barebones_tsvm_clean.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "barebones.ipynb",
      "provenance": [],
      "collapsed_sections": [],
      "authorship_tag": "ABX9TyNlMt2pK+5Z4hYQBREWe2Ax",
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/foxtrotmike/CS909/blob/master/barebones_tsvm_clean.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "TG-En3BVlBGG",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 335
        },
        "outputId": "7140142d-70bd-4174-fd6d-b5b3c4e8416a"
      },
      "source": [
        "\"\"\"\n",
        "Author: Fayyaz Minhas\n",
        "\"\"\"\n",
        "import numpy as np\n",
        "import matplotlib.pyplot as plt\n",
        "import torch\n",
        "import torch.nn.modules as nn\n",
        "import torch.optim as optim\n",
        "\n",
        "#Let's generate some data\n",
        "Xp = 1.0+np.random.randn(10,2); Xn = -1.0-np.random.randn(100,2);\n",
        "inputs = np.vstack((Xp,Xn))\t\n",
        "#labels\n",
        "targets = np.hstack((np.ones(len(Xp)),-np.ones(len(Xn))))\n",
        "ttargets = targets+0 #copy of true labels\n",
        "targets[0:]=np.nan; #setting labels to unknown\n",
        "targets[0]=1.0;targets[-1]=-1.0;#assume we only know two labels\n",
        "#kernel \n",
        "from sklearn.metrics.pairwise import rbf_kernel\n",
        "K = (np.dot(inputs,inputs.T)+1)**2\n",
        "K = rbf_kernel(inputs)\n",
        "K = torch.from_numpy(K).float() #kernel\n",
        "\n",
        "device = torch.device('cpu')\n",
        "#device = torch.device('cuda') # Uncomment this to run on GPU\n",
        "x = torch.from_numpy(inputs).float()\n",
        "y = torch.from_numpy(targets).float()\n",
        "N, D_in,D_out = x.shape[0], x.shape[1], 1\n",
        "# Loss function implementation\n",
        "def L1(z):\n",
        "    g = 10\n",
        "    zz = torch.zeros((2,len(z)))\n",
        "    zz[1] = 1-z    \n",
        "    closs = torch.logsumexp(g*zz,0)/g\n",
        "    closs = torch.max(torch.Tensor([0]) , 1 - z) \n",
        "    return closs\n",
        "def L2(z):\n",
        "    #g = 10\n",
        "    #zz = torch.zeros((2,len(z)))\n",
        "    #zz[1] = z\n",
        "    #az = (torch.logsumexp(-g*zz,0)+torch.logsumexp(g*zz,0))/g #smooth approximation of abs https://math.stackexchange.com/questions/728094/approximate-x-with-a-smooth-function\n",
        "    #az = torch.logsumexp(g*zz,0)*2/g-z-(2/g)*torch.log(torch.tensor(2)) #smooth approximation of abs \n",
        "    #zz[1] = 1-az    \n",
        "    #closs= torch.logsumexp(g*zz,0)/g #approx of max(0,1-abs(z))\n",
        "    #closs = torch.exp(-3*(z**2)) #original approximation used in the paper \n",
        "    #closs = torch.exp(-0.5*(z**2)) #original approximation used in the paper \"LARGE SCALE TRANSDUCTIVE SVMS\"  by Collobert 2006    \n",
        "    #closs= torch.max(torch.Tensor([0]) , 1 - torch.abs(z)) \n",
        "    s = -0.3\n",
        "    closs = torch.min(torch.tensor(1+s),torch.max(torch.tensor(0),1-torch.abs(z)))\n",
        "    return closs\n",
        "\n",
        "eps = 1e-6\n",
        "# Create random Tensors for weights; setting requires_grad=True means that we\n",
        "# want to compute gradients for these Tensors during the backward pass.\n",
        "wb = torch.randn(N, device=device, requires_grad=True)\n",
        "wb = torch.tensor(torch.empty(N).normal_(mean=0.1,std=0.01), requires_grad=True)\n",
        "# Parameters\n",
        "learning_rate = 1e-2\n",
        "lambda_w = 1e-1\n",
        "lambda_u = 5e-1\n",
        "lambda_b = 10e-1\n",
        "cratio = np.mean(ttargets)\n",
        "#Training \n",
        "optimizer = optim.SGD([wb], lr=learning_rate)\n",
        "L = [] #history of losses\n",
        "uidx, pidx, nidx = torch.isnan(y), y==1, y==-1\n",
        "n_unsupv, n_pos, n_neg = torch.sum(uidx), torch.sum(pidx), torch.sum(nidx)\n",
        "for epoch in range(100):\n",
        "      \"\"\"\n",
        "      # REPRESENTATION\n",
        "      \"\"\"\n",
        "      y_pred = wb@K\n",
        "      \"\"\"\n",
        "      # EVALUATION\n",
        "      \"\"\"      \n",
        "      loss_s, loss_u = 0.0,0.0\n",
        "      if n_pos:\n",
        "        loss_s = torch.mean(L1(y_pred[pidx]))\n",
        "      if n_neg:\n",
        "        loss_s+=torch.mean(L1(-1*y_pred[nidx]))\n",
        "      if n_unsupv:\n",
        "        loss_u = torch.mean(L2(y_pred[uidx]))/L2(torch.tensor([0]))\n",
        "      balance = (torch.mean(y_pred[~uidx])-cratio)**2\n",
        "      obj = loss_s+lambda_u*loss_u+lambda_w*wb@K@wb+lambda_b*balance #empirical loss + regularization     \n",
        "\n",
        "      \"\"\"\n",
        "      #OPTIMIZATION\n",
        "      \"\"\"    \n",
        "      obj.backward()\n",
        "      # Using built-in optimizer\n",
        "      optimizer.step()\n",
        "      optimizer.zero_grad()\n",
        "      L.append(balance.detach().numpy()) #save for history and plotting\n",
        "\n",
        "wbn = wb.detach().numpy()\n",
        "plt.close(\"all\")\n",
        "plt.plot(L);plt.yscale('log')\n",
        "\n",
        "plt.grid(); plt.xlabel('Epochs'); plt.ylabel('value');plt.legend(['Loss','Objective'])\n",
        "\"\"\"\n",
        "def clf(inputs): \n",
        "  return inputs@wb\n",
