Compression_Factor.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "Compression_Factor.ipynb",
      "provenance": [],
      "collapsed_sections": [],
      "toc_visible": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "U0cwnJcQYXMu",
        "colab_type": "text"
      },
      "source": [
        "## <center> Compression Factor </center>\n",
        "\n",
        "This notebook contains a small experimental validation of the role played by Compression factor. We create a toy dataset based on sampling random points from a unit circle split into two classes. Based on the classes, we compress our embeddings such that samples from the same class appear to be very close in the original space, and samples from dissimilar classes appear farther apart in the orginal space. \n",
        "\n",
        "<center>\n",
        "<img src=\"https://drive.google.com/uc?id=1-btjQParAQciNJBlqp90LgLJgd1isGUo\" alt=\"CompressionFactor\"\n",
        "\ttitle=\"Compression Factor\" width=\"480\" />\n",
        "</center>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "46fVytsWdkfw",
        "colab_type": "text"
      },
      "source": [
        "## Part 1: Installations\n",
        "\n",
        "Install required libraries in a new virtual environment.<br>\n",
        "Skip this step if you have already created a virtual environment for LEt-SNE."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "TeltQjZ9YVMQ",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 1000
        },
        "outputId": "b5fc350c-5ddf-405f-82f5-79c6a63a9135"
      },
      "source": [
        "# If using with Google Colab, restart the runtime if prompted\n",
        "# Resetting the runtime ensures correct version of a library is selected\n",
        "!sudo apt-get update\n",
        "!sudo apt install python3-venv libnvidia-common-430 unrar               # Needed if using Google Colab\n",
        "!python3 -m venv LEt-SNE_environment                            # Create a new environment and activate it \n",
        "!source LEt-SNE_environment/bin/activate; pip install -r LEt-SNE_requirements.txt"
      ],
      "execution_count": 1,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "\r0% [Working]\r            \rGet:1 https://cloud.r-project.org/bin/linux/ubuntu bionic-cran35/ InRelease [3,626 B]\n",
            "\r0% [Connecting to archive.ubuntu.com] [Connecting to security.ubuntu.com (91.18\r0% [Connecting to archive.ubuntu.com] [Connecting to security.ubuntu.com (91.18\r0% [1 InRelease gpgv 3,626 B] [Connecting to archive.ubuntu.com] [Connecting to\r0% [Connecting to archive.ubuntu.com] [Waiting for headers] [Waiting for header\r                                                                               \rIgn:2 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64  InRelease\n",
            "Get:3 https://cloud.r-project.org/bin/linux/ubuntu bionic-cran35/ Packages [88.1 kB]\n",
            "Get:4 http://security.ubuntu.com/ubuntu bionic-security InRelease [88.7 kB]\n",
            "Ign:5 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  InRelease\n",
            "Hit:6 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64  Release\n",
            "Hit:7 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  Release\n",
            "Get:8 http://ppa.launchpad.net/graphics-drivers/ppa/ubuntu bionic InRelease [21.3 kB]\n",
            "Hit:11 http://archive.ubuntu.com/ubuntu bionic InRelease\n",
            "Get:12 http://archive.ubuntu.com/ubuntu bionic-updates InRelease [88.7 kB]\n",
            "Get:13 http://security.ubuntu.com/ubuntu bionic-security/restricted amd64 Packages [44.6 kB]\n",
            "Get:14 http://ppa.launchpad.net/marutter/c2d4u3.5/ubuntu bionic InRelease [15.4 kB]\n",
            "Get:15 http://security.ubuntu.com/ubuntu bionic-security/multiverse amd64 Packages [8,213 B]\n",
            "Get:16 http://security.ubuntu.com/ubuntu bionic-security/main amd64 Packages [889 kB]\n",
            "Get:17 http://ppa.launchpad.net/graphics-drivers/ppa/ubuntu bionic/main amd64 Packages [37.4 kB]\n",
            "Get:18 http://archive.ubuntu.com/ubuntu bionic-backports InRelease [74.6 kB]\n",
            "Get:19 http://security.ubuntu.com/ubuntu bionic-security/universe amd64 Packages [839 kB]\n",
            "Get:20 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 Packages [1,184 kB]\n",
            "Get:21 http://ppa.launchpad.net/marutter/c2d4u3.5/ubuntu bionic/main Sources [1,811 kB]\n",
            "Get:22 http://archive.ubuntu.com/ubuntu bionic-updates/restricted amd64 Packages [59.0 kB]\n",
            "Get:23 http://archive.ubuntu.com/ubuntu bionic-updates/multiverse amd64 Packages [12.6 kB]\n",
            "Get:24 http://archive.ubuntu.com/ubuntu bionic-updates/universe amd64 Packages [1,372 kB]\n",
            "Get:25 http://archive.ubuntu.com/ubuntu bionic-backports/universe amd64 Packages [7,671 B]\n",
            "Get:26 http://archive.ubuntu.com/ubuntu bionic-backports/main amd64 Packages [8,286 B]\n",
            "Get:27 http://ppa.launchpad.net/marutter/c2d4u3.5/ubuntu bionic/main amd64 Packages [874 kB]\n",
            "Fetched 7,528 kB in 7s (1,100 kB/s)\n",
            "Reading package lists... Done\n",
            "Reading package lists... Done\n",
            "Building dependency tree       \n",
            "Reading state information... Done\n",
            "unrar is already the newest version (1:5.5.8-1).\n",
            "The following additional packages will be installed:\n",
            "  libpython3.6 libpython3.6-dev libpython3.6-minimal libpython3.6-stdlib\n",
            "  python-pip-whl python3.6 python3.6-dev python3.6-minimal python3.6-venv\n",
            "Suggested packages:\n",
            "  python3.6-doc binfmt-support\n",
            "The following NEW packages will be installed:\n",
            "  libnvidia-common-430 python-pip-whl python3-venv python3.6-venv\n",
            "The following packages will be upgraded:\n",
            "  libpython3.6 libpython3.6-dev libpython3.6-minimal libpython3.6-stdlib\n",
            "  python3.6 python3.6-dev python3.6-minimal\n",
            "7 upgraded, 4 newly installed, 0 to remove and 84 not upgraded.\n",
            "Need to get 52.5 MB of archives.\n",
            "After this operation, 1,975 kB of additional disk space will be used.\n",
            "Get:1 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 python3.6-dev amd64 3.6.9-1~18.04ubuntu1 [508 kB]\n",
            "Get:2 http://ppa.launchpad.net/graphics-drivers/ppa/ubuntu bionic/main amd64 libnvidia-common-430 all 440.59-0ubuntu0.18.04.1 [6,048 B]\n",
            "Get:3 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 libpython3.6-dev amd64 3.6.9-1~18.04ubuntu1 [44.9 MB]\n",
            "Get:4 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 libpython3.6 amd64 3.6.9-1~18.04ubuntu1 [1,414 kB]\n",
            "Get:5 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 python3.6 amd64 3.6.9-1~18.04ubuntu1 [203 kB]\n",
            "Get:6 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 libpython3.6-stdlib amd64 3.6.9-1~18.04ubuntu1 [1,710 kB]\n",
            "Get:7 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 python3.6-minimal amd64 3.6.9-1~18.04ubuntu1 [1,609 kB]\n",
            "Get:8 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 libpython3.6-minimal amd64 3.6.9-1~18.04ubuntu1 [533 kB]\n",
            "Get:9 http://archive.ubuntu.com/ubuntu bionic-updates/universe amd64 python-pip-whl all 9.0.1-2.3~ubuntu1.18.04.1 [1,653 kB]\n",
            "Get:10 http://archive.ubuntu.com/ubuntu bionic-updates/universe amd64 python3.6-venv amd64 3.6.9-1~18.04ubuntu1 [6,188 B]\n",
            "Get:11 http://archive.ubuntu.com/ubuntu bionic-updates/universe amd64 python3-venv amd64 3.6.7-1~18.04 [1,208 B]\n",
            "Fetched 52.5 MB in 7s (7,202 kB/s)\n",
            "debconf: unable to initialize frontend: Dialog\n",
            "debconf: (No usable dialog-like program is installed, so the dialog based frontend cannot be used. at /usr/share/perl5/Debconf/FrontEnd/Dialog.pm line 76, <> line 11.)\n",
            "debconf: falling back to frontend: Readline\n",
            "debconf: unable to initialize frontend: Readline\n",
            "debconf: (This frontend requires a controlling tty.)\n",
            "debconf: falling back to frontend: Teletype\n",
            "dpkg-preconfigure: unable to re-open stdin: \n",
            "(Reading database ... 144568 files and directories currently installed.)\n",
            "Preparing to unpack .../00-python3.6-dev_3.6.9-1~18.04ubuntu1_amd64.deb ...\n",
            "Unpacking python3.6-dev (3.6.9-1~18.04ubuntu1) over (3.6.9-1~18.04) ...\n",
            "Preparing to unpack .../01-libpython3.6-dev_3.6.9-1~18.04ubuntu1_amd64.deb ...\n",
            "Unpacking libpython3.6-dev:amd64 (3.6.9-1~18.04ubuntu1) over (3.6.9-1~18.04) ...\n",
            "Preparing to unpack .../02-libpython3.6_3.6.9-1~18.04ubuntu1_amd64.deb ...\n",
            "Unpacking libpython3.6:amd64 (3.6.9-1~18.04ubuntu1) over (3.6.9-1~18.04) ...\n",
            "Preparing to unpack .../03-python3.6_3.6.9-1~18.04ubuntu1_amd64.deb ...\n",
            "Unpacking python3.6 (3.6.9-1~18.04ubuntu1) over (3.6.9-1~18.04) ...\n",
            "Preparing to unpack .../04-libpython3.6-stdlib_3.6.9-1~18.04ubuntu1_amd64.deb ...\n",
            "Unpacking libpython3.6-stdlib:amd64 (3.6.9-1~18.04ubuntu1) over (3.6.9-1~18.04) ...\n",
            "Preparing to unpack .../05-python3.6-minimal_3.6.9-1~18.04ubuntu1_amd64.deb ...\n",
            "Unpacking python3.6-minimal (3.6.9-1~18.04ubuntu1) over (3.6.9-1~18.04) ...\n",
            "Preparing to unpack .../06-libpython3.6-minimal_3.6.9-1~18.04ubuntu1_amd64.deb ...\n",
            "Unpacking libpython3.6-minimal:amd64 (3.6.9-1~18.04ubuntu1) over (3.6.9-1~18.04) ...\n",
            "Selecting previously unselected package libnvidia-common-430.\n",
            "Preparing to unpack .../07-libnvidia-common-430_440.59-0ubuntu0.18.04.1_all.deb ...\n",
            "Unpacking libnvidia-common-430 (440.59-0ubuntu0.18.04.1) ...\n",
            "Selecting previously unselected package python-pip-whl.\n",
            "Preparing to unpack .../08-python-pip-whl_9.0.1-2.3~ubuntu1.18.04.1_all.deb ...\n",
            "Unpacking python-pip-whl (9.0.1-2.3~ubuntu1.18.04.1) ...\n",
            "Selecting previously unselected package python3.6-venv.\n",
            "Preparing to unpack .../09-python3.6-venv_3.6.9-1~18.04ubuntu1_amd64.deb ...\n",
            "Unpacking python3.6-venv (3.6.9-1~18.04ubuntu1) ...\n",
            "Selecting previously unselected package python3-venv.\n",
            "Preparing to unpack .../10-python3-venv_3.6.7-1~18.04_amd64.deb ...\n",
            "Unpacking python3-venv (3.6.7-1~18.04) ...\n",
            "Setting up python-pip-whl (9.0.1-2.3~ubuntu1.18.04.1) ...\n",
            "Setting up libnvidia-common-430 (440.59-0ubuntu0.18.04.1) ...\n",
            "Setting up libpython3.6-minimal:amd64 (3.6.9-1~18.04ubuntu1) ...\n",
            "Setting up libpython3.6-stdlib:amd64 (3.6.9-1~18.04ubuntu1) ...\n",
            "Setting up python3.6-minimal (3.6.9-1~18.04ubuntu1) ...\n",
            "Setting up libpython3.6:amd64 (3.6.9-1~18.04ubuntu1) ...\n",
            "Setting up python3.6 (3.6.9-1~18.04ubuntu1) ...\n",
            "Setting up python3.6-venv (3.6.9-1~18.04ubuntu1) ...\n",
            "Setting up libpython3.6-dev:amd64 (3.6.9-1~18.04ubuntu1) ...\n",
            "Setting up python3.6-dev (3.6.9-1~18.04ubuntu1) ...\n",
            "Setting up python3-venv (3.6.7-1~18.04) ...\n",
            "Processing triggers for libc-bin (2.27-3ubuntu1) ...\n",
            "/sbin/ldconfig.real: /usr/local/lib/python3.6/dist-packages/ideep4py/lib/libmkldnn.so.0 is not a symbolic link\n",
            "\n",
            "Processing triggers for man-db (2.8.3-2ubuntu0.1) ...\n",
            "Processing triggers for mime-support (3.60ubuntu1) ...\n",
            "Collecting notebook (from -r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/b1/f1/0a67f09ef53a342403ffa66646ee39273e0ac79ffa5de5dbe2f3e28b5bdf/notebook-6.0.3-py3-none-any.whl (9.7MB)\n",
