Upload 2 files

app.py

@@ -0,0 +1,81 @@
+import io, base64, numpy as np, torch
+from PIL import Image
+from fastapi import FastAPI, Request
+from fastapi.responses import JSONResponse, PlainTextResponse
+import gradio as gr
+from transformers import AutoImageProcessor, DepthProForDepthEstimation
+
+# --------- 전역: 서버만 띄우고, 무거운 건 로딩하지 않음(지연 로딩) ---------
+device = "cuda" if torch.cuda.is_available() else "cpu"
+_proc = None
+_model = None
+
+def _lazy_init():
+    global _proc, _model
+    if _proc is None:
+        _proc = AutoImageProcessor.from_pretrained("apple/DepthPro-hf")
+    if _model is None:
+        _model = DepthProForDepthEstimation.from_pretrained("apple/DepthPro-hf").to(device).eval()
+
+def _infer(pil_img: Image.Image):
+    _lazy_init()
+    H, W = pil_img.height, pil_img.width
+    inputs = _proc(images=pil_img.convert("RGB"), return_tensors="pt").to(device)
+    with torch.no_grad():
+        outputs = _model(**inputs)
+    post = _proc.post_process_depth_estimation(outputs, target_sizes=[(H, W)])[0]
+    depth = post["predicted_depth"].float().cpu().numpy()  # meters
+    fov   = float(post.get("field_of_view", 0.0))
+    focal = float(post.get("focal_length", 0.0))
+    return depth, fov, focal
+
+# ===================== FastAPI (API 엔드포인트) =====================
+app = FastAPI()
+
+@app.get("/health")
+async def health():
+    return PlainTextResponse("ok")
+
+@app.post("/depth")
+async def depth_endpoint(request: Request):
+    """
+    두 가지 입력 모두 지원:
+    1) Content-Type: application/octet-stream  (원본 이미지 바이트)
+    2) Content-Type: application/json          {"inputs": "<base64>"}
+    """
+    ctype = (request.headers.get("content-type") or "").lower()
+
+    if "application/json" in ctype:
+        data = await request.json()
+        b64 = data.get("inputs", "")
+        if b64.startswith("data:"):  # data URL도 허용
+            b64 = b64.split(",", 1)[1]
+        image_bytes = base64.b64decode(b64)
+    else:
+        # 기본: octet-stream
+        image_bytes = await request.body()
+
+    img = Image.open(io.BytesIO(image_bytes)).convert("RGB")
+    depth, fov, focal = _infer(img)
+    H, W = img.height, img.width
+
+    depth_b64 = base64.b64encode(depth.astype(np.float32).tobytes()).decode("ascii")
+    return JSONResponse({
+        "height": H, "width": W,
+        "focal_px": float(focal),
+        "field_of_view": float(fov),
+        "depth_flat": depth_b64
+    })
+
+# ===================== Gradio (테스트용 미니 UI) =====================
+def preview(image: Image.Image):
+    depth, _, _ = _infer(image)
+    v = depth[np.isfinite(depth)]
+    lo, hi = (np.percentile(v, 1), np.percentile(v, 99)) if v.size else (0, 1)
+    norm = np.clip((depth - lo) / max(1e-6, hi - lo), 0, 1)
+    return Image.fromarray((norm * 255).astype(np.uint8))
+
+with gr.Blocks() as demo:
+    gr.Markdown("## DepthPro-hf (CPU, 무료 Space)\n- API: **POST /depth**  (octet-stream 또는 JSON base64)")
+    inp = gr.Image(type="pil", label="Input")
+    out = gr.Image(label="Depth (preview)")

requirements.txt

@@ -0,0 +1,6 @@
+fastapi
+gradio>=4.0
+transformers>=4.40
+torch
+pillow
+numpy