uniform_data.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 5,
  "metadata": {
    "kernelspec": {
      "display_name": "diff_rendering_ml",
      "language": "python",
      "name": "diff_rendering_ml"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.8.5"
    },
    "colab": {
      "name": "uniform_data.ipynb",
      "provenance": [],
      "include_colab_link": true
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/Spandan-Madan/A-Collection-of-important-tasks-in-pytorch/blob/master/uniform_data.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "authentic-personal"
      },
      "source": [
        "import torch\n",
        "import random\n",
        "import torch.nn as nn"
      ],
      "id": "authentic-personal",
      "execution_count": 1,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "lesser-motion"
      },
      "source": [
        "import torch.distributions as D"
      ],
      "id": "lesser-motion",
      "execution_count": 2,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "C0-v8H99cJBY",
        "outputId": "f20e1651-e051-4502-bbf4-9336e193478c"
      },
      "source": [
        "from google.colab import drive\n",
        "drive.mount('/content/drive')"
      ],
      "id": "C0-v8H99cJBY",
      "execution_count": 3,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Mounted at /content/drive\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "bNXRGhARcaFZ",
        "outputId": "fdac2d3f-8acc-4f39-ed16-ec8641d71473"
      },
      "source": [
        "%cd /content/drive/MyDrive/github/differentiable_graphics_ml"
      ],
      "id": "bNXRGhARcaFZ",
      "execution_count": 4,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "/content/drive/MyDrive/github/differentiable_graphics_ml\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ceLeCbgzctB0"
      },
      "source": [
        "import numpy as np"
      ],
      "id": "ceLeCbgzctB0",
      "execution_count": 5,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "current-latex"
      },
      "source": [
        "import cma\n",
        "from tqdm.notebook import tqdm\n",
        "import pickle\n",
        "\n",
        "import matplotlib.pyplot as plt"
      ],
      "id": "current-latex",
      "execution_count": 6,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "respective-front"
      },
      "source": [
        "import torch.nn.functional as F"
      ],
      "id": "respective-front",
      "execution_count": 7,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "floppy-resolution"
      },
      "source": [
        "def sample_uniform(batch_size, N_dim, r1, r2):\n",
        "    random_matrix = (r1 - r2) * torch.rand([batch_size, N_dim]) + r2\n",
        "    return random_matrix"
      ],
      "id": "floppy-resolution",
      "execution_count": 8,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "residential-spider"
      },
      "source": [
        "class MLP(nn.Module):\n",
        "    def __init__(self, data_dimensions,name):\n",
        "        super(MLP, self).__init__()\n",
        "        self.layers = nn.Sequential(\n",
        "            nn.Linear(data_dimensions, int(data_dimensions)),\n",
        "            nn.ReLU(),\n",
        "#             nn.Linear(int(data_dimensions*5), int(data_dimensions/5)),\n",
        "#             nn.ReLU(),\n",
        "#             nn.Linear(int(data_dimensions), int(data_dimensions/5)),\n",
        "#             nn.ReLU(),\n",
        "#             nn.Linear(int(data_dimensions/5), int(data_dimensions/5)),\n",
        "#             nn.ReLU(),\n",
        "            nn.Linear(int(data_dimensions), 2)\n",
        "        )\n",
        "        self.name = name\n",
        "        self.train_accuracy = 0\n",
        "        self.test_accuracy = 0\n",
        "        self.dataset = []\n",
        "\n",
        "    def forward(self, x):\n",
        "        x = x.view(x.size(0), -1)\n",
        "        x = self.layers(x)\n",
        "        return x"
      ],
      "id": "residential-spider",
      "execution_count": 9,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "actual-exhibition"
      },
      "source": [
        "class CMA_info():\n",
        "    def __init__(self, model_name, num_samples):\n",
        "        super(CMA_info, self).__init__()\n",
        "        self.model_name = model_name\n",
        "        self.num_samples = num_samples\n",
        "        \n",
        "        self.distances = []\n",
        "        self.in_dist_advs = []\n",
        "        self.advs = []\n",
        "        self.starts = []\n",
        "    \n",
        "    def summary(self):\n",
        "        print('****************** CMA Summary *******************')\n",
        "        print('Trained on %s points:'%self.model_name.split('_')[-1])\n",
        "        print('Adversarials: %s/%s'%(len(self.advs), self.num_samples))\n",
        "        print('In-distribution: %s/%s'%(len(self.in_dist_advs), len(self.advs)))\n",
        "        \n",
        "        avg_dist = np.mean(self.distances)\n",
        "        print('Average L2 distance: %s'%(avg_dist))\n",
        "        print('*'*50)"
      ],
      "id": "actual-exhibition",
      "execution_count": 10,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "dental-belle"
      },
      "source": [
        "x1_min = -10\n",
        "x1_max = 10\n",
        "\n",
        "x2_min = 20\n",
        "x2_max = 40\n",
        "\n",
        "test_min = -1\n",
        "test_max= 1"
      ],
      "id": "dental-belle",
      "execution_count": 11,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "actual-announcement"
      },
      "source": [
        "def make_dataset(dataset_size):\n",
        "    sample_1 = sample_uniform(dataset_size, N_dim, x1_min, x1_max)\n",
        "    sample_2 = sample_uniform(dataset_size, N_dim, x2_min, x2_max)\n",
        "    \n",
        "    labels_1 = torch.zeros(len(sample_1))\n",
        "    labels_2 = torch.ones(len(sample_2))\n",
        "    \n",
        "    X = torch.vstack([sample_1, sample_2])\n",
        "    Y = torch.hstack([labels_1, labels_2])\n",
        "    ids = list(range(len(X)))\n",
        "\n",
        "    random.shuffle(ids)\n",
        "    train_ids = ids[:int(0.8*len(X))]\n",
        "    test_ids = ids[int(0.8*len(X)):]\n",
        "\n",
        "    # len(X)\n",
        "\n",
        "    X_train = X[train_ids].cuda()\n",
        "    X_test = X[test_ids].cuda()\n",
        "\n",
        "    Y_train = Y[train_ids].cuda()\n",
        "    Y_test = Y[test_ids].cuda()\n",
        "    return X_train, X_test, Y_train, Y_test"
      ],
      "id": "actual-announcement",
      "execution_count": 12,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "secure-interference"
      },
      "source": [
        "def train_model(N_dim, dataset, name):\n",
        "    X_train, X_test, Y_train, Y_test = dataset\n",