            "\u001b[K    100% |████████████████████████████████| 9.7MB 139kB/s \n",
            "\u001b[?25hCollecting tensorflow==1.13.2 (from -r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/db/d3/651f95288a6cd9094f7411cdd90ef12a3d01a268009e0e3cd66b5c8d65bd/tensorflow-1.13.2-cp36-cp36m-manylinux1_x86_64.whl (92.6MB)\n",
            "\u001b[K    100% |████████████████████████████████| 92.6MB 14kB/s \n",
            "\u001b[?25hCollecting tensorflow-gpu==1.13.2 (from -r LEt-SNE_requirements.txt (line 3))\n",
            "  Downloading https://files.pythonhosted.org/packages/63/e5/6f47e0e3b8e9215efb3f41692ab47991d96cb3ccc172cb578435cbcb4959/tensorflow_gpu-1.13.2-cp36-cp36m-manylinux1_x86_64.whl (345.0MB)\n",
            "\u001b[K    100% |████████████████████████████████| 345.0MB 3.9kB/s \n",
            "tcmalloc: large alloc 1073750016 bytes == 0x54da0000 @  0x7f73317c02a4 0x591d67 0x5b7eb3 0x50a3ea 0x50bfb4 0x509758 0x50a48d 0x50bfb4 0x507d64 0x509a90 0x50a48d 0x50bfb4 0x507d64 0x509a90 0x50a48d 0x50bfb4 0x507d64 0x509a90 0x50a48d 0x50cd96 0x509758 0x50a48d 0x50bfb4 0x507d64 0x509a90 0x50a48d 0x50cd96 0x507d64 0x509a90 0x50a48d 0x50cd96\n",
            "\u001b[?25hCollecting umap-learn==0.3.10 (from -r LEt-SNE_requirements.txt (line 4))\n",
            "  Downloading https://files.pythonhosted.org/packages/ad/92/36bac74962b424870026cb0b42cec3d5b6f4afa37d81818475d8762f9255/umap-learn-0.3.10.tar.gz (40kB)\n",
            "\u001b[K    100% |████████████████████████████████| 40kB 9.1MB/s \n",
            "\u001b[?25hCollecting scikit-learn==0.20.3 (from -r LEt-SNE_requirements.txt (line 5))\n",
            "  Downloading https://files.pythonhosted.org/packages/5e/82/c0de5839d613b82bddd088599ac0bbfbbbcbd8ca470680658352d2c435bd/scikit_learn-0.20.3-cp36-cp36m-manylinux1_x86_64.whl (5.4MB)\n",
            "\u001b[K    100% |████████████████████████████████| 5.4MB 234kB/s \n",
            "\u001b[?25hCollecting scipy==1.2.1 (from -r LEt-SNE_requirements.txt (line 6))\n",
            "  Downloading https://files.pythonhosted.org/packages/7f/5f/c48860704092933bf1c4c1574a8de1ffd16bf4fde8bab190d747598844b2/scipy-1.2.1-cp36-cp36m-manylinux1_x86_64.whl (24.8MB)\n",
            "\u001b[K    100% |████████████████████████████████| 24.8MB 54kB/s \n",
            "\u001b[?25hCollecting scikit-image==0.15.0 (from -r LEt-SNE_requirements.txt (line 7))\n",
            "  Downloading https://files.pythonhosted.org/packages/d4/ab/674e168bf7d0bc597218b3bec858d02c23fbac9ec1fec9cad878c6cee95f/scikit_image-0.15.0-cp36-cp36m-manylinux1_x86_64.whl (26.3MB)\n",
            "\u001b[K    100% |████████████████████████████████| 26.3MB 52kB/s \n",
            "\u001b[?25hCollecting pandas==0.24.2 (from -r LEt-SNE_requirements.txt (line 8))\n",
            "  Downloading https://files.pythonhosted.org/packages/19/74/e50234bc82c553fecdbd566d8650801e3fe2d6d8c8d940638e3d8a7c5522/pandas-0.24.2-cp36-cp36m-manylinux1_x86_64.whl (10.1MB)\n",
            "\u001b[K    100% |████████████████████████████████| 10.1MB 119kB/s \n",
            "\u001b[?25hCollecting numpy==1.16.3 (from -r LEt-SNE_requirements.txt (line 9))\n",
            "  Downloading https://files.pythonhosted.org/packages/c1/e2/4db8df8f6cddc98e7d7c537245ef2f4e41a1ed17bf0c3177ab3cc6beac7f/numpy-1.16.3-cp36-cp36m-manylinux1_x86_64.whl (17.3MB)\n",
            "\u001b[K    100% |████████████████████████████████| 17.3MB 77kB/s \n",
            "\u001b[?25hCollecting matplotlib==3.0.3 (from -r LEt-SNE_requirements.txt (line 10))\n",
            "  Downloading https://files.pythonhosted.org/packages/e9/69/f5e05f578585ed9935247be3788b374f90701296a70c8871bcd6d21edb00/matplotlib-3.0.3-cp36-cp36m-manylinux1_x86_64.whl (13.0MB)\n",
            "\u001b[K    100% |████████████████████████████████| 13.0MB 104kB/s \n",
            "\u001b[?25hCollecting networkx==2.3 (from -r LEt-SNE_requirements.txt (line 11))\n",
            "  Downloading https://files.pythonhosted.org/packages/85/08/f20aef11d4c343b557e5de6b9548761811eb16e438cee3d32b1c66c8566b/networkx-2.3.zip (1.7MB)\n",
            "\u001b[K    100% |████████████████████████████████| 1.8MB 628kB/s \n",
            "\u001b[?25hCollecting Send2Trash (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/49/46/c3dc27481d1cc57b9385aff41c474ceb7714f7935b1247194adae45db714/Send2Trash-1.5.0-py3-none-any.whl\n",
            "Collecting prometheus-client (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/b3/23/41a5a24b502d35a4ad50a5bb7202a5e1d9a0364d0c12f56db3dbf7aca76d/prometheus_client-0.7.1.tar.gz\n",
            "Collecting nbformat (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/d5/3c/251cfe5e3c9545ece107d342909ee3822d8cd9f3f11a591df37baeff2089/nbformat-5.0.6-py3-none-any.whl (170kB)\n",
            "\u001b[K    100% |████████████████████████████████| 174kB 4.5MB/s \n",
            "\u001b[?25hCollecting traitlets>=4.2.1 (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/ca/ab/872a23e29cec3cf2594af7e857f18b687ad21039c1f9b922fac5b9b142d5/traitlets-4.3.3-py2.py3-none-any.whl (75kB)\n",
            "\u001b[K    100% |████████████████████████████████| 81kB 11.0MB/s \n",
            "\u001b[?25hCollecting jupyter-client>=5.3.4 (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/34/0b/2ebddf775f558158ca8df23b35445fb15d4b1558a9e4a03bc7e75b13476e/jupyter_client-6.1.3-py3-none-any.whl (106kB)\n",
            "\u001b[K    100% |████████████████████████████████| 112kB 10.3MB/s \n",
            "\u001b[?25hCollecting ipykernel (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/ed/5d/cf47741fa80826f8edf435d9bcf0e84eef2d6d02953ff85e0563b3ab3f0b/ipykernel-5.2.1-py3-none-any.whl (118kB)\n",
            "\u001b[K    100% |████████████████████████████████| 122kB 9.5MB/s \n",
            "\u001b[?25hCollecting nbconvert (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/79/6c/05a569e9f703d18aacb89b7ad6075b404e8a4afde2c26b73ca77bb644b14/nbconvert-5.6.1-py2.py3-none-any.whl (455kB)\n",
            "\u001b[K    100% |████████████████████████████████| 460kB 1.5MB/s \n",
            "\u001b[?25hCollecting tornado>=5.0 (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/95/84/119a46d494f008969bf0c775cb2c6b3579d3c4cc1bb1b41a022aa93ee242/tornado-6.0.4.tar.gz (496kB)\n",
            "\u001b[K    100% |████████████████████████████████| 501kB 1.5MB/s \n",
            "\u001b[?25hCollecting pyzmq>=17 (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/41/fa/e6e10410f01b03d10ab0705717d1246f63cdbbc0676140c78f0f757db332/pyzmq-19.0.0-cp36-cp36m-manylinux1_x86_64.whl (1.1MB)\n",
            "\u001b[K    100% |████████████████████████████████| 1.1MB 925kB/s \n",
            "\u001b[?25hCollecting jinja2 (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/30/9e/f663a2aa66a09d838042ae1a2c5659828bb9b41ea3a6efa20a20fd92b121/Jinja2-2.11.2-py2.py3-none-any.whl (125kB)\n",
            "\u001b[K    100% |████████████████████████████████| 133kB 9.0MB/s \n",
            "\u001b[?25hCollecting jupyter-core>=4.6.1 (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/63/0d/df2d17cdf389cea83e2efa9a4d32f7d527ba78667e0153a8e676e957b2f7/jupyter_core-4.6.3-py2.py3-none-any.whl (83kB)\n",
            "\u001b[K    100% |████████████████████████████████| 92kB 11.1MB/s \n",
            "\u001b[?25hCollecting terminado>=0.8.1 (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/ff/96/1d9a2c23990aea8f8e0b5c3b6627d03196a73771a17a2d9860bbe9823ab6/terminado-0.8.3-py2.py3-none-any.whl\n",
            "Collecting ipython-genutils (from notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/fa/bc/9bd3b5c2b4774d5f33b2d544f1460be9df7df2fe42f352135381c347c69a/ipython_genutils-0.2.0-py2.py3-none-any.whl\n",
            "Collecting tensorflow-estimator<1.14.0rc0,>=1.13.0 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/bb/48/13f49fc3fa0fdf916aa1419013bb8f2ad09674c275b4046d5ee669a46873/tensorflow_estimator-1.13.0-py2.py3-none-any.whl (367kB)\n",
            "\u001b[K    100% |████████████████████████████████| 368kB 2.1MB/s \n",
            "\u001b[?25hCollecting wheel>=0.26 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/8c/23/848298cccf8e40f5bbb59009b32848a4c38f4e7f3364297ab3c3e2e2cd14/wheel-0.34.2-py2.py3-none-any.whl\n",
            "Collecting termcolor>=1.1.0 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/8a/48/a76be51647d0eb9f10e2a4511bf3ffb8cc1e6b14e9e4fab46173aa79f981/termcolor-1.1.0.tar.gz\n",
            "Collecting six>=1.10.0 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/65/eb/1f97cb97bfc2390a276969c6fae16075da282f5058082d4cb10c6c5c1dba/six-1.14.0-py2.py3-none-any.whl\n",
            "Collecting keras-preprocessing>=1.0.5 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/28/6a/8c1f62c37212d9fc441a7e26736df51ce6f0e38455816445471f10da4f0a/Keras_Preprocessing-1.1.0-py2.py3-none-any.whl (41kB)\n",
            "\u001b[K    100% |████████████████████████████████| 51kB 11.4MB/s \n",
            "\u001b[?25hCollecting protobuf>=3.6.1 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/57/02/5432412c162989260fab61fa65e0a490c1872739eb91a659896e4d554b26/protobuf-3.11.3-cp36-cp36m-manylinux1_x86_64.whl (1.3MB)\n",
            "\u001b[K    100% |████████████████████████████████| 1.3MB 812kB/s \n",
            "\u001b[?25hCollecting absl-py>=0.1.6 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/1a/53/9243c600e047bd4c3df9e69cfabc1e8004a82cac2e0c484580a78a94ba2a/absl-py-0.9.0.tar.gz (104kB)\n",
            "\u001b[K    100% |████████████████████████████████| 112kB 10.4MB/s \n",
            "\u001b[?25hCollecting keras-applications>=1.0.6 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/71/e3/19762fdfc62877ae9102edf6342d71b28fbfd9dea3d2f96a882ce099b03f/Keras_Applications-1.0.8-py3-none-any.whl (50kB)\n",
            "\u001b[K    100% |████████████████████████████████| 51kB 8.9MB/s \n",
            "\u001b[?25hCollecting tensorboard<1.14.0,>=1.13.0 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/0f/39/bdd75b08a6fba41f098b6cb091b9e8c7a80e1b4d679a581a0ccd17b10373/tensorboard-1.13.1-py3-none-any.whl (3.2MB)\n",
            "\u001b[K    100% |████████████████████████████████| 3.2MB 309kB/s \n",
            "\u001b[?25hCollecting astor>=0.6.0 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/c3/88/97eef84f48fa04fbd6750e62dcceafba6c63c81b7ac1420856c8dcc0a3f9/astor-0.8.1-py2.py3-none-any.whl\n",
            "Collecting gast>=0.2.0 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/d6/84/759f5dd23fec8ba71952d97bcc7e2c9d7d63bdc582421f3cd4be845f0c98/gast-0.3.3-py2.py3-none-any.whl\n",
            "Collecting grpcio>=1.8.6 (from tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/cf/7a/9744998129fce7e29c5f2d8b0f545913b7383e65d8366fc0ae98d11936af/grpcio-1.28.1.tar.gz (19.5MB)\n",
            "\u001b[K    100% |████████████████████████████████| 19.5MB 69kB/s \n",
            "\u001b[?25hCollecting numba>=0.37 (from umap-learn==0.3.10->-r LEt-SNE_requirements.txt (line 4))\n",
            "  Downloading https://files.pythonhosted.org/packages/c8/1d/3d904b2d5f2c350dbab76912e0ba5cb17469f392520a9899361a363e1741/numba-0.49.0-cp36-cp36m-manylinux1_x86_64.whl (2.6MB)\n",
            "\u001b[K    100% |████████████████████████████████| 2.6MB 465kB/s \n",
            "\u001b[?25hCollecting imageio>=2.0.1 (from scikit-image==0.15.0->-r LEt-SNE_requirements.txt (line 7))\n",
            "  Downloading https://files.pythonhosted.org/packages/4c/2b/9dd19644f871b10f7e32eb2dbd6b45149c350b4d5f2893e091b882e03ab7/imageio-2.8.0-py3-none-any.whl (3.3MB)\n",
            "\u001b[K    100% |████████████████████████████████| 3.3MB 365kB/s \n",
            "\u001b[?25hCollecting pillow>=4.3.0 (from scikit-image==0.15.0->-r LEt-SNE_requirements.txt (line 7))\n",
            "  Downloading https://files.pythonhosted.org/packages/e0/50/8e78e6f62ffa50d6ca95c281d5a2819bef66d023ac1b723e253de5bda9c5/Pillow-7.1.2-cp36-cp36m-manylinux1_x86_64.whl (2.1MB)\n",
            "\u001b[K    100% |████████████████████████████████| 2.1MB 556kB/s \n",
