coarsen_labels.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "33d4366d",
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import scipy.stats\n",
    "import torchvision\n",
    "\n",
    "import datasets\n",
    "import hier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "6606b6dc",
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset = torchvision.datasets.ImageNet('/home/jack/data/torchvision/imagenet/', 'train')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "49fdcee9",
   "metadata": {},
   "outputs": [],
   "source": [
    "with open('resources/hierarchy/imagenet_fiveai.csv') as f:\n",
    "    tree, names = hier.make_hierarchy_from_edges(hier.load_edges(f))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "c46858f1",
   "metadata": {},
   "outputs": [],
   "source": [
    "node_spec = -np.log(tree.num_leaf_descendants() / tree.num_leaf_nodes())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "8abf32db",
   "metadata": {},
   "outputs": [],
   "source": [
    "label_order = tree.leaf_subset()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "7ac1a2ed",
   "metadata": {},
   "outputs": [],
   "source": [
    "paths = tree.paths()\n",
    "paths = [paths[i] for i in label_order]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "41effdd9",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Simplify paths! Remove nodes that have 1 child.\n",
    "is_valid = (tree.num_children() != 1)\n",
    "if not np.all(is_valid):\n",
    "    len_before = np.mean(list(map(len, paths)))\n",
    "    paths = [path[is_valid[path]] for path in paths]\n",
    "    len_after = np.mean(list(map(len, paths)))\n",
    "    print('exclude trivial nodes, mean len', len_before, '->', len_after)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "6e7b5751",
   "metadata": {},
   "outputs": [],
   "source": [
    "specs = [node_spec[path] for path in paths]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "5021748d",
   "metadata": {},
   "outputs": [],
   "source": [
    "spec_fracs = [spec / spec[-1] for spec in specs]\n",
    "out = np.full([len(paths), max(map(len, paths))], np.nan)\n",
    "for i, x in enumerate(spec_fracs):\n",
    "    out[i, :len(x)] = x\n",
    "spec_fracs = out\n",
    "del out"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "1e14b28f",
   "metadata": {},
   "outputs": [],
   "source": [
    "a, b = 2, 1\n",
    "\n",
