![\"Open](\"https://colab.research.google.com/assets/colab-badge.svg\"){kind=icon}
"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 1,
+ "id": "49240255",
+ "metadata": {
+ "id": "colab-setup"
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "💻 Running locally\n",
+ "✅ Local setup complete!\n"
+ ]
+ }
+ ],
+ "source": [
+ "# @title 🚀 Environment Setup (Run this cell first!)\n",
+ "# @markdown This cell sets up the environment for both Colab and local runs.\n",
+ "\n",
+ "import sys\n",
+ "import os\n",
+ "\n",
+ "# Check if running in Google Colab\n",
+ "IN_COLAB = 'google.colab' in sys.modules\n",
+ "\n",
+ "if IN_COLAB:\n",
+ " print(\"🌐 Running in Google Colab\")\n",
+ " \n",
+ " # Clone the repository if not already cloned\n",
+ " if not os.path.exists('VisionInterpretability'):\n",
+ " !git clone https://github.com/cataluna84/VisionInterpretability.git\n",
+ " \n",
+ " # Change to project directory\n",
+ " os.chdir('VisionInterpretability')\n",
+ " \n",
+ " # Install dependencies\n",
+ " !pip install -q torch torchvision matplotlib numpy pillow tqdm opencv-python requests torch-lucent\n",
+ " \n",
+ " # Add src to Python path\n",
+ " sys.path.insert(0, 'src')\n",
+ " print(\"✅ Colab setup complete!\")\n",
+ "else:\n",
+ " print(\"💻 Running locally\")\n",
+ " # For local runs, add src to path if running from notebooks directory\n",
+ " if os.path.basename(os.getcwd()) == 'notebooks':\n",
+ " sys.path.insert(0, os.path.join(os.path.dirname(os.getcwd()), 'src'))\n",
+ " elif 'src' not in sys.path:\n",
+ " sys.path.insert(0, 'src')\n",
+ " print(\"✅ Local setup complete!\")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "36874ae6",
+ "metadata": {},
+ "source": [
+ "# Detailed Analysis: Feature Visualization & Dataset Examples\n",
+ "## InceptionV1 - Layer mixed4a\n",
+ "\n",
+ "This notebook provides a rigorous exploration of **INTERPRETABILITY** in Convolutional Neural Networks (CNNs). we focus on the `mixed4a` layer of InceptionV1 (GoogLeNet).\n",
+ "\n",
+ "### The Goal\n",
+ "We aim to answer: **\"What features is the network looking for?\"**\n",
+ "We use two complementary approaches:\n",
+ "1. **Dataset Examples**: Finding real images from the dataset (ImageNet) that maximally activate a specific neuron.\n",
+ " * *Question:* \"What real-world patterns trigger this neuron?\"\n",
+ "2. **Feature Visualization (Optimization)**: Synthesizing an \"ideal\" input image that maximizes the neuron's activation using gradient ascent.\n",
+ " * *Question:* \"What is the 'platonic ideal' of the feature this neuron detects?\"\n",
+ "\n",
+ "---\n",
+ "### Theoretical Foundations\n",
+ "\n",
+ "#### 1. Forward Pass & Activation\n",
+ "A neural network is a function $f: X \\to Y$.\n",
+ "For an input image $x$, the activation $A_{n,l}(x)$ represents the output of the $n$-th neuron at layer $l$.\n",
+ "We refer to this map $A_{n,l}(x)$ as the \"Feature Map\".\n",
+ "\n",
+ "#### 2. Optimization (Feature Visualization)\n",
+ "To visualize what a neuron wants, we perform **Optimization on the Input**.\n",
+ "Unlike training (where we update weights $w$ to minimize loss), here we freeze the weights and update the image $x$.\n",
+ "\n",
+ "**The Objective:**\n",
+ "$$ x^* = \\arg\\max_x (A_n(x) - \\lambda R(x)) $$\n",
+ "\n",
+ "Where:\n",
+ "* $A_n(x)$: The activation of the target neuron.\n",
+ "* $R(x)$: Regularization terms (e.g., total variation, frequency penalties) to ensure the image looks \"natural\" and not just high-frequency adversarial noise.\n",
+ "\n",
+ "**The Update Rule (Gradient Ascent):**\n",
+ "$$ x_{t+1} = x_t + \\eta \\nabla_x (A_n(x_t) - \\lambda R(x_t)) $$\n",
+ "\n",
+ "We use the **Lucent** library (PyTorch port of Lucid) which implements robust optimization techniques:\n",
+ "* **Fourier Parameterization**: Optimizing in frequency space (decorrelated) speeds up convergence.\n",