            "\u001b[?25hCollecting PyWavelets>=0.4.0 (from scikit-image==0.15.0->-r LEt-SNE_requirements.txt (line 7))\n",
            "  Downloading https://files.pythonhosted.org/packages/59/bb/d2b85265ec9fa3c1922210c9393d4cdf7075cc87cce6fe671d7455f80fbc/PyWavelets-1.1.1-cp36-cp36m-manylinux1_x86_64.whl (4.4MB)\n",
            "\u001b[K    100% |████████████████████████████████| 4.4MB 279kB/s \n",
            "\u001b[?25hCollecting pytz>=2011k (from pandas==0.24.2->-r LEt-SNE_requirements.txt (line 8))\n",
            "  Downloading https://files.pythonhosted.org/packages/e7/f9/f0b53f88060247251bf481fa6ea62cd0d25bf1b11a87888e53ce5b7c8ad2/pytz-2019.3-py2.py3-none-any.whl (509kB)\n",
            "\u001b[K    100% |████████████████████████████████| 512kB 1.4MB/s \n",
            "\u001b[?25hCollecting python-dateutil>=2.5.0 (from pandas==0.24.2->-r LEt-SNE_requirements.txt (line 8))\n",
            "  Downloading https://files.pythonhosted.org/packages/d4/70/d60450c3dd48ef87586924207ae8907090de0b306af2bce5d134d78615cb/python_dateutil-2.8.1-py2.py3-none-any.whl (227kB)\n",
            "\u001b[K    100% |████████████████████████████████| 235kB 2.5MB/s \n",
            "\u001b[?25hCollecting pyparsing!=2.0.4,!=2.1.2,!=2.1.6,>=2.0.1 (from matplotlib==3.0.3->-r LEt-SNE_requirements.txt (line 10))\n",
            "  Downloading https://files.pythonhosted.org/packages/8a/bb/488841f56197b13700afd5658fc279a2025a39e22449b7cf29864669b15d/pyparsing-2.4.7-py2.py3-none-any.whl (67kB)\n",
            "\u001b[K    100% |████████████████████████████████| 71kB 8.4MB/s \n",
            "\u001b[?25hCollecting cycler>=0.10 (from matplotlib==3.0.3->-r LEt-SNE_requirements.txt (line 10))\n",
            "  Downloading https://files.pythonhosted.org/packages/f7/d2/e07d3ebb2bd7af696440ce7e754c59dd546ffe1bbe732c8ab68b9c834e61/cycler-0.10.0-py2.py3-none-any.whl\n",
            "Collecting kiwisolver>=1.0.1 (from matplotlib==3.0.3->-r LEt-SNE_requirements.txt (line 10))\n",
            "  Downloading https://files.pythonhosted.org/packages/ae/23/147de658aabbf968324551ea22c0c13a00284c4ef49a77002e91f79657b7/kiwisolver-1.2.0-cp36-cp36m-manylinux1_x86_64.whl (88kB)\n",
            "\u001b[K    100% |████████████████████████████████| 92kB 10.3MB/s \n",
            "\u001b[?25hCollecting decorator>=4.3.0 (from networkx==2.3->-r LEt-SNE_requirements.txt (line 11))\n",
            "  Downloading https://files.pythonhosted.org/packages/ed/1b/72a1821152d07cf1d8b6fce298aeb06a7eb90f4d6d41acec9861e7cc6df0/decorator-4.4.2-py2.py3-none-any.whl\n",
            "Collecting jsonschema!=2.5.0,>=2.4 (from nbformat->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/c5/8f/51e89ce52a085483359217bc72cdbf6e75ee595d5b1d4b5ade40c7e018b8/jsonschema-3.2.0-py2.py3-none-any.whl (56kB)\n",
            "\u001b[K    100% |████████████████████████████████| 61kB 7.5MB/s \n",
            "\u001b[?25hCollecting ipython>=5.0.0 (from ipykernel->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/61/6f/69f1eec859ce48a86660529b166b6ea466f0f4ab98e4fc0807b835aa22c6/ipython-7.13.0-py3-none-any.whl (780kB)\n",
            "\u001b[K    100% |████████████████████████████████| 788kB 1.1MB/s \n",
            "\u001b[?25hCollecting defusedxml (from nbconvert->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/06/74/9b387472866358ebc08732de3da6dc48e44b0aacd2ddaa5cb85ab7e986a2/defusedxml-0.6.0-py2.py3-none-any.whl\n",
            "Collecting entrypoints>=0.2.2 (from nbconvert->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/ac/c6/44694103f8c221443ee6b0041f69e2740d89a25641e62fb4f2ee568f2f9c/entrypoints-0.3-py2.py3-none-any.whl\n",
            "Collecting mistune<2,>=0.8.1 (from nbconvert->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/09/ec/4b43dae793655b7d8a25f76119624350b4d65eb663459eb9603d7f1f0345/mistune-0.8.4-py2.py3-none-any.whl\n",
            "Collecting testpath (from nbconvert->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/1b/9e/1a170feaa54f22aeb5a5d16c9015e82234275a3c8ab630b552493f9cb8a9/testpath-0.4.4-py2.py3-none-any.whl (163kB)\n",
            "\u001b[K    100% |████████████████████████████████| 174kB 6.3MB/s \n",
            "\u001b[?25hCollecting pygments (from nbconvert->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/2d/68/106af3ae51daf807e9cdcba6a90e518954eb8b70341cee52995540a53ead/Pygments-2.6.1-py3-none-any.whl (914kB)\n",
            "\u001b[K    100% |████████████████████████████████| 921kB 987kB/s \n",
            "\u001b[?25hCollecting bleach (from nbconvert->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/c7/bc/9c2bbd8bff5f393e3373687109117a061db4c8ccdf6b5c70b1f8834bd67a/bleach-3.1.4-py2.py3-none-any.whl (151kB)\n",
            "\u001b[K    100% |████████████████████████████████| 153kB 3.1MB/s \n",
            "\u001b[?25hCollecting pandocfilters>=1.4.1 (from nbconvert->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/4c/ea/236e2584af67bb6df960832731a6e5325fd4441de001767da328c33368ce/pandocfilters-1.4.2.tar.gz\n",
            "Collecting MarkupSafe>=0.23 (from jinja2->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/b2/5f/23e0023be6bb885d00ffbefad2942bc51a620328ee910f64abe5a8d18dd1/MarkupSafe-1.1.1-cp36-cp36m-manylinux1_x86_64.whl\n",
            "Collecting ptyprocess; os_name != \"nt\" (from terminado>=0.8.1->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/d1/29/605c2cc68a9992d18dada28206eeada56ea4bd07a239669da41674648b6f/ptyprocess-0.6.0-py2.py3-none-any.whl\n",
            "Collecting mock>=2.0.0 (from tensorflow-estimator<1.14.0rc0,>=1.13.0->tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/cd/74/d72daf8dff5b6566db857cfd088907bb0355f5dd2914c4b3ef065c790735/mock-4.0.2-py3-none-any.whl\n",
            "Requirement already satisfied: setuptools in ./LEt-SNE_environment/lib/python3.6/site-packages (from protobuf>=3.6.1->tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "Collecting h5py (from keras-applications>=1.0.6->tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/60/06/cafdd44889200e5438b897388f3075b52a8ef01f28a17366d91de0fa2d05/h5py-2.10.0-cp36-cp36m-manylinux1_x86_64.whl (2.9MB)\n",
            "\u001b[K    100% |████████████████████████████████| 2.9MB 413kB/s \n",
            "\u001b[?25hCollecting werkzeug>=0.11.15 (from tensorboard<1.14.0,>=1.13.0->tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/cc/94/5f7079a0e00bd6863ef8f1da638721e9da21e5bacee597595b318f71d62e/Werkzeug-1.0.1-py2.py3-none-any.whl (298kB)\n",
            "\u001b[K    100% |████████████████████████████████| 307kB 1.9MB/s \n",
            "\u001b[?25hCollecting markdown>=2.6.8 (from tensorboard<1.14.0,>=1.13.0->tensorflow==1.13.2->-r LEt-SNE_requirements.txt (line 2))\n",
            "  Downloading https://files.pythonhosted.org/packages/ab/c4/ba46d44855e6eb1770a12edace5a165a0c6de13349f592b9036257f3c3d3/Markdown-3.2.1-py2.py3-none-any.whl (88kB)\n",
            "\u001b[K    100% |████████████████████████████████| 92kB 10.7MB/s \n",
            "\u001b[?25hCollecting llvmlite<=0.33.0.dev0,>=0.31.0.dev0 (from numba>=0.37->umap-learn==0.3.10->-r LEt-SNE_requirements.txt (line 4))\n",
            "  Downloading https://files.pythonhosted.org/packages/23/19/20c085fe98c8df8953a18d8ef15902b9be0438b7ee19586b9e9124da1a32/llvmlite-0.32.0-cp36-cp36m-manylinux1_x86_64.whl (20.2MB)\n",
            "\u001b[K    100% |████████████████████████████████| 20.2MB 64kB/s \n",
            "\u001b[?25hCollecting pyrsistent>=0.14.0 (from jsonschema!=2.5.0,>=2.4->nbformat->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/9f/0d/cbca4d0bbc5671822a59f270e4ce3f2195f8a899c97d0d5abb81b191efb5/pyrsistent-0.16.0.tar.gz (108kB)\n",
            "\u001b[K    100% |████████████████████████████████| 112kB 10.0MB/s \n",
            "\u001b[?25hCollecting importlib-metadata; python_version < \"3.8\" (from jsonschema!=2.5.0,>=2.4->nbformat->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/ad/e4/891bfcaf868ccabc619942f27940c77a8a4b45fd8367098955bb7e152fb1/importlib_metadata-1.6.0-py2.py3-none-any.whl\n",
            "Collecting attrs>=17.4.0 (from jsonschema!=2.5.0,>=2.4->nbformat->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/a2/db/4313ab3be961f7a763066401fb77f7748373b6094076ae2bda2806988af6/attrs-19.3.0-py2.py3-none-any.whl\n",
            "Collecting prompt-toolkit!=3.0.0,!=3.0.1,<3.1.0,>=2.0.0 (from ipython>=5.0.0->ipykernel->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/e4/a7/81b39aa50e9284fe2cb21cc7fb7de7817b224172d42793fd57451d38842b/prompt_toolkit-3.0.5-py3-none-any.whl (351kB)\n",
            "\u001b[K    100% |████████████████████████████████| 358kB 1.8MB/s \n",
            "\u001b[?25hCollecting backcall (from ipython>=5.0.0->ipykernel->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/84/71/c8ca4f5bb1e08401b916c68003acf0a0655df935d74d93bf3f3364b310e0/backcall-0.1.0.tar.gz\n",
            "Collecting jedi>=0.10 (from ipython>=5.0.0->ipykernel->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/f3/3f/67f027e18c60a800875df1a0894a2436ce9053637fa39725766e937c0a71/jedi-0.17.0-py2.py3-none-any.whl (1.1MB)\n",
            "\u001b[K    100% |████████████████████████████████| 1.1MB 978kB/s \n",
            "\u001b[?25hCollecting pexpect; sys_platform != \"win32\" (from ipython>=5.0.0->ipykernel->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/39/7b/88dbb785881c28a102619d46423cb853b46dbccc70d3ac362d99773a78ce/pexpect-4.8.0-py2.py3-none-any.whl (59kB)\n",
            "\u001b[K    100% |████████████████████████████████| 61kB 9.3MB/s \n",
            "\u001b[?25hCollecting pickleshare (from ipython>=5.0.0->ipykernel->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/9a/41/220f49aaea88bc6fa6cba8d05ecf24676326156c23b991e80b3f2fc24c77/pickleshare-0.7.5-py2.py3-none-any.whl\n",
            "Collecting webencodings (from bleach->nbconvert->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/f4/24/2a3e3df732393fed8b3ebf2ec078f05546de641fe1b667ee316ec1dcf3b7/webencodings-0.5.1-py2.py3-none-any.whl\n",
            "Collecting zipp>=0.5 (from importlib-metadata; python_version < \"3.8\"->jsonschema!=2.5.0,>=2.4->nbformat->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/b2/34/bfcb43cc0ba81f527bc4f40ef41ba2ff4080e047acb0586b56b3d017ace4/zipp-3.1.0-py3-none-any.whl\n",
            "Collecting wcwidth (from prompt-toolkit!=3.0.0,!=3.0.1,<3.1.0,>=2.0.0->ipython>=5.0.0->ipykernel->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/f6/d5/1ecdac957e3ea12c1b319fcdee8b6917ffaff8b4644d673c4d72d2f20b49/wcwidth-0.1.9-py2.py3-none-any.whl\n",
            "Collecting parso>=0.7.0 (from jedi>=0.10->ipython>=5.0.0->ipykernel->notebook->-r LEt-SNE_requirements.txt (line 1))\n",
            "  Downloading https://files.pythonhosted.org/packages/b5/61/998cce9e7476de000d031874df26a18f67cb73448164fc44a98f0c55920b/parso-0.7.0-py2.py3-none-any.whl (100kB)\n",
            "\u001b[K    100% |████████████████████████████████| 102kB 10.3MB/s \n",
            "\u001b[?25hBuilding wheels for collected packages: umap-learn, networkx, prometheus-client, tornado, termcolor, absl-py, grpcio, pandocfilters, pyrsistent, backcall\n",
            "  Running setup.py bdist_wheel for umap-learn ... \u001b[?25lerror\n",
            "\u001b[31m  Failed building wheel for umap-learn\u001b[0m\n",
            "\u001b[?25h  Running setup.py clean for umap-learn\n",
            "  Running setup.py bdist_wheel for networkx ... \u001b[?25lerror\n",
            "\u001b[31m  Failed building wheel for networkx\u001b[0m\n",
            "\u001b[?25h  Running setup.py clean for networkx\n",
            "  Running setup.py bdist_wheel for prometheus-client ... \u001b[?25lerror\n",
            "\u001b[31m  Failed building wheel for prometheus-client\u001b[0m\n",
            "\u001b[?25h  Running setup.py clean for prometheus-client\n",
            "  Running setup.py bdist_wheel for tornado ... \u001b[?25lerror\n",
            "\u001b[31m  Failed building wheel for tornado\u001b[0m\n",
            "\u001b[?25h  Running setup.py clean for tornado\n",
            "  Running setup.py bdist_wheel for termcolor ... \u001b[?25lerror\n",
            "\u001b[31m  Failed building wheel for termcolor\u001b[0m\n",
