This code repository contains all code and links to datasets required for repeated experiments for a scientific publication introducing Neural Ordered Clusters (NOC), an adaptive, input-dependent clustering and cluster ordering method capable of making predictions for sets of varying cardinality. What follows is a description of the steps necessary to set up the environment and run experiments locally.
The following section explains how to obtain the freely available datasets and prepare the virtual environment.
The datasets needed to be downloaded prior to running the experiments are available under the following links:
- Gauss2D Ordered By Distance from Origin is generated based on the default configuration provided in the linked file's call to
main(). - Synthetic Catalogs are generated based on the default configuration file already provided within this repository here.
- PROCAT is also freely available and contains human-made product catalogs.
The PROCAT dataset requires additional preprocessing as provided in the procat_preprocess_for_bert.py file, which produces 3 csv files. The final directory tree should look like this:
|-- README.md
|-- danish-bert-botxo
|-- data
| |-- PROCAT
| | |-- PROCAT.test.csv
| | |-- PROCAT.train.csv
| | |-- PROCAT.validation.csv
| | |-- procat_dataset_loader.py
| | `-- procat_preprocess_for_bert.py
├── data_gauss2D.py
├── data_procat.py
├── data_synthetic.py
├── figures
│ ├── Gauss2D
│ ├── PROCAT
│ └── SyntheticStructures
├── inspect_gauss2D.py
├── inspect_synthetic.py
├── logs
├── model_noc.py
├── requirements.txt
├── run_configs
│ └── synthetic_rulesets.json
├── run_gauss2D.py
├── run_procat.py
├── run_synthetic.py
├── saved_models
│ ├── Gauss2D
│ ├── PROCAT
│ └── SyntheticStructures
├── show.py
├── utils.py
└── visual
The PROCAT experiment employs a language specific version for the BERT model. The PROCAT model is available for download here,
and should be placed in its danish-bert-botxo directory.
The code requires python 3.6.5. All requirements are listed in the requirements.txt and can be installed in the following way (on a Linux system):
python -m venv /path/to/new/virtual/environment
source <ven_pathv>/bin/activate
python -m pip install -r requirements.txt
If you wish to store & view repeated experimental results somewhere other than in the log files saved in the /logs directory, you will also need to set up a Mongo database (instructions here) and have a local omniboard instance running (more information here).
After finishing the preceding steps, you should be able to run each experiment through its corresponding .py file. The majority of the model architecture is defined in the NOC() class from model_noc.py. For example:
# Run 2D Gauss Experiment, with the default model configuration:
python run_gauss2D.py
The synthetic catalogs and PROCAT experiments generate the preprocessed datasets using parameters specified in their corresponding scripts. When a training script is running, comprehensive progress logs will be visible in the console and saved to a log file in the ./logs directory, e.g.:
2022-09-27 14:24:40,575 | INFO | EXPERIMENT run started, id: 1340156480
2022-09-27 14:24:40,686 | INFO | 1340156480 | Time started (UTC): 2022-09-27 14:24:40,686
2022-09-27 14:24:40,689 | INFO | 1340156480 | Seeding for reproducibility with: 1234
2022-09-27 14:24:40,690 | INFO | 1340156480 | Arguments: (...)
2022-09-27 14:24:40,692 | INFO | 1340156480 | Device: gpu
2022-09-27 14:24:40,801 | INFO | 1340156480 | The model has 12,117,655 trainable parameters
2022-09-27 14:25:38,906 | INFO | 1340156480 | 1 N:83 K:5 Clustering Loss: 115.910 Ordering Loss: 6.862 Cardinality Loss: 5.815 Total Loss: 128.586 Mean Time/Iteration: 0.6
(...)