notebooks/t-SNE-visualisations.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "t-SNE-visualisations.ipynb",
      "provenance": [],
      "collapsed_sections": [],
      "toc_visible": true,
      "authorship_tag": "ABX9TyPywDv6bWpnG4rPTfRXvvxD",
      "include_colab_link": true
    },
    "kernelspec": {
      "display_name": "Python 3",
      "name": "python3"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/marta-0/medical_mnist/blob/main/notebooks/t-SNE-visualisations.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "QVyj6CK445Sm"
      },
      "source": [
        "# t-SNE visualisations"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "qQqUpWd544dN",
        "outputId": "bc35031e-28a6-46ae-edbb-efbd95818b8e"
      },
      "source": [
        "!pip install lapjv --quiet\n",
        "!pip install rasterfairy-py3 --quiet"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "  Building wheel for lapjv (setup.py) ... \u001b[?25l\u001b[?25hdone\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "aKjlmKFm460W",
        "outputId": "f7b1bc3d-8164-42d3-b31e-4b21658ddaa9"
      },
      "source": [
        "import os\n",
        "import random\n",
        "import zipfile\n",
        "\n",
        "from PIL import Image, ImageSequence\n",
        "\n",
        "import numpy as np\n",
        "import pandas as pd\n",
        "import math\n",
        "import seaborn as sns\n",
        "import matplotlib.pyplot as plt\n",
        "from matplotlib.offsetbox import OffsetImage, AnnotationBbox\n",
        "\n",
        "from sklearn.manifold import TSNE\n",
        "\n",
        "import tensorflow as tf\n",
        "from tensorflow.keras.models import Model\n",
        "from tensorflow.python.keras.preprocessing import image\n",
        "\n",
        "from lapjv import lapjv\n",
        "from scipy.spatial.distance import cdist\n",
        "import IPython\n",
        "\n",
        "import rasterfairy\n",
        "import io\n",
        "from moviepy.editor import *\n",
        "\n",
        "random.seed(1)\n",
        "np.random.seed(1)\n",
        "tf.random.set_seed(1)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Imageio: 'ffmpeg-linux64-v3.3.1' was not found on your computer; downloading it now.\n",
            "Try 1. Download from https://github.com/imageio/imageio-binaries/raw/master/ffmpeg/ffmpeg-linux64-v3.3.1 (43.8 MB)\n",
            "Downloading: 8192/45929032 bytes (0.0%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b1753088/45929032 bytes (3.8%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b4907008/45929032 bytes (10.7%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b8208384/45929032 bytes (17.9%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b11460608/45929032 bytes (25.0%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b14745600/45929032 bytes (32.1%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b17973248/45929032 bytes (39.1%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b20815872/45929032 bytes (45.3%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b24043520/45929032 bytes (52.3%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b27312128/45929032 bytes (59.5%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b30425088/45929032 bytes (66.2%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b33898496/45929032 bytes (73.8%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b36978688/45929032 bytes (80.5%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b40280064/45929032 bytes (87.7%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b43507712/45929032 bytes (94.7%)\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b45929032/45929032 bytes (100.0%)\n",
            "  Done\n",
            "File saved as /root/.imageio/ffmpeg/ffmpeg-linux64-v3.3.1.\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "I7n3Ty3B5oBC"
      },
      "source": [
        "labels = ['AbdomenCT', 'BreastMRI', 'CXR', 'ChestCT', 'Hand', 'HeadCT']"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "nOXRvH5P5oC7"
      },
      "source": [
        "with zipfile.ZipFile('/content/models.zip','r') as zippedfile:\n",
        "  zippedfile.extractall()\n",
        "\n",
        "model_baseline = tf.keras.models.load_model('models/baseline')\n",
        "model_gmp = tf.keras.models.load_model('models/gmp')\n",
        "model_gmp_hflip_rot180 = tf.keras.models.load_model('models/gmp_hflip_rot180')\n",
        "model_hflip = tf.keras.models.load_model('models/hflip')\n",
        "model_rot180 = tf.keras.models.load_model('models/rotate180')\n",
        "model_d01 = tf.keras.models.load_model('models/dropout_01')\n",
        "model_d2_01 = tf.keras.models.load_model('models/dropout_2_01')\n",
