Awesome

KRLS: Improving E2E Response Generation in TOD with Reinforced Keywords Learning

This is the repository for the paper: KRLS: Improving End-to-End Response Generation in Task Oriented Dialog with Reinforced Keywords Learning

Datasets

We use the MultiWoZ dataset for training and evaluation. You can download already preprocessed dataset files from HERE, and place them under the dataset folder like this:

dataset
├── MultiWOZ_2.0
├── MultiWOZ_2.1
└── MultiWOZ_2.2

If you want to preprocess the dataset by yourself, we used the same preprocessing scripts implemented by Zhang et al., 2020, and Lee, 2021. Please refer to their repositories for more details.

Training

Note: we track our experiments using wandb on Weights & Biases. We highly recommend you to setup wandb to make sure our code works properly.

Dependencies

If you encounter some problems with running our code with your existing setup, try the following:

transformers==4.28.0, sentencepiece==0.1.96, torch==1.13.0, numpy==1.20.3

Pretrained Weights

We use GODEL-base as pretrained weights, which you can download from the official github repo (which may change/be updated by the corresponding authors) or directly from here HERE from our own google drive. Place the downloaded weights under the model_checkpoints folder like this:

model_checkpoints
└── godel_cp
    └── GODEL-Base # here

In our experiments, we also train from a SL-finetuned checkpoint of GODEL-base, which you can download from HERE. This will be used for model training later. Place the downloaded weights and place them under the model_checkpoints folder like this:

model_checkpoints
├── godel_cp
└── ppotod
	└──baseline # here
		├── ckpt-epoch8
		├── log.log
		└── run_config.json

Training Script

To train with the best hyperparameters, run the following command:

MODEL_DIR= # where to save the model
python main.py -mode train \
	--config_path model_checkpoints/preset_configs/best.json \
	--model_dir $MODEL_DIR --log_file $MODEL_DIR/log.log \
	--exp_group_name debug \  # saves exp log to wandb
	--batch_size 4 --epochs 4 \
	--is_policy_optimization false 

We trained all of our models on a single NVIDIA RTX A4000, which takes about 1 day to train. Alternatively, you can directly download our best checkpoint from HERE, and place it under the model_checkpoints folder like this:

model_checkpoints
└── ppotod_reprod
	└── best # here
	    ├── ckpt-epoch2
	    └── preds

Evaluation

We use the standard multi-woz evaluation script to evaluate the predictions. You can find the predictions for our best checkpoints in the outputs folder:

outputs/
├── krls-e2e.json  # E2E response generation in MultiWoZ
├── krls-policy.json  # policy optimization in MultiWoZ
└── krls-gold_dst_n_act.json  # additional exp using both gold DST and SYS ACT

To generate predictions and evaluate a checkpoint manually (e.g. using the best checkpoint):

CKPT_DIR=model_checkpoints/ppotod_reprod/best/ckpt-epoch2

1. to generate E2E response from a checkpoint path:

   python main.py -mode predict \
   	--ckpt $CKPT_DIR --version 2.2 \
   	--output $CKPT_DIR/preds.json \
   	--batch_size 8 \  # larger batch size will be faster
   	--use_true_curr_aspn false \
   	--is_policy_optimization false
   

   to generate with gold DST (policy optimization):

   python main.py -mode predict \
   	--ckpt $CKPT_DIR --version 2.2 \
   	--output $CKPT_DIR/preds.json \
   	--batch_size 8 \
   	--use_true_curr_aspn false \
   	--is_policy_optimization true
   

   to generate with both gold DST and gold system ACT:

   python main.py -mode predict \
   	--ckpt $CKPT_DIR --version 2.2 \
   	--output $CKPT_DIR/preds.json \
   	--batch_size 8 \
   	--use_true_curr_aspn true \
   	--is_policy_optimization true
   

2. to evaluate the predictions:

   python evaluate.py -input $CKPT_DIR/preds.json
   

Other Experiments

Reward Function Ablation

You can swap out different reward functions which we implemented by:

python main.py -mode train \
	# other args omitted
	--reward token_prob  # here

available options include:

• "zeros": zero per-token reward, hence only using the terminal reward
• "token_error": hard penalty of ±μ for each correct/incorrect token
• "token_contextual_sim": BERTScore like contextual similarity
• "token_sim": cosine similarity from GODEL embeddings
• "token_prob": used in the paper

Training with PG instead of PPO

Uncomment the following line in main.py:

if __name__ == "__main__":
	# train with PPO
	# cfg, trainer = train_ppotod()
	# train with PG 
	cfg, trainer = train_pgtod() ### UNCOMMENT THIS LINE

Training from Scratch/Backbone

To train from the backbone (--ckpt none) instead of the SL-finetuned checkpoint:

python main.py -mode train \
	--config_path model_checkpoints/preset_configs/best.json \
	--model_dir $MODEL_DIR --log_file $MODEL_DIR/log.log \
	--exp_group_name debug \  # saves exp log to wandb
	--batch_size 10 --epochs 4 \
	--is_policy_optimization false \
	--ckpt none

Training with SL Only

You can train with only performing SL by switching off the ppo training. For example:

python main.py -mode train \
	--config_path model_checkpoints/preset_configs/best.json \
	--model_dir $MODEL_DIR --log_file $MODEL_DIR/log.log \
	--exp_group_name debug \  # saves exp log to wandb
	--batch_size 4 --epochs 10 \
	--is_policy_optimization false \
	--ckpt none --kl_loss_coeff 0.0 \
	--use_ppo false

finetunes from backbone only using the SL objective (--kl_loss_coeff 0.0 and --use_ppo false).

Training with standard RL

You can train with standard RL + auto-regressive generation by first uncommenting the following line in main.py:

if __name__ == "__main__":
	# train with PPO
	# cfg, trainer = train_ppotod()
	# train with real RL only
	cfg, trainer = train_realrl_tod()  # this line

Then train with switching off SL training, per-token reward, and use all terminal reward (see last line):

python main.py -mode train \
	--config_path model_checkpoints/preset_configs/best.json \
	--model_dir $MODEL_DIR --log_file $MODEL_DIR/log.log \
	--exp_group_name debug \  # saves exp log to wandb
	--batch_size 4 --epochs 4 \
	--is_policy_optimization false \
	--use_sl false --reward zeros --terminal_reward_fn all

Parameter Tuning

Please see sweep.py for the hyperparameter sweeps using ray tune and wandb. If you already have a wandb account and have setup ray tune on your machine, you can run the following command to start the sweep:

export CUDA_VISIBLE_DEVICES=#specify how many GPUs to use
python sweep.py