Distances_bw_FeatureMaps.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "Distances_bw_FeatureMaps.ipynb",
      "provenance": []
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "s4IPkDyiopRN"
      },
      "source": [
        "# Introduction\n",
        "\n",
        "In this exploratory notebook, we investigate how distances between inputs change as the inputs propagate through a network. This work is inspired by my Bottleneck MLP notebook where I saw that deeper classifier layers had more clustered/separable representations of inputs from the same class. Of course, that was for a toy example with a very simple classifier. We will investigate how distances change across layers of more sophisticated pretrained models.\n",
        "\n",
        "One other slight connection between previous work and this notebook is that the perceptual threat model leverages activations from different layers to imitate the true human perception. That is, the combination of distances between the represenations of an input and an adversarial example at different layers is used to create a single 'perceptual' metric. \n",
        "\n",
        "It seems intuitive that classifiers would learn this self-organizing behavior in feature extraction layers to facilitate easier classification in the final layers. For the heck of it, let's check for ourselves and see if anything surprises us. "
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "bh7yVs77onXx"
      },
      "source": [
        "import torch\n",
        "import torch.nn as nn\n",
        "import torch.nn.functional as F\n",
        "import torch.optim as optim\n",
        "from PIL import Image\n",
        "import matplotlib.pyplot as plt\n",
        "\n",
        "import torchvision.transforms as transforms\n",
        "import torchvision.models as models\n",
        "import torch.utils.data\n",
        "import torchvision.datasets as datasets\n",
        "import copy\n",
        "import numpy as np"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "FR-blUDCw7oG"
      },
      "source": [
        "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
        "vgg19 = models.vgg19_bn(pretrained=True).to(device).eval()"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "fqfCZmhTw7AQ"
      },
      "source": [
        "mean = torch.tensor([0.485, 0.456, 0.406]).to(device).view(-1, 1, 1)\n",
        "std = torch.tensor([0.229, 0.224, 0.225]).to(device).view(-1, 1, 1)\n",
        "\n",
        "def norm_input(input):\n",
        "  return (input-mean)/std"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "uiVZUvbO2u2T"
      },
      "source": [
        "First let's try on images of random noise to see how the distances between them change. My hypothesis is that these distances will not change as much as distances for images from the train set. "
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Gi8IxGey3RUB"
      },
      "source": [
        "# each VGG block has a conv, batchnorm, and then a ReLU\n",
        "# let's record the feature map after each ReLU\n",
        "x = torch.rand((64,3,32,32))\n",
        "y = torch.rand((64,3,32,32))\n",
        "def get_feature_maps(cnn, input):\n",
        "  i = 0\n",
        "  input = norm_input(input)\n",
        "  feature_maps = [input]\n",
        "  for layer in cnn.features.children():\n",
        "    input = layer(input)\n",
        "    if isinstance(layer, nn.ReLU):\n",
        "      feature_maps.append(input)\n",
        "  return feature_maps\n",
        "fm1, fm2 = [get_feature_maps(vgg19, input) for input in [x,y]]"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "qcz-pl16_q7G",
        "outputId": "0de50d9c-d259-4d3f-b75b-8c1aae4487b6",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 264
        }
      },
      "source": [
        "def display_differences_in_fms(fm1, fm2, label=''):\n",
        "  metric = nn.MSELoss()\n",
        "  distances = [metric(fm1[i], fm2[i]).item() for i in range(len(fm1))]\n",
        "  plt.plot(distances, label=label)\n",
        "  plt.legend()\n",
        "  return distances\n",
        "\n",
        "_ = display_differences_in_fms(fm1, fm2, label='random inputs')"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": [],
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "2Eb4V-BF8rad",
        "outputId": "b89faed5-f0a9-45b4-d42d-d29fa4a36702",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        }
      },
      "source": [
        "transform = transforms.Compose(\n",
        "    [transforms.ToTensor(),\n",
        "     transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))])\n",
        "dataset = datasets.CIFAR10(root='./data', train=True, transform=transform, download=True)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Files already downloaded and verified\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "juOTwZL3I2VN",
        "outputId": "05026e9f-d8a2-489d-a218-b5306e38ef00",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        }
      },
      "source": [
        "print(dataset.class_to_idx)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "{'airplane': 0, 'automobile': 1, 'bird': 2, 'cat': 3, 'deer': 4, 'dog': 5, 'frog': 6, 'horse': 7, 'ship': 8, 'truck': 9}\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "63Ay12slFe3Z"