        "model_best = tf.keras.models.load_model('models/gmp+hflip+rot180+bn128')\n",
        "model_best_var_sizes = tf.keras.models.load_model('models/best_various_sizes')\n",
        "model_gmp_var_sizes = tf.keras.models.load_model('models/gmp_various_sizes')\n",
        "model_gmp_hflip = tf.keras.models.load_model('models/gmp_hflip')\n",
        "model_baseline_occ = tf.keras.models.load_model('models/baseline_occ')\n",
        "model_best_occ = tf.keras.models.load_model('models/best_occ')\n",
        "model_hflip_occ = tf.keras.models.load_model('models/hflip_occ')\n",
        "model_d2_01_occ = tf.keras.models.load_model('models/dropout_2_01_occ')\n",
        "model_gmp_occ = tf.keras.models.load_model('models/gmp_occ')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "lApMomuw5oFF"
      },
      "source": [
        "with zipfile.ZipFile('data_24.zip','r') as zippedfile:\n",
        "  zippedfile.extractall()\n",
        "with zipfile.ZipFile('data_32_40.zip','r') as zippedfile:\n",
        "  zippedfile.extractall()\n",
        "with zipfile.ZipFile('data_48.zip','r') as zippedfile:\n",
        "  zippedfile.extractall()\n",
        "with zipfile.ZipFile('data_64.zip','r') as zippedfile:\n",
        "  zippedfile.extractall()\n",
        "\n",
        "X_test_24 = np.load('/content/data_24/X_test_24.npy')\n",
        "y_test_24 = np.load('/content/data_24/y_test_24.npy')\n",
        "X_test_32 = np.load('/content/data_32_40/X_test_32.npy')\n",
        "y_test_32 = np.load('/content/data_32_40/y_test_32.npy')\n",
        "X_test_40 = np.load('/content/data_32_40/X_test_40.npy')\n",
        "y_test_40 = np.load('/content/data_32_40/y_test_40.npy')\n",
        "X_test_48 = np.load('/content/data_48/X_test_48.npy')\n",
        "y_test_48 = np.load('/content/data_48/y_test_48.npy')\n",
        "X_test_64 = np.load('/content/data_64/X_test.npy')\n",
        "y_test_64 = np.load('/content/data_64/y_test.npy')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "LE2SmYYaIAu0"
      },
      "source": [
        "Function calculating t-SNE 2D coordinates:"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "zif1H2bu7eTD"
      },
      "source": [
        "def return_tsne(model, X_test, y_test, output_points, layer_name='global_max_pooling2d_1', low_res=False):\n",
        "  activations = []\n",
        "  \n",
        "  if low_res:\n",
        "    image = X_test\n",
        "\n",
        "    for l in model.layers:\n",
        "      if 'dense' in  l.name:\n",
        "        break\n",
        "\n",
        "      image = l(image)\n",
        "\n",
        "      if layer_name in l.name:\n",
        "        activations.append(image)\n",
        "\n",
        "    activations = activations[0]\n",
        "\n",
        "  else:\n",
        "    intermediate_layer_model = Model(inputs=model.input, outputs=model.get_layer(layer_name).output)\n",
        "\n",
        "    for i, img in enumerate(X_test):\n",
        "      if i == output_points:\n",
        "        break\n",
        "\n",
        "      img = np.expand_dims(img, axis=0)\n",
        "      activations.append(np.squeeze(intermediate_layer_model.predict(img)))\n",
        "\n",
        "  tsne_results = TSNE(random_state=1).fit_transform(np.array(activations)[0:output_points,:])\n",
        "\n",
        "  return tsne_results"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "_sI8H9CvIHRH"
      },
      "source": [
        "Function showing only plot legend:"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "HGixZU1o6apj"
      },
      "source": [
        "def get_legend(legend_colors, legend_labels):\n",
        "  plt.figure(figsize=(3.9,3.65))\n",
        "  f = lambda m,c: plt.plot([], [], marker=m, color=c, ls=\"none\")[0]\n",
        "  handles = [f(\"s\", legend_colors[i]) for i in range(len(legend_labels))]\n",
        "  legend = plt.legend(handles, legend_labels, loc=3, framealpha=1, frameon=True, fontsize=20, markerscale=3)\n",
        "\n",
        "  plt.gca().set_axis_off()\n",
        "  plt.show()"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "hHNXC2V36vYI"
      },
      "source": [
        "## Mosaics"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "2Gqkv-kX543d"
      },
      "source": [
        "def tsne_mosaic(tsne_results, X_test, y_test, output_rows, out_res=64, show_img=False, gray_scale=False, color_scale=False):\n",
        "  output_num_img = output_rows * output_rows\n",
        "  \n",
        "  tsne_results -= tsne_results.min(axis=0)\n",
        "  tsne_results /= tsne_results.max(axis=0)\n",
        "\n",
        "\n",
        "  grid = np.dstack(np.meshgrid(np.linspace(0, 1, output_rows), np.linspace(0, 1, output_rows))).reshape(-1, 2)\n",
