ICLR 2019 paper: "textTOvec: DEEP CONTEXTUALIZED NEURAL AUTOREGRESSIVE TOPIC MODELS OF LANGUAGE WITH DISTRIBUTED COMPOSITIONAL PRIOR"
This code consists of the implementations for the model proposed in the paper published at ICLR 2019: "textTOvec: DEEP CONTEXTUALIZED NEURAL AUTOREGRESSIVE TOPIC MODELS OF LANGUAGE WITH DISTRIBUTED COMPOSITIONAL PRIOR".
Paper: https://arxiv.org/pdf/1810.03947.pdf
Requires Python 3 (tested with 3.6.1). The remaining dependencies can then be installed via:
$ pip install -r requirements.txt
$ python -c "import nltk; nltk.download('all')"Datasets: A directory containing CSV files. There is expected to be 1 CSV file per set or collection, with separate sets for training, validation and test. The CSV files in the directory must be named accordingly for DocNADE model: training_docnade.csv, validation_docnade.csv, test_docnade.csv. The CSV files in the directory must be named accordingly for lstm portion of the model ctx-DocNADE(e): training_lstm.csv, validation_lstm.csv, test_lstm.csv. For this task, each CSV file (prior to preprocessing) consists of 2 string fields with a comma delimiter - the first is the label and the second is the document body.
Vocabulary files: A plain text file, with 1 vocabulary token per line (note that this must be created in advance, we do not provide a script for creating vocabularies), each for DocNADE and lstm portions.
mapping_dict.pkl: A dictionary that maps the indices of words in DocNADE to indices of words in LSTM data format (i.e., CSVs).
The script train_DATANAME_docnade_PPL.sh or train_DATANAME_docnade_IR.sh invokes train_model.py to train the baseline DocNADE model, compute PPL/IR and save it in a repository. It will also log all the information with the PPL and IR models in the seperate directories. Here's how to use the script:
$ ./train_20NSshort_docnade_PPL.sh
# to compute PPL
$ ./train_20NSshort_docnade_IR.sh
# to compute IR
--dataset is the path to the input dataset.
--docnadeVocab is the path to the vocabulary of the input dataset in DocNADE portion.
--model is the path to the save the best model.
--initialize-docnade Init DocNADE weights. False for DocNADE; True/False for ctx-DocNADE(e) model
--bidirectional True, if bidirectional settings reqquired in DocNADE. Default: False
--activation *sigmoid* for PPL and *tanh* for IR computations.
--learning-rate 0.001
--batch-size training batch szie, for instance, 100
--num-steps the number of training steps
--log-every
--validation-bs validation batch size, set to 1
--test-bs test batch size, set to 1
--validation-ppl-freq computate PPL of validation set at this frequency
--validation-ir-freq computate IR of validation set at this frequency
--test-ir-freq computate IR of test set at this frequency
--test-ppl-freq computate PPL of test set at this frequency
--num-classes number of class labels; not used.
--patience stopping criteria on validation scores
--supervised If training in supervised setting. Set to False.
--hidden-size The number of hidden units in a hiden vector
--combination-type The mode of combining hidden vectors from DocNADE and LSTM portions. Set to 'sum'
--vocab-size Voabulary size in DocNADE portion.
--deep True, if additional layers on both DocNADE and LSTM portions. Set False for non-deep versions.
--deep-hidden-sizes List of hidden sizes. For instance, for a two layered network, set: 200 200. Used, if deep = True
--trainfile is path to training text file. (required in case of topic coherence)
--valfile is path to validation text file. (required in case of topic coherence)
--testfile is path to testing text file. (required in case of topic coherence)
--reload True, if reloading of a model required. Set to False otherwise.
--reload-model-dir Path to the model to reload. Used only if reload set to True.
*** TO DO ***: Improve Documentation. The script train_20NSshort_docnade_lstm_PPL.sh or train_20NSshort_docnade_lstm_IR.sh invokes train_model_lstm.py to train the ctx-DocNADE or ctx-DocNADEe model, compute PPL/IR and save it in a repository. It will also log all the information with the PPL and IR models in the seperate directories. Here's how to use the script:
$ ./train_20NSshort_docnade_lstm_PPL.sh
# to compute PPL using ctx-DocNADE or crx-DocNADEe model i.e., textTOvec models
$ ./train_20NSshort_docnade_lstm_IR.sh
# to compute IR using ctx-DocNADE or crx-DocNADEe model i.e., textTOvec models
--dataset is the path to the input dataset.
--mapping-dict is the path to mapping_dict.pkl file.
--rnnVocab is the path to the vocabulary of the input dataset in LSTM portion.
--docnadeVocab is the path to the vocabulary of the input dataset in DocNADE portion.
--model is the path to the save the best model.
--initialize-docnade Init DocNADE weights. False for DocNADE; True/False for ctx-DocNADE(e) model
--bidirectional True, if bidirectional settings reqquired in DocNADE. Default: False
--activation sigmoid for PPL and tanh for IR computations.
--learning-rate 0.001
--batch-size training batch szie, for instance, 100
--num-steps the number of training steps
--log-every
--validation-bs validation batch size, set to 1
--test-bs test batch size, set to 1
--validation-ppl-freq computate PPL of validation set at this frequency
--validation-ir-freq computate IR of validation set at this frequency
--test-ir-freq computate IR of test set at this frequency
--test-ppl-freq computate PPL of test set at this frequency
--num-classes number of class labels; not used.
--patience stopping criteria on validation scores
--supervised If training in supervised setting. Set to False.
--hidden-size The number of hidden units in a hiden vector
--combination-type The mode of combining hidden vectors from DocNADE and LSTM portions. Set to 'sum'
--vocab-size Voabulary size in DocNADE portion.
--deep True, if additional layers on both DocNADE and LSTM portions. Set False for non-deep versions.
--deep-hidden-sizes List of hidden sizes. For instance, for a two layered network, set: 200 200. Used, if deep = True
--use-docnade-for-ir True, to log IR due to DocNADE portion *only* in the ctx-DocNADE ctx-DocNADEe model
--use-lstm-for-ir True, to log IR due to LSTM portion *only* in the ctx-DocNADE ctx-DocNADEe model
--use-combination-for-ir True, to log IR due to DocNADE+LSTM portion together in the ctx-DocNADE ctx-DocNADEe model
--initialize-rnn Init LSTM with Glove embeddings, i.e., ctx-DocNADEe
--update-docnade-w True, to update weights in DocNADE portion.
--update-rnn-w False, to not update embeddings in LSTM portion.
--lambda-hidden-lstm mixture weight, lambda in [0.0-1.0]
--trainfile is path to training text file. (required in case of topic coherence)
--valfile is path to validation text file. (required in case of topic coherence)
--testfile is path to testing text file. (required in case of topic coherence)
--reload True, if reloading of a model required. Set to False otherwise.
--reload-model-dir Path to the model to reload. Used only if reload set to True.
--reload-docnade-embeddings True, to init DocNADE portion of the ctx-DocNADE from a pretrained DocNADE model
--docnade-embeddings-path Path to topic-embedding matrix W from a pre-trained DocNADE model
*** TO DO ***: Improve Documentation. Datasets directory: ./datasets/
Pre-trained embeddings dir: /home/usr/resources/pretrained_embeddings/
Results directory: ./model/
Saved logs model dir: ./model/MODELNAME/logs/
NOTE: We will improve the readme and code documentation soon.