        "    model = MLP(N_dim, name).cuda()\n",
        "    optimizer = torch.optim.SGD(model.parameters(), lr=0.0001)\n",
        "    loss_fn = nn.CrossEntropyLoss()\n",
        "    model.dataset = dataset\n",
        "    \n",
        "    for epoch in tqdm(range(1000)): \n",
        "        outputs = model(X_train)\n",
        "        loss = loss_fn(outputs, Y_train.long())\n",
        "        loss.backward()\n",
        "        optimizer.step()\n",
        "\n",
        "        predictions = torch.argmax(model(X_test), dim=1)\n",
        "        accuracy = torch.sum(predictions == Y_test)/len(Y_test)\n",
        "        \n",
        "        train_predictions = torch.argmax(model(X_train), dim=1)\n",
        "        train_accuracy = torch.sum(train_predictions == Y_train)/len(Y_train)\n",
        "\n",
        "        if epoch == 999:\n",
        "            print(\"Epoch:%s, Train Acc:%s\"%(epoch, train_accuracy))\n",
        "            print(\"Epoch:%s, Test Acc:%s\"%(epoch, accuracy))\n",
        "        \n",
        "        model.test_accuracy = accuracy\n",
        "        model.train_accuracy = train_accuracy\n",
        "    \n",
        "    return model"
      ],
      "id": "secure-interference",
      "execution_count": 13,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "literary-wyoming"
      },
      "source": [
        "def batched_train_model(N_dim, dataset, name):\n",
        "    X_train, X_test, Y_train, Y_test = dataset\n",
        "    model = MLP(N_dim, name).cuda()\n",
        "    optimizer = torch.optim.SGD(model.parameters(), lr=0.0001)\n",
        "    loss_fn = nn.CrossEntropyLoss()\n",
        "    model.dataset = dataset\n",
        "    \n",
        "    BATCH_SIZE = 64000\n",
        "    \n",
        "    for epoch in tqdm(range(1000)):\n",
        "        num_batches = int(X_train.shape[0]/(BATCH_SIZE))\n",
        "        \n",
        "        batch_accuracies = []\n",
        "        batch_train_accuracies = []\n",
        "        for batch_num in range(num_batches):\n",
        "            start_pt = BATCH_SIZE * batch_num\n",
        "            end_pt = start_pt + BATCH_SIZE\n",
        "            train_X = X_train[start_pt:end_pt]\n",
        "            train_Y = Y_train[start_pt:end_pt]\n",
        "            test_X = X_test\n",
        "            test_Y = Y_test\n",
        "\n",
        "            outputs = model(train_X)\n",
        "            loss = loss_fn(outputs, train_Y.long())\n",
        "            loss.backward()\n",
        "            optimizer.step()\n",
        "\n",
        "            batch_predictions = torch.argmax(model(test_X), dim=1)\n",
        "            batch_accuracy = torch.sum(batch_predictions == test_Y)/len(test_Y)\n",
        "            batch_accuracies.append(batch_accuracy.item())\n",
        "            \n",
        "            batch_train_predictions = torch.argmax(model(train_X), dim=1)\n",
        "            batch_train_accuracy = torch.sum(batch_train_predictions == train_Y)/len(train_Y)\n",
        "            batch_train_accuracies.append(batch_train_accuracy.item())        \n",
        "        \n",
        "        accuracy = np.mean(batch_accuracies)\n",
        "        train_accuracy = np.mean(batch_train_accuracies)\n",
        "\n",
        "        \n",
        "        if epoch%100 == 0:\n",
        "            print(\"Epoch:%s, Train Acc:%s\"%(epoch, train_accuracy))\n",
        "            print(\"Epoch:%s, Test Acc:%s\"%(epoch, accuracy))\n",
        "        \n",
        "        if epoch == 999:\n",
        "            print(\"Epoch:%s, Train Acc:%s\"%(epoch, train_accuracy))\n",
        "            print(\"Epoch:%s, Test Acc:%s\"%(epoch, accuracy))\n",
        "        \n",
        "        model.test_accuracy = accuracy\n",
        "        model.train_accuracy = train_accuracy\n",
        "    \n",
        "    return model"
      ],
      "id": "literary-wyoming",
      "execution_count": 14,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "magnetic-circuit"
      },
      "source": [
        "#### Tests for generating and training image models####\n",
        "\n",
        "# image_test_data = make_dataset(1000, True)\n",
        "# image_test_model = train_model(N_dim, image_test_data, 'dset_size_%s'%dsize, True)"
      ],
      "id": "magnetic-circuit",
      "execution_count": 15,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "individual-apache"
      },
      "source": [
        "# CMA failures"
      ],
      "id": "individual-apache"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "difficult-flood"
      },
      "source": [
        "# def cma_objective(x_input):\n",
        "#     torch_x = torch.from_numpy(x_input).unsqueeze(0).float()\n",
        "#     if 'LeNet' in CURRENT_MODEL.name\n",
        "#     output = CURRENT_MODEL(torch_x)\n",
        "#     pred_prob = output[0][CATEGORY_NUM].item()\n",
        "#     prediction = torch.argmax(output[0]).item()\n",
        "#     return pred_prob, prediction"
      ],
      "id": "difficult-flood",
      "execution_count": 16,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "phantom-cancellation"
      },
      "source": [
        "# torch.vstack([start_pos[0].unsqueeze(0), start_pos[0].unsqueeze(0), start_pos[0].unsqueeze(0)]).shape"
      ],
      "id": "phantom-cancellation",
      "execution_count": 17,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "opened-louisville"
      },
      "source": [
        "def cma_objective(x_input):\n",
        "    torch_x = torch.from_numpy(x_input).unsqueeze(0).float()\n",
        "#     try:\n",
        "#         if 'LeNet' in CURRENT_MODEL.name:\n",
        "#             side_size = int(np.sqrt(N_dim))\n",
        "#             torch_x = torch.from_numpy(x_input).reshape(1,1,side_size,side_size).float()\n",
        "#         else:\n",
        "#             torch_x = torch.from_numpy(x_input).unsqueeze(0).float()\n",
        "#     except:\n",
        "#         side_size = int(np.sqrt(int(N_dim/3)))\n",
        "#         torch_x_channel = torch.from_numpy(x_input).reshape(1,3,side_size,side_size).float()\n",
        "#         torch_x = torch.vstack([torch_x_channel, torch_x_channel, torch_x_channel]).cuda()\n",
        "        \n",
        "    output = CURRENT_MODEL(torch_x)\n",
        "    pred_prob = output[0][CATEGORY_NUM].item()\n",
        "    prediction = torch.argmax(output[0]).item()\n",
        "    return pred_prob, prediction"
      ],
      "id": "opened-louisville",
      "execution_count": 18,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "renewable-constitutional"
      },
      "source": [
        "def cma_experiment(attacked_model, num_samples = 50):\n",
        "    cma_search_output = {}\n",
        "    global CURRENT_MODEL\n",
        "    CURRENT_MODEL = attacked_model\n",
