app.py

import os, sys, types, subprocess, tempfile
import torch, gradio as gr
from transformers import (
    VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
)
from PIL import Image

# ── 환경 변수 ────────────────────────────────────────────────
os.environ["HF_FORCE_SAFE_SERIALIZATION"] = "1"
os.environ["XFORMERS_FORCE_DISABLE"] = "1"    # audiocraft 내부 플래그

# ── xformers 더미 모듈 ───────────────────────────────────────
dummy = types.ModuleType("xformers")
dummy.__version__ = "0.0.0"
ops = types.ModuleType("xformers.ops")

def _fake_mem_eff_attn(q, k, v, *_, dropout_p: float = 0.0, **__):
    return torch.nn.functional.scaled_dot_product_attention(
        q, k, v, dropout_p=dropout_p, is_causal=False
    )
class _FakeLowerTriangularMask: pass

ops.memory_efficient_attention = _fake_mem_eff_attn
ops.LowerTriangularMask = _FakeLowerTriangularMask
dummy.ops = ops
sys.modules["xformers"] = dummy
sys.modules["xformers.ops"] = ops
# ────────────────────────────────────────────────────────────

# ── audiocraft 로드 (postInstall에서 이미 설치됐음) ───────────
try:
    from audiocraft.models import MusicGen
    from audiocraft.data.audio import audio_write
except ModuleNotFoundError:                 # 예외적 로컬 실행 대비
    subprocess.check_call([
        sys.executable, "-m", "pip", "install",
        "git+https://github.com/facebookresearch/audiocraft@main",
        "--no-deps", "--use-pep517"
    ])
    from audiocraft.models import MusicGen
    from audiocraft.data.audio import audio_write

# ── 이미지 캡셔닝 모델 ─────────────────────────────────────
caption_model = VisionEncoderDecoderModel.from_pretrained(
    "nlpconnect/vit-gpt2-image-captioning",
    use_safetensors=True, low_cpu_mem_usage=True
)
feature_extractor = ViTImageProcessor.from_pretrained(
    "nlpconnect/vit-gpt2-image-captioning"
)
tokenizer = AutoTokenizer.from_pretrained(
    "nlpconnect/vit-gpt2-image-captioning"
)

# ── MusicGen 모델 ──────────────────────────────────────────
musicgen = MusicGen.get_pretrained("facebook/musicgen-small")
musicgen.set_generation_params(duration=10)

# ── 파이프라인 함수들 ──────────────────────────────────────
def generate_caption(image: Image.Image) -> str:
    pixel_values = feature_extractor(images=image, return_tensors="pt").pixel_values
    ids = caption_model.generate(pixel_values, max_length=50)
    return tokenizer.decode(ids[0], skip_special_tokens=True)

def generate_music(prompt: str) -> str:
    wav = musicgen.generate([prompt])
    tmpdir = tempfile.mkdtemp()
    path = os.path.join(tmpdir, "musicgen.wav")
    audio_write(path, wav[0], musicgen.sample_rate, strategy="loudness")
    return path

def process(image: Image.Image):
    caption = generate_caption(image)
    path = generate_music(f"A cheerful melody inspired by: {caption}")
    return caption, path

# ── Gradio UI ──────────────────────────────────────────────
demo = gr.Interface(
    fn=process,
    inputs=gr.Image(type="pil"),
    outputs=[
        gr.Text(label="AI가 생성한 그림 설명"),
        gr.Audio(label="생성된 AI 음악 (MusicGen)")
    ],
    title="🎨 AI 그림‑음악 생성기",
    description="그림을 업로드하면 AI가 설명을 만들고, 설명을 바탕으로 음악을 10초간 생성해 들려줍니다."
)

if __name__ == "__main__":
    demo.launch()