Fail to run train.lua

[shuait]

This is maybe a trivial question but I'm completely new to torch, I tried to search on Google but no luck. I'm working with a Ubuntu 14.04 machine, cuda 7.0 and cudnn R4 version. I prepared all training files and when running train.lua it gives me this error:

{
input_img_train_h5 : "data/vqa_data_img_vgg_train.h5"
learning_rate_decay_every : 300
optim : "rmsprop"
hidden_size : 512
optim_epsilon : 1e-08
output_size : 1000
rnn_layers : 2
input_img_test_h5 : "data/vqa_data_img_vgg_test.h5"
losses_log_every : 600
id : "0"
input_ques_h5 : "data/vqa_data_prepro.h5"
learning_rate_decay_start : 0
start_from : ""
gpuid : 6
seed : 123
input_json : "data/vqa_data_prepro.json"
optim_beta : 0.995
batch_size : 20
iterPerEpoch : 1200
rnn_size : 512
max_iters : -1
checkpoint_path : "save/train_vgg"
save_checkpoint_every : 6000
learning_rate : 0.0004
co_atten_type : "Alternating"
feature_type : "VGG"
backend : "cudnn"
optim_alpha : 0.99
}
DataLoader loading h5 image file: data/vqa_data_img_vgg_train.h5
DataLoader loading h5 image file: data/vqa_data_img_vgg_test.h5
DataLoader loading h5 question file: data/vqa_data_prepro.h5
DataLoader loading json file: data/vqa_data_prepro.json
assigned 215375 images to split 0
assigned 121512 images to split 2
Building the model...
total number of parameters in word_level: 8031747
total number of parameters in phrase_level: 2889219
total number of parameters in ques_level: 5517315
constructing clones inside the ques_level
total number of parameters in recursive_attention: 2862056
/home/raamac/torch/install/bin/luajit: ./misc/word_level.lua:86: the class torch.CudaByteTensor cannot be indexed
stack traceback:
[C]: in function '__newindex'
./misc/word_level.lua:86: in function 'forward'
train.lua:253: in function 'lossFun'
train.lua:310: in main chunk
[C]: in function 'dofile'
...amac/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:145: in main chunk
[C]: at 0x00406670

[haibin894609937]

I have the same error , have you solved the problem?

[shuait]

@haibin894609937 I still has no clue what caused the problem, tried to reinstall torch and that failed too.

[haibin894609937]

I run it on centos7 cuda8.0 来自 魅族 PRO 5

…

-------- 原始邮件 -------- 发件人：Shuai Tang <[email protected]> 时间：周二 5月9日 21:29 收件人：jiasenlu/HieCoAttenVQA <[email protected]> 抄送：haibin894609937 <[email protected]>,Mention <[email protected]> 主题：Re: [jiasenlu/HieCoAttenVQA] Fail to run train.lua (#30) @haibin894609937<https://github.com/haibin894609937> I still has no clue what caused the problem, tried to reinstall torch and that failed too. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<#30 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AGon7qYRZuyyGl2-8czEbOf2NMIGJ5OFks5r4GongaJpZM4NFrCe>.

[Jhhuangkay]

You guys are working on VQA dataset, right?
If yes, I guess the problem is on your vqa_data_prepro.json and vqa_data_prepro.h5.
You can try to use other dataset the author provided, cocoqa.
If you replace the above two files by cocoqa_data_prepro.json and cocoqa_data_prepro.h5, all the code should run well. When I replace those two files, everything works well.
So, you also can try this, then you will know the problem is the generation of prepro files.