doubledaibo/clcaption_nips2017

Contrastive Learning for Image Captioning

clcaption_nips2017's Issues

Can't reproduce the result.

Hello Bo Dai,

I use cl_train_adaptiveattention.lua,
and I found that I can't reproduce the result in your paper.

After the 1st epoch training, I found that the CIDEr score is lower than the original model(before training) by ~0.15,
the training loss is about 4.01 and val loss is about 1.40.

As the training continues, training loss goes down slowly to 3.8
and val loss goes up and down between 1.31.4,
although the CIDEr score is improving, it still can't go higher than the original score.

I use the pre-trained model in AdaptiveAtteion as the target model and reference model, resnet-152 as CNN model and learning rate follows the value in your paper, which is 1e-6.

Could you please help me with this?
Thanks.

luajit: ./misc/utils.lua:28: attempt to index local 'file' (a nil value)

When several epochs' training finished and the code start to save the checkpoint, the error message shows:

/home/wcai/torch/install/bin/luajit: ./misc/utils.lua:28: attempt to index local 'file' (a nil value)
stack traceback:
./misc/utils.lua:28: in function 'write_json'
./misc/net_utils.lua:398: in function 'language_eval'
test.lua:261: in function 'evaluate_split'
test.lua:405: in main chunk
[C]: in function 'dofile'
...wcai/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
[C]: at 0x00405d50

But when I use the code of neuraltalk2, I can successfully save the checkpoint:

evaluating validation performance... 3200/3200 (9.179654)
validation loss: 9.1724067846911
wrote json checkpoint to model_id.json
iter 1: 9.124193
iter 2: 9.075657
...

class nn.OneHot has been already assigned a parent class

wcai@tdtd-desktop ~/git/clcaption_nips2017 $ th cl_train_cross.lua
/home/wcai/torch/install/bin/luajit: /home/wcai/torch/install/share/lua/5.1/trepl/init.lua:389: /home/wcai/torch/install/share/lua/5.1/trepl/init.lua:389: /home/wcai/torch/install/share/lua/5.1/torch/init.lua:102: class nn.OneHot has been already assigned a parent class

stack traceback:
[C]: in function 'error'
/home/wcai/torch/install/share/lua/5.1/trepl/init.lua:389: in function 'require'
cl_train_cross.lua:11: in main chunk
[C]: in function 'dofile'
...wcai/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
[C]: at 0x00405d80

I am new to torch, did I miss any package for running this code?Thanks.