Upload 10 files

update app and weights

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+examples/2.JPG filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,14 +1,14 @@
----
-title: Pear Playground
-emoji: 🐨
-colorFrom: gray
-colorTo: red
-sdk: gradio
-sdk_version: 5.8.0
-app_file: app.py
-pinned: false
-license: apache-2.0
-short_description: Defects Detection on Pear
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: Pear Playground
+emoji: 🐨
+colorFrom: gray
+colorTo: red
+sdk: gradio
+sdk_version: 5.8.0
+app_file: app.py
+pinned: false
+license: apache-2.0
+short_description: Defects Detection on Pear
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -1,88 +1,88 @@
-import cv2
-import gradio as gr
-from detection import PearDetectionModel
-from classification import predict
-
-# make streaming interface that reads from camera and displays the output with bounding boxes
-config = {"model_path": "./weights/best.pt", "classes": ['burn_bbox', 'defected_pear', 'defected_pear_bbox', 'normal_pear', 'normal_pear_bbox']}
-model = PearDetectionModel(config)
-
-
-def classify(image):
-    """
-    Gradio에서 PIL 이미지를 입력받아 추론 결과를 반환.
-
-    Args:
-        image (PIL.Image): 업로드된 이미지.
-
-    Returns:
-        str: 모델 예측 결과.
-    """
-    # 임시 파일 저장 후 처리
-    image_path = "temp_image.jpg"
-    image.save(image_path)
-    return predict(image_path)
-
-def detect(img):
-    cls, xyxy, conf = model.inference(img)
-    for box, conf in zip(xyxy, conf):
-        cv2.rectangle(
-            img,
-            (int(box[0]), int(box[1])),
-            (int(box[2]), int(box[3])),
-            (0, 255, 0),
-            2,
-        )
-        cv2.putText(
-            img,
-            f"{conf:.2f}",
-            (int(box[0]), int(box[1])),
-            cv2.FONT_HERSHEY_SIMPLEX,
-            1,
-            (0, 255, 0),
-            2,
-        )
-    cv2.putText(
-        img,
-        "Class: Normal Pear" if cls == 0 else "Class: Abnormal Pear",
-        (0, 50),
-        cv2.FONT_HERSHEY_SIMPLEX,
-        1,
-        (0, 255, 0),
-        2,
-    )
-    return img
-
-
-css = """.my-group {max-width: 500px !important; max-height: 500px !important;}
-            .my-column {display: flex !important; justify-content: center !important; align-items: center !important};"""
-
-with gr.Blocks(css=css) as demo:
-    demo.title = "Pear Playground"
-    # add markdown
-    gr.Markdown("## This is a demo for Pear Playground by AISeed.")
-    with gr.Tab(label="Classification"):
-        gr.Interface(
-            fn=classify,
-            inputs=gr.Image(type="pil", label="Upload an image"),
-            outputs=gr.Label(num_top_classes=9),
-            examples=["examples/1.jpg", "examples/2.jpg"],
-            title="비정상 과수 분류기",
-            description="경량 모델 ResNet101e 을 활용하여 비정상배 분류"
-        )
-    with gr.Tab(label="Detection"): 
-        with gr.Column(elem_classes=["my-column"]):
