brain-score · KartikP · May 14, 2025 · May 14, 2025 · May 19, 2025 · May 19, 2025
diff --git a/PROD_EXPORT_FINAL.sqlite b/PROD_EXPORT_FINAL.sqlite
diff --git a/benchmarks/templates/benchmarks/model.html b/benchmarks/templates/benchmarks/model.html
@@ -154,6 +154,19 @@ <h1 class="title">Anonymous Model #{{ model.id }}</h1>
             <span class="has-text-weight-bold">{{ visual_degrees }}</span> degrees
         </div>
     </div>
+
+    <div class="box" id="selected-metadata-box">
+        <div class="card-content">
+            <p id="selected-metadata-title" class="subtitle is-5" style="display: none;">
+                Selected Layers Metadata
+            </p>
+            <div id="selected-metadata-content">
+                <span class="has-text-grey">Shift-click layers to view metadata.</span>
+            </div>
+        </div>
+    </div>
+
+
 {% endblock %}
 
 
@@ -287,6 +300,10 @@ <h3 id="scores" class="title is-3">Scores on benchmarks</h3>
             </div>
         </div>
 
+        <!-- ADD VISUALIZATION HERE -->
+        {% include "benchmarks/model_visual.html" with model_architecture=model_architecture %}
+
+
         {# Usage #}
         <div class="box">
             <h4 class="subtitle is-4">

diff --git a/benchmarks/templates/benchmarks/model_visual.html b/benchmarks/templates/benchmarks/model_visual.html
@@ -0,0 +1,45 @@
+{% load static %}
+
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <title>Model Visualization - {{ model.id }}</title>
+    <script src="https://d3js.org/d3.v7.min.js"></script>
+
+</head>
+<script>
+    document.addEventListener("DOMContentLoaded", function() {
+        console.log("D3.js is loaded:", d3);
+    });
+</script>
+
+
+<body>
+    <h1>Model Architecture Visualization (Model ID: {{ model.id }})</h1>
+
+    <div id="svg-wrapper" style="width: 100%; overflow-x: auto;">
+        <svg width="1600" height="600"></svg>
+    </div>
+
+    <!-- Floating info box for layer metadata -->
+    <div id="layer-info-box"
+        style="display:none; position:absolute; z-index:1000; background:white; border:1px solid #ccc;
+            padding:10px; border-radius:8px; box-shadow:0 4px 8px rgba(0,0,0,0.2); font-size:14px;">
+    </div>
+
+
+    {{ visualization_metadata|json_script:"layers-json" }}
+
+
+
+
+<script src="{% static 'benchmarks/js/draw_layers.js' %}"></script>
+
+
+
+</body>
+</html>
+
+
+
diff --git a/benchmarks/views/extract_model_architecture.py b/benchmarks/views/extract_model_architecture.py
@@ -0,0 +1,131 @@
+# import json
+# import os
+# import torch
+# from torchsummary import summary
+# from io import StringIO
+# from contextlib import redirect_stdout
+#
+# # Directory where model JSON files will be stored
+# ARCHITECTURE_JSON_DIR = "static/model_architecture_json/"
+#
+#
+# def extract_trainable_layers(model, model_id):
+#     """
+#     Extracts trainable layers from a PyTorch model and saves them as a JSON file.
+#     - Trainable layers are those where `Param # != 0`
+#     - Uses `torchsummary.summary()` to get layer details.
+#     """
+#     os.makedirs(ARCHITECTURE_JSON_DIR, exist_ok=True)
+#
+#     # Redirect torchsummary output to a string buffer
+#     buffer = StringIO()
+#     with redirect_stdout(buffer):
+#         summary(model, input_size=(3, 224, 224))  # Assuming an image model with 3x224x224 input
+#
+#     output = buffer.getvalue()
+#     lines = output.split("\n")  # Convert output to lines
+#
+#     model_params = {
+#         "Visualization-Layer-Parameters": {}
+#     }
+#
+#     # Process each line and extract Layer Type, Output Shape, and Param #
+#     for line in lines[3:]:  # Skip the header
+#         parts = line.split()
+#         if len(parts) < 4:  # Ignore invalid lines
+#             continue
+#
+#         layer_name = parts[0]  # First column is Layer Type
+#         param_count = parts[-1]  # Last column is Param #
+#
+#         try:
+#             param_count = int(param_count)  # Convert to integer
+#         except ValueError:
+#             continue  # Skip if it's not a number
+#
+#         if param_count > 0:  # **Only keep trainable layers**
+#             model_params["Visualization-Layer-Parameters"][layer_name] = [param_count, param_count + 5]  # Example range
+#
+#     # Define JSON file path
+#     json_file_path = os.path.join(ARCHITECTURE_JSON_DIR, f"model_{model_id}.json")
+#
+#     # Save JSON data
+#     with open(json_file_path, "w") as json_file:
+#         json.dump(model_params, json_file, indent=4)
+#
+#     print(f"Model architecture saved at: {json_file_path}")
+#
+#
+# if __name__ == "__main__":
+#     # Example: Use a pre-trained ResNet model for testing
+#     model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet50', pretrained=True)
+#     extract_trainable_layers(model, model_id="resnet50")
+
+import json
+import os
+import torch
+from torchsummary import summary
+from io import StringIO
+from contextlib import redirect_stdout
+
+ARCHITECTURE_JSON_DIR = "static/model_architecture_json/"
+
+import json
+import os
+import torch
+from torchsummary import summary
+from io import StringIO
+from contextlib import redirect_stdout
+
+ARCHITECTURE_JSON_DIR = "static/model_architecture_json/"
+
+
+def extract_model_parameters(model, model_id):
+    """
+    Extracts model layer details (Layer Name, Output Shape, Param #) and saves as JSON.
+    """
+    os.makedirs(ARCHITECTURE_JSON_DIR, exist_ok=True)
+
+    model_params = {"layers": []}
+
+    buffer = StringIO()
+    with redirect_stdout(buffer):
+        summary(model, input_size=(3, 224, 224))
+
+    output = buffer.getvalue()
+    lines = output.split("\n")
+
+    for line in lines[3:]:
+        parts = line.split()
+        if len(parts) < 4:
+            continue
+
+        print(parts)
+        layer_name = parts[0]
+        output_shape = parts[1]
+        param_count = parts[-1]
+
+        try:
+            param_count = int(param_count)
+        except ValueError:
+            continue
+
+        model_params["layers"].append({
+            "layer_name": layer_name,
+            "output_shape": output_shape,
+            "param_count": param_count
+        })
+
+
+    json_file_path = os.path.join(ARCHITECTURE_JSON_DIR, f"model_{model_id}.json")
+
+    with open(json_file_path, "w") as json_file:
+        json.dump(model_params, json_file, indent=4)
+
+    print(f"✅ Model architecture saved at: {json_file_path}")
+    print(json.dumps(model_params, indent=4))
+
+
+if __name__ == "__main__":
+    model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet50', pretrained=True)
+    extract_model_parameters(model, model_id="1226")
diff --git a/benchmarks/views/index.py b/benchmarks/views/index.py
@@ -99,6 +99,7 @@ def view(request, domain: str):
     leaderboard_context = get_context(domain=domain)
     return render(request, 'benchmarks/leaderboard/leaderboard.html', leaderboard_context)
 