    "rng = np.random.default_rng(0)\n",
    "# r = rng.beta(a, b, size=[len(dataset)])\n",
    "u = rng.uniform(size=[len(dataset)])\n",
    "r = scipy.stats.beta.ppf(u, a, b)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "02d575db",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.hist(r, range=(0, 1), bins=100);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "9321a55c",
   "metadata": {},
   "outputs": [],
   "source": [
    "def first(condition, axis=-1):\n",
    "    assert np.all(np.any(condition, axis=axis))\n",
    "    return np.argmax(condition, axis=axis)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "8f6c9b95",
   "metadata": {},
   "outputs": [],
   "source": [
    "# first_ge = np.array([first(spec_fracs[y_i] >= r_i) for y_i, r_i in zip(dataset.targets, r)])\n",
    "# coarse_target = np.array([path[index] for path, index in zip(paths, first_ge)])\n",
    "\n",
    "index_first_ge = first(spec_fracs[dataset.targets] >= r[:, None])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "ae046df5",
   "metadata": {},
   "outputs": [],
   "source": [
    "coarse_target = np.asarray([paths[label][index] for label, index in zip(dataset.targets, index_first_ge)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "6dba1096",
   "metadata": {},
   "outputs": [],
   "source": [
    "depths = np.array(list(map(len, paths)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "76b1f9a9",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "def hist(int_vals, **kwargs):\n",
    "    a = np.min(int_vals)\n",
    "    b = np.max(int_vals)\n",
    "    plt.hist(int_vals, bins=np.arange(a, b+2), **kwargs)\n",
    "\n",
    "hist(depths[dataset.targets], alpha=0.5)\n",
    "hist(index_first_ge + 1, alpha=0.5)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "181e444a",
   "metadata": {},
   "outputs": [],
   "source": [
    "%reset -f"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "13c03e1b",
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import scipy.stats\n",
    "import torchvision\n",
    "\n",
    "import datasets\n",
    "import hier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "ef96f9b5",
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset = torchvision.datasets.ImageNet('/home/jack/data/torchvision/imagenet/', 'train')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "8f9c28ab",
   "metadata": {},
   "outputs": [],
   "source": [
