Adding tinyllama form critic analisys

.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
+*.gguf filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -1,20 +1,40 @@
 FROM python:3.11-slim
 
-# Dependências do sistema para opencv
+# ─────────────────────────────────────────────
+# System deps
+# ─────────────────────────────────────────────
 RUN apt-get update && apt-get install -y \
-    libgl1 libglib2.0-0 \
+    libgl1 \
+    libglib2.0-0 \
+    wget \
     && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
 
+# ─────────────────────────────────────────────
+# Python deps
+# ─────────────────────────────────────────────
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
-# Copia código e modelo
+# ─────────────────────────────────────────────
+# App files
+# ─────────────────────────────────────────────
 COPY main.py .
 COPY yolo_deart.onnx .
 COPY style_classifier.onnx .
+COPY tinyllama.gguf .
 
-# HuggingFace Spaces expõe a porta 7860
+# ─────────────────────────────────────────────
+# HF / LLM performance config
+# ─────────────────────────────────────────────
+ENV HF_HOME=/tmp/huggingface
+ENV TRANSFORMERS_CACHE=/tmp/huggingface
+ENV TOKENIZERS_PARALLELISM=false
+
+# ─────────────────────────────────────────────
+# Port
+# ─────────────────────────────────────────────
 EXPOSE 7860
-CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "7860"]
+
+CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "7860"]

main.py

@@ -13,14 +13,15 @@ from fastapi.responses import JSONResponse
 from ultralytics import YOLO
 from torchvision import transforms
 
+from ctransformers import AutoModelForCausalLM
+
 # ─────────────────────────────────────────────
 # CONFIG
 # ─────────────────────────────────────────────
 
 YOLO_PATH = Path("yolo_deart.onnx")
 STYLE_MODEL_PATH = Path("style_classifier.onnx")
-
-DEVICE = "cpu"
+LLM_PATH = "tinyllama.gguf"
 
 FRONTEND_ORIGIN = os.environ.get(
     "FRONTEND_ORIGIN",
@@ -35,22 +36,19 @@ app = FastAPI()
 
 app.add_middleware(
     CORSMiddleware,
-    allow_origins=[
-        FRONTEND_ORIGIN,
-        "http://localhost:4200",
-    ],
+    allow_origins=[FRONTEND_ORIGIN, "http://localhost:4200"],
     allow_methods=["*"],
     allow_headers=["*"],
 )
 
 # ─────────────────────────────────────────────
-# YOLO MODEL
+# YOLO
 # ─────────────────────────────────────────────
 
 yolo_model = YOLO(str(YOLO_PATH))
 
 # ─────────────────────────────────────────────
-# ONNX STYLE MODEL
+# STYLE MODEL (ONNX)
 # ─────────────────────────────────────────────
 
 ort_session = ort.InferenceSession(
@@ -58,7 +56,6 @@ ort_session = ort.InferenceSession(
     providers=["CPUExecutionProvider"]
 )
 
-# ⚠️ ajuste isso se quiser persistir no export
 STYLES = [
     "Abstract Expressionism","Action painting","Analytical Cubism","Art Nouveau",
     "Baroque","Color Field Painting","Contemporary Realism","Cubism",
@@ -70,17 +67,43 @@ STYLES = [
 ]
 
 # ─────────────────────────────────────────────
-# TRANSFORM (igual ao treino)
+# LLM (TinyLlama GGUF)
+# ─────────────────────────────────────────────
+
+llm = AutoModelForCausalLM.from_single_file(
+    LLM_PATH,
+    model_type="llama"
+)
+
+def generate_art_analysis(detections, style):
+    objects = ", ".join([d["class_name"] for d in detections]) or "no objects"
+
+    prompt = f"""
+You are an expert art critic and poet.
+
+Detected objects: {objects}
+Art style: {style["style_name"]}
+
+Write:
+1. Visual description
+2. Metaphorical interpretation
+3. Poetic reflection
+
+Be concise and artistic.
+"""
+
+    return llm(prompt, max_new_tokens=180, temperature=0.7)
+
+# ─────────────────────────────────────────────
+# TRANSFORM
 # ─────────────────────────────────────────────
 
 style_tf = transforms.Compose([
     transforms.ToPILImage(),
     transforms.Resize((224, 224)),
     transforms.ToTensor(),
-    transforms.Normalize(
-        [0.485, 0.456, 0.406],
-        [0.229, 0.224, 0.225]
-    ),
+    transforms.Normalize([0.485, 0.456, 0.406],
+                         [0.229, 0.224, 0.225]),
 ])
 