        "\n",
        "print(\"Predictions: \",clf(inputs))\n",
        "print(\"Weights: \",wbn)\n",
        "plt.figure()\n",
        "\n",
        "\n",
        "plotit(inputs,targets,clf=clf,conts=[0],extent = [-5,+5,-5,+5])\n",
        "\"\"\""
      ],
      "execution_count": 129,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "/usr/local/lib/python3.7/dist-packages/ipykernel_launcher.py:56: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n"
          ]
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "'\\ndef clf(inputs): \\n  return inputs@wb\\n\\nprint(\"Predictions: \",clf(inputs))\\nprint(\"Weights: \",wbn)\\nplt.figure()\\n\\n\\nplotit(inputs,targets,clf=clf,conts=[0],extent = [-5,+5,-5,+5])\\n'"
            ],
            "application/vnd.google.colaboratory.intrinsic+json": {
              "type": "string"
            }
          },
          "metadata": {},
          "execution_count": 129
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "z = (wb@K).detach().numpy()\n",
        "print(\"Accuracy\",np.mean(z*ttargets>0))\n",
        "from sklearn.metrics import roc_auc_score\n",
        "print(\"AUROC\",roc_auc_score(ttargets, z))"
      ],
      "metadata": {
        "id": "ERafpLRRK0Dm",
        "outputId": "ac8d0267-508a-49dd-a954-0ac143e2076c",
        "colab": {
          "base_uri": "https://localhost:8080/"
        }
      },
      "execution_count": 130,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Accuracy 0.6181818181818182\n",
            "AUROC 0.794\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "plt.hist(z);plt.title('Prediction Scores')"
      ],
      "metadata": {
        "id": "S5XtmMsjNVab",
        "outputId": "8d68fc09-72ec-44c2-c8c4-26d09e17a26e",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 298
        }
      },
      "execution_count": 131,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "Text(0.5, 1.0, 'Prediction Scores')"
            ]
          },
          "metadata": {},
          "execution_count": 131
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "plt.hist(wbn);plt.title('Sample Weighths')"
      ],
      "metadata": {
        "id": "hOgJslRuN2mb",
        "outputId": "952b3cb3-47c1-4efc-99e7-971c50b74029",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 298
        }
      },
      "execution_count": 132,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "Text(0.5, 1.0, 'Sample Weighths')"
            ]
          },
          "metadata": {},
          "execution_count": 132
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "y"
      ],
      "metadata": {
        "id": "UPkKJ1WCOW0K",
        "outputId": "abb535a6-af0c-40ed-e4c3-d74d869566ad",
        "colab": {
          "base_uri": "https://localhost:8080/"
        }
      },
      "execution_count": 133,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor([ 1., nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n",
              "        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n",
              "        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n",
              "        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n",
              "        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n",
              "        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n",
              "        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,\n",
              "        nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, -1.])"
            ]
          },
          "metadata": {},
          "execution_count": 133
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "ttargets"
      ],
      "metadata": {
        "id": "Sd-mQyAfbHSx",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "11f7abf8-0346-4c13-d626-054120d95391"
      },
      "execution_count": 134,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "array([ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1., -1., -1., -1.,\n",
              "       -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,\n",
              "       -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,\n",
              "       -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,\n",
              "       -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,\n",
              "       -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,\n",
              "       -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,\n",
              "       -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,\n",
              "       -1., -1., -1., -1., -1., -1.])"
            ]
          },
          "metadata": {},
          "execution_count": 134
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "torch.mean(y_pred[~uidx]),cratio"
      ],
      "metadata": {
        "id": "8g0YTvhjkSGp",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "428ff7db-6d24-4bc7-a3b0-9137b19330ec"
      },
      "execution_count": 135,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(tensor(-0.4070, grad_fn=<MeanBackward0>), -0.8181818181818182)"
            ]
          },
          "metadata": {},
          "execution_count": 135
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        ""
      ],
      "metadata": {
        "id": "9XfH2Z-Whwxv"
      },
      "execution_count": 135,
      "outputs": []
    }
  ]
}