• Python 3.6
• Pytorch 0.1.12
• CUDA 8.0

The dataset of HabeasCorpus should be existed under the folder data, which is download from ReviewNet. Then we pre-process the dataset as follows:

python3.6 prepro_gen_data.py

We can obtain the following data:

• index2token.pkl
• token2index.pkl
• train_data.pkl
• val_data.pkl
• test_data.pkl

To train the encoder-decoder model with MLE, run

./bash_code_caption_ende_xe.sh

Then, we fune-tuning the network with our ARNet,

./bash_code_caption_ende_rcst.sh

To test the model with greedy search, run

./bach_code_caption_ende_inf.sh

To visualize the hidden states with t-SNE, run

./bash_code_caption_ende_vis.sh

We will get the hidden states of generated sequence. Then, run

python3.6 prepro_tsne_reduction_vis.py --vis_batch_size 80 \
                                       --truncation 0  \
                                       --hidden_path '...' \
                                       --hidden_reduction_save_path '...'

We reimplement the ReviewNet proposed in Yang with PyTorch. To train the ReviewNet with MLE, run

./bash_code_caption_reviewnet_xe.sh

To test the ReviewNet with greedy search, run

./bash_code_caption_reviewnet_inf.sh

To train the attention-based model with MLE, run

./bash_code_caption_soft_att_xe.sh

To fine-tuning the model with our ARNet based on the MLE model, run

./bash_code_caption_soft_att_rcst.sh

To test the model, run

./bash_code_caption_soft_att_inf.sh

To get the hidden states of the whole sequence, run

./bash_code_caption_soft_att_vis.sh

Then, visualize the hidden state as follows:

python3.6 prepro_tsne_reduction_vis.py --vis_batch_size 80 \
                                       --truncation 0  \
                                       --hidden_path '...' \
                                       --hidden_reduction_save_path '...'