import gradio as gr

import torch
from transformers import pipeline
from transformers import PegasusForConditionalGeneration, PegasusTokenizer

classifier = pipeline(
    "question-answering", 
    model="deepset/roberta-base-squad2",
    tokenizer="deepset/roberta-base-squad2"
)

model_name = 'tuner007/pegasus_paraphrase'
torch_device = 'cuda' if torch.cuda.is_available() else 'cpu'
tokenizer3 = PegasusTokenizer.from_pretrained(model_name)
model3 = PegasusForConditionalGeneration.from_pretrained(model_name).to(torch_device)


def qa_paraphrase(text_input, question):
    prediction = classifier(
        context=text_input,
        question=question,
        truncation=True,
        max_length=512,
        padding=True,
    )
    answer = prediction['answer']
    answer_start = prediction['start']
    answer_end = prediction['end']
    context = text_input.split(".")
    for i in range(len(context)):
        if answer in context[i]:
            sentence = context[i].strip() + "."
            break
    batch = tokenizer3([sentence],truncation=True,padding='longest',max_length=60, return_tensors="pt").to(torch_device)
    translated = model3.generate(**batch,max_length=60,num_beams=10, num_return_sequences=1, temperature=1.5)
    paraphrase = tokenizer3.batch_decode(translated, skip_special_tokens=True)[0]
    return f"Answer: {answer}\nLong Form Answer: {paraphrase}"


iface = gr.Interface(
    fn=qa_paraphrase,
    inputs=[
        gr.inputs.Textbox(label="Text Input"),
        gr.inputs.Textbox(label="Question")
    ],
    outputs=gr.outputs.Textbox(label="Output"),
    title="Long Form Question Answering",
    description="mimics long form question answering by extracting the sentence containing the answer and paraphrasing it"
)

iface.launch()