    "with open('resources/hierarchy/imagenet_fiveai.csv') as f:\n",
    "    tree, node_names = hier.make_hierarchy_from_edges(hier.load_edges(f))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "d0a60d55",
   "metadata": {},
   "outputs": [],
   "source": [
    "def first(condition, axis=-1):\n",
    "    assert np.all(np.any(condition, axis=axis))\n",
    "    return np.argmax(condition, axis=axis)\n",
    "\n",
    "\n",
    "def hist(int_vals, **kwargs):\n",
    "    a = np.min(int_vals)\n",
    "    b = np.max(int_vals)\n",
    "    plt.hist(int_vals, bins=np.arange(a, b+2), **kwargs)\n",
    "\n",
    "\n",
    "def degrade_labels(tree, label_nodes, targets, a, b, seed=0):\n",
    "    node_spec = -np.log(tree.num_leaf_descendants() / tree.num_leaf_nodes())\n",
    "    paths = tree.paths()\n",
    "\n",
    "    # Take subset of paths for labels.\n",
    "    paths = [paths[i] for i in label_nodes]\n",
    "\n",
    "    # Simplify paths! Remove nodes that have 1 child.\n",
    "    is_valid = (tree.num_children() != 1)\n",
    "    if not np.all(is_valid):\n",
    "        len_before = np.mean(list(map(len, paths)))\n",
    "        paths = [path[is_valid[path]] for path in paths]\n",
    "        len_after = np.mean(list(map(len, paths)))\n",
    "        print('exclude trivial nodes, mean len', len_before, '->', len_after)\n",
    "\n",
    "    specs = [node_spec[path] for path in paths]\n",
    "    spec_fracs = [spec / spec[-1] for spec in specs]\n",
    "    out = np.full([len(paths), max(map(len, paths))], np.nan)\n",
    "    for i, x in enumerate(spec_fracs):\n",
    "        out[i, :len(x)] = x\n",
    "    spec_fracs = out\n",
    "\n",
    "    rng = np.random.default_rng(seed)\n",
    "    u = rng.uniform(size=[len(targets)])\n",
    "    r = scipy.stats.beta.ppf(u, a, b)\n",
    "\n",
    "    # plt.hist(r, range=(0, 1), bins=100)\n",
    "    # plt.show()\n",
    "\n",
    "    depths = np.array(list(map(len, paths)))\n",
    "\n",
    "    index_first_ge = first(spec_fracs[targets] >= r[:, None])\n",
    "    coarse = np.asarray([paths[label][index] for label, index in zip(targets, index_first_ge)])\n",
    "\n",
    "    hist(depths[targets], alpha=0.5)\n",
    "    hist(index_first_ge + 1, alpha=0.5)\n",
    "    plt.show()\n",
    "    \n",
