#TWEAC: Transformer with Extendable QA Agent Classifiers We use a easily extendable Transformer model which has a dedicated classification head for each agent. This makes adding (or removing) new agents as easy as adding a new head.
Run python run_transformer.py $config.yaml to train and/ or test a model.
See configs/example_config.yaml for an example config detailing the parameters
and see configs/standard_train_eval.yaml for a config to train & evaluate a model on 10 agents.
To extend a model, you will need to update all_agents in the config along with
setting agents_extended: 1 and providing a base_model which should be extended.
If you want to extend an extended model, you will first have to merge the old classification head with the extended one - see merge_extended in iterative_add.py for details.
We provide our code for iterative extension (with optional down-sampling of training examples for not-extended agents to achieve constant training time regardless of agents - see half-and-half sampling in our publication for more details).
We also provide code for our leave-one-out evaluation setup.
We iteratively extend a TWEAK model initially trained with some agents one by one with multiple agents.
Run python iterative_add.py $config --data_config=$data_config --stop=90 --start=0.
See configs/iterative_extend_reddit.yaml for an example config for reddit
and see ../data/dataconfig_reddit.yaml for an example data config.
Note, you will have to remove the first 10 agents from the data config if you want to use it with the config for iterative extension and you will have to train a TWEAK model with those initial 10 agents (see above on how to train your model).
We evaluate extension in a leave-one-out approach where given k agents, we train models with k-1 agents and extend with the remaining one, repeating this for all k agents.
Run python leave_one_out.py --modes "finetune,extend" --econfig=$extend_config --fconfig=$finetune_config
For the finetune config, you can use a standard training config (e.g. configs/standard_train_eval.yaml).
Note, the folders for the finetuned models will be placed in folders in out_dir and will contain the left-out agent.
For the extend config, see configs/loo_extend.yaml.
Note, base_model has to point out_dir of the finetune config.