            "\u001b[?25h  Running setup.py clean for termcolor\n",
            "  Running setup.py bdist_wheel for absl-py ... \u001b[?25lerror\n",
            "\u001b[31m  Failed building wheel for absl-py\u001b[0m\n",
            "\u001b[?25h  Running setup.py clean for absl-py\n",
            "  Running setup.py bdist_wheel for grpcio ... \u001b[?25lerror\n",
            "\u001b[31m  Failed building wheel for grpcio\u001b[0m\n",
            "\u001b[?25h  Running setup.py clean for grpcio\n",
            "  Running setup.py bdist_wheel for pandocfilters ... \u001b[?25lerror\n",
            "\u001b[31m  Failed building wheel for pandocfilters\u001b[0m\n",
            "\u001b[?25h  Running setup.py clean for pandocfilters\n",
            "  Running setup.py bdist_wheel for pyrsistent ... \u001b[?25lerror\n",
            "\u001b[31m  Failed building wheel for pyrsistent\u001b[0m\n",
            "\u001b[?25h  Running setup.py clean for pyrsistent\n",
            "  Running setup.py bdist_wheel for backcall ... \u001b[?25lerror\n",
            "\u001b[31m  Failed building wheel for backcall\u001b[0m\n",
            "\u001b[?25h  Running setup.py clean for backcall\n",
            "Failed to build umap-learn networkx prometheus-client tornado termcolor absl-py grpcio pandocfilters pyrsistent backcall\n",
            "Installing collected packages: Send2Trash, prometheus-client, six, pyrsistent, zipp, importlib-metadata, attrs, jsonschema, decorator, ipython-genutils, traitlets, jupyter-core, nbformat, tornado, pyzmq, python-dateutil, jupyter-client, wcwidth, prompt-toolkit, backcall, pygments, parso, jedi, ptyprocess, pexpect, pickleshare, ipython, ipykernel, defusedxml, entrypoints, mistune, testpath, webencodings, bleach, MarkupSafe, jinja2, pandocfilters, nbconvert, terminado, notebook, mock, absl-py, numpy, tensorflow-estimator, wheel, termcolor, keras-preprocessing, protobuf, h5py, keras-applications, grpcio, werkzeug, markdown, tensorboard, astor, gast, tensorflow, tensorflow-gpu, llvmlite, numba, scipy, scikit-learn, umap-learn, pillow, imageio, PyWavelets, pyparsing, cycler, kiwisolver, matplotlib, networkx, scikit-image, pytz, pandas\n",
            "  Running setup.py install for prometheus-client ... \u001b[?25ldone\n",
            "\u001b[?25h  Running setup.py install for pyrsistent ... \u001b[?25ldone\n",
            "\u001b[?25h  Running setup.py install for tornado ... \u001b[?25ldone\n",
            "\u001b[?25h  Running setup.py install for backcall ... \u001b[?25ldone\n",
            "\u001b[?25h  Running setup.py install for pandocfilters ... \u001b[?25ldone\n",
            "\u001b[?25h  Running setup.py install for absl-py ... \u001b[?25ldone\n",
            "\u001b[?25h  Running setup.py install for termcolor ... \u001b[?25ldone\n",
            "\u001b[?25h  Running setup.py install for grpcio ... \u001b[?25ldone\n",
            "\u001b[?25h  Running setup.py install for umap-learn ... \u001b[?25ldone\n",
            "\u001b[?25h  Running setup.py install for networkx ... \u001b[?25ldone\n",
            "\u001b[?25hSuccessfully installed MarkupSafe-1.1.1 PyWavelets-1.1.1 Send2Trash-1.5.0 absl-py-0.9.0 astor-0.8.1 attrs-19.3.0 backcall-0.1.0 bleach-3.1.4 cycler-0.10.0 decorator-4.4.2 defusedxml-0.6.0 entrypoints-0.3 gast-0.3.3 grpcio-1.28.1 h5py-2.10.0 imageio-2.8.0 importlib-metadata-1.6.0 ipykernel-5.2.1 ipython-7.13.0 ipython-genutils-0.2.0 jedi-0.17.0 jinja2-2.11.2 jsonschema-3.2.0 jupyter-client-6.1.3 jupyter-core-4.6.3 keras-applications-1.0.8 keras-preprocessing-1.1.0 kiwisolver-1.2.0 llvmlite-0.32.0 markdown-3.2.1 matplotlib-3.0.3 mistune-0.8.4 mock-4.0.2 nbconvert-5.6.1 nbformat-5.0.6 networkx-2.3 notebook-6.0.3 numba-0.49.0 numpy-1.16.3 pandas-0.24.2 pandocfilters-1.4.2 parso-0.7.0 pexpect-4.8.0 pickleshare-0.7.5 pillow-7.1.2 prometheus-client-0.7.1 prompt-toolkit-3.0.5 protobuf-3.11.3 ptyprocess-0.6.0 pygments-2.6.1 pyparsing-2.4.7 pyrsistent-0.16.0 python-dateutil-2.8.1 pytz-2019.3 pyzmq-19.0.0 scikit-image-0.15.0 scikit-learn-0.20.3 scipy-1.2.1 six-1.14.0 tensorboard-1.13.1 tensorflow-1.13.2 tensorflow-estimator-1.13.0 tensorflow-gpu-1.13.2 termcolor-1.1.0 terminado-0.8.3 testpath-0.4.4 tornado-6.0.4 traitlets-4.3.3 umap-learn-0.3.10 wcwidth-0.1.9 webencodings-0.5.1 werkzeug-1.0.1 wheel-0.34.2 zipp-3.1.0\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "HA8RH0R6d2rT",
        "colab_type": "text"
      },
      "source": [
        "## Part 2: Experimental setup\n",
        "\n",
        "Import the required libraries with some tweaks to ensure aesthetic appearance of our plots. Defining helper functions and LEt-SNE."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ro-kyqiXeScG",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "014445e9-9b68-4966-dec1-b13daa20b984"
      },
      "source": [
        "# If using with Google Colab, set TensorFlow to version 1.x\n",
        "%tensorflow_version 1.x\n",
        "\n",
        "import os\n",
        "import random\n",
        "import warnings\n",
        "\n",
        "import numpy as np\n",
        "import tensorflow as tf\n",
        "from matplotlib import pyplot as plt\n",
        "from tensorflow.contrib.layers import l2_regularizer as l2_reg\n",
        "\n",
        "# allow growth = True eliminates TF 1.13 cuDNN error\n",
        "if tf.__version__[:4] == '1.13':\n",
        "    config = tf.ConfigProto()\n",
        "    config.gpu_options.allow_growth = True\n",
        "else:\n",
        "    config = None\n",
        "\n",
        "plt.style.use('ggplot')\n",
        "plt.rcParams['figure.figsize'] = [15/2.54, 15/2.54]\n",
        "plt.rcParams['figure.dpi'] = 150"
      ],
      "execution_count": 2,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "TensorFlow 1.x selected.\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "DS69ecCte0oO",
        "colab_type": "text"
      },
      "source": [
        "Setting up the directory of TensorBoard for Compression Factor"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "SDtBmNpEerN-",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# Setting the seed to a fixed value for reproducibility purpose\n",
        "# Based on https://medium.com/@ODSC/properly-setting-the-random-seed-in-ml-experiments-not-as-simple-as-you-might-imagine-219969c84752\n",
        "seed_value = 0\n",
        "os.environ['PYTHONHASHSEED']=str(seed_value)\n",
        "random.seed(seed_value)\n",
        "np.random.seed(seed_value)\n",
        "tf.set_random_seed(seed_value)\n",
        "\n",
        "# Location of current file\n",
        "FILE_DIR = os.path.dirname(os.path.abspath(\"__file__\"))\n",
        "TENSORBOARD_DIR = os.path.join(FILE_DIR, 'TensorBoard_CF')\n",
        "\n",
        "os.mkdir(TENSORBOARD_DIR)\n",
        "\n",
        "# Setting runtime options\n",
        "np.set_printoptions(threshold=np.inf)                            # To print all elements of np array without truncation\n",
        "tf.logging.set_verbosity(tf.logging.INFO)\n",
        "warnings.simplefilter(action='ignore', category=FutureWarning)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "G1G02Mm0e7le",
        "colab_type": "text"
      },
      "source": [
        "Defining experiment configurations"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "XioHloQOe-w_",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "LABEL = True                         # Either MANIFOLD = True or LABEL = True or SEGMENTATION = True\n",
        "MANIFOLD = False\n",
        "SEGMENTATION = False\n",
        "\n",
        "NUM_BANDS = 2                        # Dimensionality of our original data\n",
        "NUM_CLASSES = 2                      # The dataset will contain two classes\n",
        "NUM_EPOCHS = 500                     # Training epochs. One epoch is passing all training data through the model\n",
        "\n",
        "# Inverted t-SNE\n",
        "PERPLEXITY = 45                      # A soft measure of the number of neighbours for a sample. High variance, high perp\n",
        "TOLERANCE = 0.1                      # Acceptable deviation from the perplexity\n",
        "COMPRESSION_FACTOR = 50              # Multiplicative factor to boost probabilities of samples having same class\n",
        "\n",
        "# Model Parameters\n",
        "BATCH_SIZE = 1024\n",
        "DIM_REDUCED = 2\n",
        "HIDDEN = [DIM_REDUCED]               # If single element list, convolutional layers flattened to produce encoding\n",
        "LOSS_REGULARIZATION = 0.0\n",
        "LEARNING_RATE = 0.0001\n",
        "\n",
        "# Laplacian\n",
        "NUM_NEIGHBOURS = 0                   \n",
        "# = BATCH_SIZE // (5 * NUM_CLASSES), Updated if MANIFOLD == True in load_data()\n",
        "# = 0, SEGMENTATION OR LABEL uses segmented regions/labels to compute adjacency\n",
        "\n",
        "LOSS_WEIGHTS = np.array([0.05, 1.0, 0.0], dtype=np.float32)\n",
        "# Laplacian, inv_tSNE = 1 (SEG, LAB) = 5 (MANIFOLD),  Reconstruction, CAE, Encoding Norm Constraint\n",
        "\n",
        "# Global constants for the purpose of debugging\n",
        "LOG_TENSOR = [False, False, False, False]     # Debug: Graph_Adjacency, Perplexity, Embedding pairwise l2norm, embedding\n",
        "LOG_TENSOR_DICT = {}                          # Will contain various attributes to log"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "qndW21eJf-zV",
        "colab_type": "text"
      },
      "source": [
        "Defining a function to convert dataset into TensorFlow compatible format"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Gms_bnFCgDg3",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def dataset_tf(bands, labels, train=True):\n",
        "    '''\n",
        "    TENSORFLOW\n",
        "    Creates a TensorFlow dataset object which is given as input to the Estimator object\n",
        "    Converts the labels into one-hot encoded targets.\n",
        "    :param bands: 2D array (numpy) of hyperspectral image bands\n",
        "    :param labels: 1D array (numpy) of labels corresponding to bands\n",
        "    :param train: boolean (True) If true, dataset repeats for NUM_EPOCHS, no repeatation otherwise (prediction phase)\n",
        "    :return: TensorFlow Dataset object\n",
        "    '''\n",
        "\n",
        "    with tf.variable_scope(\"Dataset_Batches\"):\n",
        "        # Dataset dimensions should be (NUM_BANDS, 1) as tf.layers.conv1d() requires channel depth to be known\n",
        "        # Labels are converted to one_hot inside classification_loss (LEt_SNE function)\n",
        "        band_dataset = tf.data.Dataset.from_tensor_slices((bands, labels)).map(lambda x, y: (tf.reshape(x, (-1, 1)), y))\n",
        "        if train:\n",
        "            band_dataset = band_dataset.shuffle(100000).repeat(NUM_EPOCHS).batch(BATCH_SIZE, drop_remainder=True)\n",
        "        else:\n",
        "            band_dataset = band_dataset.batch(BATCH_SIZE)\n",
        "        tf.logging.info('Batches created')\n",
        "        tf.logging.info('1. Dataset element \\nType: %s \\nShape: %s',\n",
        "                        str(band_dataset.output_types),\n",
        "                        str(band_dataset.output_shapes))\n",
        "        return band_dataset"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "eM8ow21QgJYH",
        "colab_type": "text"
      },
      "source": [
        "We define the LEt-SNE model"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "trofgCohgPNz",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def let_sne(features, labels, mode, params):\n",
        "    '''\n",
        "    TENSORFLOW\n",
        "    This acts as model function which is provided to object of Estimator class\n",
        "    The model consists of an autoencoder and a LINEAR classification branch, which\n",
        "    originates from the hidden layer representation. We include a linear branch\n",
        "    to guide the network to learn linearly separable lower dimensional representation\n",
