floeaster / kern_reproduce

This is the repo for paper: Knowledge Enhanced Neural Fashion Trend Forecasting

Requirements

1. OS: Ubuntu 16.04 or higher version
2. python3
3. Supported(tested) CUDA Versions: V7.5, V9.1, V10.2
4. python modules: refer to the modules in requirements.txt

Code Structure

1. The entry script for training is: train.py
2. Test the model using the saved model: test.py
3. An auxiliary script to generate all the testing results (i.e., the results in Table 3&4 shown in the companinon paper) together: reproduce.py
4. The config file is: config.yaml
5. The script for data preprocess and dataloader: utility.py
6. The model folder: ./model/

How to Run

1. Download the dataset, decompress it and put it on the top directory: tar -zxvf dataset.tgz Note that, the downloaded files include both datasets of GeoStyle and FIT.

2. Change the hyper-parameters in the configure file config.yaml. First, you should specify which dataset you want to use in the first line of config.yaml; second, you can use any pre-defined settings under each dataset (to make sure only keep one set of settings and comment all the others). Note that, all the settings correpsonding to the results in the companion paper's table 3/4 have been provided in the config.yaml, of which one is active and all the others are commented.

3. Run the training: train.py

4. Log/Runs: after running a eperiment under a certain setting, it will generate two log files: one is under ./log, which consists of the predictions, learned embeddings, screen outputs, and the learned model parameters; the other is ./runs, which is the log file that can be visualized by Tensorboard. You can go into the ./runs directory, execute "tensorboard --host="your host ip" --logdir=./", then you can see the training curves by a browser. Note that, you need to install the tensorboard to use this function.

5. Test using saved model: test.py.

• The best model will be saved under the folder ./log. You can run this script to load the saved model and produce the metrics score on the test set. Note that this script shares the same config file with train.py, i.e., config.yaml. In other words, you can change the config.yaml to decide which model to load and test. Specifically, we include in this repo the saved models whose performance are reported in this companion paper. You can directly uncomment each corresponding part in the config.yaml and run the test.py to reproduce the results.

6. Generate all the results in Table 3&4 of the reproduce paper: reproduce.py

Acknowledgement

This project is supported by the National Research Foundation, Prime Minister's Office, Singapore under its IRC@Singapore Funding Initiative.