        "    print(CURRENT_MODEL.name)\n",
        "    cma_output = CMA_info(CURRENT_MODEL.name, num_samples)        \n",
        "    \n",
        "    for i in tqdm(range(num_samples)):\n",
        "        initial_pred = 1\n",
        "        while initial_pred == 1:\n",
        "            start_pos = sample_uniform(1, N_dim, test_min, test_max)\n",
        "            output = CURRENT_MODEL(start_pos)\n",
        "            initial_pred = torch.argmax(output[0]).item()\n",
        "        cma_output.starts.append(start_pos)\n",
        "        start_pos = start_pos[0]\n",
        "        \n",
        "        es = cma.CMAEvolutionStrategy(start_pos, 0.00005)\n",
        "        es.optimize(cma_objective, verb_disp = False, iterations=1500, correct_prediction = CATEGORY_NUM)\n",
        "        adv_offspring_ids = np.where(np.array(es.predictions) != 0)\n",
        "        \n",
        "        if len(adv_offspring_ids[0]) > 0:\n",
        "            random_adv_offspring_id = random.choice(list(adv_offspring_ids[0]))\n",
        "            random_adv_offspring = es.prediction_settings[random_adv_offspring_id]\n",
        "            cma_output.advs.append(random_adv_offspring)        \n",
        "\n",
        "            max_val = np.max(random_adv_offspring)\n",
        "            if max_val < x1_max:\n",
        "                cma_output.in_dist_advs.append(random_adv_offspring)\n",
        "                distance = np.linalg.norm(start_pos - random_adv_offspring)\n",
        "                cma_output.distances.append(distance)\n",
        "    return cma_output"
      ],
      "id": "renewable-constitutional",
      "execution_count": 19,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "affected-trade"
      },
      "source": [
        "CATEGORY_NUM = 0\n",
        "CURRENT_MODEL = None\n",
        "\n",
        "N_dim = 100"
      ],
      "id": "affected-trade",
      "execution_count": 20,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "national-shock"
      },
      "source": [
        "# all_info[-1]"
      ],
      "id": "national-shock",
      "execution_count": 21,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "distant-hobby"
      },
      "source": [
        "# all_info[0]"
      ],
      "id": "distant-hobby",
      "execution_count": 22,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "outside-startup"
      },
      "source": [
        "### CMA attack accuracy vs Dataset size"
      ],
      "id": "outside-startup"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "utility-league"
      },
      "source": [
        "all_info = []"
      ],
      "id": "utility-league",
      "execution_count": 23,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "comic-landscape",
        "outputId": "df89c926-457d-4481-c9db-074e32c3786e"
      },
      "source": [
        "for iteration in range(10):\n",
        "    print('*'*50)\n",
        "    print('Iteration %s'%iteration)\n",
        "    print('*'*50)\n",
        "    \n",
        "    trained_models = {}\n",
        "    attack_output = {}\n",
        "    for dsize in [10000, 100000]:\n",
        "        print('Working with with %s'%dsize)\n",
        "        dset = make_dataset(dsize)\n",
        "        if dsize > 64000:\n",
        "            trained_models[dsize] = batched_train_model(N_dim, dset, 'dset_size_%s'%dsize)\n",
        "        else:\n",
        "            trained_models[dsize] = train_model(N_dim, dset, 'dset_size_%s'%dsize)\n",
        "        model_to_attack = trained_models[dsize].cpu()\n",
        "        attack_output[dsize] = cma_experiment(model_to_attack, 50)\n",
        "    iter_info = [trained_models, attack_output]\n",
        "    all_info.append(iter_info)"
      ],
      "id": "comic-landscape",
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "**************************************************\n",
            "Iteration 0\n",
            "**************************************************\n",
            "Working with with 10000\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "o8u77aT_fhl4"
      },
      "source": [
        "with open('dataset_size_experiments.p','wb') as F:\n",
        "    pickle.dump(all_info, F)"
      ],
      "id": "o8u77aT_fhl4",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "conscious-spirit"
      },
      "source": [
        "# trained_models[1024000].test_accuracy"
      ],
      "id": "conscious-spirit",
      "execution_count": 49,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "warming-position"
      },
      "source": [
        "# all_info[0][0][1024000] = all_info[1][0][1024000]"
      ],
      "id": "warming-position",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "numerical-economics"
      },
      "source": [
        "# all_info[0][1][1024000] = all_info[1][1][1024000]"
      ],
      "id": "numerical-economics",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "satellite-labor"
      },
      "source": [
        "# all_info[1]"
      ],
      "id": "satellite-labor",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "intended-blackjack"
      },
      "source": [
        "all_info[0][1]"
      ],
      "id": "intended-blackjack",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "bibliographic-methodology"
      },
      "source": [
        "attack_per_dset_size = {}\n",
        "for key in all_info[0][1].keys():\n",
        "    adv_count = len(all_info[0][1][key].advs)\n",
        "    start_count = len(all_info[0][1][key].starts)\n",
        "    attack_rate = adv_count/start_count*100\n",
        "    attack_per_dset_size[key] = attack_rate"
      ],
      "id": "bibliographic-methodology",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "usual-observation",
        "outputId": "32585821-2bd5-49fe-b28e-cd61905546e6"
      },
      "source": [
        "attack_per_dset_size"
      ],
      "id": "usual-observation",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "{4000: 100.0, 16000: 100.0, 64000: 100.0, 256000: 100.0, 1024000: 100.0}"
            ]
          },
          "execution_count": 105,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "requested-albania",
        "outputId": "d684fd86-77be-4c4f-8d7a-aa57b9080bd2"
      },
      "source": [
        "plt.plot(attack_per_dset_size.values(), marker='o')\n",
        "plt.xticks(list(range(len(attack_per_dset_size.keys()))), labels = attack_per_dset_size.keys())\n",
        "plt.ylim(0,110)\n",
        "plt.title('CMA attack rate as dataset size changes')\n",
        "plt.xlabel('Dataset Size')\n",
        "plt.ylabel('CMA attack rate')\n",
        "plt.show()"
      ],
      "id": "requested-albania",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "monetary-greeting"
      },
      "source": [
        "##### Execute because used extensively below #####\n",
        "def predict(model, point):\n",
