Awesome

Creating the Environment

To create the environment, run

conda env create --file enviornment.yaml

And before running any code, you have to add src to your pythonpath and activate the environment, which you can do with:

source init.sh

Running Experiments

Models can be found in configs/models Data can be found in configs/data

To run a model on a single dataset, run:

python src/run_config.py model=<model_config_name> data=<data_config_name> [train=True] {...}

Include train=True to train the model on the dataset, otherwise the model will be loaded and evaluated. If there is no trained model a new one will be initialized and evaluated (which is likely undesirable!). All other config options can be overriden at the command line. Model options should be prefixed with "model." and data options hould be prefixed with "data.".

To run a model on all datasets, run:

python scripts/run_experiments.py model=<model_config_name> data=<data_config_name> [train=True] {...}

For example:

python scripts/run_experiments.py model=adaptive-prompt_prefix-lstm_bert-base-cased model.seed=37 train=True

Will train a prefix-lstm model with BERT-base-cased used as the large language model. It will train on the id_subsample dataset, and evaluate on the ood datasets.

Mixture of Experts and Oracle are a bit different because their configs specify paths to their component models--relation classifier (relclf) and ptuning. Do not train these. An example for MOE is:

python scripts/run_experiments.py model=moe_roberta-large model.relclf.model_path=configs/model/relclf_bert-base-cased-sd36.yaml model.relclf.data_path=configs/data/id_subsample model.ptuning.model_path=configs/model/p-tuning_sd36_roberta-large.yaml

And oracle only needs the ptuning path

python scripts/run_experiments.py model=oracle_roberta-large model.ptuning.model_path=configs/model/p-tuning_sd36_roberta-large.yaml

For training the P-tuning embeddings, we only use the subject, object pairs, not the filled-in templates, so to greatly speed up training, we use just the relations from the LAMA dataset, so each pair is used only once per epoch:

python src/run_config.py data=id model=p-tuning_bert-base-cased train=True data.train.template_path=data/templates/relations_lama.json data.dev.template_path=data/templates/relations_lama.json data.test.template_path=data/templates/relations_lama.json

References

This repo is built off the repo found here