      },
      "source": [
        "def get_one_class_loader(ind=5, batch_size=128):\n",
        "  dataset = datasets.CIFAR10(root='./data', train=True, transform=transform, download=True)\n",
        "  inds = np.where(np.array(dataset.targets) == ind)[0]\n",
        "  dataset.data = np.take(dataset.data, inds, axis=0)\n",
        "  dataset.targets = [ind]*len(inds)\n",
        "  loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True)\n",
        "  batch,_ = next(iter(loader))\n",
        "  return loader"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "68tHZjbNIyVW",
        "outputId": "79d52d6b-f4c5-44e6-df01-7b8888e733fd",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 70
        }
      },
      "source": [
        "dog_loader = get_one_class_loader(ind=5)\n",
        "cat_loader = get_one_class_loader(ind=3)\n",
        "plane_loader = get_one_class_loader(ind=0)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Files already downloaded and verified\n",
            "Files already downloaded and verified\n",
            "Files already downloaded and verified\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "5o1QtwD8J9cO"
      },
      "source": [
        "Now let's inspect data that is closer to the training distribution. My idea is that distances will be warped more aggressively over subsets of the image domain that were involved in training. We should note that these images are far lower resolution than ImageNet (training distr), but similar in content. \n",
        "We first see how they differ class to class. "
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "YpDVEVURI-c3",
        "outputId": "7fb04a6a-89ef-4919-aafb-2cb76185f614",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 264
        }
      },
      "source": [
        "dog_batch = next(iter(dog_loader))[0]\n",
        "cat_batch = next(iter(cat_loader))[0]\n",
        "plane_batch = next(iter(plane_loader))[0]\n",
        "dog_fm = get_feature_maps(vgg19, dog_batch)\n",
        "cat_fm = get_feature_maps(vgg19, cat_batch)\n",
        "plane_fm = get_feature_maps(vgg19, plane_batch)\n",
        "dog_cat = display_differences_in_fms(dog_fm, cat_fm, label='Dog vs. Cat')\n",
        "dog_plane = display_differences_in_fms(dog_fm, plane_fm, label='Dog vs. Plane')\n",
        "plane_cat = display_differences_in_fms(plane_fm, cat_fm, label='Plane vs. Cat')\n",
        "rand_rand = display_differences_in_fms(fm1, fm2, label='random inputs')"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": [],
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "RlUyN6XeKUdb"
      },
      "source": [
        "Interestingly, the pattern is similar for all three comparisons. Also, the scale of differences is larger for inputs from trained classes as opposed to random inputs. One curious thing is that the differences trend to zero in the most of the deep layers, except in the final layer where they begin to separate slightly (only for inputs from classes the classsifier was trained on). Let's see how feature maps of inputs from the same class differ to give ourselves a better baseline."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "LuQNFlCNJP0I",
        "outputId": "49c623b5-44d8-4d0c-c062-cbefcd77cbe1",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 264
        }
      },
      "source": [
        "dog_batch2 = (next(iter(dog_loader)))[0]\n",
        "cat_batch2 = next(iter(cat_loader))[0]\n",
        "plane_batch2 = next(iter(plane_loader))[0]\n",
        "dog_fm2 = get_feature_maps(vgg19, dog_batch2)\n",
        "cat_fm2 = get_feature_maps(vgg19, cat_batch2)\n",
        "plane_fm2 = get_feature_maps(vgg19, plane_batch2)\n",
        "dog_dog = display_differences_in_fms(dog_fm, dog_fm2, label='Dog vs. Dog')\n",
        "cat_cat = display_differences_in_fms(cat_fm, cat_fm2, label='Cat vs. Cat')\n",
        "plane_plane = display_differences_in_fms(plane_fm, plane_fm2, label='Plane vs. Plane')"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": [],
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "slspFr2PNKVy"
      },
      "source": [
        "Still very similar to what we saw before. The real test is to see what happens when we normalize cross-class distances to the mean of the intra-class distances, though this does not seem promising."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "GHI9FuBANGXL",
        "outputId": "6fe54aa0-a5e8-4a66-de81-264c4c1b91d6",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 282
        }
      },
      "source": [
        "dog_cat_normed = [2*dog_cat[i]/(dog_dog[i]+cat_cat[i]) for i in range(len(dog_cat))]\n",
        "dog_plane_normed = [2*dog_plane[i]/(dog_dog[i]+plane_plane[i]) for i in range(len(dog_cat))]\n",
        "plane_cat_normed = [2*plane_cat[i]/(plane_plane[i]+cat_cat[i]) for i in range(len(dog_cat))]\n",
        "plt.plot(dog_cat_normed, label='dog vs cat')\n",
        "plt.plot(dog_plane_normed, label='dog vs plane')\n",
        "plt.plot(plane_cat_normed, label='plane vs cat')\n",