        "    prepped_point = torch.from_numpy(point).cuda().float().unsqueeze(0)\n",
        "    prediction = torch.argmax(model(prepped_point)[0]).item()\n",
        "    return prediction"
      ],
      "id": "monetary-greeting",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "specified-hungary"
      },
      "source": [
        "# Impact of Dimensionality"
      ],
      "id": "specified-hungary"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "dated-bargain"
      },
      "source": [
        "all_info = []"
      ],
      "id": "dated-bargain",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "congressional-sauce",
        "outputId": "01554161-b041-4c2f-d637-94a88f25a000"
      },
      "source": [
        "for iteration in range(1):\n",
        "    print('*'*50)\n",
        "    print('Iteration %s'%iteration)\n",
        "    print('*'*50)\n",
        "    \n",
        "    trained_models = {}\n",
        "    attack_output = {}\n",
        "    for N_dim in [5000]:\n",
        "        print('Working with with %s dimensions'%N_dim)\n",
        "        dsize = 10000\n",
        "        dset = make_dataset(dsize)\n",
        "        trained_models[N_dim] = train_model(N_dim, dset, 'dset_size_%s'%dsize)\n",
        "        model_to_attack = trained_models[N_dim].cpu()\n",
        "        attack_output[N_dim] = cma_experiment(model_to_attack, 10)\n",
        "    iter_info = [trained_models, attack_output]\n",
        "    all_info.append(iter_info)"
      ],
      "id": "congressional-sauce",
      "execution_count": null,
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "**************************************************\n",
            "Iteration 0\n",
            "**************************************************\n",
            "Working with with 5000 dimensions\n"
          ]
        },
        {
          "data": {
            "application/vnd.jupyter.widget-view+json": {
              "model_id": "514cacd2d8c54bed8cad478062d88ce1",
              "version_major": 2,
              "version_minor": 0
            },
            "text/plain": [
              "  0%|          | 0/100 [00:00<?, ?it/s]"
            ]
          },
          "metadata": {},
          "output_type": "display_data"
        },
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "dset_size_10000\n"
          ]
        },
        {
          "data": {
            "application/vnd.jupyter.widget-view+json": {
              "model_id": "07b29b1eff744859a158be414c2a3fa4",
              "version_major": 2,
              "version_minor": 0
            },
            "text/plain": [
              "  0%|          | 0/10 [00:00<?, ?it/s]"
            ]
          },
          "metadata": {},
          "output_type": "display_data"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "damaged-hobby",
        "outputId": "86709c38-88a2-478e-8921-bf3bf43fcde4"
      },
      "source": [
        "len(attack_output[5000].in_dist_advs)"
      ],
      "id": "damaged-hobby",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "0"
            ]
          },
          "execution_count": 25,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "rolled-camcorder",
        "outputId": "626ba95f-cccc-4ceb-8545-07a8d768ae38"
      },
      "source": [
        "len(attack_output[1000].in_dist_advs)"
      ],
      "id": "rolled-camcorder",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "10"
            ]
          },
          "execution_count": 23,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "continent-custody",
        "outputId": "08b09815-cd49-4141-d626-eaffa8e1b517"
      },
      "source": [
        "len(attack_output[100].in_dist_advs)"
      ],
      "id": "continent-custody",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "10"
            ]
          },
          "execution_count": 21,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "iraqi-guess",
        "outputId": "e8d0a77e-011c-44d9-b4b0-90925a9494d1"
      },
      "source": [
        "len(attack_output[50].in_dist_advs)"
      ],
      "id": "iraqi-guess",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "10"
            ]
          },
          "execution_count": 22,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "secret-outside",
        "outputId": "96c4fda5-da88-4d91-dd8d-df916d11fb47"
      },
      "source": [
        "len(attack_output[100].in_dist_advs)"
      ],
      "id": "secret-outside",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "10"
            ]
          },
          "execution_count": 23,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "conscious-clinton",
        "outputId": "fdd852f9-152a-462d-d6c2-bf1ce5a978a6"
      },
      "source": [
        "len(attack_output[250].in_dist_advs)"
      ],
      "id": "conscious-clinton",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "10"
            ]
          },
          "execution_count": 28,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "wanted-glory",
        "outputId": "3dac26fe-4df1-46cb-e5f0-5c46f6106b83"
      },
      "source": [
        "len(attack_output[500].in_dist_advs)"
      ],
      "id": "wanted-glory",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "6"
            ]
          },
          "execution_count": 29,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "grave-remains",
        "outputId": "e11f141b-3f9a-4f88-a257-81206aa3acf7"
      },
      "source": [
        "len(attack_output[1000].in_dist_advs)"
      ],
      "id": "grave-remains",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "10"
            ]
          },
          "execution_count": 31,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "internal-germany"
      },
      "source": [
        "len(attack_output[2500].in_dist_advs)"
      ],
      "id": "internal-germany",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "aggregate-resolution"
      },
      "source": [
        "attack_per_dset_dim = {}\n",
        "for key in iter_info[1].keys():\n",
        "    adv_count = len(iter_info[1][key].advs)\n",
        "    start_count = len(iter_info[1][key].starts)\n",
        "    attack_rate = adv_count/start_count*100\n",
        "    attack_per_dset_dim[key] = attack_rate"
      ],
      "id": "aggregate-resolution",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "flying-economics"
      },
      "source": [
        "plt.plot(attack_per_dset_dim.values(), marker='o')\n",
        "plt.xticks(list(range(len(attack_per_dset_dim.keys()))), labels = attack_per_dset_dim.keys())\n",
        "plt.ylim(0,110)\n",
        "plt.title('CMA attack rate as dataset size changes')\n",
        "plt.xlabel('Dataset Size')\n",
        "plt.ylabel('CMA attack rate')\n",
        "plt.show()"
      ],
      "id": "flying-economics",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "sophisticated-gregory"
      },
      "source": [
        "# Does Pooling solve the problem?"
