Update app.py

app.py

@@ -1,43 +1,57 @@
 import gradio as gr
-from PIL import Image
-from transformers import pipeline
+from PIL import Image, ImageOps
+import torch
+from transformers import CLIPProcessor, CLIPModel
 
-# 1) Cria o classificador zero‐shot com CLIP
-classifier = pipeline(
-    task="zero-shot-image-classification",
-    model="openai/clip-vit-base-patch32"
-)
+# ─── 1) Carrega modelo e processor CLIP fine-tuned ───
+MODEL_ID = "Keetawan/clip-vit-large-patch14-plant-disease-finetuned"
+processor = CLIPProcessor.from_pretrained(MODEL_ID)
+model     = CLIPModel.from_pretrained(MODEL_ID)
+
+# ─── 2) Labels que o modelo conhece ───
+HF_LABELS = [
+    "Grape leaf with Black rot",
+    "Grape leaf with Esca (Black Measles)",
+    "Grape leaf with Leaf blight (Isariopsis Leaf Spot)",
+    "Healthy Grape leaf"
+]
+# Mapeamento para as tuas classes curtas
+MAP = {
+    "Grape leaf with Black rot": "Black Rot",
+    "Grape leaf with Esca (Black Measles)": "ESCA",
+    "Grape leaf with Leaf blight (Isariopsis Leaf Spot)": "Leaf Blight",
+    "Healthy Grape leaf": "Healthy"
+}
+
+def predict(img: Image.Image):
+    # Pré-processamento igual ao notebook
+    img = ImageOps.exif_transpose(img).convert("RGB")
+    img = img.resize((224,224))
+
+    # Zero-shot inference CLIP
+    inputs = processor(text=HF_LABELS, images=img, return_tensors="pt", padding=True)
+    outputs = model(**inputs)
+    probs   = outputs.logits_per_image.softmax(dim=1)[0].tolist()
+
+    # Constrói dicionário label→prob
+    mapping = { MAP[HF_LABELS[i]]: probs[i] for i in range(len(probs)) }
+    # Escolhe a classe de maior probabilidade
+    best = max(mapping, key=mapping.get)
 
-# 2) Define as labels
-LABELS = ["Healthy", "Leaf Blight", "Black Rot", "ESCA"]
-
-def predict(image: Image.Image):
-    # Opcional: corrige EXIF e redimensiona como antes
-    image = image.convert("RGB").resize((224,224))
-    
-    # Zero‐shot classification
-    res = classifier(image, candidate_labels=LABELS)
-    # res é lista de dicts: [{"label":..., "score":...}, ...]
-    
-    # Mapeia para texto ordenado
-    probs = {item["label"]: float(item["score"]) for item in res}
-    # Escolhe o mais provável
-    best = max(probs, key=probs.get)
-    
-    # Formata saída
-    prob_lines = "\n".join(f"{lbl}: {probs[lbl]:.2f}" for lbl in LABELS)
+    # Formata as probabilidades
+    prob_lines = "\n".join(f"{lbl}: {mapping[lbl]:.2f}" for lbl in ["Healthy","Leaf Blight","Black Rot","ESCA"])
     return best, prob_lines
 
-# 3) Interface Gradio
+# ─── 3) UI Gradio ───────────────────────────────────────
 demo = gr.Interface(
     fn=predict,
-    inputs=gr.Image(type="pil", label="Carrega a folha"),
+    inputs=gr.Image(type="pil", label="Carrega uma folha"),
     outputs=[
       gr.Textbox(label="Classe predita"),
-      gr.Textbox(label="Probabilidades (0–1)")
+      gr.Textbox(label="Probabilidades")
     ],
-    title="CropVision (Backup CLIP Zero-Shot)",
-    description="Usa CLIP zero-shot para classificar folhas em Healthy, Leaf Blight, Black Rot ou ESCA"
+    title="CropVision – CLIP Zero-Shot Fine-Tuned",
+    description="Healthy / Leaf Blight / Black Rot / ESCA"
 )
 
 if __name__ == "__main__":