        "plt.legend()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<matplotlib.legend.Legend at 0x7f6c361ae5f8>"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 21
        },
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": [],
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "NhuHfsCVRVTx"
      },
      "source": [
        "So, we do see that the distances between inputs from different classes are indeed further apart (though not by much) on average than inputs from the same class throughout all layers in the VGG network. However, they do not get further separated across the layers. \n",
        "\n",
        "What can be the explanation for this? I don't know. Perhaps the feature extraction layers are not concerned with making the data more separable, and this instead only happens in the final fully connected classifying layers. \n",
        "\n",
        "For the heck of it, we'll try again with resnet.\n",
        "\n",
        "\n",
        "EDIT: I tried again with more samples (32 to 128), and the separation became much less pronounced. We do see that separation mostly increases across layers 2 through 14, and then there is a surprisingly uniform sharp decrease in the final couple layers."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "lPvo1w_EQ-rc",
        "outputId": "43230c01-fd4c-4233-8bd2-534cc966ad9a",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 1000
        }
      },
      "source": [
        "resnet18 = models.resnet18(pretrained=True).eval()\n",
        "print(resnet18)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "ResNet(\n",
            "  (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)\n",
            "  (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "  (relu): ReLU(inplace=True)\n",
            "  (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)\n",
            "  (layer1): Sequential(\n",
            "    (0): BasicBlock(\n",
            "      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (relu): ReLU(inplace=True)\n",
            "      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "    )\n",
            "    (1): BasicBlock(\n",
            "      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (relu): ReLU(inplace=True)\n",
            "      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "    )\n",
            "  )\n",
            "  (layer2): Sequential(\n",
            "    (0): BasicBlock(\n",
            "      (conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)\n",
            "      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (relu): ReLU(inplace=True)\n",
            "      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (downsample): Sequential(\n",
            "        (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)\n",
            "        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      )\n",
            "    )\n",
            "    (1): BasicBlock(\n",
            "      (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (relu): ReLU(inplace=True)\n",
            "      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "    )\n",
            "  )\n",
            "  (layer3): Sequential(\n",
            "    (0): BasicBlock(\n",
            "      (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)\n",
            "      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (relu): ReLU(inplace=True)\n",
            "      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (downsample): Sequential(\n",
            "        (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)\n",
            "        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      )\n",
            "    )\n",
            "    (1): BasicBlock(\n",
            "      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (relu): ReLU(inplace=True)\n",
            "      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "    )\n",
            "  )\n",
            "  (layer4): Sequential(\n",
            "    (0): BasicBlock(\n",
            "      (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)\n",
            "      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (relu): ReLU(inplace=True)\n",
            "      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (downsample): Sequential(\n",
            "        (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)\n",
            "        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      )\n",
            "    )\n",
            "    (1): BasicBlock(\n",
            "      (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "      (relu): ReLU(inplace=True)\n",
            "      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)\n",
            "      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
            "    )\n",
            "  )\n",
            "  (avgpool): AdaptiveAvgPool2d(output_size=(1, 1))\n",
            "  (fc): Linear(in_features=512, out_features=1000, bias=True)\n",
            ")\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "F5nfD0LVfRWY"
      },
      "source": [
