Create app.py

app.py

@@ -0,0 +1,66 @@
+import gradio as gr
+from transformers import AutoModelForObjectDetection, AutoImageProcessor
+import cv2
+import numpy as np
+import torch
+
+# Lade das Modell und den Image Processor
+model = AutoModelForObjectDetection.from_pretrained("facebook/dinov3-convnext-small-pretrain-lvd1689m")
+image_processor = AutoImageProcessor.from_pretrained("facebook/dinov3-convnext-small-pretrain-lvd1689m")
+
+def analyze_image(image, prompt):
+    # Konvertiere PIL-Bild zu OpenCV-Format
+    image_np = np.array(image)
+    image_cv = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
+
+    # Vorbereitung für DINOv3
+    inputs = image_processor(images=image_np, return_tensors="pt")
+    with torch.no_grad():
+        outputs = model(**inputs)
+
+    # Extrahiere Bounding-Boxes und Labels
+    detections = []
+    for score, label, box in zip(outputs.logits.softmax(-1).max(-1)[0], outputs.logits.softmax(-1).max(-1)[1], outputs.pred_boxes):
+        if score > 0.5:  # Confidence-Schwellenwert
+            # Konvertiere Box-Koordinaten
+            box = box * torch.tensor([image_np.shape[1], image_np.shape[0], image_np.shape[1], image_np.shape[0]])
+            xmin, ymin, xmax, ymax = box.int().tolist()
+
+            # Schneide die Kerze für Farbanalyse
+            candle_roi = image_cv[ymin:ymax, xmin:xmax]
+            if candle_roi.size == 0:  # Vermeide leere ROIs
+                continue
+            mean_color = np.mean(candle_roi, axis=(0, 1)).astype(int)
+            color_rgb = f"RGB({mean_color[2]},{mean_color[1]},{mean_color[0]})"
+
+            # Generische Labels, da DINOv3 nicht candlestick-spezifisch ist
+            detections.append({
+                "pattern": f"Candlestick_{label.item()}",  # Placeholder
+                "confidence": score.item(),
+                "color": color_rgb,
+                "x_center": (xmin + xmax) / 2,
+                "prompt_used": prompt
+            })
+
+    # Sortiere nach x-Position (rechts nach links = neueste Kerzen)
+    detections = sorted(detections, key=lambda x: x["x_center"], reverse=True)
+
+    # Begrenze auf die letzten 10 Kerzen, wenn im Prompt gefordert
+    if "last 10 candles" in prompt.lower():
+        detections = detections[:10]
+
+    return detections
+
+# Erstelle Gradio-Schnittstelle
+iface = gr.Interface(
+    fn=analyze_image,
+    inputs=[
+        gr.Image(type="pil", label="Upload TradingView Screenshot"),
+        gr.Textbox(label="Prompt", placeholder="Enter your prompt, e.g., 'List last 10 candles with their colors'")
+    ],
+    outputs="json",
+    title="Candlestick Pattern Detection with DINOv3",
+    description="Upload a TradingView screenshot and provide a prompt to detect candlestick patterns and colors."
+)
+
+iface.launch()