        "  cost_matrix = cdist(grid, tsne_results, \"sqeuclidean\").astype(np.float32)\n",
        "  cost_matrix = cost_matrix * (100000 / cost_matrix.max())\n",
        "  row_asses, col_asses, _ = lapjv(cost_matrix)\n",
        "  grid_jv = grid[col_asses]\n",
        "  out = np.ones((output_rows*out_res, output_rows*out_res, 3))\n",
        "\n",
        "\n",
        "  i = 0\n",
        "  for pos, img in zip(grid_jv, X_test[0:output_num_img]):\n",
        "      h_range = int(round(pos[0] * (output_rows - 1) * out_res))\n",
        "      w_range = int(round(pos[1] * (output_rows - 1) * out_res))\n",
        "\n",
        "      if gray_scale:\n",
        "        _img = image.img_to_array(img)\n",
        "        _img[:,:] = np.argmax(y_test[i]) * 50\n",
        "        i = i + 1\n",
        "        scale = '_gray'\n",
        "        out = out.astype(np.uint8)\n",
        "\n",
        "      elif color_scale:\n",
        "        colors = [sns.color_palette('Set2', 6)[5],\n",
        "                  sns.color_palette('Set2', 6)[3],\n",
        "                  sns.color_palette('Set2', 6)[2],\n",
        "                  sns.color_palette('Set2', 6)[1],\n",
        "                  sns.color_palette('Set2', 6)[4],\n",
        "                  sns.color_palette('Set2', 6)[0]]\n",
        "\n",
        "        _img = np.full((out_res,out_res,3),colors[np.argmax(y_test[i])])\n",
        "        i = i + 1\n",
        "        scale = '_color'\n",
        "      \n",
        "      else:\n",
        "        scale = ''\n",
        "        _img = image.img_to_array(img)\n",
        "\n",
        "      out[h_range:h_range + out_res, w_range:w_range + out_res] = _img\n",
        "\n",
        "\n",
        "  filename = 'tsne_mosaic_' + str(output_rows) + scale + '.jpg'\n",
        "  out = image.array_to_img(out)\n",
        "  out.save(filename)\n",
        "\n",
        "  if show_img:\n",
        "    IPython.display.display(IPython.display.Image(data=filename, width=550, height=550))"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "jJz4RocrgsVB"
      },
      "source": [
        "20 x 20 images:"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "iz4Zi0_l5454"
      },
      "source": [
        "tsne_results_400 = return_tsne(model_best, X_test_64, y_test_64, 20*20)\n",
        "tsne_mosaic(tsne_results_400, X_test_64, y_test_64, 20, show_img=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "GRQ8JaqRguql"
      },
      "source": [
        "30 x 30 images:"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "LLLC7y-f548Q"
      },
      "source": [
        "tsne_results_900 = return_tsne(model_best, X_test_64, y_test_64, 30*30)\n",
        "tsne_mosaic(tsne_results_900, X_test_64, y_test_64, 30, show_img=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "zHNm51bZ54-m"
      },
      "source": [
        "tsne_mosaic(tsne_results_900, X_test_64, y_test_64, 30, show_img=True, color_scale=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "VfKrrTI_gwwF"
      },
      "source": [
        "40 x 40 images:"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "qGnKvvkb6N4n"
      },
      "source": [
        "tsne_results_1600 = return_tsne(model_best, X_test_64, y_test_64, 40*40)\n",
        "tsne_mosaic(tsne_results_1600, X_test_64, y_test_64, 40, show_img=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "LfdgyPHygy4H"
      },
      "source": [
        "132 x 132 images (test set contains 17687 images; square root is 132 rounding down - so that there are no empty spaces):"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "R3M6BJC155BI"
      },
      "source": [
        "tsne_results_17424 = return_tsne(model_best, X_test_64, y_test_64, 132*132)\n",
        "tsne_mosaic(tsne_results_17424, X_test_64, y_test_64, 132, show_img=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "o-g6CK7G6R69"
      },
      "source": [
        "tsne_mosaic(tsne_results_17424, X_test_64, y_test_64, 132, show_img=True, color_scale=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 229
        },
        "id": "QOstf0a76R8_",
        "outputId": "6cc0ccb2-0d5e-4622-ffe0-b3ef9be4ac6e"
      },
      "source": [
        "new_legend_colors = [sns.color_palette('Set2', 6)[5],\n",
        "                     sns.color_palette('Set2', 6)[3],\n",
        "                     sns.color_palette('Set2', 6)[2],\n",
        "                     sns.color_palette('Set2', 6)[1],\n",
        "                     sns.color_palette('Set2', 6)[4],\n",
        "                     sns.color_palette('Set2', 6)[0]]\n",
        "\n",
        "get_legend(new_legend_colors, labels)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 280.8x262.8 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "UhMT7mJc6R_W"