        "So the resnet has a bunch of Sequential modules, each consisting of 'BasicBlock's. Let's get the activations after a bunch of these I know this is super hacky -- I apologize :/. The try-catch at the end there (once again, forgive me) is because we are using CIFAR inputs instead of ImageNet -- we'll just ignore that layer."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "8SjFKeoWfH0v"
      },
      "source": [
        "def get_resnet_features(input, resnet=resnet18):\n",
        "  feature_maps = []\n",
        "  input = norm_input(input)\n",
        "  i = 0\n",
        "  for layer in resnet.children():\n",
        "    for c in layer.children():\n",
        "      i = i+1\n",
        "      input = c(input)\n",
        "      feature_maps.append(input)\n",
        "    if i == 0:\n",
        "      try:\n",
        "        input = layer(input)\n",
        "        feature_maps.append(input)\n",
        "      except Exception as e:\n",
        "        print('failed at layer: ', layer)\n",
        "    i=0\n",
        "  return feature_maps"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Nlz9j5i0jqGA",
        "outputId": "a7629a80-8fc7-4060-d400-cac75c8b9d0f",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 123
        }
      },
      "source": [
        "dog_resnet_fm = get_resnet_features(dog_batch)\n",
        "dog_resnet_fm2 = get_resnet_features(dog_batch2)\n",
        "cat_resnet_fm = get_resnet_features(cat_batch)\n",
        "cat_resnet_fm2 = get_resnet_features(cat_batch2)\n",
        "plane_resnet_fm = get_resnet_features(plane_batch)\n",
        "plane_resnet_fm2 = get_resnet_features(plane_batch2)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "failed at layer:  Linear(in_features=512, out_features=1000, bias=True)\n",
            "failed at layer:  Linear(in_features=512, out_features=1000, bias=True)\n",
            "failed at layer:  Linear(in_features=512, out_features=1000, bias=True)\n",
            "failed at layer:  Linear(in_features=512, out_features=1000, bias=True)\n",
            "failed at layer:  Linear(in_features=512, out_features=1000, bias=True)\n",
            "failed at layer:  Linear(in_features=512, out_features=1000, bias=True)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "5zD_cbmFjthD",
        "outputId": "83b9140d-547c-462f-a90d-89a568764a71",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 264
        }
      },
      "source": [
        "dog_dog_resnet = display_differences_in_fms(dog_resnet_fm, dog_resnet_fm2, label='dog v dog')\n",
        "cat_cat_resnet = display_differences_in_fms(cat_resnet_fm, cat_resnet_fm2, label='cat v cat')\n",
        "plane_plane_resnet = display_differences_in_fms(plane_resnet_fm, plane_resnet_fm2, label='plane v plane')"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": [],
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "vQYauwOsj87f",
        "outputId": "3dee45df-9366-467c-bf91-66af865e8cfe",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 268
        }
      },
      "source": [
        "dog_cat_resnet = display_differences_in_fms(dog_resnet_fm, cat_resnet_fm, label='dog v cat')\n",
        "plane_cat_resnet = display_differences_in_fms(plane_resnet_fm, cat_resnet_fm2, label='plane v cat')\n",
        "dog_plane_resnet = display_differences_in_fms(dog_resnet_fm, plane_resnet_fm2, label='dog v plane')"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": [],
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "nhCY1uQfmZoE",
        "outputId": "5d712b6f-8639-45fe-e5ed-868179e93ac6",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 282
        }
      },
      "source": [
        "dog_cat_normed_resnet = [2*dog_cat_resnet[i]/(dog_dog_resnet[i]+cat_cat_resnet[i]) for i in range(len(dog_cat_resnet))]\n",
        "dog_plane_normed_resnet = [2*dog_plane_resnet[i]/(dog_dog_resnet[i]+plane_plane_resnet[i]) for i in range(len(dog_cat_resnet))]\n",
        "plane_cat_normed_resnet = [2*plane_cat_resnet[i]/(plane_plane_resnet[i]+cat_cat_resnet[i]) for i in range(len(dog_cat_resnet))]\n",
        "plt.plot(dog_cat_normed_resnet, label='dog vs cat')\n",
        "plt.plot(dog_plane_normed_resnet, label='dog vs plane')\n",
        "plt.plot(plane_cat_normed_resnet, label='plane vs cat')\n",
        "plt.legend()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<matplotlib.legend.Legend at 0x7f6c355a4e10>"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 27
        },
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": [],
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "sAgmM_K-qIPo"
      },
      "source": [
        "This is much more difficult to interpret. \n",
        "\n",
        "If I had faith in this idea, what I would do next is actually train a CIFAR10 classifier and use that to do the above analyses. But, I don't have much faith in this idea so I'm going to move on to something else. \n",
        "\n",
        "If you're wondering, why didn't I train a classifier from the get go? The answer is that the perceptual metric is model agnostic -- in theory, you could use feature maps extracted from hidden layers of any kind of model. "
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "D_h_7Iyxremr"
      },
      "source": [
        ""
      ],
      "execution_count": null,
      "outputs": []
    }
  ]
}