      ],
      "id": "sophisticated-gregory"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "median-hazard"
      },
      "source": [
        "def predict(model, point):\n",
        "    prepped_point = torch.from_numpy(point).cuda().float().unsqueeze(0)\n",
        "    prediction = torch.argmax(model(prepped_point)[0]).item()\n",
        "    return prediction"
      ],
      "id": "median-hazard",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "latin-bryan",
        "outputId": "27011c03-5416-4f44-b5c1-2cff2d13d8f1"
      },
      "source": [
        "print('****** This is a sanity check *******')\n",
        "for i in range(5):\n",
        "    random_point = sample_uniform(1, N_dim, test_min, test_max)\n",
        "    pred_out = predict(model_to_attack.cuda(), np.array(random_point[0]))\n",
        "    print('Randomly sampled a point from the test_min, test_max range....')\n",
        "    print('Prediction is %s'%pred_out)\n",
        "print('****** Should be a lot 0s predicted above*****')"
      ],
      "id": "latin-bryan",
      "execution_count": null,
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "****** This is a sanity check *******\n",
            "Randomly sampled a point from the test_min, test_max range....\n",
            "Prediction is 0\n",
            "Randomly sampled a point from the test_min, test_max range....\n",
            "Prediction is 0\n",
            "Randomly sampled a point from the test_min, test_max range....\n",
            "Prediction is 0\n",
            "Randomly sampled a point from the test_min, test_max range....\n",
            "Prediction is 0\n",
            "Randomly sampled a point from the test_min, test_max range....\n",
            "Prediction is 0\n",
            "****** Should be a lot 0s predicted above*****\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "moral-musician",
        "outputId": "ece800d0-eb14-4c96-94f9-fb13b79363ab"
      },
      "source": [
        "predict(model_to_attack.cuda(), all_info[0][1][10000].advs[1])"
      ],
      "id": "moral-musician",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "1"
            ]
          },
          "execution_count": 305,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "endless-macro"
      },
      "source": [
        "##### Let's try to fix this by randomly perturbing the adversarial"
      ],
      "id": "endless-macro"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "interstate-summary"
      },
      "source": [
        "pert_to_recovery = {}\n",
        "for repeats in range(5):\n",
        "    for p in np.arange(0.05,0.3,0.05):\n",
        "        all_top_preds = []\n",
        "        for i in range(50):\n",
        "            adv_sample = all_info[0][1][10000].advs[i]\n",
        "\n",
        "            all_preds = []\n",
        "            for i in range(10):\n",
        "                perturb_min = -p\n",
        "                perturb_max = p\n",
        "                perturb_vector = np.array(sample_uniform(1, N_dim, perturb_min, perturb_max)[0])\n",
        "                perturbed_sample = adv_sample + perturb_vector\n",
        "                perturbed_prediction = predict(model_to_attack.cuda(), perturbed_sample)\n",
        "                all_preds.append(perturbed_prediction)\n",
        "            top_pred = max(set(all_preds), key = all_preds.count)\n",
        "\n",
        "            all_top_preds.append(top_pred)\n",
        "\n",
        "        recovered_acc = (1 - np.sum(all_top_preds)/50)*100\n",
        "        if \"%.02f\"%p in pert_to_recovery.keys():\n",
        "            pert_to_recovery[\"%.02f\"%p].append(recovered_acc)\n",
        "        else:\n",
        "            pert_to_recovery[\"%.02f\"%p] = [recovered_acc]\n",
        "#         print('Recovered %s%% samples by random perturbation.'%recovered_acc)"
      ],
      "id": "interstate-summary",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "antique-preparation",
        "outputId": "bebe6613-c25d-49f0-8b62-a8e553fa2e88"
      },
      "source": [
        "np.linalg.norm(perturb_vector)"
      ],
      "id": "antique-preparation",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "1.4335177"
            ]
          },
          "execution_count": 325,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "brief-chase"
      },
      "source": [
        "mean_perfs = np.mean(np.array(list(pert_to_recovery.values())),axis=1)\n",
        "std_perfs = np.std(np.array(list(pert_to_recovery.values())),axis=1)"
      ],
      "id": "brief-chase",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "delayed-chain",
        "outputId": "aa2cac46-ea97-46e4-e960-2f5cdd0a491b"
      },
      "source": [
        "plt.errorbar([0,1,2,3,4], mean_perfs, std_perfs, marker='o')\n",
        "plt.title('Recovery as function of perturbation norm')\n",
        "plt.xlabel('amount of random perturbation on adv sample')\n",
        "plt.ylabel('Accuracy after voting on multiple predictions')\n",
        "plt.xticks([0,1,2,3,4],labels=list(pert_to_recovery.keys()))\n",
        "plt.show()"
      ],
      "id": "delayed-chain",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "weekly-netscape"
      },
      "source": [
        "#### Lets see the norm of adversarial perturbation: how much is needed if the direction was known."