 # ─────────────────────────────────────────────
@@ -88,42 +111,30 @@ style_tf = transforms.Compose([
 # ─────────────────────────────────────────────
 
 def encode_image(img_bgr: np.ndarray) -> str:
-    _, buffer = cv2.imencode(
-        ".jpg",
-        img_bgr,
-        [cv2.IMWRITE_JPEG_QUALITY, 85]
-    )
+    _, buffer = cv2.imencode(".jpg", img_bgr, [cv2.IMWRITE_JPEG_QUALITY, 85])
     return base64.b64encode(buffer).decode("utf-8")
 
-
 # ─────────────────────────────────────────────
-# STYLE INFERENCE (ONNX)
+# STYLE INFERENCE
 # ─────────────────────────────────────────────
 
 def predict_style(img_bgr: np.ndarray):
     img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
-
     tensor = style_tf(img_rgb).unsqueeze(0).numpy()
 
-    outputs = ort_session.run(
-        ["logits"],
-        {"image": tensor}
-    )[0]
+    outputs = ort_session.run(["logits"], {"image": tensor})[0]
 
-    # softmax estável
     exp = np.exp(outputs - np.max(outputs))
     probs = exp / exp.sum(axis=1, keepdims=True)
 
     idx = int(np.argmax(probs))
-    conf = float(np.max(probs))
 
     return {
         "style_id": idx,
-        "style_name": STYLES[idx] if idx < len(STYLES) else "unknown",
-        "confidence": conf
+        "style_name": STYLES[idx],
+        "confidence": float(np.max(probs))
     }
 
-
 # ─────────────────────────────────────────────
 # ROUTES
 # ─────────────────────────────────────────────
@@ -133,10 +144,10 @@ def health():
     return {
         "status": "ok",
         "yolo_classes": len(yolo_model.names),
-        "style_classes": len(STYLES)
+        "style_classes": len(STYLES),
+        "llm": "tinyllama-gguf"
     }
 
-
 @app.post("/detect")
 async def detect(file: UploadFile = File(...)):
     contents = await file.read()
@@ -146,7 +157,7 @@ async def detect(file: UploadFile = File(...)):
         cv2.IMREAD_COLOR
     )
 
-    # ── YOLO inference ───────────────────────
+    # ── YOLO ───────────────────────────────
     results = yolo_model(img_np)[0]
 
     detections = [
@@ -154,22 +165,23 @@ async def detect(file: UploadFile = File(...)):
             "class_id": int(box.cls[0]),
             "class_name": yolo_model.names[int(box.cls[0])],
             "confidence": round(float(box.conf[0]), 3),
-            "bbox": {
-                "x1": box.xyxy[0][0].item(),
-                "y1": box.xyxy[0][1].item(),
-                "x2": box.xyxy[0][2].item(),
-                "y2": box.xyxy[0][3].item(),
-            },
         }
         for box in results.boxes
     ]
 
-    # ── STYLE inference (ONNX) ──────────────
+    # ── STYLE ──────────────────────────────
     style_pred = predict_style(img_np)
 
+    # ── LLM ────────────────────────────────
+    try:
+        analysis = generate_art_analysis(detections, style_pred)
+    except Exception as e:
+        analysis = f"LLM error: {str(e)}"
+
     return JSONResponse({
         "detections": sorted(detections, key=lambda d: -d["confidence"]),
         "style": style_pred,
+        "analysis": analysis,
         "annotated_image": encode_image(results.plot()),
         "image_width": int(img_np.shape[1]),
         "image_height": int(img_np.shape[0]),

requirements.txt

@@ -2,10 +2,12 @@ fastapi==0.111.0
 uvicorn[standard]==0.29.0
 
 ultralytics==8.3.0
-
 onnxruntime==1.18.0
 
 opencv-python-headless==4.9.0.80
 pillow==10.3.0
 python-multipart==0.0.9
 numpy==1.26.4
+
+# ── LLM (TinyLlama GGUF) ─────────────────────
+ctransformers==0.2.27

tinyllama.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:030a469a63576d59f601ef5608846b7718eaa884dd820e9aa7493efec1788afa
+size 483116416