# Generates positive movie reviews by tuning a pretrained model on IMDB dataset
# with a sentiment reward function
import json
import os
import sys
from typing import List

import torch
from datasets import load_dataset
from peft import LoraConfig
from peft.utils.config import TaskType
from transformers import pipeline

import trlx
from trlx.data.default_configs import TRLConfig, default_ppo_config


def get_positive_score(scores):
    "Extract value associated with a positive sentiment from pipeline's output"
    return dict(map(lambda x: tuple(x.values()), scores))["POSITIVE"]


def main(hparams={}):
    # Merge sweep config with default config if given
    config = TRLConfig.update(default_ppo_config().to_dict(), hparams)

    if torch.cuda.is_available():
        device = int(os.environ.get("LOCAL_RANK", 0))
    else:
        device = -1

    sentiment_fn = pipeline(
        "sentiment-analysis",
        "lvwerra/distilbert-imdb",
        top_k=2,
        truncation=True,
        batch_size=256,
        device=device,
    )

    # Just insert your peft config here (the type must be an instance of peft.PeftConfig or a dict).
    config.model.peft_config = LoraConfig(
        r=8,
        task_type=TaskType.CAUSAL_LM,
        lora_alpha=32,
        lora_dropout=0.1,
    )

    def reward_fn(samples: List[str], **kwargs) -> List[float]:
        sentiments = list(map(get_positive_score, sentiment_fn(samples)))
        return sentiments

    # Take few words off of movies reviews as prompts
    imdb = load_dataset("imdb", split="train+test")
    prompts = [" ".join(review.split()[:4]) for review in imdb["text"]]

    trlx.train(
        reward_fn=reward_fn,
        prompts=prompts,
        eval_prompts=["I don't know much about Hungarian underground"] * 256,
        config=config,
    )


if __name__ == "__main__":
    hparams = {} if len(sys.argv) == 1 else json.loads(sys.argv[1])
    main(hparams)