-            with gr.Group(elem_classes=["my-group"]):
-
-                input_img = gr.Image(sources=["webcam"], type="numpy", streaming=True)
-        input_img.stream(
-            detect,
-            [input_img],
-            [input_img],
-            time_limit=30,
-            stream_every=0.1,
-        )
-
-if __name__ == "__main__":
-
-    demo.launch()
+import cv2
+import gradio as gr
+from detection import PearDetectionModel
+from classification import predict
+
+# make streaming interface that reads from camera and displays the output with bounding boxes
+config = {"model_path": "./weights/best.pt", "classes": ['burn_bbox', 'defected_pear', 'defected_pear_bbox', 'normal_pear', 'normal_pear_bbox']}
+model = PearDetectionModel(config)
+
+
+def classify(image):
+    """
+    Gradio에서 PIL 이미지를 입력받아 추론 결과를 반환.
+
+    Args:
+        image (PIL.Image): 업로드된 이미지.
+
+    Returns:
+        str: 모델 예측 결과.
+    """
+    # 임시 파일 저장 후 처리
+    image_path = "temp_image.jpg"
+    image.save(image_path)
+    return predict(image_path)
+
+def detect(img):
+    cls, xyxy, conf = model.inference(img)
+    for box, conf in zip(xyxy, conf):
+        cv2.rectangle(
+            img,
+            (int(box[0]), int(box[1])),
+            (int(box[2]), int(box[3])),
+            (0, 255, 0),
+            2,
+        )
+        cv2.putText(
+            img,
+            f"{conf:.2f}",
+            (int(box[0]), int(box[1])),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            1,
+            (0, 255, 0),
+            2,
+        )
+    cv2.putText(
+        img,
+        "Class: Normal Pear" if cls == 0 else "Class: Abnormal Pear",
+        (0, 50),
+        cv2.FONT_HERSHEY_SIMPLEX,
+        1,
+        (0, 255, 0),
+        2,
+    )
+    return img
+
+
+css = """.my-group {max-width: 500px !important; max-height: 500px !important;}
+            .my-column {display: flex !important; justify-content: center !important; align-items: center !important};"""
+
+with gr.Blocks(css=css) as demo:
+    demo.title = "Pear Playground"
+    # add markdown
+    gr.Markdown("## This is a demo for Pear Playground by AISeed.")
+    with gr.Tab(label="Classification"):
+        gr.Interface(
+            fn=classify,
+            inputs=gr.Image(type="pil", label="Upload an image"),
+            outputs=gr.Label(num_top_classes=9),
+            examples=["examples/1.jpg", "examples/2.jpg"],
+            title="비정상 과수 분류기",
+            description="경량 모델 ResNet101e 을 활용하여 비정상배 분류"
+        )
+    with gr.Tab(label="Detection"): 
+        with gr.Column(elem_classes=["my-column"]):
+            with gr.Group(elem_classes=["my-group"]):
+
+                input_img = gr.Image(sources=["webcam"], type="numpy", streaming=True)
+        input_img.stream(
+            detect,
+            [input_img],
+            [input_img],
+            time_limit=30,
+            stream_every=0.1,
+        )
+
+if __name__ == "__main__":
+
+    demo.launch()

