app.py

import gradio as gr
import torch
from torchvision.models import resnet50, ResNet50_Weights
from PIL import Image
import tempfile
from gtts import gTTS
import whisper
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

# ----- 画像認識用モデル (ResNet-50) -----
weights = ResNet50_Weights.IMAGENET1K_V2
img_model = resnet50(weights=weights)
img_model.eval()
img_transform = weights.transforms()
imagenet_classes = weights.meta["categories"]


def image_classify(img: Image.Image):
    img_tensor = img_transform(img).unsqueeze(0)
    with torch.no_grad():
        outputs = img_model(img_tensor)
    probabilities = torch.nn.functional.softmax(outputs[0], dim=0)
    top5_prob, top5_catid = torch.topk(probabilities, 5)
    result = {imagenet_classes[top5_catid[i]]: float(top5_prob[i]) for i in range(5)}
    return result


model_name = "cyberagent/open-calm-1b"
model = AutoModelForCausalLM.from_pretrained(
    model_name, device_map="auto", torch_dtype=torch.float16
)
tokenizer = AutoTokenizer.from_pretrained(
    model_name, use_fast=True, trust_remote_code=True
)

text_gen_pipeline = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    max_length=128,
    temperature=0.7,
    top_p=0.9,
    pad_token_id=tokenizer.eos_token_id,
)


# ----- 言語モデル (LM) -----
def generate_text(prompt):
    # promptに基づき続きのテキストを生成
    result = text_gen_pipeline(prompt, do_sample=True, num_return_sequences=1)
    generated_text = result[0]["generated_text"]
    # prompt部分を含めた全文が返るので、prompt部分はそのままでOK
    return generated_text


# ----- 音声合成 (TTS) -----
def text_to_speech(text, lang="ja"):
    tts = gTTS(text=text, lang=lang)
    with tempfile.NamedTemporaryFile(suffix=".mp3", delete=False) as fp:
        tts.save(fp.name)
        return fp.name


# ----- 音声認識 (ASR) -----
whisper_model = whisper.load_model("small")


def speech_to_text(audio_file):
    result = whisper_model.transcribe(audio_file)
    return result["text"]


# ----- Gradio UI -----
def run():
    with gr.Blocks() as demo:
        gr.Markdown("# 画像認識・言語モデル・音声合成・音声認識")

        with gr.Tabs():
            with gr.TabItem("画像認識"):
                gr.Markdown("### 画像認識 (ResNet-50)")
                gr.Interface(
                    fn=image_classify,
                    inputs=gr.Image(type="pil"),
                    outputs=gr.Label(num_top_classes=5),
                    description="画像をアップロードして分類します。(ImageNet)",
                )

            with gr.TabItem("言語モデル"):
                gr.Markdown("### 言語モデル")
                lm_output = gr.Textbox(label="生成結果")
                user_input = gr.Textbox(label="入力テキスト")
                send_btn = gr.Button("送信")
                send_btn.click(generate_text, inputs=user_input, outputs=lm_output)

            with gr.TabItem("音声合成"):
                gr.Markdown("### 音声合成 (gTTS)")
                tts_input = gr.Textbox(label="音声にしたいテキスト")
                tts_output = gr.Audio(label="合成音声")
                tts_button = gr.Button("合成")
                tts_button.click(text_to_speech, inputs=tts_input, outputs=tts_output)

            with gr.TabItem("音声認識"):
                gr.Markdown("### 音声認識 (Whisper)")
                gr.Interface(
                    fn=speech_to_text,
                    inputs=gr.Audio(sources=["microphone", "upload"], type="filepath"),
                    outputs="text",
                    description="マイクから録音して文字起こし",
                )

    demo.launch()

if __name__ == "__main__":
    run()