+
 # get_context is used for both leaderboard and model views. We can cache the results of it so that after the first
 # request, the cached results is served faster
 @cache_get_context()
@@ -192,7 +193,7 @@ def get_context(user=None, domain: str = "vision", benchmark_filter=None, model_
             'citation_domain_url': citation_domain_url,
             'citation_domain_title': citation_domain_title,
             'citation_domain_bibtex': citation_domain_bibtex,
-            'csv_downloadable': csv_data
+            'csv_downloadable': csv_data,
             }
 
 

diff --git a/benchmarks/views/model.py b/benchmarks/views/model.py
@@ -10,8 +10,17 @@
 from .index import get_context
 from ..models import BenchmarkType, Model
 
+# FOR DUMMY VISUALIZATION #
+import json
+import os
+
+
+
+
 _logger = logging.getLogger(__name__)
 
+# Load JSON file for model architecture visualization
+ARCHITECTURE_JSON_DIR = "static/model_architecture_json/"
 
 def view(request, id: int, domain: str):
     model, model_context, reference_context = determine_context(id, request, domain)
@@ -31,12 +40,29 @@ def view(request, id: int, domain: str):
     layers = get_layers(model)
     model_context['layers'] = layers
 
+    # _logger.info(f"Extracted Layers Data: {json.dumps(layers, indent=4)}")
+
+    DUMMY_LAYERS_FILE = os.path.join(ARCHITECTURE_JSON_DIR, "dummy_layers.json")
+
+    # Add Visualization-Layer-Parameters to model_context
+    if os.path.exists(DUMMY_LAYERS_FILE):
+        with open(DUMMY_LAYERS_FILE, 'r') as f:
+            dummy_layers_data = json.load(f)
+
+        model_context['visualization_metadata'] = dummy_layers_data  # Store full JSON
+
+        if "Visualization-Layer-Parameters" in dummy_layers_data:
+            vis_layers = dummy_layers_data["Visualization-Layer-Parameters"]
+            model_context['visualization_layers'] = vis_layers
+
+
     # only show detailed model info to superuser or owner:
     if request.user.is_superuser or model.user.id == request.user.id:
         model_context['submission_details_visible'] = True
     return render(request, 'benchmarks/model.html', model_context)
 
 
+
 def determine_context(id, request, domain: str):
     # this is a bit hacky: we're loading scores for *all* public models as well as *all* of the user's models
     # so we're loading a lot of unnecessary detail. But it lets us re-use already existing code.

diff --git a/requirements.txt b/requirements.txt
@@ -1,5 +1,4 @@
 asgiref==3.7.2
-backports.zoneinfo==0.2.1
 boto3==1.28.15
 botocore==1.31.15
 certifi==2023.7.22