        "    of the input space. Thus, class labels are used to learn encoding of the bands.\n",
        "    :param features: A 2D tensor of size [BATCH_SIZE, NUM_BANDS]; input to the network\n",
        "    :param labels: A 2D tensor of size [BATCH_SIZE, NUM_CLASSES]; one-hot representation of the labels\n",
        "    :param mode: (tf.keys) Predict, Evaluate, Train. Value depends on the keys passed to the estimator\n",
        "    :param params: Dictionary, consisting of the number of hidden nodes, and the ratio of the losses\n",
        "    :return: tf.estimator.EstimatorSpec() object, return value depends on the ModeKeys used to call the estimator.\n",
        "    '''\n",
        "\n",
        "    def model(input_features, bn_train):\n",
        "        '''\n",
        "        1-D convolutional autoencoder with weight regularization for the decoder and classification layers.\n",
        "        :param input_features: A 3D tensor of size [BATCH_SIZE, NUM_BANDS, 1]; input to the network\n",
        "        :return: 1. encoder: A 2D tensor of size [BATCH_SIZE, DIM_REDUCED]; lower dimension encoding of data\n",
        "                 2. decoder: A 2D tensor of size [BATCH_SIZE, NUM_BANDS]; reconstruction from the encoder_bands\n",
        "                 3. weights_encoder: A 2D tensor of size [HIDDEN[0] , HIDDEN[1]], linear mapping to the embedding\n",
        "        '''\n",
        "\n",
        "        with tf.variable_scope(\"LEt_SNE\"):\n",
        "            # We do not use convolutional features to provide a fair comparison with other approaches.\n",
        "            # However, convolutional feature extraction should provide superior results over the Fully Connected variant\n",
        "            input_features_flat = tf.layers.Flatten()(input_features)\n",
        "            encoder_mid = tf.layers.Dense(units=64, activation=None,\n",
        "                                          name='encoder_mid')(input_features_flat)\n",
        "            batch_norm_1 = tf.nn.relu(tf.layers.batch_normalization(inputs=encoder_mid, momentum=0.99, epsilon=1e-5,\n",
        "                                                                    name=\"batch_norm_1\", training=bn_train))\n",
        "            encoder = tf.layers.Dense(units=params['hidden'][0], activation=None,\n",
        "                                      name='encoder')(batch_norm_1)\n",
        "\n",
        "        with tf.variable_scope(\"LEt_SNE/Reconstruction\"):\n",
        "            # Decoder network\n",
        "            decoder_mid = tf.nn.relu(tf.layers.Dense(units=64,\n",
        "                                                     name='decoder_mid',\n",
        "                                                     kernel_regularizer=l2_reg(scale=LOSS_REGULARIZATION))(encoder))\n",
        "            decoder = tf.layers.Dense(units=NUM_BANDS,\n",
        "                                      name='decoder',\n",
        "                                      kernel_regularizer=l2_reg(scale=LOSS_REGULARIZATION))(decoder_mid)\n",
        "\n",
        "        graph = tf.get_default_graph()\n",
        "        p = graph.get_operations()\n",
        "        tf.logging.debug('\\nGlobal Variables: ' + str(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)))\n",
        "        tf.logging.debug(\"Graph operations: \" + str(p[-1].values()) + \"\\n\")\n",
        "        tf.summary.histogram('encoder', graph.get_tensor_by_name('LEt_SNE/encoder/kernel:0'))\n",
        "        tf.summary.histogram('decoder', graph.get_tensor_by_name('LEt_SNE/Reconstruction/decoder/kernel:0'))\n",
        "\n",
        "        # Plot the l2-norm for debugging purposes\n",
        "        encode_tensor = graph.get_tensor_by_name('LEt_SNE/encoder/BiasAdd:0')\n",
        "        encode_l2_norm = tf.reduce_mean(tf.norm(encode_tensor, ord='euclidean', axis=1))\n",
        "        tf.summary.scalar('Encoding_l2_norm', encode_l2_norm)\n",
        "        if params['convolution']:\n",
        "            batch_norm_2_tensor = graph.get_tensor_by_name('LEt_SNE/batch_norm_2/batchnorm/add_1:0')\n",
        "            bn_2_l2_norm = tf.reduce_mean(tf.norm(batch_norm_2_tensor, ord='euclidean', axis=1))\n",
        "            tf.summary.scalar('BatchNorm_2_l2_norm', bn_2_l2_norm)\n",
        "\n",
        "        weights_encoder = graph.get_tensor_by_name('LEt_SNE/encoder/kernel:0')\n",
        "\n",
        "        return encoder, decoder, weights_encoder\n",
        "\n",
        "    def graph_distance(input_features):\n",
        "        '''\n",
        "        Computes the pairwise SQUARED distance between 2 samples in the mini-batch.\n",
        "        This is given by D[i,j] = (a[i]-a[j])(a[i]-a[j])', which simplifies to D = r - 2 A A' + r' in the matrix form.\n",
        "        Similar to https://stackoverflow.com/a/37040451\n",
        "        :return: (BATCH_SIZE x BATCH_SIZE) Tensor containing pairwise distance between the samples.\n",
        "        '''\n",
        "        # Similar to https://stackoverflow.com/a/37040451\n",
        "        with tf.name_scope(\"Graph_Distance\"):\n",
        "\n",
        "            input_features = tf.layers.Flatten()(input_features)\n",
        "            feature_square_l2 = tf.reduce_sum(input_features * input_features, 1)\n",
        "            feature_square_l2 = tf.reshape(feature_square_l2, [-1, 1])\n",
        "            feature = tf.reshape(input_features, [BATCH_SIZE, -1])\n",
        "            dist = feature_square_l2 - (2 * tf.matmul(feature, tf.transpose(feature))) + tf.transpose(feature_square_l2)\n",
        "\n",
        "            return dist\n",
        "\n",
        "    def permutation(indices):\n",
        "        '''\n",
        "        Genrates all possible input permutations, taken two at a time: nP2\n",
        "        :param indices: 1D Tensor, contains indices of samples having same class\n",
        "        :return: 2D Tensor, returns permutations.\n",
        "        '''\n",
        "        with tf.name_scope('Permutation'):\n",
        "            # Similar to https://stackoverflow.com/a/48225468\n",
        "            tile_a = tf.tile(tf.expand_dims(indices, 1), [1, tf.shape(indices)[0]])\n",
        "            tile_a = tf.expand_dims(tile_a, 2)\n",
        "            tile_b = tf.tile(tf.expand_dims(indices, 0), [tf.shape(indices)[0], 1])\n",
        "            tile_b = tf.expand_dims(tile_b, 2)\n",
        "            cart = tf.concat([tile_a, tile_b], axis=2)\n",
        "            cart = tf.reshape(cart, [-1, 2])\n",
        "            return cart\n",
        "\n",
        "    def graph_adjacency(input_features):\n",
        "        '''\n",
        "        Computes the adjacency matrix of the graph by connecting the K Nearest neighbours to the sample.\n",
        "        The adjacency matrix is symmetric in nature, computes element-wise maximum of Adj_Matrix and Adj_Matrix_T\n",
        "        :param input_features: (3D Tensor) input to the model\n",
        "        :return: (BATCH_SIZE x BATCH_SIZE) Tensor which is the Adjacency matrix\n",
        "        '''\n",
        "        with tf.name_scope(\"Adjacency\"):\n",
        "            if NUM_NEIGHBOURS:\n",
        "                # Compute adjacency using euclidean nearest neigbours\n",
        "                tf.logging.info(\"Computing Graph Laplacian using k-Nearest Neighbours\")\n",
        "                distance_matrix = graph_distance(input_features)\n",
        "\n",
        "                # Similar to https://stackoverflow.com/a/49966843\n",
        "                values, index = tf.math.top_k(input=(-1 * distance_matrix), k=(NUM_NEIGHBOURS + 1), sorted=True)\n",
        "                index_one_hot = tf.one_hot(index, depth=BATCH_SIZE)\n",
        "                index_reduced = tf.reduce_sum(index_one_hot, axis=-2)\n",
        "                index = tf.where(tf.not_equal(index_reduced, 0))\n",
        "\n",
        "                adjacency_matrix = tf.scatter_nd(indices=index,\n",
        "                                                 updates=tf.ones([(NUM_NEIGHBOURS + 1) * BATCH_SIZE]),\n",
        "                                                 shape=(BATCH_SIZE, BATCH_SIZE),\n",
        "                                                 name='Graph_Adjacency_kNN')\n",
        "\n",
        "                return tf.maximum(adjacency_matrix, tf.transpose(adjacency_matrix))\n",
        "\n",
        "            else:\n",
        "                # Compute adjacency using LABEL or SEGMENTATION information\n",
        "                tf.logging.info(\"Computing Graph Laplacian using class labels\")\n",
        "\n",
        "                # We split up the labels into segments, each having the same label\n",
        "                graph_labels = tf.reshape(labels, [-1])\n",
        "                graph_index = tf.argsort(graph_labels, axis=0)\n",
        "                graph_labels_sort = tf.sort(graph_labels)\n",
        "                _, __, graph_count = tf.unique_with_counts(graph_labels_sort)\n",
        "                # There is a small chance that the mini batch may have some classes absent.\n",
        "                # Since TensorFlow has static graph, we pad the missing classes with '0'.\n",
        "                # NOTE: We do not know which classes are missing. Our algorithm splits labels having same class.\n",
        "                if SEGMENTATION:\n",
        "                    padding = tf.constant(NUM_REGIONS) - tf.shape(graph_count)[0]\n",
        "                else:\n",
        "                    assert LABEL, \"Segmentation and Label both are False\"\n",
        "                    padding = tf.constant(NUM_CLASSES) - tf.shape(graph_count)[0]\n",
        "                count_pad = tf.pad(graph_count, [[0, padding]])\n",
        "                # Split_indices returns a list containing Tensors, thus we can use python map, since we are mapping\n",
        "                # elements of list to function, and not elements of tensor to function.\n",
        "                if SEGMENTATION:\n",
        "                    split_indices = tf.split(graph_index, count_pad, num=NUM_REGIONS)\n",
        "                else:\n",
        "                    split_indices = tf.split(graph_index, count_pad, num=NUM_CLASSES)\n",
        "\n",
        "                # Label argmax are split into tensors of indices having same class.\n",
        "                # Generate nP2 combinations of indices that denote pointers in the adjacency matrix; which are connected\n",
        "                graph_connected_indices = list(map(permutation, split_indices))\n",
        "                graph_connected_indices = tf.concat(graph_connected_indices, axis=0)\n",
        "                # The shape of graph_connected_indices will defer from batch to batch; which isn't allowed in TensorFlow\n",
        "                # Eg: Batch having one class: shape = BATCH_SIZE ** 2\n",
        "                #     Batch having two class (equal): shape = 2(((BATCH_SIZE) / 2) ** 2)\n",
        "                # Thus we pad this tensor by BATCH_SIZE ** 2 (which is max possible length of the tensor)and truncate it\n",
        "                # Value for padding is chosen as BATCH_SIZE\n",
        "                graph_connected_indices = tf.pad(tensor=graph_connected_indices,\n",
        "                                                 paddings=[[0, (BATCH_SIZE ** 2)], [0, 0]],\n",
        "                                                 constant_values=BATCH_SIZE)  # pad along rows\n",
        "                graph_connected_indices = graph_connected_indices[:(BATCH_SIZE ** 2), :]\n",
        "\n",
        "                adjacency_matrix = tf.scatter_nd(indices=graph_connected_indices,\n",
        "                                                 updates=tf.ones([BATCH_SIZE ** 2]),\n",
        "                                                 shape=(BATCH_SIZE + 1, BATCH_SIZE + 1),\n",
