Subspace Multinomial Model

• Learning document representations using subspace multinomial model. See paper
• This version of the code implements the same model, but with Adagrad optimization. This results in a slightly faster convergence with relatively lower memory requirements.

Requirements

• python3.6
• pytorch, numpy, scipy, scikit-learn

Data preparation

• python TwentyNewsDataset.py
• This will download the data from the web and converts it into scipy.sparse matrix.

Training

• Input data: scipy.sparse matrix of shape n_words x n_docs

• python run_smm_20news.py train -o exp/ -trn 100 -lw 1e-04 -rt l1 -lt 1e-4 -k 100

• The trained model is saved as exp/lw_1e-40_l1_1e-04_100/model_T100.pt

Positional parameters:

• phase: train or extract

Hyper parameters:

• -lw : l2 regularization const for i-vectors
• -rt : type of regularization for bases (l1 or l2)
• -lt : regularization const for bases
• -k : i-vector dimension

Other options:

• -o : path to output directory
• -trn: training iterations
• --ovr: over-write existing experiment directory

Extracting i-vectors:

• python run_smm_20news.py extract -m exp/lw_1e-04_l1_1e-04_100/model_T100.pt -xtr 30 --nth 2

• The document i-vectors are saved in exp/lw_1e-40_l1_1e-04_100/ivecs/

Other options:

• -xtr: extraction iterations.
• --nth: save every n-th i-vector while extraction.

Classification using GLC

• python train_and_clf.py exp/lw_1e-40_l1_1e-04_100/train_model_T100_e30.npy
• Test data and labels are automatically read.

On GPU

• prefix with CUDA_VISIBLE_DEVICES=<device_id> followed by python run_smm_20news.py