new response

app.py

@@ -25,7 +25,7 @@ FRESHNESS_ELIGIBLE = {'apple', 'banana', 'orange', 'lemon'}
 
 @app.get("/")
 def greet_json():
-    return {"Hello": "World!"}
+    return {"swagger https://ivanm151-fruits.hf.space/docs#"}
 
 @app.post("/predict_full")
 async def predict_full(
@@ -61,10 +61,7 @@ async def predict_full(
     fruit_area_ratio = np.mean(mask)
     if fruit_area_ratio < 0.01:
         return {
-            "status": "no_fruit_detected",
-            "fruit_area_ratio": round(fruit_area_ratio, 4),
-            "fruit": None,
-            "fruit_confidence": None,
+            "fruit_top3": [],
             "freshness": None,
             "freshness_confidence": None,
             "cropped_base64": None
@@ -73,27 +70,34 @@ async def predict_full(
     # Обрезка под 100×100 для сорта
     cropped_100 = crop_fruit_contour_letterbox(orig_np, mask, out_size=100)
     input_tensor2 = preprocess_for_classifier(cropped_100).unsqueeze(0).to(DEVICE)
+
     with torch.no_grad():
         logits2 = model2(input_tensor2)
         probs2 = torch.softmax(logits2, dim=1).squeeze().cpu().numpy()
 
-    fruit_idx = int(np.argmax(probs2))
-    fruit_name = FRUIT_CLASSES[fruit_idx]
-    fruit_conf = float(probs2[fruit_idx])
+    # ТОП-3 фрукта
+    top3_indices = np.argsort(probs2)[-3:][::-1]  # индексы от самого уверенного
+    top3 = [
+        {
+            "fruit": FRUIT_CLASSES[idx],
+            "confidence": round(float(probs2[idx]), 4)
+        }
+        for idx in top3_indices
+    ]
+
+    # Проверяем, есть ли хотя бы один фрукт из FRESHNESS_ELIGIBLE в топ-3
+    eligible_in_top3 = any(item["fruit"] in FRESHNESS_ELIGIBLE for item in top3)
 
     result = {
-        "status": "success",
-        "fruit_area_ratio": round(fruit_area_ratio, 4),
-        "fruit": fruit_name,
-        "fruit_confidence": round(fruit_conf, 4),
+        "fruit_top3": top3,
         "freshness": None,
         "freshness_confidence": None,
         "cropped_base64": None
     }
 
-    # Свежесть, если подходит
-    if fruit_name in FRESHNESS_ELIGIBLE:
-        cropped_224 = crop_fruit_contour_letterbox(orig_np, mask, out_size=224)
+    # Свежесть, если есть eligible фрукт в топ-3
+    if eligible_in_top3:
+        cropped_224 = crop_fruit_contour_letterbox(orig_np, mask, out_size=100)
         input_tensor3 = preprocess_for_classifier(cropped_224).unsqueeze(0).to(DEVICE)
         with torch.no_grad():
             logits3 = model3(input_tensor3)
@@ -113,6 +117,5 @@ async def predict_full(
         buffered = io.BytesIO()
         pil_img.save(buffered, format="PNG")
         result["cropped_base64"] = base64.b64encode(buffered.getvalue()).decode('utf-8')
-        result["cropped_size"] = f"{cropped_size}x{cropped_size}"
 
     return result