      },
      "source": [
        "tsne_mosaic(tsne_results_17424, X_test_64, y_test_64, 132, show_img=True, gray_scale=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 229
        },
        "id": "iuDqnodQ6SBv",
        "outputId": "0792e372-94fa-40b0-f76d-e61a758f95d5"
      },
      "source": [
        "get_legend([(i*50/255.0, i*50/255.0, i*50/255.0, 1.0) for i in range(len(labels))], labels)"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 280.8x262.8 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "qqefU8FaILFg"
      },
      "source": [
        "Function blending two mosaics:"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "eq98zKlu6SHF"
      },
      "source": [
        "def blend_mosaics(img1, img2, filename, alpha=0.3):\n",
        "  res = Image.blend(img1, img2, alpha)\n",
        "\n",
        "  res.save(filename)\n",
        "  IPython.display.display(IPython.display.Image(data=filename, width=550, height=550))"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "U08aUF7k6jLw"
      },
      "source": [
        "blend_mosaics(Image.open('/content/tsne_mosaic_132.jpg'), Image.open('/content/tsne_mosaic_132_color.jpg'), 'tsne_mosaics_added_132.jpg')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "D091uVSW6jNx"
      },
      "source": [
        "blend_mosaics(Image.open('/content/tsne_mosaic_30.jpg'), Image.open('/content/tsne_mosaic_30_color.jpg'), 'tsne_mosaics_added_30.jpg')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "dGKBJp71BR7I"
      },
      "source": [
        "## Scatter plots"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "c9mnHJlF6jUL"
      },
      "source": [
        "def tsne_scatter(tsne_results, X_test, y_test, title, loc=4, as_dots=True):\n",
        "  tsne_results -= tsne_results.min(axis=0)\n",
        "  tsne_results /= tsne_results.max(axis=0)\n",
        "\n",
        "  y_labels = list(map(lambda x: labels[x], np.argmax(y_test, axis=1)))\n",
        "\n",
        "\n",
        "  if as_dots:\n",
        "    df = pd.DataFrame(y_labels, columns=['label'])\n",
        "    df['tsne-2d-one'] = tsne_results[:,0]\n",
        "    df['tsne-2d-two'] = tsne_results[:,1]\n",
        "\n",
        "    plt.figure(figsize=(16,10))\n",
        "    ax = sns.scatterplot(x='tsne-2d-one', y='tsne-2d-two', hue='label', palette=sns.color_palette('Set2', 6), data=df, legend='full', alpha=0.3)\n",
        "    plt.legend()\n",
        "    handles, legend_labels = ax.get_legend_handles_labels()\n",
        "    legend_labels, handles = zip(*sorted(zip(legend_labels, handles), key=lambda t: t[0]))\n",
        "    ax.legend(handles=handles, labels=legend_labels, loc=loc, fontsize=15)\n",
        "\n",
        "\n",
        "  else:\n",
        "    fig, ax = plt.subplots(figsize=(20,15))\n",
        "    artists = []\n",
        "    size = X_test.shape[1]\n",
        "    index = 0\n",
        "\n",
        "    for xy, i in zip(tsne_results, np.squeeze(X_test)):\n",
        "      \n",
        "      img_ = np.dstack((i, i, i))\n",
        "\n",
        "      color_index = np.argmax(y_test[index])\n",
        "      if np.argmax(y_test[index]) == 0:\n",
        "        color_index = 5\n",
        "      elif np.argmax(y_test[index]) == 1:\n",
        "        color_index = 3\n",
        "      elif np.argmax(y_test[index]) == 3:\n",
        "        color_index = 1 \n",
        "      elif np.argmax(y_test[index]) == 5:\n",
        "        color_index = 0\n",
        "\n",
        "      for w in range(size):\n",
        "        for h in range(size):\n",
        "          if (h < 1) or (h > size-3):\n",
        "            img_[h:h+2,:,0] = sns.color_palette('Set2',6)[color_index][0]\n",
        "            img_[h:h+2,:,1] = sns.color_palette('Set2',6)[color_index][1]\n",
        "            img_[h:h+2,:,2] = sns.color_palette('Set2',6)[color_index][2]\n",
        "\n",
        "          if (w < 1) or (w > size-3):\n",
        "            img_[:,w:w+2,0] = sns.color_palette('Set2',6)[color_index][0]\n",
        "            img_[:,w:w+2,1] = sns.color_palette('Set2',6)[color_index][1]\n",
        "            img_[:,w:w+2,2] = sns.color_palette('Set2',6)[color_index][2]\n",
        "\n",
        "      x0, y0 = xy\n",
        "      img = OffsetImage(img_)\n",
        "      ab = AnnotationBbox(img, (x0, y0), xycoords='data', frameon=False)\n",
        "      artists.append(ax.add_artist(ab))\n",
        "      index = index + 1\n",
        "\n",
        "    ax.update_datalim(tsne_results)\n",
        "    ax.autoscale()\n",
        "\n",
        "\n",
        "  plt.title(title, fontsize=25)\n",