        "                                                 name='Graph_Adjacency_Classes')\n",
        "                corrected_adjacency = adjacency_matrix[:BATCH_SIZE, :BATCH_SIZE]\n",
        "\n",
        "                if LOG_TENSOR[0]:\n",
        "                    global LOG_TENSOR_DICT\n",
        "                    LOG_TENSOR_DICT[\"Graph_Adjacency\"] = corrected_adjacency\n",
        "\n",
        "                return corrected_adjacency\n",
        "\n",
        "    def graph_laplacian(input_features):\n",
        "        '''\n",
        "        Computes the Graph Laplacian L = D - A where L is the Laplacian, D is the degree of the graph, A is adjacency\n",
        "        :param input_features: (3D Tensor) input to the model\n",
        "        :return: (BATCH_SIZE x BATCH_SIZE) Tensor which is the Laplacian matrix\n",
        "        '''\n",
        "        with tf.variable_scope(\"Laplacian\", reuse=False):\n",
        "            adjacency = graph_adjacency(input_features)\n",
        "            degree = tf.linalg.diag(tf.reduce_sum(adjacency, -1))\n",
        "            laplacian = degree - adjacency\n",
        "\n",
        "            adjacency = tf.stop_gradient(adjacency)\n",
        "            degree = tf.stop_gradient(degree)\n",
        "            laplacian = tf.stop_gradient(laplacian)\n",
        "\n",
        "            return laplacian, degree, adjacency\n",
        "\n",
        "    def kl_divergence(num, den):\n",
        "        '''\n",
        "        Computes the KL divergence between two distributions A and B\n",
        "        :param a: 2D-Tensor; First probability distribution\n",
        "        :param b: 2D-Tensor; Second probability distribution\n",
        "        :return: Scalar; The KL divergence between a and b\n",
        "        '''\n",
        "        kl = tf.math.reduce_sum(tf.multiply(num, tf.log(tf.divide(num, den))))\n",
        "        return kl\n",
        "\n",
        "    def shannon_entropy(probs_array):\n",
        "        '''\n",
        "        Computes the Shannon Entropy for a distribution\n",
        "        :param probs_array: 2D-Tensor; Probability distribution along axis=1\n",
        "        :return: Scalar; H(p)\n",
        "        '''\n",
        "        h_shannon = tf.cast(tf.reduce_sum(-probs_array * (tf.log(probs_array)), axis=1), dtype=tf.float32)\n",
        "        return h_shannon\n",
        "\n",
        "    def js_divergence(probs_array):\n",
        "        '''\n",
        "        Computes the JS divergence between the compressed representation of the neurons\n",
        "        Our goal is to maximize the divergence causing every neuron to encode\n",
        "        different information from the rest.\n",
        "        :param probs_array: (BARCH_SIZE x DIM_REDUCED) representation of hidden neurons\n",
        "        :return: Scalar; JS divergence between the bands\n",
        "        '''\n",
        "        # Convert them to probability values, all elements in P_array are +ve due to sigmoid activation\n",
        "        normalizer = tf.reduce_sum(probs_array, axis=0)\n",
        "        probs_normalized = probs_array / normalizer\n",
        "\n",
        "        probs_mean = tf.reduce_mean(probs_normalized, axis=1)\n",
        "\n",
        "        # Index to access individual neurons in the hidden layer\n",
        "        i = tf.constant(0)\n",
        "        # Tensor to store the accumulated value of individual hidden neuron entropy\n",
        "        entropy_sum = tf.constant(0.0, dtype=tf.float32)\n",
        "\n",
        "        # Iterate over all the hidden bands\n",
        "        def loop_condition(i, entropy_sum):\n",
        "            return tf.less(i, DIM_REDUCED)\n",
        "\n",
        "        # Parallel accumulation of individual band entropy into entropy_sum\n",
        "        def loop_body(i, entropy_sum):\n",
        "            entropy_sum = entropy_sum + shannon_entropy(probs_normalized[:, i])\n",
        "            return i + 1, entropy_sum\n",
        "\n",
        "        _, entropy_sum_while = tf.while_loop(loop_condition, loop_body, [i, entropy_sum], back_prop=True)\n",
        "        js = shannon_entropy(probs_mean) - (entropy_sum_while / DIM_REDUCED)\n",
        "        return js\n",
        "\n",
        "    def p_conditional(pairwise_dist, adjacency):\n",
        "        '''\n",
        "        Computes P(j|i) using Binary Search\n",
        "        :param pairwise_dist: (2D Tensor) pairwise distances between samples\n",
        "        :return: 2D Tensor containing conditional probabilities\n",
        "        '''\n",
        "        # Estimating P i|j in the original space\n",
        "        with tf.variable_scope('P_ij_conditional'):\n",
        "            lower = tf.zeros([BATCH_SIZE])\n",
        "            upper = 500 * tf.ones([BATCH_SIZE])\n",
        "            pair_dist_neg = -pairwise_dist\n",
        "            perplexity = PERPLEXITY * tf.ones([BATCH_SIZE])\n",
        "\n",
        "            def calc_probs_perp(lower_bound, upper_bound):\n",
        "                with tf.variable_scope('t-SNE/calc_perplexity'):\n",
        "                    # Binary search\n",
        "                    sigmas = (lower_bound + upper_bound) / 2\n",
        "                    variance = tf.reshape(tf.square(sigmas), [-1, 1])\n",
        "                    scaled_pair_dist_neg = tf.divide(pair_dist_neg, 2 * variance)\n",
        "                    probs_unnormalized = tf.math.exp(scaled_pair_dist_neg)\n",
        "\n",
        "                    # probability of point choosing itself as neighbour ~ 0\n",
        "                    probs_unnormalized = tf.matrix_set_diag(probs_unnormalized, tf.zeros([BATCH_SIZE]))\n",
        "                    softmax = tf.divide(probs_unnormalized, tf.reduce_sum(probs_unnormalized, axis=1, keepdims=True))\n",
        "\n",
        "                    # clipping to avoid NaN values during entropy computation\n",
        "                    softmax_clip = tf.clip_by_value(softmax, clip_value_min=1e-15, clip_value_max=1.)\n",
        "\n",
        "                    entropy = shannon_entropy(softmax_clip)\n",
        "                    perplexity_hat = tf.math.pow(2 * tf.ones([BATCH_SIZE]), entropy)\n",
        "\n",
        "                    return perplexity_hat, softmax_clip\n",
        "\n",
        "            perplexity_initial, probs = calc_probs_perp(lower, upper)\n",
        "\n",
        "            def cond(_, __, perplexity_hat, ___):\n",
        "                mask = tf.math.less(tf.abs(perplexity_hat - perplexity), TOLERANCE * tf.ones([BATCH_SIZE]))\n",
        "                return tf.math.logical_not(tf.reduce_all(mask))\n",
        "\n",
        "            def body(upper, lower, estimate_perplexity, _):\n",
        "                # tf.cast(): True --> 1    False --> 0\n",
        "                mask_gt = tf.math.greater(estimate_perplexity, perplexity)\n",
        "                mask_gt = tf.cast(mask_gt, tf.float32)\n",
        "                upper_update = upper - tf.multiply(mask_gt, (upper - lower) / 2)\n",
        "\n",
        "                mask_lt = tf.math.less(estimate_perplexity, perplexity)\n",
        "                mask_lt = tf.cast(mask_lt, tf.float32)\n",
        "                lower_update = lower + tf.multiply(mask_lt, (upper - lower) / 2)\n",
        "\n",
        "                upper = upper_update\n",
        "                lower = lower_update\n",
        "\n",
        "                perplexity_hat, probs = calc_probs_perp(lower, upper)\n",
        "\n",
        "                return upper, lower, perplexity_hat, probs\n",
        "\n",
        "            # Keeps computing till all values converge with the tolerance limit for perplexity or max iter reached\n",
        "            lower_bound, upper_bound, perplexity_hat, probs_while = tf.while_loop(cond=cond,\n",
        "                                                                                  body=body,\n",
        "                                                                                  loop_vars=(upper,\n",
        "                                                                                             lower,\n",
        "                                                                                             perplexity_initial,\n",
        "                                                                                             probs),\n",
        "                                                                                  parallel_iterations=1,\n",
        "                                                                                  maximum_iterations=1000)\n",
        "\n",
        "            if LOG_TENSOR[1]:\n",
        "                global LOG_TENSOR_DICT\n",
        "                LOG_TENSOR_DICT[\"Perplexity\"] = perplexity_hat\n",
        "\n",
        "            # Compressing space around a point by boosting probability values of neighbours having same class\n",
        "            if COMPRESSION_FACTOR != 1:\n",
        "                probs_while = tf.matrix_set_diag(probs_while, tf.zeros([BATCH_SIZE]))\n",
        "                probs_while = tf.multiply(probs_while, ((COMPRESSION_FACTOR - 1) * adjacency) + 1)\n",
        "                probs_while = tf.divide(probs_while, tf.reduce_sum(probs_while, axis=1, keepdims=True))\n",
        "                probs_while = tf.clip_by_value(probs_while, clip_value_min=1e-15, clip_value_max=1.)\n",
        "            probs_while = tf.stop_gradient(probs_while)\n",
        "            return probs_while\n",
        "\n",
        "    def q_conditional(pairwise_dist, cauchy_dist=True, gaussian_dist=False):\n",
        "        '''\n",
        "        Computes Q(j|i) using Binary Search\n",
        "        :param pairwise_dist: (2D Tensor) pairwise distances between samples\n",
        "        :return: 2D Tensor containing conditional probabilities\n",
        "        '''\n",
        "        # Estimating Q i|j in the latent space\n",
        "        with tf.variable_scope('Q_ij_conditional'):\n",
        "            if cauchy_dist:\n",
        "                cauchy = 1 + pairwise_dist\n",
        "                cauchy = tf.math.pow(x=cauchy, y=(-1 * tf.ones([BATCH_SIZE, BATCH_SIZE])))\n",
        "                cauchy = tf.matrix_set_diag(cauchy, tf.zeros([BATCH_SIZE]))\n",
        "                cauchy = tf.divide(cauchy, tf.reduce_sum(cauchy, axis=1, keepdims=True))\n",
        "                cauchy_clip = tf.clip_by_value(cauchy, clip_value_min=1e-15, clip_value_max=1.)\n",
        "                return cauchy_clip\n",
        "\n",
        "            if gaussian_dist:\n",
        "                pair_dist_neg = -pairwise_dist\n",
        "                variance = 1 * tf.ones(shape=[BATCH_SIZE, 1])\n",
        "                scaled_pair_dist_neg = tf.divide(pair_dist_neg, 2 * variance)\n",
        "                probs_unnormalized = tf.math.exp(scaled_pair_dist_neg)\n",
        "\n",
        "                # probability of point choosing itself as neighbour ~ 0\n",
        "                probs_unnormalized = tf.matrix_set_diag(probs_unnormalized, tf.zeros([BATCH_SIZE]))\n",
        "                softmax = tf.divide(probs_unnormalized, tf.reduce_sum(probs_unnormalized, axis=1, keepdims=True))\n",
        "\n",
        "                # clipping to avoid NaN values during entropy computation\n",
        "                softmax_clip = tf.clip_by_value(softmax, clip_value_min=1e-15, clip_value_max=1.)\n",
        "                return softmax_clip\n",
        "\n",
        "    def inv_tsne(input_features, embedding, adjacency):\n",
        "        '''\n",
        "        KL(p||q) for manifold, KL(q|p) for segmentation + labelling\n",
        "        :param input_features: (3D Tensor) input batch\n",
