app.py

import gradio as gr
from transformers import (
    AutoTokenizer,
    AutoModelForSeq2SeqLM,
    AutoProcessor,
    AutoModelForDocumentQuestionAnswering,
    pipeline,
)
import torch
import torchaudio

processor = AutoProcessor.from_pretrained(
    "MariaK/layoutlmv2-base-uncased_finetuned_docvqa_v2"
)
model = AutoModelForDocumentQuestionAnswering.from_pretrained(
    "MariaK/layoutlmv2-base-uncased_finetuned_docvqa_v2"
)

tokenizer_ru2en = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-ru-en")
model_ru2en = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-ru-en")
tokenizer_en2ru = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-en-ru")
model_en2ru = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-en-ru")

transcriber = pipeline(
    "automatic-speech-recognition", model="artyomboyko/whisper-base-fine_tuned-ru"
)


def translate_ru2en(text):
    inputs = tokenizer_ru2en(text, return_tensors="pt")
    outputs = model_ru2en.generate(**inputs)
    translated_text = tokenizer_ru2en.decode(outputs[0], skip_special_tokens=True)
    return translated_text


def translate_en2ru(text):
    inputs = tokenizer_en2ru(text, return_tensors="pt")
    outputs = model_en2ru.generate(**inputs)
    translated_text = tokenizer_en2ru.decode(outputs[0], skip_special_tokens=True)
    return translated_text


def generate_answer_git(image, question):
    with torch.no_grad():
        encoding = processor(
            images=image,
            text=question,
            return_tensors="pt",
            max_length=512,
            truncation=True,
        )
        outputs = model(**encoding)
        start_logits = outputs.start_logits
        end_logits = outputs.end_logits
        predicted_start_idx = start_logits.argmax(-1).item()
        predicted_end_idx = end_logits.argmax(-1).item()

    return processor.tokenizer.decode(
        encoding.input_ids.squeeze()[predicted_start_idx : predicted_end_idx + 1]
    )


def generate_answer(image, question):
    question_en = translate_ru2en(question)
    print(f"Вопрос на английском: {question_en}")

    answer_en = generate_answer_git(image, question_en)
    print(f"Ответ на английском: {answer_en}")

    answer_ru = translate_en2ru(answer_en)

    return answer_ru


def transcribe(image, audio):
    if not image or not audio:
        return

    sr, y = audio

    if y.ndim > 1:
        y = y.mean(axis=1)

    y_tensor = torch.tensor(y, dtype=torch.float32)
    print(y.shape)

    if sr != 16000:
        resampler = torchaudio.transforms.Resample(orig_freq=sr, new_freq=16000)
        y_tensor = resampler(y_tensor)
        sr = 16000

    y_tensor /= torch.max(torch.abs(y_tensor))

    y = y_tensor.numpy()
    print(y.shape)

    input_features = transcriber.feature_extractor(
        y, sampling_rate=sr, return_tensors="pt"
    ).input_features
    print(input_features.shape)
    print(input_features)

    transcription = transcriber.model.generate(input_features)
    transcription_text = transcriber.tokenizer.decode(
        transcription[0], skip_special_tokens=True
    )

    return generate_answer(image, transcription_text)


qa_interface = gr.Interface(
    fn=generate_answer,
    inputs=[
        gr.Image(type="pil"),
        gr.Textbox(label="Вопрос (на русском)", placeholder="Ваш вопрос"),
    ],
    outputs=gr.Textbox(label="Ответ (на русском)"),
    examples=[["doc.png", "О чем данный документ?"]],
    live=False,
)

speech_interface = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Image(type="pil"),
        gr.Audio(sources="microphone", label="Голосовой ввод"),
    ],
    outputs=gr.Textbox(label="Распознанный текст"),
    live=True,
)
interface = gr.TabbedInterface(
    [qa_interface, speech_interface],
    ["Текстовый вопрос", "Голосовой вопрос"],
    title="Демо визуального ответчика на вопросы (на русском)",
)

interface.launch(debug=True, share=True)