      ],
      "id": "weekly-netscape"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "subsequent-visibility"
      },
      "source": [
        "adv_point = all_info[0][1][10000].advs[1]\n",
        "start_point = np.array(all_info[0][1][10000].starts[1])"
      ],
      "id": "subsequent-visibility",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "deluxe-booth"
      },
      "source": [
        "adv_vec = adv_point - start_point"
      ],
      "id": "deluxe-booth",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "printable-snapshot",
        "outputId": "600b8297-5388-49be-a7a8-664486f09d5d"
      },
      "source": [
        "np.linalg.norm(adv_vec)"
      ],
      "id": "printable-snapshot",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "2.9962379296577284"
            ]
          },
          "execution_count": 321,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "liked-evans",
        "outputId": "a844d44d-397b-465e-d956-ab93163d965e"
      },
      "source": [
        "np.linalg.norm(adv_point)"
      ],
      "id": "liked-evans",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "5.389204731238513"
            ]
          },
          "execution_count": 322,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "beneficial-supervisor",
        "outputId": "9e83f99f-815c-44ad-8f0d-bb9ceda01373"
      },
      "source": [
        "np.linalg.norm(start_point)"
      ],
      "id": "beneficial-supervisor",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "5.800697"
            ]
          },
          "execution_count": 323,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "nominated-output"
      },
      "source": [
        "# Church Window Plots"
      ],
      "id": "nominated-output"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "appointed-absolute"
      },
      "source": [
        "def perpendicular_vector(v):\n",
        "    random_vector = np.random.randn(N_dim)\n",
        "    orthogonal = random_vector - random_vector.dot(v)*v/np.linalg.norm(v)**2\n",
        "    orthogonal_normalized = orthogonal/np.linalg.norm(orthogonal) * np.linalg.norm(v)\n",
        "    return orthogonal_normalized\n",
        "\n",
        "def plot_church_window(start_point, adv_point):\n",
        "    adv_vector = adv_point - start_point\n",
        "    orthogonal_vector = perpendicular_vector(adv_vector)\n",
        "\n",
        "\n",
        "    ##### Collect Predictions ######\n",
        "    predictions = np.zeros((40,40))\n",
        "\n",
        "    row_num = 0\n",
        "    col_num = 0\n",
        "\n",
        "    for alpha in np.arange(0, 2.0, 0.05):\n",
        "        for beta in np.arange(0, 2.0, 0.05):\n",
        "            curr_point = start_point + beta*adv_vector + alpha*orthogonal_vector\n",
        "            curr_pred = predict(model, curr_point)\n",
        "            predictions[row_num][col_num] = curr_pred\n",
        "            col_num += 1\n",
        "        row_num += 1\n",
        "        col_num = 0\n",
        "\n",
        "    ##### Plot ######\n",
        "    plt.imshow(predictions, cmap=custom_cmap)\n",
        "    plt.xlim([0, 40])\n",
        "    plt.ylim([0, 40])\n",
        "    plt.title('Church Window Plot')\n",
        "    plt.xticks(range(0,40,5), labels = np.arange(0,2,0.25))\n",
        "    plt.yticks(range(0,40,5), labels = np.arange(0,2,0.25))\n",
        "    plt.xlabel('Adversarial Direction')\n",
        "    plt.ylabel('Orthogonal Direction')\n",
        "    plt.show()"
      ],
      "id": "appointed-absolute",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": false,
        "id": "sized-contrast",
        "outputId": "1ced093d-b6f9-46b6-9f85-a660779597f4"
      },
      "source": [
        "DSIZE = 10000\n",
        "\n",
        "\n",
        "from matplotlib import colors\n",
        "\n",
        "custom_cmap = colors.ListedColormap(['white', 'red'])\n",
        "bounds=[0,1]\n",
        "norm = colors.BoundaryNorm(bounds, custom_cmap.N)\n",
        "\n",
        "model = all_info[0][0][DSIZE].cuda()\n",
        "\n",
        "for i in range(5):\n",
        "    adv_point = all_info[0][1][DSIZE].in_dist_advs[i]\n",
        "    start_point = all_info[0][1][DSIZE].starts[i]\n",
        "    start_point = np.array(start_point)[0]\n",
        "    plot_church_window(start_point, adv_point)"
      ],
      "id": "sized-contrast",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        },
        {
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        },
        {
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        },
        {
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        },
        {
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "parallel-forestry"
      },
      "source": [
        "# Nearest neighbor analysis"
      ],
      "id": "parallel-forestry"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "piano-expense"
      },
      "source": [
        "starting_points = all_info[0][1][10000].starts\n",
        "in_dist_advs_points = all_info[0][1][10000].in_dist_advs"
      ],
      "id": "piano-expense",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "waiting-rates"
      },
      "source": [
        "training_points = all_info[0][0][10000].dataset[0]"
      ],
      "id": "waiting-rates",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "suspected-subject"
      },
      "source": [
        "start_neighbor_distances = []\n",
        "for start_p in starting_points:\n",
        "    distances = torch.sum((training_points - start_p.cuda())**2, dim = 1)\n",
        "    neighbors = torch.argsort(distances)[:10].cpu()\n",
        "    neighbor_distance = torch.sum(distances[neighbors]).item()\n",
        "    start_neighbor_distances.append(neighbor_distance)"
      ],
      "id": "suspected-subject",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "irish-newsletter"
      },
      "source": [
        "adv_neighbor_distances = []\n",
        "for adv_p in in_dist_advs_points:\n",
        "    distances = torch.sum((training_points - torch.from_numpy(adv_p).cuda())**2, dim = 1)\n",
        "    neighbors = torch.argsort(distances)[:10].cpu()\n",
        "    neighbor_distance = torch.sum(distances[neighbors]).item()\n",
        "    adv_neighbor_distances.append(neighbor_distance)"
      ],
      "id": "irish-newsletter",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "czech-sterling",
        "outputId": "c67a00d1-3200-4cbf-b01c-92dadacf57a8"
      },
      "source": [
        "plt.hist(start_neighbor_distances)"
      ],
      "id": "czech-sterling",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "(array([ 2.,  3., 24., 40., 45., 47., 24., 10.,  4.,  1.]),\n",
              " array([ 9915.65136719,  9974.9921875 , 10034.33300781, 10093.67382812,\n",
              "        10153.01464844, 10212.35546875, 10271.69628906, 10331.03710938,\n",
              "        10390.37792969, 10449.71875   , 10509.05957031]),\n",
              " <BarContainer object of 10 artists>)"
            ]
          },
          "execution_count": 32,
          "metadata": {},
          "output_type": "execute_result"
        },
        {
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "foreign-minnesota",
        "outputId": "bfa4398a-9d3e-411e-d265-1f153782e076"
      },
      "source": [
        "plt.hist(adv_neighbor_distances)"
      ],
      "id": "foreign-minnesota",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "(array([ 3.,  7., 27., 39., 52., 34., 23., 10.,  4.,  1.]),\n",
              " array([ 9939.33247206,  9988.23968375, 10037.14689545, 10086.05410715,\n",
              "        10134.96131885, 10183.86853054, 10232.77574224, 10281.68295394,\n",
              "        10330.59016564, 10379.49737733, 10428.40458903]),\n",
              " <BarContainer object of 10 artists>)"
