Response-Aware-Hybrid-Response-Generator

Only for inference

Overall Architecture

Note
Not published yet

Datasets

Datasets	Training	Validation	Testing
AI-Hub	656,332	28,361	28,464

Results of AI-Hub Dataset

Retrieval Model

Model	Hit@10	Hit@100	MRR@10	MRR@100
Q-CanAns	14.80	29.29	7.56	8.13
Q-CanQue	28.03	46.61	16.15	16.53
QR'-CanQueAns	36.61	55.28	19.50	21.41

Hybrid Response Generator

Model	BLEU-1	BLEU-2	BLEU-3	BLEU-4	Dist-1	Dist-2
NoRet	18.48	12.41	8.65	6.19	1.04	17.41
Q-CanAns	18.32	12.16	8.31	5.81	1.10	20.17
Q-CanQue	18.61	12.47	8.63	6.12	1.11	20.56
QR'-CanQueAns	19.96	13.72	9.84	7.31	1.16	22.12

API Inference

pip install -r requirements.txt

CUDA_VISIBLE_DEVICES=1 python main.py \
  --lang ko \
  --model hybrid \
  --corpus_v1 data/database_v1.tsv \
  --corpus_v2 data/database_v2.tsv \
  --retrieval_ckpt models/outputs/single_retrieval_model.pt \
  --generative_ckpt models/outputs/single_generative_model.pt \
  --retrieval_mode base \
  --num_centroids 64 \
  --n_beams 5 \
  --min_length 3 \
  --db_embedding_bsz 256 \
  --num_ks 3 \
  --max_len 80 

Hyperparameters

• end_command: user inference시 대화 중단을 위한 command key
  
• num_ks: 사용할 retrieved response 개수
• ret_max_len: 검색시 최대 token 개수
• mips: True는 maximum inner product 계산, False는 FAISS Lib 사용
• num_centroids: FAISS centroids 개수
• n_beams: 디코딩시 beam 개수
• min_length: 디코딩시 출력 최소 token 개수
• retrieval_ckpt: 검색 모델 checkpoint 위치
• generative_ckpt: 생성 모델 checkpoint 위치

Demo with FastAPI