Update app.py

app.py

@@ -1,105 +1,108 @@
-/// <reference types="vite/client" />
-
-import { client } from "@gradio/client";
-
-export const HF_SPACE_URL = "https://thompson001-crack-api.hf.space/";
-
-export interface InferenceResult {
-  hasCrack: boolean;
-  confidence: number;
-  label: string;
-  details: string;
-  isCustomModel: boolean;
-}
-
-// Gradio Result 구조
-interface GradioEvent {
-  data: any[];
-  [key: string]: any;
-}
-
-interface HFJsonItem {
-  label?: string;
-  confidence?: number;
-  message?: string;
-}
-
-interface HFJsonPayload {
-  data: HFJsonItem[];
-}
-
-// -------------------------------------------------------
-// HF Space 호출
-// -------------------------------------------------------
-export async function predictCrack(file: File): Promise<InferenceResult> {
-  console.log("🔵 Connecting to HF Space…");
-
-  let app;
-  try {
-    app = await client(HF_SPACE_URL);
-  } catch (err) {
-    console.error("❌ HF Space 연결 실패:", err);
-    throw new Error("HF Space 연결 실패");
-  }
-
-  console.log("🔵 Sending image to /predict…");
-
-  let rawResult: GradioEvent;
-  try {
-    rawResult = (await app.predict("/predict", {
-      img: file, // 반드시 img
-    })) as GradioEvent;
-  } catch (err: any) {
-    console.error("❌ /predict 호출 실패:", err);
-    throw new Error("HF Space 예측 호출 실패");
-  }
-
-  console.log("🔵 Raw Response:", rawResult);
-
-  const jsonPayload = rawResult?.data?.[0] as HFJsonPayload;
-
-  console.log("🔵 JSON Payload:", jsonPayload);
-
-  // JSON 없는 경우 → 정상 취급
-  if (!jsonPayload?.data || jsonPayload.data.length === 0) {
-    return {
-      hasCrack: false,
-      confidence: 1,
-      label: "정상",
-      details: "",
-      isCustomModel: true,
-    };
-  }
-
-  const first: HFJsonItem = jsonPayload.data[0];
-
-  const rawLabel = typeof first.label === "string" ? first.label : "normal";
-  const confidence =
-    typeof first.confidence === "number" ? first.confidence : 1;
-
-  const normalized = rawLabel.replace(/\s+/g, "").toLowerCase();
-
-  // 에러 처리
-  if (normalized === "error") {
-    throw new Error(first.message || "모델 내부 오류");
-  }
-
-  // 균열 판단: crack, damaged 등 포함
-  const isCrack =
-    normalized.includes("crack") ||
-    normalized.includes("positive") ||
-    normalized.includes("damaged");
-
-  // 🔥 너가 원한 로직: 정상일 때는 % 숨김
-  const details = isCrack
-    ? `균열 확률 ${(confidence * 100).toFixed(1)}%`
-    : "";
-
-  return {
-    hasCrack: isCrack,
-    confidence,
-    label: isCrack ? "균열 감지" : "정상",
-    details,
-    isCustomModel: true,
-  };
-}
+import gradio as gr
+import numpy as np
+from PIL import Image
+import tensorflow as tf
+
+# ------------------------------
+# 1) 모델 파일 경로
+# ------------------------------
+MODEL_PATH = "crack_detection.h5"
+
+# ------------------------------
+# 2) 모델 로드
+# ------------------------------
+model = tf.keras.models.load_model(MODEL_PATH)
+print("🔥 Loaded TensorFlow crack classifier")
+print("   Input shape :", model.input_shape)
+print("   Output shape:", model.output_shape)
+
+_, H, W, C = model.input_shape
+
+
+# ------------------------------
+# 3) 예측 함수
+# ------------------------------
+def predict(img: Image.Image):
+    try:
+        img = img.convert("RGB")
+        img_resized = img.resize((W, H))
+
+        arr = np.array(img_resized).astype("float32") / 255.0
+        X = np.expand_dims(arr, axis=0)  # (1, H, W, C)
+
+        raw = model.predict(X)[0]
+        probs = np.array(raw, dtype="float32").flatten()
+
+        # ------------------------------
+        # 1-output (sigmoid) 모델
+        # ------------------------------
+        if probs.shape[0] == 1:
+            p_crack = float(probs[0])
+            p_normal = 1.0 - p_crack
+
+            if p_crack >= 0.5:
+                label = "crack"
+                conf = p_crack
+            else:
+                label = "normal"
+                conf = p_normal
+
+        # ------------------------------
+        # 2-output (softmax) 모델
+        # probs = [p_normal, p_crack]
+        # ------------------------------
+        elif probs.shape[0] >= 2:
+            p_normal = float(probs[0])
+            p_crack = float(probs[1])
+
+            if p_crack >= p_normal:
+                label = "crack"
+                conf = p_crack
+            else:
+                label = "normal"
+                conf = p_normal
+
+        else:
+            idx = int(np.argmax(probs))
+            label = f"class_{idx}"
+            conf = float(probs[idx])
+
+        # ------------------------------
+        # 프론트 요구 구조
+        # ------------------------------
+        return {
+            "data": [
+                {
+                    "label": label,
+                    "confidence": float(conf),
+                }
+            ]
+        }
+
+    except Exception as e:
+        print("❌ Error:", e)
+        return {
+            "data": [
+                {
+                    "label": "error",
+                    "confidence": 0.0,
+                    "message": str(e),
+                }
+            ]
+        }
+
+
+# ------------------------------
+# 4) Gradio API UI
+# ------------------------------
+demo = gr.Interface(
+    fn=predict,
+    inputs=gr.Image(type="pil", label="Input image"),
+    outputs=gr.JSON(label="Detection Result"),
+    title="Crack Detection Classifier (Keras .h5)",
+    description="사진을 업로드하면 균열/정상 여부와 확률(%)을 반환합니다.",
+    flagging_mode="never",
+)
+
+if __name__ == "__main__":
+    demo.launch()