        "        :param embedding: (3D Tensor) network output\n",
        "        :param adjacency: (2D Tensor) Adjacency map\n",
        "        :return: (scalar Tensor) KL Divergence\n",
        "        '''\n",
        "        # Eliminates extra dimension introduced by Conv1D\n",
        "        # input_features = tf.layers.Flatten()(input_features)\n",
        "\n",
        "        # Computes the pairwise distance between samples in the original and reduced space\n",
        "        input_l2_pair = graph_distance(input_features)\n",
        "        embedding_l2_pair = graph_distance(embedding)\n",
        "\n",
        "        if LOG_TENSOR[2]:\n",
        "            global LOG_TENSOR_DICT\n",
        "            LOG_TENSOR_DICT[\"Embedding_l2_pair\"] = embedding_l2_pair\n",
        "\n",
        "        probs_p = p_conditional(tf.cast(input_l2_pair, dtype=tf.float32), adjacency)\n",
        "        probs_q = q_conditional(embedding_l2_pair, cauchy_dist=True)\n",
        "\n",
        "        if MANIFOLD:\n",
        "            kl_p_q = kl_divergence(num=probs_p, den=probs_q)\n",
        "            return kl_p_q\n",
        "        else:\n",
        "            kl_q_p = kl_divergence(num=probs_q, den=probs_p)\n",
        "            return kl_q_p\n",
        "\n",
        "    if mode == tf.estimator.ModeKeys.TRAIN:\n",
        "        (embedding, reconstruction, weights_encoder) = model(input_features=features, bn_train=True)\n",
        "    else:\n",
        "        (embedding, reconstruction, weights_encoder) = model(input_features=features, bn_train=False)\n",
        "\n",
        "    # Compute predictions.\n",
        "    with tf.variable_scope(\"LEt_SNE/Prediction\"):\n",
        "        if mode == tf.estimator.ModeKeys.PREDICT:\n",
        "            predictions = {'encoded': embedding}\n",
        "\n",
        "            return tf.estimator.EstimatorSpec(mode, predictions=predictions)\n",
        "\n",
        "    with tf.variable_scope(\"LEt_SNE/Losses\"):\n",
        "\n",
        "        embedding = tf.cast(embedding, dtype=tf.float32)\n",
        "\n",
        "        # Loss as defined by the Laplacian Eigenmaps\n",
        "        laplacian, degree, adjacency = graph_laplacian(input_features=features)\n",
        "        loss_laplacian = tf.trace(tf.matmul(tf.matmul(tf.transpose(embedding), laplacian), embedding))\n",
        "\n",
        "        # Loss for Reconstruction (Autoencoder)\n",
        "        loss_reconstruction = tf.cast(\n",
        "            tf.losses.mean_squared_error(labels=features,\n",
        "                                         predictions=tf.reshape(reconstruction, [-1, NUM_BANDS, 1]),\n",
        "                                         reduction=tf.losses.Reduction.SUM),\n",
        "            dtype=tf.float32)\n",
        "\n",
        "        # Modified t-SNE loss\n",
        "        loss_inv_tsne = inv_tsne(features, embedding, adjacency)\n",
        "\n",
        "        total_loss = tf.convert_to_tensor([loss_laplacian, loss_inv_tsne, loss_reconstruction])\n",
        "\n",
        "        if LOG_TENSOR[3]:\n",
        "            global LOG_TENSOR_DICT\n",
        "            LOG_TENSOR_DICT[\"Encoding\"] = embedding\n",
        "\n",
        "        tf.summary.scalar('Laplacian_loss', loss_laplacian * params['loss_weights'][0])\n",
        "        tf.summary.scalar('Inverse_tSNE_loss', loss_inv_tsne * params['loss_weights'][1])\n",
        "        tf.summary.scalar('Reconstruction_loss', loss_reconstruction * params['loss_weights'][2])\n",
        "\n",
        "    with tf.variable_scope(\"LEt_SNE/Train\"):\n",
        "        optimizer = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE, name='Adam')\n",
        "        loss = tf.reduce_sum(tf.multiply(params['loss_weights'], total_loss))\n",
        "        train = optimizer.minimize(loss=loss,\n",
        "                                   global_step=tf.train.get_or_create_global_step(),\n",
        "                                   name='Adam_minimizer')\n",
        "        if mode == tf.estimator.ModeKeys.TRAIN:\n",
        "            update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n",
        "            with tf.control_dependencies(update_ops):\n",
        "                loss = tf.identity(loss)\n",
        "            if True in LOG_TENSOR:\n",
        "                logging_hook = tf.train.LoggingTensorHook(LOG_TENSOR_DICT, every_n_iter=500)\n",
        "                return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train, training_hooks=[logging_hook])\n",
        "            else:\n",
        "                return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train)\n"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "PJgnhNaFgZi5",
        "colab_type": "text"
      },
      "source": [
        "We define a new function to obtain predictions from LEt-SNE custom estimator"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ttJXdub0gY4N",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def predictor(bands, labels, model, type):\n",
        "    '''\n",
        "    NUMPY AND TENSORFLOW\n",
        "    '''\n",
        "    assert type == 'encoded', \"Choose valid Prediction\"\n",
        "    generator = model.predict(input_fn=lambda: dataset_tf(bands, labels, train=False))\n",
        "    prediction = np.empty(shape=(0, DIM_REDUCED), dtype=np.float32)\n",
        "    while True:\n",
        "        try:\n",
        "            prediction = np.append(prediction,\n",
        "                                   np.reshape(next(generator)[type], [1, -1]),\n",
        "                                   axis=0)\n",
        "        except StopIteration:\n",
        "            break\n",
        "    assert prediction.shape[0] == bands.shape[0], \"Predictions not present for all bands\"\n",
        "    return prediction\n"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "34jQLa9TgifH",
        "colab_type": "text"
      },
      "source": [
        "## Part 3: Creating the Dataset \n",
        "\n",
        "We simulate a two-class dataset drawn from the equation of a circle"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "xqDvtZ__gyNI",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 785
        },
        "outputId": "273a2909-9f29-41ad-d49d-63d66607cb8e"
      },
      "source": [
        "def artificial_dataset():\n",
        "    '''\n",
        "\n",
        "    :return:\n",
        "    '''\n",
        "    import logging\n",
        "    global NUM_BANDS, NUM_CLASSES\n",
        "    NUM_BANDS = 2\n",
        "    NUM_CLASSES = 2\n",
        "\n",
        "    positive_class = []\n",
        "    negative_class = []\n",
        "\n",
        "    @np.vectorize\n",
        "    def circle(x):\n",
        "        return np.sqrt(1 - (x**2))\n",
        "\n",
        "    x_range = np.arange(-0.8, 0.800001, 0.000001)\n",
        "\n",
        "    # Positive Class\n",
        "    x = np.random.choice(a=x_range, size=10000)\n",
        "    positive_x = x[:5000]\n",
        "    negative_x = x[5000:]\n",
        "\n",
        "    positive_y = circle(positive_x)\n",
        "    negative_y = -circle(negative_x)\n",
        "\n",
        "    labels = np.concatenate([np.ones(5000), np.zeros(5000)], axis=0).reshape(-1, 1)\n",
        "    y = np.concatenate([positive_y, negative_y]).reshape(-1, 1)\n",
        "    x = x.reshape(-1, 1)\n",
        "\n",
        "    dataset = np.concatenate([x, y, labels], axis=-1)\n",
        "\n",
        "    color=['r', 'b']\n",
        "\n",
        "    for label in [0, 1]:\n",
        "        indices = np.where(dataset[:, -1] == label)\n",
        "        x = dataset[indices, 0]\n",
        "        y = dataset[indices, 1]\n",
        "        plt.plot(x, y, marker='o', linestyle=\"\", c=color[label])\n",
        "        plt.title('Before Compression')\n",
        "\n",
        "    plt.show()\n",
        "    return dataset[:, :2], dataset[:, 2]\n",
        "\n",
        "train_data, train_label = artificial_dataset()\n",
        "train_label = train_label.astype(np.int32)"
      ],
      "execution_count": 8,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 885.827x885.827 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "WLsreZJwiFep",
        "colab_type": "text"
      },
      "source": [
        "## Part 4: Training LEt-SNE"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Py-_C25WiJLP",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 1000
        },
        "outputId": "8685576e-7ce8-4b4b-9620-d0d3a093623e"
      },
      "source": [
        "let_sne_ = tf.estimator.Estimator(model_fn=let_sne,\n",
        "                                  model_dir=os.path.join(TENSORBOARD_DIR, 'LEt_SNE'),\n",
        "                                  config=tf.estimator.RunConfig(session_config=config,\n",
        "                                                                keep_checkpoint_max=20),\n",
        "                                  params={'hidden': HIDDEN,\n",
        "                                          'loss_weights': LOSS_WEIGHTS,\n",
        "                                          'convolution': False})\n",
        "let_sne_.train(input_fn=lambda: dataset_tf(train_data, train_label, train=True))"
      ],
      "execution_count": 9,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "INFO:tensorflow:Using config: {'_model_dir': '/content/TensorBoard_CF/LEt_SNE', '_tf_random_seed': None, '_save_summary_steps': 100, '_save_checkpoints_steps': None, '_save_checkpoints_secs': 600, '_session_config': allow_soft_placement: true\n",
            "graph_options {\n",
            "  rewrite_options {\n",
            "    meta_optimizer_iterations: ONE\n",
            "  }\n",
            "}\n",
            ", '_keep_checkpoint_max': 20, '_keep_checkpoint_every_n_hours': 10000, '_log_step_count_steps': 100, '_train_distribute': None, '_device_fn': None, '_protocol': None, '_eval_distribute': None, '_experimental_distribute': None, '_experimental_max_worker_delay_secs': None, '_session_creation_timeout_secs': 7200, '_service': None, '_cluster_spec': <tensorflow.python.training.server_lib.ClusterSpec object at 0x7fb4aa455400>, '_task_type': 'worker', '_task_id': 0, '_global_id_in_cluster': 0, '_master': '', '_evaluation_master': '', '_is_chief': True, '_num_ps_replicas': 0, '_num_worker_replicas': 1}\n",
            "WARNING:tensorflow:From /tensorflow-1.15.2/python3.6/tensorflow_core/python/training/training_util.py:236: Variable.initialized_value (from tensorflow.python.ops.variables) is deprecated and will be removed in a future version.\n",
            "Instructions for updating:\n",
            "Use Variable.read_value. Variables in 2.X are initialized automatically both in eager and graph (inside tf.defun) contexts.\n",
            "WARNING:tensorflow:Entity <function dataset_tf.<locals>.<lambda> at 0x7fb4431c40d0> could not be transformed and will be executed as-is. Please report this to the AutoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: module 'gast' has no attribute 'Str'\n",
            "WARNING: Entity <function dataset_tf.<locals>.<lambda> at 0x7fb4431c40d0> could not be transformed and will be executed as-is. Please report this to the AutoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: module 'gast' has no attribute 'Str'\n",
            "INFO:tensorflow:Batches created\n",
            "WARNING:tensorflow:From <ipython-input-5-a7b0babd229b>:22: DatasetV1.output_types (from tensorflow.python.data.ops.dataset_ops) is deprecated and will be removed in a future version.\n",
            "Instructions for updating:\n",
            "Use `tf.compat.v1.data.get_output_types(dataset)`.\n",
            "WARNING:tensorflow:From <ipython-input-5-a7b0babd229b>:23: DatasetV1.output_shapes (from tensorflow.python.data.ops.dataset_ops) is deprecated and will be removed in a future version.\n",
            "Instructions for updating:\n",
            "Use `tf.compat.v1.data.get_output_shapes(dataset)`.\n",