            ]
          },
          "execution_count": 33,
          "metadata": {},
          "output_type": "execute_result"
        },
        {
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "yellow-prague",
        "outputId": "b27ae27c-372d-4bf1-9a10-23c16c9e5961"
      },
      "source": [
        "np.mean(adv_neighbor_distances)"
      ],
      "id": "yellow-prague",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "10160.343138349179"
            ]
          },
          "execution_count": 34,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "marine-packaging",
        "outputId": "3e82fbee-8233-4ab8-b803-a235bc1f2b58"
      },
      "source": [
        "np.mean(start_neighbor_distances)"
      ],
      "id": "marine-packaging",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "10196.771958007812"
            ]
          },
          "execution_count": 35,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "coated-burns",
        "outputId": "8bc1735e-9e98-4cbb-c0ee-6123bce55aeb"
      },
      "source": [
        "np.std(adv_neighbor_distances)"
      ],
      "id": "coated-burns",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "81.89637551160129"
            ]
          },
          "execution_count": 36,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "confirmed-rabbit",
        "outputId": "631a9ad5-d5c7-43e0-b48f-66a5cdc0d068"
      },
      "source": [
        "np.std(start_neighbor_distances)"
      ],
      "id": "confirmed-rabbit",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "text/plain": [
              "94.20267086666723"
            ]
          },
          "execution_count": 37,
          "metadata": {},
          "output_type": "execute_result"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "governmental-leone"
      },
      "source": [
        "# Adversarial Training analysis"
      ],
      "id": "governmental-leone"
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "baking-reynolds"
      },
      "source": [
        "### Collect adv points (25% of training data size)"
      ],
      "id": "baking-reynolds"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "casual-atlantic"
      },
      "source": [
        "attack_output_2 = {}"
      ],
      "id": "casual-atlantic",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": false,
        "id": "former-winter",
        "outputId": "b1336177-b27a-4ad1-86a3-bdfe7172ac18"
      },
      "source": [
        "attack_output_2[dsize] = cma_experiment(model_to_attack, 4000)"
      ],
      "id": "former-winter",
      "execution_count": null,
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "dset_size_10000\n"
          ]
        },
        {
          "data": {
            "application/vnd.jupyter.widget-view+json": {
              "model_id": "a1bc90d8466d49cca92ac278ceb321a7",
              "version_major": 2,
              "version_minor": 0
            },
            "text/plain": [
              "  0%|          | 0/4000 [00:00<?, ?it/s]"
            ]
          },
          "metadata": {},
          "output_type": "display_data"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "specified-lender"
      },
      "source": [
        "# with open('inof.p', 'wb') as F:\n",
        "#     pickle.dump(attack_output_2, F)"
      ],
      "id": "specified-lender",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": false,
        "id": "coated-retro"
      },
      "source": [
        "augmented_training_data = torch.vstack([training_points, torch.tensor(attack_output_2[dsize].in_dist_advs).float().cuda()])"
      ],
      "id": "coated-retro",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "adult-finance"
      },
      "source": [
        "### Train with them"
      ],
      "id": "adult-finance"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "assigned-circular"
      },
      "source": [
        "training_labels = dset[2]"
      ],
      "id": "assigned-circular",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "controlling-quantity"
      },
      "source": [
        "augmented_training_labels = torch.hstack([training_labels, torch.zeros(4000).long().cuda()])"
      ],
      "id": "controlling-quantity",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "southern-floor"
      },
      "source": [
        "# augmented_training_labels[-1]"
      ],
      "id": "southern-floor",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "institutional-attraction"
      },
      "source": [
        "dset_new = [augmented_training_data, dset[1], augmented_training_labels, dset[3]] "
      ],
      "id": "institutional-attraction",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "similar-perry"
      },
      "source": [
        "adv_trained_models = {}"
      ],
      "id": "similar-perry",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "furnished-particle"
      },
      "source": [
        "adv_attack_output = {}"
      ],
      "id": "furnished-particle",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": false,
        "id": "minus-split",
        "outputId": "8aa1bc03-8e33-4c34-f581-cdc045a65e52"
      },
      "source": [
        "# dset = make_dataset(dsize)\n",
        "adv_trained_models[dsize] = train_model(N_dim, dset_new, 'dset_size_%s_adv'%dsize)"
      ],
      "id": "minus-split",
      "execution_count": null,
      "outputs": [
        {
          "data": {
            "application/vnd.jupyter.widget-view+json": {
              "model_id": "d292af428f1846458c1fb9a9fa88716d",
              "version_major": 2,
              "version_minor": 0
            },
            "text/plain": [
              "  0%|          | 0/1000 [00:00<?, ?it/s]"
            ]
          },
          "metadata": {},
          "output_type": "display_data"
        },
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Epoch:999, Train Acc:tensor(1., device='cuda:0')\n",
            "Epoch:999, Test Acc:tensor(1., device='cuda:0')\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "dedicated-howard"
      },
      "source": [
        "# adv_trained_models[10000].dataset[0].shape"
      ],
      "id": "dedicated-howard",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "double-yesterday",
        "outputId": "b0a45c52-beea-4cd1-cbf9-a2ceea74df40"
      },
      "source": [
        "adv_model_to_attack = adv_trained_models[dsize].cpu()\n",
        "adv_attack_output[dsize] = cma_experiment(adv_model_to_attack, 50)"
      ],
      "id": "double-yesterday",
      "execution_count": null,
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "dset_size_10000_adv\n"
          ]
        },
        {
          "data": {
            "application/vnd.jupyter.widget-view+json": {
              "model_id": "4df79fac3a3448ab9cbb50045058d882",
              "version_major": 2,
              "version_minor": 0
            },
            "text/plain": [
              "  0%|          | 0/50 [00:00<?, ?it/s]"
            ]
          },
          "metadata": {},
          "output_type": "display_data"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "familiar-visibility"
      },
      "source": [
        "# len(adv_attack_output[dsize].in_dist_advs)"
      ],
      "id": "familiar-visibility",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "educated-amplifier"
      },
      "source": [
        "# Aggregate evaluation"
      ],
      "id": "educated-amplifier"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "committed-clause"
      },
      "source": [
        "all_distances = {}\n",
        "in_dists = {}"
      ],
      "id": "committed-clause",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "portuguese-security"
      },
      "source": [
        "# all_info[0][1][dsize].in_dist_advs"
      ],
      "id": "portuguese-security",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "diagnostic-chicago"