        "  plt.xticks(fontsize=20)\n",
        "  plt.yticks(fontsize=20)\n",
        "  plt.xlabel('tsne-2d-one', fontsize=20)\n",
        "  plt.ylabel('tsne-2d-two', fontsize=20)\n",
        "  plt.show()"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "m8TR1mww6jWe"
      },
      "source": [
        "tsne_results_best = return_tsne(model_best, X_test_64, y_test_64, X_test_64.shape[0])\n",
        "tsne_scatter(tsne_results_best, X_test_64, y_test_64, 'Best model', loc=2)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "vLPA-TR36jYs"
      },
      "source": [
        "tsne_results_gmp = return_tsne(model_gmp, X_test_64, y_test_64, X_test_64.shape[0])\n",
        "tsne_scatter(tsne_results_gmp, X_test_64, y_test_64, 'Global Max Pooling', loc=2)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ElWsQShNBiBn"
      },
      "source": [
        "tsne_results_gmp_hflip = return_tsne(model_gmp_hflip, X_test_64, y_test_64, X_test_64.shape[0], layer_name='global_max_pooling2d_2')\n",
        "tsne_scatter(tsne_results_gmp_hflip, X_test_64, y_test_64, 'Global Max Pooling + Horizontal flip', loc=2)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "I-oY1W9jBiED"
      },
      "source": [
        "tsne_results_gmp_hflip_rot180 = return_tsne(model_gmp_hflip_rot180, X_test_64, y_test_64, X_test_64.shape[0])\n",
        "tsne_scatter(tsne_results_gmp_hflip_rot180, X_test_64, y_test_64, 'Global Max Pooling + Horizontal flip + Rotation <= 180°', loc=2)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "WjwOPlovBiGd"
      },
      "source": [
        "tsne_results_var_sizes = return_tsne(model_best_var_sizes, X_test_64, y_test_64, X_test_64.shape[0], layer_name='global_max_pooling2d_2')\n",
        "tsne_scatter(tsne_results_var_sizes, X_test_64, y_test_64, 'Best model - trained on different size images generator', loc=2)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "LlkLTzn3BiIq"
      },
      "source": [
        "tsne_results_gmp_var_sizes = return_tsne(model_gmp_var_sizes, X_test_64, y_test_64, X_test_64.shape[0])\n",
        "tsne_scatter(tsne_results_gmp_var_sizes, X_test_64, y_test_64, 'Global Max Pooling - trained on different size images generator', loc=2)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "PO4b4Sq7B8kK"
      },
      "source": [
        "## Scatter plots - smaller input images"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Ze-KKhBJBiKT"
      },
      "source": [
        "tsne_results_best_24 = return_tsne(model_best, X_test_24, y_test_24, X_test_24.shape[0], low_res=True)\n",
        "tsne_scatter(tsne_results_best_24, X_test_24, y_test_24, 'Best model - image size 24 x 24 px')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ZwQipI-6Bqfr"
      },
      "source": [
        "tsne_results_best_32 = return_tsne(model_best, X_test_32, y_test_32, X_test_32.shape[0], low_res=True)\n",
        "tsne_scatter(tsne_results_best_32, X_test_32, y_test_32, 'Best model - image size 32 x 32 px')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "5nVtop1IBqhd"
      },
      "source": [
        "tsne_results_best_40 = return_tsne(model_best, X_test_40, y_test_40, X_test_40.shape[0], low_res=True)\n",
        "tsne_scatter(tsne_results_best_40, X_test_40, y_test_40, 'Best model - image size 40 x 40 px')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "DFBxU3R4Bqjr"
      },
      "source": [
        "tsne_results_best_48 = return_tsne(model_best, X_test_48, y_test_48, X_test_48.shape[0], low_res=True)\n",
        "tsne_scatter(tsne_results_best_48, X_test_48, y_test_48, 'Best model - image size 48 x 48 px')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "GC4v_YUOBql8"
      },
      "source": [
        "tsne_results_best_var_24 = return_tsne(model_best_var_sizes, X_test_24, y_test_24, X_test_24.shape[0], layer_name='global_max_pooling2d_2', low_res=True)\n",
        "tsne_scatter(tsne_results_best_var_24, X_test_24, y_test_24, 'Best model - trained on different size images generator - image size 24 x 24 px')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "BPZNxJiVBqoT"
      },
      "source": [
        "tsne_results_best_var_32 = return_tsne(model_best_var_sizes, X_test_32, y_test_32, X_test_32.shape[0], layer_name='global_max_pooling2d_2', low_res=True)\n",
        "tsne_scatter(tsne_results_best_var_32, X_test_32, y_test_32, 'Best model - trained on different size images generator - image size 32 x 32 px')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "7LbrnRjDBz7K"
      },
      "source": [
        "tsne_results_best_var_40 = return_tsne(model_best_var_sizes, X_test_40, y_test_40, X_test_40.shape[0], layer_name='global_max_pooling2d_2', low_res=True)\n",