            "INFO:tensorflow:1. Dataset element \n",
            "Type: (tf.float64, tf.int32) \n",
            "Shape: (TensorShape([Dimension(1024), Dimension(2), Dimension(1)]), TensorShape([Dimension(1024)]))\n",
            "INFO:tensorflow:Calling model_fn.\n",
            "WARNING:tensorflow:From <ipython-input-6-e1e4fa0a95d7>:32: batch_normalization (from tensorflow.python.layers.normalization) is deprecated and will be removed in a future version.\n",
            "Instructions for updating:\n",
            "Use keras.layers.BatchNormalization instead.  In particular, `tf.control_dependencies(tf.GraphKeys.UPDATE_OPS)` should not be used (consult the `tf.keras.layers.batch_normalization` documentation).\n",
            "WARNING:tensorflow:From /tensorflow-1.15.2/python3.6/tensorflow_core/python/layers/normalization.py:327: Layer.apply (from tensorflow.python.keras.engine.base_layer) is deprecated and will be removed in a future version.\n",
            "Instructions for updating:\n",
            "Please use `layer.__call__` method instead.\n",
            "INFO:tensorflow:Scale of 0 disables regularizer.\n",
            "INFO:tensorflow:Scale of 0 disables regularizer.\n",
            "INFO:tensorflow:Computing Graph Laplacian using class labels\n",
            "WARNING:tensorflow:From /tensorflow-1.15.2/python3.6/tensorflow_core/python/ops/math_grad.py:1424: where (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\n",
            "Instructions for updating:\n",
            "Use tf.where in 2.0, which has the same broadcast rule as np.where\n",
            "INFO:tensorflow:Done calling model_fn.\n",
            "INFO:tensorflow:Create CheckpointSaverHook.\n",
            "INFO:tensorflow:Graph was finalized.\n",
            "INFO:tensorflow:Running local_init_op.\n",
            "INFO:tensorflow:Done running local_init_op.\n",
            "INFO:tensorflow:Saving checkpoints for 0 into /content/TensorBoard_CF/LEt_SNE/model.ckpt.\n",
            "INFO:tensorflow:loss = 18992.008, step = 0\n",
            "INFO:tensorflow:global_step/sec: 6.67023\n",
            "INFO:tensorflow:loss = 9755.826, step = 100 (14.996 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.77838\n",
            "INFO:tensorflow:loss = 6573.068, step = 200 (14.751 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.74308\n",
            "INFO:tensorflow:loss = 5626.892, step = 300 (14.832 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.80518\n",
            "INFO:tensorflow:loss = 5192.543, step = 400 (14.696 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.73233\n",
            "INFO:tensorflow:loss = 5233.556, step = 500 (14.851 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.72248\n",
            "INFO:tensorflow:loss = 5104.8193, step = 600 (14.875 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.75401\n",
            "INFO:tensorflow:loss = 5101.712, step = 700 (14.807 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.82755\n",
            "INFO:tensorflow:loss = 4864.368, step = 800 (14.646 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.79763\n",
            "INFO:tensorflow:loss = 4872.524, step = 900 (14.713 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.9301\n",
            "INFO:tensorflow:loss = 4775.3857, step = 1000 (14.431 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.84689\n",
            "INFO:tensorflow:loss = 4670.9653, step = 1100 (14.603 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.92513\n",
            "INFO:tensorflow:loss = 4741.012, step = 1200 (14.440 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.79328\n",
            "INFO:tensorflow:loss = 4544.957, step = 1300 (14.723 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.74373\n",
            "INFO:tensorflow:loss = 4530.3315, step = 1400 (14.826 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.79146\n",
            "INFO:tensorflow:loss = 4418.61, step = 1500 (14.727 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.73425\n",
            "INFO:tensorflow:loss = 4322.845, step = 1600 (14.850 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.78197\n",
            "INFO:tensorflow:loss = 4220.448, step = 1700 (14.744 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.77549\n",
            "INFO:tensorflow:loss = 4232.0137, step = 1800 (14.761 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.78543\n",
            "INFO:tensorflow:loss = 4141.2344, step = 1900 (14.736 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.78507\n",
            "INFO:tensorflow:loss = 4129.114, step = 2000 (14.736 sec)\n",
            "INFO:tensorflow:global_step/sec: 7.02234\n",
            "INFO:tensorflow:loss = 4042.1895, step = 2100 (14.242 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.81234\n",
            "INFO:tensorflow:loss = 3978.5986, step = 2200 (14.681 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.8136\n",
            "INFO:tensorflow:loss = 4015.4272, step = 2300 (14.676 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.71955\n",
            "INFO:tensorflow:loss = 3895.264, step = 2400 (14.879 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.81783\n",
            "INFO:tensorflow:loss = 3864.3787, step = 2500 (14.667 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.91334\n",
            "INFO:tensorflow:loss = 3726.9272, step = 2600 (14.465 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.76099\n",
            "INFO:tensorflow:loss = 3704.546, step = 2700 (14.793 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.81974\n",
            "INFO:tensorflow:loss = 3664.4934, step = 2800 (14.664 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.94113\n",
            "INFO:tensorflow:loss = 3652.0244, step = 2900 (14.406 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.92805\n",
            "INFO:tensorflow:loss = 3588.1099, step = 3000 (14.434 sec)\n",
            "INFO:tensorflow:global_step/sec: 7.56936\n",
            "INFO:tensorflow:loss = 3539.5332, step = 3100 (13.208 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.95931\n",
            "INFO:tensorflow:loss = 3636.4036, step = 3200 (14.369 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.99319\n",
            "INFO:tensorflow:loss = 3449.7405, step = 3300 (14.300 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.93601\n",
            "INFO:tensorflow:loss = 3387.69, step = 3400 (14.420 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.86424\n",
            "INFO:tensorflow:loss = 3482.724, step = 3500 (14.568 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.86987\n",
            "INFO:tensorflow:loss = 3342.6077, step = 3600 (14.558 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.67851\n",
            "INFO:tensorflow:loss = 3321.3933, step = 3700 (14.972 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.76012\n",
            "INFO:tensorflow:loss = 3298.5784, step = 3800 (14.790 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.74351\n",
            "INFO:tensorflow:loss = 3239.4946, step = 3900 (14.832 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.72938\n",
            "INFO:tensorflow:loss = 3230.463, step = 4000 (14.861 sec)\n",
            "INFO:tensorflow:Saving checkpoints for 4062 into /content/TensorBoard_CF/LEt_SNE/model.ckpt.\n",
            "INFO:tensorflow:global_step/sec: 6.7\n",
            "INFO:tensorflow:loss = 3230.4016, step = 4100 (14.925 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.96163\n",
            "INFO:tensorflow:loss = 3176.999, step = 4200 (14.364 sec)\n",
            "INFO:tensorflow:global_step/sec: 7.35454\n",
            "INFO:tensorflow:loss = 3158.763, step = 4300 (13.596 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.68297\n",
            "INFO:tensorflow:loss = 3143.201, step = 4400 (14.965 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.70541\n",
            "INFO:tensorflow:loss = 3115.1497, step = 4500 (14.915 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.60422\n",
            "INFO:tensorflow:loss = 3067.1763, step = 4600 (15.138 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.74595\n",
            "INFO:tensorflow:loss = 3012.9343, step = 4700 (14.827 sec)\n",
            "INFO:tensorflow:global_step/sec: 6.6216\n",
            "INFO:tensorflow:loss = 3099.2495, step = 4800 (15.101 sec)\n",
            "INFO:tensorflow:Saving checkpoints for 4882 into /content/TensorBoard_CF/LEt_SNE/model.ckpt.\n",
            "INFO:tensorflow:Loss for final step: 3002.3801.\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<tensorflow_estimator.python.estimator.estimator.Estimator at 0x7fb44ab75d68>"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 9
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "F5rxspNJiYgp",
        "colab_type": "text"
      },
      "source": [
        "## Part 5: Evaluating embeddings"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "5y8A2_dEiXwY",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 283
        },
        "outputId": "00b8f563-6115-4e2c-895c-e44730812963"
      },
      "source": [
        "# It does't matter what labels you provide for prediction, since graph pertaining to loss is not executed\n",
        "train_encoding = predictor(train_data, train_label, let_sne_, 'encoded')\n",
        "\n",
        "labels = np.concatenate([np.ones(5000), np.zeros(5000)], axis=0).reshape(-1, 1)\n",
        "dataset = np.concatenate([train_encoding, labels], axis=-1)"
      ],
      "execution_count": 10,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "WARNING:tensorflow:Entity <function dataset_tf.<locals>.<lambda> at 0x7fb4431c41e0> could not be transformed and will be executed as-is. Please report this to the AutoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: module 'gast' has no attribute 'Str'\n",
            "WARNING: Entity <function dataset_tf.<locals>.<lambda> at 0x7fb4431c41e0> could not be transformed and will be executed as-is. Please report this to the AutoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: module 'gast' has no attribute 'Str'\n",
            "INFO:tensorflow:Batches created\n",
            "INFO:tensorflow:1. Dataset element \n",
            "Type: (tf.float64, tf.int32) \n",
            "Shape: (TensorShape([Dimension(None), Dimension(2), Dimension(1)]), TensorShape([Dimension(None)]))\n",
            "INFO:tensorflow:Calling model_fn.\n",
            "INFO:tensorflow:Scale of 0 disables regularizer.\n",
            "INFO:tensorflow:Scale of 0 disables regularizer.\n",
            "INFO:tensorflow:Done calling model_fn.\n",
            "INFO:tensorflow:Graph was finalized.\n",
            "INFO:tensorflow:Restoring parameters from /content/TensorBoard_CF/LEt_SNE/model.ckpt-4882\n",
            "INFO:tensorflow:Running local_init_op.\n",
            "INFO:tensorflow:Done running local_init_op.\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "zrKATiuZiw9X",
        "colab_type": "text"
      },
      "source": [
        "## Part 6: Visualizing the embeddings\n",
        "\n",
        "Note the clustering of samples in the embeddings. Also, the scale of the original and embedded samples is different, implying a greater separation between dissimilar classes."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "DIsnEgi8i1aI",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 785
        },
        "outputId": "55e72090-9dcb-446c-a84b-f72e4606b606"
      },
      "source": [
        "color = ['r', 'b']\n",
        "\n",
        "for label in [0, 1]:\n",
        "    indices = np.where(dataset[:, -1] == label)\n",
        "    x = dataset[indices, 0]\n",
        "    y = dataset[indices, 1]\n",
        "    plt.plot(x, y, marker='o', linestyle=\"\", c=color[label])\n",
        "    plt.title('After Compression')\n",
        "\n",
        "plt.show()"
      ],
      "execution_count": 11,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 885.827x885.827 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    }
  ]
}