kimiyoung / review_net Goto Github PK

2017-06-19 14:59:55 [INFO ] Image channels are 3, Image width is 256 and Image height is 256
2017-06-19 14:59:55 [INFO ] found 1550 images in train db/root/digits/digits/jobs/20170601-111739-f754/val_db
2017-06-19 14:59:55 [INFO ] Loading network definition from /root/digits/digits/jobs/20170619-145952-06b8/model
Using CuDNN backend
2017-06-19 14:59:56 [FAIL] bad argument #1 to '?' (expecting number or torch.DoubleTensor or torch.DoubleStorage at /tmp/luarocks_torch-scm-1-1041/torch7/generic/Tensor.c:1153)
DIGITS Lua Error
stack traceback:
[C]: at 0x7fb4bdb2cda0
[C]: in function 'Tensor'
/root/torch/install/share/lua/5.1/nn/Linear.lua:6: in function '__init'
/root/torch/install/share/lua/5.1/torch/init.lua:91: in function </root/torch/install/share/lua/5.1/torch/init.lua:87>
[C]: in function 'Linear'
/root/digits/digits/jobs/20170619-145952-06b8/model.lua:113: in function 'createModel'
/root/digits/digits/jobs/20170619-145952-06b8/model.lua:133: in function 'network_func'
/root/digits/digits/tools/torch/main.lua:288: in main chunk
[C]: in function 'dofile'
/root/digits/digits/tools/torch/wrapper.lua:25: in function </root/digits/digits/tools/torch/wrapper.lua:25>
[C]: in function 'xpcall'
/root/digits/digits/tools/torch/wrapper.lua:25: in main chunk
[C]: in function 'dofile'
/root/torch/install/

what is the problem. Is the problem model? or different thing. Can you help me?

Thank you..

It seems that the max_common (https://github.com/kimiyoung/review_net/blob/master/code_caption/comp_prefix.lua#L2) is not correct. To access the ith element in the string s1, s1:sub(i,i) should be used. Not s1[i]. s1[i] is always nil. The correct CS-k metrics should be:

MODEL_FILE	reason.step8.litrans0.after1.merge0.model
#train batches	1074	#test batches	133
word_cnt	12859	token_cnt	20395
copy done
1	0.58108726620882
2	0.67359954807709
3	0.71993582684564
4	0.7437227668029
5	0.76213157058093
copy done
test loss	5.0341047093877

Hello, I have a little question here:
https://github.com/kimiyoung/review_net/blob/master/image_caption_offline/utils/model_utils.lua#L5
I know the function combines of many models, but it is so complex...
Why not set up a container and put all models into it and use the get parameters of the container? For example, If I want to combine parameters of model1,model2,model3
fake_net=nn.Sequential():add(model1):add(model2):add(model3)
parameters,gradients=fake_net:getParameters()

Hi,it is known to all that cnn-rnn model can benefit a lot from finetuning cnn. I want to ask why review network and many attention based model don't finetune cnn.thank you~

Thanks for sharing the code, it' s a wonderful job.
I have extracted the features and save them in data/
Howerer, I got an error with the following command :
th main.lua -model_pack reason_att_copy.lua -save_file -save_file_name model/ra.model

cloning 30
cloning soft_att_lstm
cloning reasoning lstm
/home/julius/torch/install/bin/luajit: bad argument #1 to '?' (string e
xpected, got nil)
stack traceback:
[C]: at 0x7f054d482ad0
[C]: in function 'DiskFile'
/home/julius/torch/install/share/lua/5.1/torch/File.lua:405: in
function 'load'
./dataloader.lua:127: in function 'gen_train_data'
...ect/review_net/image_caption_offline/reason_att_copy.lua:356
: in function 'train'
main.lua:33: in main chunk
[C]: in function 'dofile'
...lius/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:145
: in main chunk
[C]: at 0x00406670

Hello,

I ran your baseline system with weight sharing and without discriminative loss, I only got BLEU-4 around 25.6 instead of 28.2 reported in the paper. May I ask you how many epochs did you train your model ? I trained it for 3-4 epochs and then it converges somehow.

Just a small suggestion. You can use the Bottle module to do the same thing.

Is there any explanation for the computation of CS-k (https://github.com/kimiyoung/review_net/blob/master/code_caption/rev_reason.lua#L271-L284)?