        "tsne_scatter(tsne_results_best_var_40, X_test_40, y_test_40, 'Best model - trained on different size images generator - image size 40 x 40 px')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "KqVXiVxEBz9n"
      },
      "source": [
        "tsne_results_best_var_48 = return_tsne(model_best_var_sizes, X_test_48, y_test_48, X_test_48.shape[0], layer_name='global_max_pooling2d_2', low_res=True)\n",
        "tsne_scatter(tsne_results_best_var_48, X_test_48, y_test_48, 'Best model - trained on different size images generator- image size 48 x 48 px')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "SylLzhx6B3od"
      },
      "source": [
        "## Scatter plots with images"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "vPclPwxRBz_5"
      },
      "source": [
        "tsne_scatter(tsne_results_best, X_test_64, y_test_64, 'Best model', as_dots=False)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "lJAqzvkUB0CR"
      },
      "source": [
        "tsne_scatter(tsne_results_best_24, X_test_24, y_test_24, 'Best model - image size 24 x 24 px', as_dots=False)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "YdODT4zvB0EP"
      },
      "source": [
        "tsne_scatter(tsne_results_gmp, X_test_64, y_test_64, 'Global Max Pooling', as_dots=False)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "j4QlQyTGB0GB"
      },
      "source": [
        "tsne_results_gmp_24 = return_tsne(model_gmp, X_test_24, y_test_24, X_test_24.shape[0], low_res=True)\n",
        "tsne_scatter(tsne_results_gmp_24, X_test_24, y_test_24, 'Global Max Pooling - image size 24 x 24 px', as_dots=False)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "jaTDdZ_vXXB_"
      },
      "source": [
        "## RasterFairy"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "i6z4o2YlB0Hw"
      },
      "source": [
        "def raster_fairy(tsne_results, X_test, y_test, title, rect=True, circle=False, as_img=False, target_grid=None, return_params=False):\n",
        "  if rect:\n",
        "    grid_xy, (width, height) = rasterfairy.transformPointCloud2D(tsne_results)\n",
        "\n",
        "  elif circle:\n",
        "    radius, adjustmentFactor, count = rasterfairy.getBestCircularMatch(X_test.shape[0])\n",
        "    print('The smallest circle that can hold', X_test.shape[0], 'has a radius of', radius, 'and will fit', count, 'points')\n",
        "\n",
        "    arrangement = rasterfairy.getCircularArrangement(radius, adjustmentFactor)\n",
        "    rasterMask = rasterfairy.arrangementToRasterMask(arrangement)\n",
        "\n",
        "    grid_xy, (width, height) = rasterfairy.transformPointCloud2D(tsne_results, target=rasterMask, autoAdjustCount = False)\n",
        "\n",
        "  else:\n",
        "      rasterMask = rasterfairy.getRasterMaskFromImage(target_grid)\n",
        "      print(rasterMask['count'], 'black pixels in map')\n",
        "\n",
        "      grid_xy, (width, height) = rasterfairy.transformPointCloud2D(tsne_results, target=rasterMask, autoAdjustCount=True)\n",
        "\n",
        "      dim = max(width, height)\n",
        "      grid_xy[:,0] += 0.5*(dim-width)\n",
        "      grid_xy[:,1] += 0.5*(dim-height)\n",
        "\n",
        "\n",
        "  if as_img:\n",
        "    fig, ax = plt.subplots(figsize=(20,15))\n",
        "    artists = []\n",
        "    size = X_test.shape[1]\n",
        "    index = 0\n",
        "\n",
        "    for xy, i in zip(grid_xy, np.squeeze(X_test)):\n",
        "      img_ = np.dstack((i, i, i))\n",
        "\n",
        "      color_index = np.argmax(y_test[index])\n",
        "      if np.argmax(y_test[index]) == 0:\n",
        "        color_index = 5\n",
        "      elif np.argmax(y_test[index]) == 1:\n",
        "        color_index = 3\n",
        "      elif np.argmax(y_test[index]) == 3:\n",
        "        color_index = 1 \n",
        "      elif np.argmax(y_test[index]) == 5:\n",
        "        color_index = 0\n",
        "\n",
        "      for w in range(size):\n",
        "        for h in range(size):\n",
        "          if (h < 1) or (h > size-3):\n",
        "            img_[h:h+2,:,0] = sns.color_palette('Set2',6)[color_index][0]\n",
        "            img_[h:h+2,:,1] = sns.color_palette('Set2',6)[color_index][1]\n",
        "            img_[h:h+2,:,2] = sns.color_palette('Set2',6)[color_index][2]\n",
        "\n",
        "          if (w < 1) or (w > size-3):\n",
        "            img_[:,w:w+2,0] = sns.color_palette('Set2',6)[color_index][0]\n",
        "            img_[:,w:w+2,1] = sns.color_palette('Set2',6)[color_index][1]\n",
        "            img_[:,w:w+2,2] = sns.color_palette('Set2',6)[color_index][2]\n",
        "        \n",
        "      x0, y0 = xy\n",