      },
      "source": [
        "for dsize in [1000, 100000]:\n",
        "    all_distances[dsize] = []\n",
        "    in_dists[dsize] = []\n",
        "    for i in range(1):\n",
        "        in_dists[dsize].append(len(all_info[i][1][dsize].in_dist_advs))\n",
        "        all_distances[dsize].append(np.nanmean(all_info[i][1][dsize].distances))"
      ],
      "id": "diagnostic-chicago",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "congressional-smoke"
      },
      "source": [
        "# all_distances.keys()"
      ],
      "id": "congressional-smoke",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": false,
        "id": "entitled-bridges",
        "outputId": "47b8c9e7-7d50-4fe7-f503-e56324df9f69"
      },
      "source": [
        "for key in all_distances.keys():\n",
        "    print(np.nanmean(all_distances[key]))"
      ],
      "id": "entitled-bridges",
      "execution_count": null,
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "3.2171293443927995\n",
            "5.13962748088732\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": false,
        "id": "destroyed-airfare",
        "outputId": "a79caf70-abb6-462c-eccd-ee6b92780360"
      },
      "source": [
        "for key in in_dists.keys():\n",
        "    print(np.mean(in_dists[key]))"
      ],
      "id": "destroyed-airfare",
      "execution_count": null,
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "10.0\n",
            "10.0\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "comparable-better"
      },
      "source": [
        "def predict(model, point):\n",
        "    prepped_point = torch.from_numpy(point).cuda().float().unsqueeze(0)\n",
        "    prediction = torch.argmax(model(prepped_point)[0]).item()\n",
        "    return prediction"
      ],
      "id": "comparable-better",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "mechanical-guest"
      },
      "source": [
        "# Sensitivity Analysis"
      ],
      "id": "mechanical-guest"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "imperial-terminology"
      },
      "source": [
        "DSIZE = 1000"
      ],
      "id": "imperial-terminology",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "foreign-pepper"
      },
      "source": [
        "mod = all_info[0][0][DSIZE].cuda()"
      ],
      "id": "foreign-pepper",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "scenic-fraction"
      },
      "source": [
        "def min_subset(start, adv, mod, start_id = 0):\n",
        "    for i in range(start_id, start_id + 50):\n",
        "        index = i%50\n",
        "        start[:index] = adv[:index]\n",
        "        start_pred = predict(mod, start)\n",
        "        adv_pred = predict(mod, adv)\n",
        "        if start_pred == adv_pred:\n",
        "            break\n",
        "    return index        \n",
        "#     print('Broke at %i'%index)"
      ],
      "id": "scenic-fraction",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "growing-steal"
      },
      "source": [
        "broken_changes = []\n",
        "for i in range(10):\n",
        "    adv_point = all_info[0][1][DSIZE].in_dist_advs[i]\n",
        "    start_point = all_info[0][1][DSIZE].starts[i]\n",
        "    start_point = np.array(start_point)[0]\n",
        "    broken_change = min_subset(start_point, adv_point, mod, random.choice(list(range(50))))\n",
        "    broken_changes.append(broken_change)"
      ],
      "id": "growing-steal",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": true,
        "id": "graduate-grounds",
        "outputId": "0485abb7-e281-4e4d-b6fa-66d2f86e2285"
      },
      "source": [
        "print(np.mean(broken_changes))\n",
        "print(np.std(broken_changes))"
      ],
      "id": "graduate-grounds",
      "execution_count": null,
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "44.1\n",
            "6.057227088363123\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "phantom-radar"
      },
      "source": [
        "# Opposite change"
      ],
      "id": "phantom-radar"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "intended-albany"
      },
      "source": [
        "with open('info_iters.p','rb') as F:\n",
        "    all_info = pickle.load(F)"
      ],
      "id": "intended-albany",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "specialized-yeast"
      },
      "source": [
        "def min_subset_reform(start, adv, mod, start_id = 0):\n",
        "    for i in range(start_id, start_id + 50):\n",
        "        index = i%50\n",
        "        adv[:index] = start[:index]\n",
        "        start_pred = predict(mod, start)\n",
        "        adv_pred = predict(mod, adv)\n",
        "        if start_pred == adv_pred:\n",
        "            break\n",
        "    return index"
      ],
      "id": "specialized-yeast",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "copyrighted-validity"
      },
      "source": [
        "broken_changes = []\n",
        "for i in range(50):\n",
        "    adv_point = all_info[0][1][DSIZE].in_dist_advs[i]\n",
        "    start_point = all_info[0][1][DSIZE].starts[i]\n",
        "    start_point = np.array(start_point)[0]\n",
        "    broken_change = min_subset_reform(start_point, adv_point, mod, random.choice(list(range(50))))\n",
        "    broken_changes.append(broken_change)"
      ],
      "id": "copyrighted-validity",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "scrolled": false,
        "id": "referenced-spice",
        "outputId": "0f191176-3a5e-4047-a72b-0d3459f36282"
      },
      "source": [
        "print(np.mean(broken_changes))\n",
        "print(np.std(broken_changes))"
      ],
      "id": "referenced-spice",
      "execution_count": null,
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "24.18\n",
            "13.122027282398097\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "narrative-walker"
      },
      "source": [
        "# Random Perturb"
      ],
      "id": "narrative-walker"
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "inclusive-softball"
      },
      "source": [
        "with open('info_iters.p','rb') as F:\n",
        "    all_info = pickle.load(F)"
      ],
      "id": "inclusive-softball",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "filled-stick"
      },
      "source": [
        "def min_subset_random(start, adv, mod, start_id = 0):\n",
        "    for i in range(start_id, start_id + 50):\n",
        "        index = i%50\n",
        "        start[index] = np.random.normal(loc=np.mean(start[index]),scale=np.std(start[index]))\n",
        "        start_pred = predict(mod, start)\n",
        "        adv_pred = predict(mod, adv)\n",
        "        if start_pred == adv_pred:\n",
        "            break\n",
        "    return index"
      ],
      "id": "filled-stick",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "burning-welsh"
      },
      "source": [
        "broken_changes = []\n",
        "for i in range(50):\n",
        "    adv_point = all_info[0][1][DSIZE].in_dist_advs[i]\n",
        "    start_point = all_info[0][1][DSIZE].starts[i]\n",
        "    start_point = np.array(start_point)[0]\n",
        "    broken_change = min_subset_random(start_point, adv_point, mod, random.choice(list(range(50))))\n",
        "    broken_changes.append(broken_change)"
      ],
      "id": "burning-welsh",
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "reverse-climate",
        "outputId": "8e1aaeb4-e54d-4923-84ce-f09ae4841047"
      },
      "source": [
        "print(np.mean(broken_changes))\n",
        "print(np.std(broken_changes))"
      ],
      "id": "reverse-climate",
      "execution_count": null,
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "28.98\n",
            "12.657788116412757\n"
          ]
        }
      ]
    }
  ]
}