huggingface

This is a clone of https://github.com/huggingface/transfer-learning-conv-ai

Train Command

Following command trains model n_epochs epochs, initialized with pretrained OpenAI-GPT, and saves checkpoint in log_dir.

python train.py --n_epochs 20 --model_checkpoint openai-gpt --log_dir runs/run_e20

Hyper-Parameter Settings

In our re-implementation, we used following hyper-parameters.

num_candidates: 2
max_history: 2
train_batch_size: 4
valid_batch_size: 4
gradient_accumulation_steps: 8
lr: 6.25e-5
lm_coef: 1.0
mc_coef: 1.0
max_norm: 1.0
n_epochs: 20
personality_permutations: 1

Evaluation Command

Each commands evaluates learned model using F1-score, hits@1 and perplexity.

python eval.py --eval_type f1 --model_checkpoint runs/run_e20
python eval.py --eval_type hits@1 --model_checkpoint runs/run_e20
python eval.py --eval_type ppl --model_checkpoint runs/run_e20

Sample Command

We implemented sampling script to sample utterances from learned model. It creates n_samples samples, iterating from the beginning of the dataset with step sample_term.

python sample.py --model_checkpoint runs/run_e20 --n_samples 10 --sample_term 1

Interaction Command

This command is to conversate interactively with the learned model.

python interact.py --model_checkpoint runs/run_e20 --temperature 0.7 --top_p 0.9 --max_length 20