Implementation of "Order-free Learning Alleviating Exposure Bias in Multi-label Classification".

Feel free to use/modify them, any bug report or improvement suggestion will be appreciated. If you find this project helpful for your research, please do consider to cite our paper, thanks!

1. Python packages:
   • Python 3.5 or higher
   • Pytorch 1.0 or higher
   • Numpy

• json
• yaml

1. Download data:

   • AAPD,RCV1: https://github.com/illcat/SGM-for-Multi-label-Classification

   • Reuters: Run data/preprocess_reuters.py

   • Move data to data/$dataset, which is a directory contains text_train, label_train.....

2. Preprocess data:

   bash data/preprocess.sh
   

   • This script contains the codes for preprocessing the three datasets.

   • The script will transcibe the labels into vector format.

   • The order of labels is from frequent to rare.

   • Modify the path of dataset in data/preprocess.sh.

3. Train model:

   python3 train_rnn.py -gpus 0 -config config/config_$dataset.yaml
   

   • Hyperparameters can be modified in config/config_$dataset.yaml

   • Log can be found in the log directory.

   python3 train_logistic_baseline.py -gpus 0 -config config/config_$dataset.yaml
   

   • Codes for training binary relevance model.

4. test model:

   python3 train_rnn.py -gpus 0 -config config/config_$dataset.yaml -restore $expdir/best_in_train_micro_f1_checkpoint.pt -notrain
   

models : codes for model structure

metrics.py : metrics for multi-label classification

optims.py : code for optimizer and defining gradient clipping.

preprocess.py : code for preprocessing

data: The path of file, save_data. e.g. './data/AAPD/save_data'

epoch: Number of epoch for training.

train_batch_size: Batch size for training.

test_batch_size: Batch size for testing for beam search.

log: Directory for log files. e.g. './exp/aapd/ocd'

emb_size: Size of word embedding.

load_emb: Load pretrained word vectors. (We set false for random initialization)

emb_path: Path of pretrained word embedding. (if load_emb is true)

hidden_size: Hidden size for LSTM cell.

encoder_n_layers: Number of layers for LSTM encoder.

decoder_n_layers: Number of layers for LSTM decoder.

input_dropout_p: Probability of dropout for input of encoder.

dropout_p: Probability of dropout for RNN encoder and decoder.

bidirectional: BLSTM or not.

logistic_weight: weight of loss between BR decoder and RNN decoder.

max_tgt_len: maximun number of decoding steps.

loss_type: vallina, OCD ,or order_free

beam_size: beam size for decoding

add_mask: mask to prevent rnn decoder generate same labels.

OCD_temperature_start: Start temperature for OCD.

OCD_temperature_end: End temperature for OCD.

OCD_final_hard_epoch: Number of epoches for ocd temperature to reach OCD_temperature_end (Linear decay).

eval_interval: Number of updates to check the performance in validation set.

print_interval: Number of updates to print the current average loss.

@article{tsai2019order,
  title={Order-free Learning Alleviating Exposure Bias in Multi-label Classification},
  author={Tsai, Che-Ping and Lee, Hung-Yi},
  journal={arXiv preprint arXiv:1909.03434},
  year={2019}
}