        "      img = OffsetImage(img_)\n",
        "      ab = AnnotationBbox(img, (x0, y0), xycoords='data', frameon=False)\n",
        "      artists.append(ax.add_artist(ab))\n",
        "      index = index + 1\n",
        "\n",
        "    ax.update_datalim(grid_xy)\n",
        "    ax.autoscale()\n",
        "\n",
        "\n",
        "  else:\n",
        "    y_labels = list(map(lambda x: labels[x], np.argmax(y_test, axis=1)))\n",
        "    df = pd.DataFrame(y_labels, columns=['label'])\n",
        "    df['grid-xy-one'] = grid_xy[:,0]\n",
        "    df['grid-xy-two'] = grid_xy[:,1]\n",
        "\n",
        "    fig = plt.figure(figsize=(10,10))\n",
        "    ax = sns.scatterplot(x='grid-xy-one', y='grid-xy-two', hue='label', palette=sns.color_palette('Set2', 6), data=df, legend=False)\n",
        "\n",
        "\n",
        "  ax.spines['top'].set_visible(False)\n",
        "  ax.spines['right'].set_visible(False)\n",
        "  ax.spines['bottom'].set_visible(False)\n",
        "  ax.spines['left'].set_visible(False)\n",
        "  ax.axes.get_xaxis().set_visible(False)\n",
        "  ax.axes.get_yaxis().set_visible(False)\n",
        "  ax.autoscale_view(True,True,True)\n",
        "  ax.invert_yaxis()\n",
        "  plt.title(title, fontsize=20)\n",
        "  plt.show()\n",
        "\n",
        "  if return_params:\n",
        "    return grid_xy, width, height"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "8jGUlXCIB0KB"
      },
      "source": [
        "tsne_results_1_20th = return_tsne(model_best, X_test_64[::20], y_test_64[::20], X_test_64[::20].shape[0])\n",
        "grid_xy_rect, width_rect, height_rect = raster_fairy(tsne_results_1_20th, X_test_64[::20], y_test_64[::20], 'Best model', return_params=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "W1P2HjnYXhtF"
      },
      "source": [
        "tsne_results_1_20th_32 = return_tsne(model_best, X_test_32[::20], y_test_32[::20], X_test_32[::20].shape[0], low_res=True)\n",
        "raster_fairy(tsne_results_1_20th_32, X_test_32[::20], y_test_32[::20], 'Best model - image size 32 x 32 px')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "_hvpZYCghp1h"
      },
      "source": [
        "Plotting with 32 x 32 px images, so they do not overlap on the plot:"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "bsZsv0gvXhv6"
      },
      "source": [
        "raster_fairy(tsne_results_1_20th_32, X_test_32[::20], y_test_32[::20], 'Best model - image size 32 x 32 px', as_img=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "6uAsVkPTXhyw"
      },
      "source": [
        "grid_xy_circle, width_circle, height_circle = raster_fairy(tsne_results_1_20th, X_test_64[::20], y_test_64[::20], 'Best model', rect=False, circle=True, return_params=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "xY86xArSXh1R"
      },
      "source": [
        "raster_fairy(tsne_results_1_20th_32, X_test_32[::20], y_test_32[::20], 'Best model - image size 32 x 32 px', rect=False, circle=True, as_img=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "5t-E5XB3hZBM"
      },
      "source": [
        "Using custom binary mask:"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "BQs_fiLRXh3V"
      },
      "source": [
        "target_tree = Image.open('/content/christmas_tree.jpg')\n",
        "target_tree"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "XwJFJJ8tZ5rf"
      },
      "source": [
        "raster_fairy(tsne_results_1_20th, X_test_64[::20], y_test_64[::20], 'Best model', rect=False, circle=False, target_grid=target_tree)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "-LvasKEfZ5uG"
      },
      "source": [
        "raster_fairy(tsne_results_1_20th_32, X_test_32[::20], y_test_32[::20], 'Best model - image size 32 x 32 px', rect=False, circle=False, as_img=True, target_grid=target_tree)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "mZtPDUGSZ5w_"
      },
      "source": [
        "target_hand = Image.open('/content/hand_mask.jpeg')\n",
        "target_hand"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "_0anYH2mZ5zt"
      },
      "source": [
        "grid_xy_hand, width_hand, height_hand = raster_fairy(tsne_results_1_20th, X_test_64[::20], y_test_64[::20], 'Best model', rect=False, circle=False, target_grid=target_hand, return_params=True)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "h3jrgNC5Z52N"
      },
      "source": [
        "raster_fairy(tsne_results_1_20th_32, X_test_32[::20], y_test_32[::20], 'Best model - image size 32 x 32 px', rect=False, circle=False, as_img=True, target_grid=target_hand)"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "csUbWKgiav0P"
      },
      "source": [
