This repo implements the modified conditional restricted Boltzmann machines (M-CRBMs) model in paper Modified Conditional Restricted Boltzmann Machines for Query
Recommendation in Digital Archives. The paper is in submitting progress.
Due to the huge size of the original training/test dataset, we only provide a small part of them in this repo.
To protect user information, we do not provide real search keywords, but only the search keyword IDs in training/test dataset.
dataset_train_<category>.npy
dataset_test_<category>.npy
This repo is tested on pytorch version 1.4.0 (CPU ONLY) on macOS, please ensure that correct pytorch version with CUDA version is installed on your computer.
Install the same pytorch version and other dependencies as the developer by directly run this conda command:
conda env create -f environment.yml
python train_crbm.py
Trained models are saved to model.pth, and all the metrics are saved to metrics.csv
If you want to set training/test parameter values manually, set them in ./crbm/config.py. The descriptions of the main parameters are listed below.
| Parameter Name | Description |
|---|---|
HIDDEN_UNITS |
Number of nodes in the hidden layer. |
BATCH_SIZE |
Batch size for both training data and test data. |
EPOCHS |
Training epoches. |
LEARNING_RATE |
Learning rate of the Adam optimizer (ignore this if use different optimizer). |
WEIGHT_DECAY |
Weight decay of the Adam optimizer (ignore this if use different optimizer). |
CD_K |
Number of steps of Gibbs sampling in contrastive divergence, default is 1. |
OPTIM |
Set optimizer, 'adam' or 'rms'. |
CATEGORY |
What categories will be used, this will be shown in trained model's file name. |