classification.py

@@ -0,0 +1,86 @@
+import math
+import torch
+from timm import create_model
+from torchvision import transforms
+from PIL import Image
+
+# 기본 크롭 비율
+DEFAULT_CROP_PCT = 0.875
+
+# EfficientNet-B0 모델 설정 및 로드
+weights_path =  "./weights/resnest101e.in1k_weight_Pear_classification.pt"  # 로컬 가중치 파일 경로
+model_name = "resnest101e"
+model = create_model(model_name, pretrained=False,num_classes=9)  # 사전 학습 로드 생략
+#model.classifier = torch.nn.Linear(model.classifier.in_features, 2)  # 이진 분류로 수정
+model.load_state_dict(torch.load(weights_path))  # 로컬 가중치 로드
+model.eval()  # 평가 모드 설정
+
+# 클래스 이름 수동 지정
+# class_labels = ["Abormal Pear", "Normal Pear"]  # 데이터셋에 맞게 수정
+class_labels = ["정상", "흑성병","과피얼룩","복숭아 순나방","복숭아 삼식나방","배 깍지벌레","잎말이 나방류", "기타","과피흑변"]
+
+# 전처리 함수
+def transforms_imagenet_eval(
+    img_path: str,
+    img_size: int = 224,
+    crop_pct: float = DEFAULT_CROP_PCT,
+    mean: tuple = (0.485, 0.456, 0.406),
+    std: tuple = (0.229, 0.224, 0.225),
+    normalize: bool = True,
+):
+    """
+    ImageNet 스타일 이미지 전처리 함수.
+
+    Args:
+        img_path (str): 이미지 경로.
+        img_size (int): 크롭 크기.
+        crop_pct (float): 크롭 비율.
+        mean (tuple): 정규화 평균.
+        std (tuple): 정규화 표준편차.
+        normalize (bool): 정규화 여부.
+
+    Returns:
+        torch.Tensor: 전처리된 이미지 텐서.
+    """
+    img = Image.open(img_path).convert("RGB")  # 이미지 로드 및 RGB 변환
+    scale_size = math.floor(img_size / crop_pct)  # 리사이즈 크기 계산
+
+    # Transform 설정
+    tfl = [
+        transforms.Resize((scale_size, scale_size), interpolation=transforms.InterpolationMode.BILINEAR),
+        transforms.CenterCrop(img_size),
+        transforms.ToTensor(),
+    ]
+    if normalize:
+        tfl += [transforms.Normalize(mean=torch.tensor(mean), std=torch.tensor(std))]
+
+    transform = transforms.Compose(tfl)
+    return transform(img)
+
+# 추론 함수
+def predict(image_path: str):
+    """
+    주어진 이미지 파일 경로를 받아 모델 추론을 수행.
+
+    Args:
+        image_path (str): 입력 이미지 경로.
+
+    Returns:
+        str: 모델 예측 결과.
+    """
+    # 이미지 전처리
+    input_tensor = transforms_imagenet_eval(
+        img_path=image_path,
+        img_size=224,
+        normalize=True
+    ).unsqueeze(0)  # 배치 차원 추가
+
+    # 모델 추론
+    with torch.no_grad():
+        prediction = model(input_tensor)
+    
+    probs = torch.nn.functional.softmax(prediction[0], dim=-1)
+    confidences = {class_labels[i]: float(probs[i]) for i in range(9)}
+
+    # 예측 결과 반환
+    return confidences

detection.py

@@ -0,0 +1,35 @@
+import torch
+from ultralytics import YOLO
+
+class PearDetectionModel:
+    def __init__(self, config) -> None:
+        self.device = (
+            torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+        )
+        self.model = YOLO(config["model_path"], task="detect")
+
+        self.names = config["classes"]
+
+    def detect(self, img):
+        results = self.model.predict(img)
+        return results[0].boxes.cpu().numpy()
+
+    def inference(self, img):
+        pred = self.detect(img)
+
+        # remove the box with confidence lower than 0.9 if no "burn_bbox" is detected, else 0.8
+        pred = (
+            pred[pred.conf > 0.8]
+            if all([pred != "burn_bbox" for pred in self.names])
+            else pred[pred.conf > 0.5]
+        )
+        labels = [self.names[int(cat)] for cat in pred.cls]
+
+        # if any classes rather than "normal_pear_box" is detected, return 0 else return 1
+        if any([label == "burn_bbox" for label in labels]):
+            return 1, pred.xyxy, pred.conf
+        else:
+            return 0, pred.xyxy, pred.conf
+
+    def _preporcess(self, img):
+        pass

requirements.txt

@@ -0,0 +1,3 @@
+ultralytics>=8.2.34
+opencv-python>=4.1.2
+timm>=1.0.1

weights/EfficientNetb0_weight_Pear_classification.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f7b383cf8f5d904d85f052264f220020a3bbaf5234f11647c514490f9ad176e
+size 16347397

weights/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4554fe5814e40ca157b622e61793fda17be7559ba5733d339e205414c38c1aa1
+size 22522979

weights/resnest101e.in1k_weight_Pear_classification.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b43ed2b7473a0fc89756da00c6b4c58eca0689bba43c0b5e9679d8c12775b00a
+size 185853431