        "## RasterFairy - animation"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "lT7BV-iPioJT"
      },
      "source": [
        "Images sequence may be saved as gif directly by the function (as_gif=True) or the images may be saved as jpgs and processed later externally (eg. using ffmpeg)."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "v6P6wzbzbb0S"
      },
      "source": [
        "def raster_fairy_gif(tsne_results, grid_xy, width, height, X_test, y_test, name, as_gif=True):\n",
        "  images = []\n",
        "  pointCount = len(X_test)\n",
        "\n",
        "  px_from = grid_xy[:,0] - min(grid_xy[:,0])\n",
        "  py_from = grid_xy[:,1] - min(grid_xy[:,1])\n",
        "  px_from /= max(px_from)\n",
        "  py_from /= max(py_from)\n",
        "\n",
        "  px_to = tsne_results[:,0] - min(tsne_results[:,0])\n",
        "  py_to = tsne_results[:,1] - min(tsne_results[:,1])\n",
        "  px_to /= max(px_to)\n",
        "  py_to /= max(py_to)\n",
        "\n",
        "  if (width > height):\n",
        "      py_to = py_to * float(height) / float(width) + 0.5 * (1.0 - float(height) / float(width))\n",
        "  else:\n",
        "      px_to = px_to * float(width) / float(height) + 0.5 * (1.0 - float(width) / float(height))\n",
        "\n",
        "  y_labels = list(map(lambda x: labels[x], np.argmax(y_test, axis=1)))\n",
        "\n",
        "  steps = 50\n",
        "  for i in range(steps*2):\n",
        "      dx = []\n",
        "      dy = []\n",
        "      t = 1.0- abs(1.0 * i / (steps-1) - 1.0)\n",
        "\n",
        "      for j in range(pointCount):\n",
        "          dx.append((1.0-t)*px_from[j] + t*px_to[j])\n",
        "          dy.append((1.0-t)*py_from[j] + t*py_to[j])\n",
        "      \n",
        "      df = pd.DataFrame(y_labels, columns=['label'])\n",
        "      df['dx'] = dx\n",
        "      df['dy'] = dy\n",
        "      fig2 = plt.figure(figsize=(10,10))\n",
        "      ax = sns.scatterplot(x='dx', y='dy', hue='label', palette=sns.color_palette('Set2', 6), data=df, legend=False)\n",
        "      ax.spines['top'].set_visible(False)\n",
        "      ax.spines['right'].set_visible(False)\n",
        "      ax.spines['bottom'].set_visible(False)\n",
        "      ax.spines['left'].set_visible(False)\n",
        "      ax.axes.get_xaxis().set_visible(False)\n",
        "      ax.axes.get_yaxis().set_visible(False)\n",
        "      ax.axes.set_xlim([-0.01, 1.01])\n",
        "      ax.axes.set_ylim([-0.01, 1.01])\n",
        "      ax.invert_yaxis()\n",
        "      IPython.display.clear_output(wait=True)\n",
        "      \n",
        "      buf = io.BytesIO()\n",
        "      plt.savefig(buf, format='jpg')\n",
        "      buf.seek(0)\n",
        "\n",
        "      if as_gif:\n",
        "        images.append(np.array(Image.open(buf)))\n",
        "      else:\n",
        "        (Image.open(buf)).save(name + '/frame%04d.jpg' % i)\n",
        "      \n",
        "  if as_gif:\n",
        "    clip = ImageSequenceClip(images,fps=25)\n",
        "    clip.write_gif(name + '.gif')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "background_save": true
        },
        "id": "tIMzE2Iyb0-_"
      },
      "source": [
        "raster_fairy_gif(tsne_results_1_20th, grid_xy_rect, width_rect, height_rect, X_test_64[::20], y_test_64[::20], 'rectangle')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "background_save": true
        },
        "id": "_BFQ4pXzb1Bt"
      },
      "source": [
        "with open('/content/rectangle.gif','rb') as f:\n",
        "  IPython.display.display(IPython.display.Image(data=f.read(), format='png'))"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "96s65-esb1Es"
      },
      "source": [
        "raster_fairy_gif(tsne_results_1_20th, grid_xy_circle, width_circle, height_circle, X_test_64[::20], y_test_64[::20], 'circle')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "colab": {
          "background_save": true
        },
        "id": "KUlf9Dz1b1HH"
      },
      "source": [
        "with open('/content/circle.gif','rb') as f:\n",
        "  IPython.display.display(IPython.display.Image(data=f.read(), format='png'))"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "gYQONUypb1JS"
      },
      "source": [
        "raster_fairy_gif(tsne_results_1_20th, grid_xy_hand, width_hand, height_hand, X_test_64[::20], y_test_64[::20], 'hand')"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "i21QD7twcUBG"
      },
      "source": [
        "with open('/content/hand.gif','rb') as f:\n",
        "  IPython.display.display(IPython.display.Image(data=f.read(), format='png'))"
      ],
      "execution_count": null,
      "outputs": []
    }
  ]
}