models.py

import os
from typing import Any, Dict

from transformers import (Pipeline, T5ForConditionalGeneration, T5Tokenizer,
                          pipeline, AutoModelForSeq2SeqLM, AutoModelForCausalLM, AutoTokenizer)

auth_token = os.environ.get("CLARIN_KNEXT")


DEFAULT_DST_INPUTS: Dict[str, str] = {
    "polish": (
        "[U] Chciałbym zarezerwować stolik na 4 osoby na piątek o godzinie 18:30. "
        "[Dziedzina] Restauracje: Popularna usługa wyszukiwania i rezerwacji restauracji "
        "[Atrybut] Czas: Wstępny czas rezerwacji restauracji"
    ),
    "english": (
        "[U] I want to book a table for 4 people on Friday, 6:30 pm. "
        "[Domain] Restaurants: A popular restaurant search and reservation service "
        "[Slot] Time: Tentative time of restaurant reservation"
    ),
}


DST_MODELS: Dict[str, Dict[str, Any]] = {
    "plt5-small": {
        "model": T5ForConditionalGeneration.from_pretrained("clarin-knext/plt5-small-dst", use_auth_token=auth_token),
        "tokenizer": T5Tokenizer.from_pretrained("clarin-knext/plt5-small-dst", use_auth_token=auth_token),
        "default_input": DEFAULT_DST_INPUTS["polish"],
    },
    "plt5-base": {
        "model": T5ForConditionalGeneration.from_pretrained("clarin-knext/plt5-base-dst", use_auth_token=auth_token),
        "tokenizer": T5Tokenizer.from_pretrained("clarin-knext/plt5-base-dst", use_auth_token=auth_token),
        "default_input": DEFAULT_DST_INPUTS["polish"],
    },
    "plt5-base-poquad-dst-v2": {
        "model": T5ForConditionalGeneration.from_pretrained("clarin-knext/plt5-base-poquad-dst-v2", use_auth_token=auth_token),
        "tokenizer": T5Tokenizer.from_pretrained("clarin-knext/plt5-base-poquad-dst-v2", use_auth_token=auth_token),
        "default_input": DEFAULT_DST_INPUTS["polish"],
    },
    "t5-small": {
        "model": T5ForConditionalGeneration.from_pretrained("clarin-knext/t5-small-dst", use_auth_token=auth_token),
        "tokenizer": T5Tokenizer.from_pretrained("clarin-knext/t5-small-dst", use_auth_token=auth_token),
        "default_input": DEFAULT_DST_INPUTS["english"],
    },
    "t5-base": {
        "model": T5ForConditionalGeneration.from_pretrained("clarin-knext/t5-base-dst", use_auth_token=auth_token),
        "tokenizer": T5Tokenizer.from_pretrained("clarin-knext/t5-base-dst", use_auth_token=auth_token),
        "default_input": DEFAULT_DST_INPUTS["english"],
    },
    "flant5-small [EN/PL]": {
        "model": T5ForConditionalGeneration.from_pretrained("clarin-knext/flant5-small-dst", use_auth_token=auth_token),
        "tokenizer": T5Tokenizer.from_pretrained("clarin-knext/flant5-small-dst", use_auth_token=auth_token),
        "default_input": DEFAULT_DST_INPUTS["english"],
    },
    "flant5-base [EN/PL]": {
        "model": T5ForConditionalGeneration.from_pretrained("clarin-knext/flant5-base-dst", use_auth_token=auth_token),
        "tokenizer": T5Tokenizer.from_pretrained("clarin-knext/flant5-base-dst", use_auth_token=auth_token),
        "default_input": DEFAULT_DST_INPUTS["english"],
    },
}


DEFAULT_ENCODER_DECODER_INPUT_EN = "The alarm is set for 6 am. The alarm's name is name \"Get up\"."
DEFAULT_DECODER_ONLY_INPUT_EN = f"[BOS]{DEFAULT_ENCODER_DECODER_INPUT_EN}[SEP]"
DEFAULT_ENCODER_DECODER_INPUT_PL = "Alarm jest o godzinie 6 rano. Alarm ma nazwę \"Obudź się\"."
DEFAULT_DECODER_ONLY_INPUT_PL = f"[BOS]{DEFAULT_ENCODER_DECODER_INPUT_PL}[SEP]"



NLG_MODELS: Dict[str, Dict[str, Any]] = {
    # English
    "t5-large": {
        "model": AutoModelForSeq2SeqLM.from_pretrained("clarin-knext/utterance-rewriting-t5-large", use_auth_token=auth_token),
        "tokenizer": AutoTokenizer.from_pretrained("clarin-knext/utterance-rewriting-t5-large", use_auth_token=auth_token),
        "default_input": DEFAULT_ENCODER_DECODER_INPUT_EN,
    },
    "en-mt5-large": {
        "model": AutoModelForSeq2SeqLM.from_pretrained("clarin-knext/utterance-rewriting-en-mt5-large", use_auth_token=auth_token),
        "tokenizer": AutoTokenizer.from_pretrained("clarin-knext/utterance-rewriting-en-mt5-large", use_auth_token=auth_token),
        "default_input": DEFAULT_ENCODER_DECODER_INPUT_EN,
    },
    "gpt2": {
        "model": AutoModelForCausalLM.from_pretrained("clarin-knext/utterance-rewriting-gpt2", use_auth_token=auth_token),
        "tokenizer": AutoTokenizer.from_pretrained("clarin-knext/utterance-rewriting-gpt2", use_auth_token=auth_token),
        "default_input": DEFAULT_DECODER_ONLY_INPUT_EN,
    },

    "pt5-large": {
        "model": AutoModelForSeq2SeqLM.from_pretrained("clarin-knext/utterance-rewriting-pt5-large", use_auth_token=auth_token),
        "tokenizer": AutoTokenizer.from_pretrained("clarin-knext/utterance-rewriting-pt5-large", use_auth_token=auth_token),
        "default_input": DEFAULT_ENCODER_DECODER_INPUT_PL,
    },
    "pl-mt5-large": {
        "model": AutoModelForSeq2SeqLM.from_pretrained("clarin-knext/utterance-rewriting-pl-mt5-large", use_auth_token=auth_token),
        "tokenizer": AutoTokenizer.from_pretrained("clarin-knext/utterance-rewriting-pl-mt5-large", use_auth_token=auth_token),
        "default_input": DEFAULT_ENCODER_DECODER_INPUT_PL,
    },
    "polish-gpt2": {
        "model": AutoModelForCausalLM.from_pretrained("clarin-knext/utterance-rewriting-polish-gpt2", use_auth_token=auth_token),
        "tokenizer": AutoTokenizer.from_pretrained("clarin-knext/utterance-rewriting-polish-gpt2", use_auth_token=auth_token),
        "default_input": DEFAULT_DECODER_ONLY_INPUT_PL,
    },
}

PIPELINES: Dict[str, Pipeline] = {
    model_name: pipeline(
        "text2text-generation", model=DST_MODELS[model_name]["model"], tokenizer=DST_MODELS[model_name]["tokenizer"]
    )
    for model_name in DST_MODELS
}