-----------------This is the pairwise argument ranking model developed by Kevin Shi.-----------------
Credit for Dataset: https://mailman.uib.no/public/corpora/2016-June/024842.html
I have not uploaded the dataset to this repository, although I might at some later point.
main.py has now been updated to match proto_main.py after confirmation that proto_main.py functions as desired. To run main.py, please follow the same instructions given to run proto_main.py.
To run proto_main.py, run the following twice in terminal:
python proto_main.py
First execution runs training and the second execution runs evaluation.
IMPORTANT
Most of the attributes affecting training are in the config.ini file. These are the important variables hard-coded into the main program that SHOULD BE MOVED OVER TO THE CONFIG FILE by me at some point:
- train_test_split(0.1): Split hard coded to be 90% train and 10% validation
- logging_strategy: Set to steps
- model_path: Directory to save and load model to and from is hard coded here.
- os.environ['CUDA_VISIBLE_DEVICES']: Currently set to 4 for my machine, but should be variable in config.
- rank and world_size: Both are set to 4 to match the number of visible GPUs. Should be made dependent on length of visible GPU list.
proto_main.py is my development file; main.py is the last stable version.