of nn.ConcatTable:
In 1 module of nn.Sequential:
/home/zhangjiqing/torch/install/share/lua/5.1/nn/THNN.lua:110: cuda runtime error (2) : out of memory at /home/zhangjiqing/torch/extra/cutorch/lib/THC/generic/THCStorage.cu:66 }}}

Is it just out of memory of GPU ? Can I turn the batchsize to a small value yet using your model.t7 ? Or could you provide a samll batchsize for me?
Thanks!

Hi,
In your paper,multi-resolution attention maps are generated from features of different scales.But in your train script , only original resolution generates attention map ， is it my fault?

I see that in train.sh, it use 4x12G GPUs for training, but i only have one 12G GPU(1080 Ti), so how can i reduce the GPU cost?

i try change '-nGPU 4' to '-nGPU 1', and add '-trainBatch 4' in train.sh, but it seems doesn't work because the model 'converting the model to CUDA' again and again, just like:

So how can i do?Thanks.

hi,
I'm confused about the refined feature map. Should it be h2^att rather than h1^att?
In your paper, Equ.10 describes the refined feature h2^att is generated from h1^att, but in section 5.2.4, only the h1^att is used as refined feature and to do final prediction.

This is also conflict with Figure 4 in which the h2^att is the refined feature and used for heatmap prediction.

If h1^att is the refined feature, what does the h2^att used for?

Thank you~

Hi,

How to install qlua in ubuntu?

Great work!
Thank you for sharing this code.
I encounter a problem when I test your code named "main.lua". I input this command "qlua main.lua demo" in command of Ubuntu. The bug is as follows:

HDF5-DIAG: Error detected in HDF5 (1.8.11) thread 139704096257792:
#000: H5F.c line 586 in H5Fopen(): interface initialization failed
major: Function entry/exit
minor: Unable to initialize object
#1: H5F.c line 112 in H5F__init_pub_interface(): unable to initialize interface
major: File accessibilty
minor: Unable to initialize object
#2: ../../../src/H5I.c line 344 in H5I_register_type(): invalid hash size
major: Object atom
minor: Out of range
qlua: /home/yzl/torch/install/share/lua/5.2/hdf5/file.lua:12: HDF5File: fileID -1 is not valid
stack traceback:
[C]: in ?
[C]: in function 'error'
/home/yzl/torch/install/share/lua/5.2/hdf5/file.lua:12: in function '__init'
/home/yzl/torch/install/share/lua/5.2/torch/init.lua:91: in function </home/yzl/torch/install/share/lua/5.2/torch/init.lua:87>
[C]: in function 'HDF5File'
/home/yzl/torch/install/share/lua/5.2/hdf5/file.lua:147: in function </home/yzl/torch/install/share/lua/5.2/hdf5/file.lua:145>
(...tail calls...)
.../AI_Challenge/skeleton/Code/pose-attention/test/util.lua:19: in function 'loadAnnotations'
main.lua:113: in main chunk
HDF5: infinite loop closing library
D,T,FD,P,FD,P,FD,P,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E,E

Thank you!

Hi @bearpaw
Very nice work!
I have not found the batch size for MPII and LSP dataset in your CVPR 2017 paper.
Would you please provide your train.log valid.log for reference , the batch size for MPII and LSP dataset and the train/valid iterations per epoch for MPII and LSP dataset in your paper.

Thank you.
Feng.

Hi @bearpaw
I see the function postprocess in 'test/util.lua' just add 0.25 or 0.75 of the original coods, and could give about 0.5([email protected]) improvement.

Can you explain it? Thanks in advance.

Hi,
I had a few doubts in model creation

• Firstly is there any specific reason for having more residual modules before stacked hourglasses compared to Newell's Model Creation. (He only uses 3 residuals while hg-attention model uses 6 residuals). Do these extra residual modules provide better performance?

• Secondly if I want to create a model with less stacks (say nStacks 2 would I need to make changes to the implementation since you have different attention mechanism for later stacks as in following lines
  if i>4 then att = AttentionPartsCRF(opt.nFeats, ll2, opt.LRNKer, 3, 0) tmpOut = AttentionPartsCRF(opt.nFeats, att, opt.LRNKer, 3, 1)

Thank You