app.py

import gradio as gr
from gradio.components import Textbox
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, T5ForConditionalGeneration
from peft import PeftModel
import torch
import datasets
from sentence_transformers import CrossEncoder

# Load cross encoder
top_k = 10
cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')

# Load your fine-tuned model and tokenizer
model_name = "google/flan-t5-large"
peft_name = "legacy107/flan-t5-large-ia3-covidqa"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-large")
model = PeftModel.from_pretrained(model, peft_name)

peft_name = "legacy107/flan-t5-large-ia3-bioasq-paraphrase"
paraphrase_model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
paraphrase_model = PeftModel.from_pretrained(paraphrase_model, peft_name)

max_length = 512
max_target_length = 200

# Load your dataset
dataset = datasets.load_dataset("minh21/COVID-QA-Chunk-64-testset-biencoder-data-90_10", split="train")
dataset = dataset.shuffle()
dataset = dataset.select(range(5))


def paraphrase_answer(question, answer):
    # Combine question and context
    input_text = f"question: {question}. Paraphrase the answer to make it more natural answer: {answer}"

    # Tokenize the input text
    input_ids = tokenizer(
        input_text,
        return_tensors="pt",
        padding="max_length",
        truncation=True,
        max_length=max_length,
    ).input_ids

    # Generate the answer
    with torch.no_grad():
        generated_ids = paraphrase_model.generate(input_ids=input_ids, max_new_tokens=max_target_length)

    # Decode and return the generated answer
    paraphrased_answer = tokenizer.decode(generated_ids[0], skip_special_tokens=True)

    return paraphrased_answer


def retrieve_context(question, contexts):
    # cross-encoder
    hits = [{"corpus_id": i} for i in range(len(contexts))]
    cross_inp = [[question, contexts[hit["corpus_id"]]] for hit in hits]
    cross_scores = cross_encoder.predict(cross_inp, show_progress_bar=False)

    for idx in range(len(cross_scores)):
        hits[idx]["cross-score"] = cross_scores[idx]

    hits = sorted(hits, key=lambda x: x["cross-score"], reverse=True)

    return " ".join(
        [contexts[hit["corpus_id"]] for hit in hits[0:top_k]]
    ).replace("\n", " ")


# Define your function to generate answers
def generate_answer(question, context, contexts):
    if type(contexts) is str:
        contexts = contexts.split(',')
    context = retrieve_context(question, contexts)
    
    # Combine question and context
    input_text = f"question: {question} context: {context}"

    # Tokenize the input text
    input_ids = tokenizer(
        input_text,
        return_tensors="pt",
        padding="max_length",
        truncation=True,
        max_length=max_length,
    ).input_ids

    # Generate the answer
    with torch.no_grad():
        generated_ids = model.generate(input_ids=input_ids, max_new_tokens=max_target_length)

    # Decode and return the generated answer
    generated_answer = tokenizer.decode(generated_ids[0], skip_special_tokens=True)

    # Paraphrase answer
    paraphrased_answer = paraphrase_answer(question, generated_answer)

    return generated_answer, paraphrased_answer


# Define a function to list examples from the dataset
def list_examples():
    examples = []
    for example in dataset:
        context = example["context"]
        contexts = example["context_chunks"]
        question = example["question"]
        examples.append([question, context, contexts])
    return examples


# Create a Gradio interface
iface = gr.Interface(
    fn=generate_answer,
    inputs=[
        Textbox(label="Question"),
        Textbox(label="Context"),
        Textbox(label="Contexts")
    ],
    outputs=[
        Textbox(label="Generated Answer"),
        Textbox(label="Natural Answer")
    ],
    examples=list_examples()
)

# Launch the Gradio interface
iface.launch()