Upload 207 files

README.md

@@ -1,16 +1,3 @@
----
-title: Towards a Safer Construction Environment
-emoji: 🏆
-colorFrom: yellow
-colorTo: green
-sdk: docker
-sdk_version: '20.10'
-app_file: safety_classifier.py
-pinned: true
-license: mit
-short_description: workplace safety compliance in construction sites
----
-
 # Towards a Safer Construction Environment: Evaluating a Simple CNN for Safety Classification
 ### By: Darius Vincent C. Ardales
 

safety_classifier_model.ipynb

@@ -4029,26 +4029,7 @@
     "    - Weights in the second and third input channels (e.g., **Filter 3, Input Channel 1**) show varying intensities compared to the first channel. This demonstrates the model's adaptation to RGB data and its ability to differentiate between channels.\n",
     "\n",
     "5. **Potential Over-Sensitivity**:\n",
-    "    - Some filters, such as **Filter 10, Input Channel 2**, have highly negative values like `-0.1833`, `-0.1962`, and `-0.1772`. These might indicate over-sensitivity to certain patterns and could potentially lead to instability during inference.\n",
-    "\n",
-    "### Insights for Improvement\n",
-    "1. **Regularization Techniques**:\n",
-    "    - Given the strong negative weights in certain filters, adding weight regularization (e.g., L2 regularization) might help balance learning and avoid over-sensitivity to specific features.\n",
-    "\n",
-    "2. **Weight Initialization**:\n",
-    "    - The variability among filters suggests the model is learning, but re-evaluating initialization strategies (e.g., Xavier or He initialization) could ensure better convergence.\n",
-    "\n",
-    "3. **Feature Visualization**:\n",
-    "    - Filters with extreme weight values (e.g., Filter 11, Input Channel 2) should be visualized to confirm they are detecting meaningful patterns and not just noise.\n",
-    "\n",
-    "### Supporting Evidence\n",
-    "- **Range of Weights**:\n",
-    "    - Filter 1, Input Channel 0: `-0.1391` to `0.1372`\n",
-    "    - Filter 10, Input Channel 2: `-0.1962` to `0.1198`\n",
-    "- **Symmetry**:\n",
-    "    - Filter 4, Input Channel 0: Weights are symmetric around zero, e.g., `[-0.1101, -0.0526, -0.0251, 0.0648]`.\n",
-    "\n",
-    "These observations demonstrate the diversity of learned spatial features in the first convolutional layer, reflecting the model's capability to extract meaningful patterns."
+    "    - Some filters, such as **Filter 10, Input Channel 2**, have highly negative values like `-0.1833`, `-0.1962`, and `-0.1772`. These might indicate over-sensitivity to certain patterns and could potentially lead to instability during inference."
    ]
   },
   {
@@ -4133,23 +4114,12 @@
     "## Insights from Convolutional Layer 2 Weights (Filters 0 and 1)\n",
     "\n",
     "### Key Observations\n",
-    "1. **Wide Range of Values Across Channels**:\n",
-    "   - Filter 0, Input Channel 2: Values range from `-0.0608` to `0.0448`, indicating sensitivity to both positive and negative features in this channel.\n",
-    "   - Filter 1, Input Channel 6: Strong negative values dominate, such as `-0.0621` and `-0.0640`, suggesting a focus on detecting dark or low-intensity regions.\n",
-    "\n",
-    "2. **Specialization in Feature Extraction**:\n",
-    "   - Filter 0, Input Channel 3: Symmetrical weight distribution (e.g., `-0.0511`, `0.0120`, `0.0161`) may contribute to detecting uniform patterns or textures.\n",
-    "   - Filter 1, Input Channel 0: Strong variations (`-0.0610` to `0.0423`) imply the ability to capture edges and gradients effectively.\n",
-    "\n",
-    "3. **Potential Over-Sensitivity**:\n",
-    "   - Filter 1, Input Channel 11: Extreme negative values like `-0.0605` and `-0.0639` might indicate an over-sensitivity to specific patterns, which could require further regularization.\n",
     "\n",
-    "### Supporting Evidence\n",
-    "- **Filter 0, Input Channel 2**: Values range from `-0.0608` to `0.0448`.\n",
-    "- **Filter 1, Input Channel 0**: Values range from `-0.0610` to `0.0423`.\n",
-    "- **Filter 1, Input Channel 11**: Extreme values like `-0.0605`, `-0.0639`.\n",
+    "1. **Intensity-Based Filters**:\n",
+    "   - Filter 0, Input Channel 8 and Filter 1, Input Channel 6 exhibit predominantly negative weights, indicating an emphasis on detecting dark or low-intensity regions. Such focus may assist in identifying shadowed or less illuminated patterns in construction site images\n",
     "\n",
-    "These insights highlight how the second convolutional layer focuses on diverse patterns, with some channels emphasizing edges and gradients while others exhibit strong negative weights potentially requiring regularization.\n"
+    "2. **Filter Collaboration in Edge Detection**:\n",
+    "   - Filter 0, Input Channel 3 (-0.0511, 0.0138, 0.0161) and Filter 1, Input Channel 3 (-0.0329, 0.0102, 0.0183) both display symmetry in weights, focusing on uniform patterns and edge structures. These filters likely work in tandem to detect edges at various scales or orientations."
    ]
   },
   {