added model (v5) desc

app.py

@@ -7,6 +7,49 @@ yolov8_result = os.path.join(os.getcwd(), "data/xai/yolov8.png")
 yolov5_dff = os.path.join(os.getcwd(), "data/xai/yolov5_dff.png")
 yolov8_dff = os.path.join(os.getcwd(), "data/xai/yolov8_dff.png")
 
+
+architecture_description = """
+# YOLOv5 Architecture Overview
+
+YOLOv5 consists of three main components: the **backbone**, the **neck**, and the **head**.
+
+### 1. **Backbone**: Feature Extraction
+- **CSPDarknet53**: A modified version of **Darknet53**, leveraging **CSPNet** for improved gradient flow and memory usage.
+- **Residual Connections**: Utilizes **ResNet**-like residual connections to enable deeper learning without vanishing gradients.
+- **Focus Layer**: Performs convolutional downsampling to focus on key image features before feature extraction.
+
+### 2. **Neck**: Aggregation of Features
+- **PANet (Path Aggregation Network)**: Used for better feature aggregation and enhanced information flow across scales.
+- **FPN (Feature Pyramid Network)**: Helps in detecting objects at multiple scales by generating a pyramid of feature maps.
+- **Upsample and Downsample**: Combines low-level and high-level features for accurate localization and detection.
+
+### 3. **Head**: Detection and Output Generation
+- **Bounding Box Prediction**: Predicts **center (x, y)**, **width (w)**, and **height (h)** for each bounding box.
+- **Class Prediction**: Outputs the **class probabilities** for each detected object.
+- **Objectness Score**: Predicts whether the bounding box contains an object.
+- **Anchor Boxes**: Uses predefined anchor boxes to assist with bounding box prediction.
+
+### 4. **Detection Layer**: Grid-Based Prediction
+- **Grid-based Prediction**: Divides the image into a grid where each cell predicts multiple bounding boxes.
+- **Non-Maximum Suppression (NMS)**: Filters out redundant bounding boxes with high overlap based on confidence scores.
+
+### 5. **Loss Function**:
+- **CIoU Loss**: Used for bounding box regression to measure the overlap, aspect ratio, and center distance.
+- **Binary Cross-Entropy**: For **objectness score** prediction.
+- **Cross-Entropy**: For **classification** of the detected objects.
+
+### 6. **Post-Processing**:
+- **Non-Maximum Suppression (NMS)**: Eliminates duplicate boxes with high overlap, keeping only the most confident predictions.
+
+### Summary:
+1. **Input Image**: Pre-processed and fed into the model.
+2. **Backbone**: Feature extraction with **CSPDarknet53** and **Residual Connections**.
+3. **Neck**: Feature aggregation with **PANet** and **FPN**.
+4. **Head**: Outputs bounding boxes, objectness scores, and class probabilities.
+5. **Loss Functions**: CIoU, binary cross-entropy, and cross-entropy.
+6. **Post-Processing**: **NMS** to filter overlapping detections.
+"""
+
 # Netron HTML templates
 def get_netron_html(model_url):
     return f"""
@@ -60,6 +103,7 @@ with gr.Blocks(css=custom_css) as demo:
     
     with gr.Row():
         with gr.Column():
+            gr.Markdown(architecture_description)
             gr.HTML(get_netron_html(yolov5_url))
             gr.Image(yolov5_result, label="Detections & Interpretability Map")
             gr.Image(yolov5_dff, label="Feature Factorization & discovered concept")