+ "* **Transformation Robustness**: We maximize the expected activation under random jitters/scales: $E_{t \\sim T}[A_n(t(x))]$. This forces the feature to be robust and identifiable.\n",
+ "\n",
+ "---\n"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "4b26d528",
+ "metadata": {},
+ "source": [
+ "## 1. Setup & Configuration\n",
+ "\n",
+ "We initialize the environment, load the pre-trained InceptionV1 model, and configure our experiment parameters.\n",
+ "We use `wandb` to log our experiments for reproducibility.\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 2,
+ "id": "setup",
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Using device: cuda\n",
+ "GPU: NVIDIA GeForce RTX 2070\n"
+ ]
+ }
+ ],
+ "source": [
+ "import os\n",
+ "import sys\n",
+ "from pathlib import Path\n",
+ "\n",
+ "# Add src to path for local imports\n",
+ "project_root = Path.cwd().parent\n",
+ "src_path = project_root / \"src\"\n",
+ "if str(src_path) not in sys.path:\n",
+ " sys.path.insert(0, str(src_path))\n",
+ "\n",
+ "# Load environment variables\n",
+ "from dotenv import load_dotenv\n",
+ "load_dotenv(project_root / \".env\")\n",
+ "\n",
+ "import torch\n",
+ "import numpy as np\n",
+ "import matplotlib.pyplot as plt\n",
+ "from PIL import Image\n",
+ "from tqdm.auto import tqdm\n",
+ "\n",
+ "# Check device\n",
+ "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
+ "print(f\"Using device: {device}\")\n",
+ "if device.type == \"cuda\":\n",
+ " print(f\"GPU: {torch.cuda.get_device_name(0)}\")"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 3,
+ "id": "config",
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Configuration:\n",
+ " num_neurons: 10\n",
+ " samples_per_category: 9\n",
+ " max_samples: None\n",
+ " batch_size: 128\n",
+ " layer_name: mixed4a\n",
+ " wandb_project: vision-interpretability\n",
+ " wandb_run_name: mixed4a-activation-spectrum-v3\n",
+ " generate_optimized: True\n"
+ ]
+ }
+ ],
+ "source": [
+ "# Configuration\n",
+ "CONFIG = {\n",
+ " \"num_neurons\": 10, # First 10 neurons of mixed4a (512 total)\n",
+ " \"samples_per_category\": 9, # 9 samples per category (min, slight-, slight+, max)\n",
+ " \"max_samples\": None, # Process ALL samples (WARNING: ~1.28M images)\n",
+ " # Set to 1000 or 10000 for faster testing.\n",
+ " \"batch_size\": 128, # Batch size for streaming inference\n",
+ " \"layer_name\": \"mixed4a\", # Target layer in InceptionV1\n",
+ " \"wandb_project\": \"vision-interpretability\",\n",
+ " \"wandb_run_name\": \"mixed4a-activation-spectrum-v3\",\n",
+ " \"generate_optimized\": True, # Generate gradient-ascent visualizations\n",
+ "}\n",
+ "\n",
+ "print(\"Configuration:\")\n",
+ "for k, v in CONFIG.items():\n",
+ " print(f\" {k}: {v}\")\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 4,
+ "id": "wandb-init",
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stderr",
+ "output_type": "stream",
+ "text": [
+ "\u001b[34m\u001b[1mwandb\u001b[0m: [wandb.login()] Loaded credentials for https://api.wandb.ai from WANDB_API_KEY.\n",
+ "\u001b[34m\u001b[1mwandb\u001b[0m: Currently logged in as: \u001b[33mcataluna84\u001b[0m to \u001b[32mhttps://api.wandb.ai\u001b[0m. Use \u001b[1m`wandb login --relogin`\u001b[0m to force relogin\n"
+ ]
+ },
+ {
+ "data": {
+ "text/html": [],
+ "text/plain": [
+ "c:\\Users\\cataluna84\\Documents\\Workspace\\VisionInterpretability\\notebooks\\wandb\\run-20260131_004230-ybjw2oe7"
+ ],
+ "text/plain": [
+ "| batch | ▁▁▁▁▂▂▂▂▂▂▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇▇███ |
| batch_duration_sec | ▃█▂▂▂▃▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▅▁▁ |
| samples_processed | ▁▁▁▁▁▂▂▂▃▃▃▃▃▃▃▄▄▄▄▄▄▄▄▄▅▅▅▅▅▆▆▆▆▇▇▇▇▇▇█ |
| throughput_img_per_sec | ▁▁▁▁▂▂▂▂▃▂▅▄▁▄▄▄█▇▆▆▆▁▄▇▆▅█▆█▄▆▅▄▂▄▅▄▆▄▁ |
| batch | 10010 |
| batch_duration_sec | 0.44153 |
| samples_processed | 1281167 |
| throughput_img_per_sec | 33.97274 |
.\\wandb\\run-20260131_004230-ybjw2oe7\\logs"
+ ],
+ "text/plain": [
+ "