    "    return coarse"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "509860c7",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYkAAAD4CAYAAAAZ1BptAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjUuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/YYfK9AAAACXBIWXMAAAsTAAALEwEAmpwYAAARUklEQVR4nO3df7DldV3H8edL0PJHCgRuxEJrtdggk4obblmNSsBKjksz5uCUbEpuk2jKOBXYTDSaQqVSlNFssrEUQQxq7DQo3kEmpxkXWfwBIgo7qLAbsCuLYDlp6Ls/zmfzcDmfe9d7z95zln0+Zs6c73l/v9/P9333x33dz/f7PeemqpAkaZQnTboBSdL0MiQkSV2GhCSpy5CQJHUZEpKkroMn3cC4HX744bVixYpJtyFJ+5Vbbrnl61V1xOz6Ey4kVqxYwdatWyfdhiTtV5J8bVTd002SpC5DQpLUZUhIkroMCUlSlyEhSeoyJCRJXYaEJKnLkJAkdRkSkqSuJ9w7rqWldtHMnWMd75yTjx3reNJiOJOQJHUZEpKkLkNCktRlSEiSugwJSVKXISFJ6vIWWE2/Gy8Y73gvO2+8443ZuG+pBW+r1cI5k5AkdRkSkqQuQ0KS1GVISJK6DAlJUpchIUnqMiQkSV2GhCSpy5CQJHUZEpKkLkNCktRlSEiSugwJSVKXISFJ6jIkJEld/j4JHXD2xe9rkJ6onElIkroMCUlSlyEhSeqaNySSHJ3kxiRfTHJ7kre2+mFJZpLc1Z4PbfUkuTjJtiS3JjlhaKx1bfu7kqwbqr8oyW1tn4uTZK5jSJKWxt7MJB4F3l5VxwGrgbOTHAecC9xQVSuBG9prgFcAK9tjPXAJDL7hA+cDLwZOBM4f+qZ/CfDGof3WtHrvGJKkJTBvSFTVfVX1mbb8TeAO4ChgLbCpbbYJOL0trwUur4EtwCFJjgROBWaqandVPQTMAGvaumdW1ZaqKuDyWWONOoYkaQn8QNckkqwAXgjcBCyrqvvaqvuBZW35KODeod22t9pc9e0j6sxxDEnSEtjrkEjyDOBDwNuq6pHhdW0GUGPu7THmOkaS9Um2Jtm6a9eufdmGJB1Q9iokkjyZQUBcUVUfbuUH2qki2vPOVt8BHD20+/JWm6u+fER9rmM8RlVtqKpVVbXqiCOO2JsvSZK0F/bm7qYAlwJ3VNX7h1ZtBvbcobQOuHaofma7y2k18HA7ZXQ9cEqSQ9sF61OA69u6R5Ksbsc6c9ZYo44hSVoCe/OxHC8BXgfcluRzrfYO4ELg6iRnAV8DXtPWXQecBmwDvgW8HqCqdid5F3Bz2+6dVbW7Lb8JuAx4KvDR9mCOY0hTY/U9G8Y63pZj1o91PGkx5g2JqvoPIJ3VJ43YvoCzO2NtBDaOqG8Fjh9Rf3DUMSRJS8N3XEuSugwJSVKXISFJ6vL3SeiAM+4LzdITmTMJSVKXISFJ6jIkJEldhoQkqcuQkCR1GRKSpC5DQpLUZUhIkroMCUlSlyEhSeoyJCRJXYaEJKnLkJAkdRkSkqQuQ0KS1GVISJK6DAlJUpchIUnqMiQkSV2GhCSpy5CQJHUZEpKkLkNCktRlSEiSugwJSVKXISFJ6jIkJEldhoQkqcuQkCR1GRKSpC5DQpLUNW9IJNmYZGeSLwzV/iTJjiSfa4/Thtadl2Rbki8nOXWovqbVtiU5d6j+nCQ3tfq/JHlKq/9Qe72trV8xtq9akrRX9mYmcRmwZkT9oqp6QXtcB5DkOOAM4Hltn79NclCSg4APAK8AjgNe27YF+LM21k8DDwFntfpZwEOtflHbTpK0hOYNiar6JLB7L8dbC1xVVd+uqq8A24AT22NbVd1dVd8BrgLWJgnwcuCatv8m4PShsTa15WuAk9r2kqQlsphrEm9Ocms7HXVoqx0F3Du0zfZW69V/FPhGVT06q/6Ysdr6h9v2kqQlcvAC97sEeBdQ7fl9wBvG1dQPKsl6YD3AMcccM6k2BHDjBZPuQNIYLWgmUVUPVNV3q+p7wN8zOJ0EsAM4emjT5a3Wqz8IHJLk4Fn1x4zV1j+rbT+qnw1VtaqqVh1xxBEL+ZIkSSMsKCSSHDn08teAPXc+bQbOaHcmPQdYCXwauBlY2e5kegqDi9ubq6qAG4FXt/3XAdcOjbWuLb8a+ETbXpK0ROY93ZTkSuClwOFJtgPnAy9N8gIGp5u+CvwOQFXdnuRq4IvAo8DZVfXdNs6bgeuBg4CNVXV7O8QfAlcl+VPgs8ClrX4p8I9JtjG4cH7GYr9YSdIPZt6QqKrXjihfOqK2Z/t3A+8eUb8OuG5E/W6+f7pquP4/wK/P158kad/xHdeSpC5DQpLUZUhIkroMCUlSlyEhSeoyJCRJXYaEJKlroZ/dJGk/ctHMnWMd75yTjx3reJpeziQkSV2GhCSpy5CQJHUZEpKkLi9cS1Nm9T0bxj7mlmPWj31MHRicSUiSugwJSVKXISFJ6jIkJEldhoQkqcuQkCR1GRKSpC5DQpLUZUhIkroMCUlSlyEhSerys5s09T5194OTbkE6YDmTkCR1GRKSpC5DQpLUZUhIkroMCUlSlyEhSeoyJCRJXYaEJKnLkJAkdRkSkqQuQ0KS1DVvSCTZmGRnki8M1Q5LMpPkrvZ8aKsnycVJtiW5NckJQ/usa9vflWTdUP1FSW5r+1ycJHMdQ5K0dPZmJnEZsGZW7VzghqpaCdzQXgO8AljZHuuBS2DwDR84H3gxcCJw/tA3/UuANw7tt2aeY0iSlsi8IVFVnwR2zyqvBTa15U3A6UP1y2tgC3BIkiOBU4GZqtpdVQ8BM8Catu6ZVbWlqgq4fNZYo44hSVoiC70msayq7mvL9wPL2vJRwL1D221vtbnq20fU5zrG4yRZn2Rrkq27du1awJcjSRpl0b9PoqoqSY2jmYUeo6o2ABsAVq1atU97kQQXzdw51vHOOfnYsY6n8VnoTOKBdqqI9ryz1XcARw9tt7zV5qovH1Gf6xiSpCWy0JDYDOy5Q2kdcO1Q/cx2l9Nq4OF2yuh64JQkh7YL1qcA17d1jyRZ3e5qOnPWWKOOIUlaIvOebkpyJfBS4PAk2xncpXQhcHWSs4CvAa9pm18HnAZsA74FvB6gqnYneRdwc9vunVW152L4mxjcQfVU4KPtwRzHkCQtkXlDoqpe21l10ohtCzi7M85GYOOI+lbg+BH1B0cdQ5K0dHzHtSSpy5CQJHUZEpKkLkNCktRlSEiSugwJSVKXISFJ6jIkJEldhoQkqcuQkCR1GRKSpC5DQpLUZUhIkroMCUlS16J/fan2czdeMOkOJE0xZxKSpC5DQpLUZUhIkroMCUlSlyEhSeoyJCRJXd4CKx0AVt+zYdItzOO9k25AHYaEpCeki2buHOt455x87FjH218YEhqrT9394KRbkDRGXpOQJHUZEpKkLkNCktRlSEiSugwJSVKXISFJ6jIkJEldhoQkqcuQkCR1+Y5rSU9I4/+8qgPz86WcSUiSuhYVEkm+muS2JJ9LsrXVDksyk+Su9nxoqyfJxUm2Jbk1yQlD46xr29+VZN1Q/UVt/G1t3yymX0nSD2YcM4mXVdULqmpVe30ucENVrQRuaK8BXgGsbI/1wCUwCBXgfODFwInA+XuCpW3zxqH91oyhX0nSXtoXp5vWApva8ibg9KH65TWwBTgkyZHAqcBMVe2uqoeAGWBNW/fMqtpSVQVcPjSWJGkJLDYkCvh4kluSrG+1ZVV1X1u+H1jWlo8C7h3ad3urzVXfPqL+OEnWJ9maZOuuXbsW8/VIkoYs9u6mX6yqHUmeDcwk+dLwyqqqJLXIY8yrqjYAGwBWrVq1z48nSQeKRc0kqmpHe94JfITBNYUH2qki2vPOtvkO4Oih3Ze32lz15SPqkqQlsuCZRJKnA0+qqm+25VOAdwKbgXXAhe352rbLZuDNSa5icJH64aq6L8n1wHuGLlafApxXVbuTPJJkNXATcCbw1wvtV9L0GvevGgVYPfYRD0yLOd20DPhIuyv1YOCfq+pjSW4Grk5yFvA14DVt++uA04BtwLeA1wO0MHgXcHPb7p1Vtbstvwm4DHgq8NH2kCQtkQWHRFXdDTx/RP1B4KQR9QLO7oy1Edg4or4VOH6hPUqSFsd3XEuSuvzsJkkTN/7PWRq/cV83OefkY8c63r7iTEKS1OVMYn9z4wWT7kDSAcSZhCSpy5CQJHUZEpKkLkNCktRlSEiSugwJSVKXISFJ6jIkJEldhoQkqcuQkCR1GRKSpC5DQpLUZUhIkroMCUlSlyEhSeoyJCRJXYaEJKnL30x3gPvU3Q9OugVJU8yZhCSpy5CQJHV5umk/4+khSUvJmYQkqcuZxL504wWT7kDSmKy+Z8OYR3zvmMfbN5xJSJK6nEkM8yd/SXoMZxKSpC5DQpLUZUhIkroMCUlSlyEhSeoyJCRJXVMfEknWJPlykm1Jzp10P5J0IJnq90kkOQj4AHAysB24OcnmqvriZDuTpMW5aObOsY95zsnHjn3MqQ4J4ERgW1XdDZDkKmAtYEhI2q+N/2M+YF981Me0h8RRwL1Dr7cDL569UZL1wPr28r+SfHmBxzsc+PoC910q097jtPcH09/jtPcH9jgO4+/vt9+3mL1/YlRx2kNir1TVBmDRsZxka1WtGkNL+8y09zjt/cH09zjt/YE9jsO097fHtF+43gEcPfR6eatJkpbAtIfEzcDKJM9J8hTgDGDzhHuSpAPGVJ9uqqpHk7wZuB44CNhYVbfvw0PuiytJ4zbtPU57fzD9PU57f2CP4zDt/QGQqpp0D5KkKTXtp5skSRNkSEiSugwJIMnRSW5M8sUktyd566R7GiXJQUk+m+TfJt3LKEkOSXJNki8luSPJz0+6p2FJzml/v19IcmWSH56CnjYm2ZnkC0O1w5LMJLmrPR86hT3+Rft7vjXJR5IcMk39Da17e5JKcvgkehvqY2SPSd7S/hxvT/Lnk+pvLobEwKPA26vqOGA1cHaS4ybc0yhvBe6YdBNz+CvgY1X1M8DzmaJekxwF/B6wqqqOZ3AjxBmT7QqAy4A1s2rnAjdU1UrghvZ6ki7j8T3OAMdX1c8CdwLnLXVTQy7j8f2R5GjgFOCepW5ohMuY1WOSlzH4BInnV9Xz2Bdvlx4DQwKoqvuq6jNt+ZsMvrkdNdmuHivJcuBXgQ9OupdRkjwL+GXgUoCq+k5VfWOiTT3ewcBTkxwMPA34zwn3Q1V9Etg9q7wW2NSWNwGnL2VPs43qsao+XlWPtpdbGLyHaSI6f4YAFwF/AEz87pxOj78LXFhV327b7FzyxvaCITFLkhXAC4GbJtzKbH/J4B/89ybcR89zgF3AP7RTYh9M8vRJN7VHVe1g8JPaPcB9wMNV9fHJdtW1rKrua8v3A8sm2cxeeAPw0Uk3MSzJWmBHVX1+0r3M4Vjgl5LclOTfk/zcpBsaxZAYkuQZwIeAt1XVI5PuZ48krwR2VtUtk+5lDgcDJwCXVNULgf9m8qdJ/l87r7+WQZj9OPD0JL852a7mV4N71Cf+k3BPkj9icLr2ikn3skeSpwHvAP540r3M42DgMAanuH8fuDpJJtvS4xkSTZInMwiIK6rqw5PuZ5aXAK9K8lXgKuDlSf5psi09znZge1XtmYFdwyA0psWvAF+pql1V9b/Ah4FfmHBPPQ8kORKgPU/laYgkvwW8EviNmq43XP0Ugx8GPt/+zywHPpPkxyba1eNtBz5cA59mcJZgohfYRzEkgJbelwJ3VNX7J93PbFV1XlUtr6oVDC62fqKqpuqn4Kq6H7g3yXNb6SSm6yPd7wFWJ3la+/s+iSm6sD7LZmBdW14HXDvBXkZKsobB6c9XVdW3Jt3PsKq6raqeXVUr2v+Z7cAJ7d/oNPlX4GUASY4FnsIUfmqtITHwEuB1DH5C/1x7nDbppvZDbwGuSHIr8ALgPZNt5/vaDOca4DPAbQz+7U/8YxGSXAl8Cnhuku1JzgIuBE5OcheDGdCFU9jj3wA/Asy0/y9/N2X9TZVOjxuBn2y3xV4FrJuyGRngx3JIkubgTEKS1GVISJK6DAlJUpchIUnqMiQkSV2GhCSpy5CQJHX9H54b4H54Xe7UAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "data": {
      "text/plain": [
       "array([  10,   10,    6, ..., 1073, 1281,  309])"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "degrade_labels(tree, tree.leaf_subset(), dataset.targets, 2, 1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "id": "fd36d856",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "exclude trivial nodes, mean len 9.0 -> 7.85\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "write: resources/override_labels/tiny_imagenet-train-beta-3-1-seed-0.csv.gz\n",
      "exclude trivial nodes, mean len 9.0 -> 7.85\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYMAAAD4CAYAAAAO9oqkAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjUuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/YYfK9AAAACXBIWXMAAAsTAAALEwEAmpwYAAAQw0lEQVR4nO3df6zddX3H8edrrTh/jiJdw9qyEq0ulcyiN9DNZXEyoDBjMTEEskl1zJpYNiQmE9wSjMrADGWSKaZKR8kYSBBDY6r1pmtiTAQpSICCwE3lR7tCK0VgI5HVvffH+TQ51nvb23vPPaf33ucjOTnf8z7fH+9P0t7X/X6/n3NuqgpJ0uz2W4NuQJI0eIaBJMkwkCQZBpIkDANJEjB30A1M1PHHH19LliwZdBuSNK3ce++9P6+q+QfXp20YLFmyhG3btg26DUmaVpI8OVrdy0SSJMNAkmQYSJIwDCRJGAaSJAwDSRKGgSQJw0CShGEgSWIafwJZmmmuHX6sr8e79Iy39vV4Orp5ZiBJMgwkSYaBJAnDQJKEYSBJwjCQJGEYSJIwDCRJGAaSJAwDSRKGgSQJw0CShGEgScIwkCRhGEiSMAwkSRgGkiQMA0kShoEkCcNAkoRhIEliHGGQZHGSrUkeTrI9ySWt/pkku5Lc3x7ndG1zeZKRJI8mOaurvrLVRpJc1lU/Kcndrf7NJMf0eqCSpLGN58xgP/DJqloGrADWJlnW3ru2qpa3xyaA9t75wNuBlcBXk8xJMgf4CnA2sAy4oGs/X2j7egvwPHBRj8YnSRqHw4ZBVe2uqvva8kvAI8DCQ2yyCri1qn5ZVT8DRoBT22OkqnZU1SvArcCqJAHeC9zett8AnDvB8UiSJuCI7hkkWQKcAtzdShcneSDJ+iTzWm0h8HTXZjtbbaz6m4BfVNX+g+qjHX9Nkm1Jtu3du/dIWpckHcK4wyDJ64FvAZ+oqheB64E3A8uB3cAXp6LBblW1rqqGqmpo/vz5U304SZo15o5npSSvohMEN1fVHQBV9WzX+18HvtNe7gIWd22+qNUYo/4ccGySue3soHt9SVIfjGc2UYAbgEeq6ktd9RO6VvsA8FBb3gicn+TVSU4ClgI/Bu4BlraZQ8fQucm8saoK2Ap8sG2/GrhzcsOSJB2J8ZwZvBv4EPBgkvtb7dN0ZgMtBwp4AvgYQFVtT3Ib8DCdmUhrq+pXAEkuBjYDc4D1VbW97e9TwK1JPg/8hE74SAN17fBjg25B6pvDhkFV/RDIKG9tOsQ2VwJXjlLfNNp2VbWDzmwjSdIA+AlkSZJhIEkyDCRJGAaSJAwDSRLj/NCZNButeGpdX49314lr+no8qZtnBpIkw0CSZBhIkjAMJEkYBpIkDANJEoaBJAnDQJKEHzrTdLP1qkF3IM1InhlIkgwDSZJhIEnCMJAkYRhIkjAMJEkYBpIkDANJEoaBJAnDQJKEYSBJwjCQJDGOMEiyOMnWJA8n2Z7kklY/Lslwksfb87xWT5LrkowkeSDJO7v2tbqt/3iS1V31dyV5sG1zXZJMxWAlSaMbz5nBfuCTVbUMWAGsTbIMuAzYUlVLgS3tNcDZwNL2WANcD53wAK4ATgNOBa44ECBtnY92bbdy8kOTJI3XYcOgqnZX1X1t+SXgEWAhsArY0FbbAJzbllcBN1XHXcCxSU4AzgKGq2pfVT0PDAMr23tvrKq7qqqAm7r2JUnqgyO6Z5BkCXAKcDewoKp2t7eeARa05YXA012b7Wy1Q9V3jlIf7fhrkmxLsm3v3r1H0rok6RDGHQZJXg98C/hEVb3Y/V77jb563NtvqKp1VTVUVUPz58+f6sNJ0qwxrjBI8io6QXBzVd3Rys+2Szy05z2tvgtY3LX5olY7VH3RKHVJUp+MZzZRgBuAR6rqS11vbQQOzAhaDdzZVb+wzSpaAbzQLidtBs5MMq/dOD4T2NzeezHJinasC7v2JUnqg/H8DeR3Ax8CHkxyf6t9GrgauC3JRcCTwHntvU3AOcAI8DLwEYCq2pfkc8A9bb3PVtW+tvxx4EbgNcB320OS1CeHDYOq+iEw1rz/00dZv4C1Y+xrPbB+lPo24OTD9SJJmhp+AlmSNK7LRJL6YMVT6/p8xGv6fDwdzTwzkCQZBpIkw0CShGEgScIwkCRhGEiSMAwkSRgGkiQMA0kShoEkCcNAkoRhIEnCMJAkYRhIkjAMJEkYBpIkDANJEoaBJAnDQJKEYSBJwjCQJGEYSJKAuYNuQNJgXDv8WN+OdekZb+3bsTQxnhlIkgwDSZJhIEliHGGQZH2SPUke6qp9JsmuJPe3xzld712eZCTJo0nO6qqvbLWRJJd11U9KcnerfzPJMb0coCTp8MZzZnAjsHKU+rVVtbw9NgEkWQacD7y9bfPVJHOSzAG+ApwNLAMuaOsCfKHt6y3A88BFkxmQJOnIHTYMquoHwL5x7m8VcGtV/bKqfgaMAKe2x0hV7aiqV4BbgVVJArwXuL1tvwE498iGIEmarMncM7g4yQPtMtK8VlsIPN21zs5WG6v+JuAXVbX/oPqokqxJsi3Jtr17906idUlSt4mGwfXAm4HlwG7gi71q6FCqal1VDVXV0Pz58/txSEmaFSb0obOqevbAcpKvA99pL3cBi7tWXdRqjFF/Djg2ydx2dtC9viSpTyZ0ZpDkhK6XHwAOzDTaCJyf5NVJTgKWAj8G7gGWtplDx9C5ybyxqgrYCnywbb8auHMiPUmSJu6wZwZJbgHeAxyfZCdwBfCeJMuBAp4APgZQVduT3AY8DOwH1lbVr9p+LgY2A3OA9VW1vR3iU8CtST4P/AS4oVeDkySNz2HDoKouGKU85g/sqroSuHKU+iZg0yj1HXRmG0mSBsRPIEuSDANJkmEgScIwkCRhGEiSMAwkSRgGkiQMA0kShoEkCcNAkoRhIEnCMJAkYRhIkjAMJEkYBpIkDANJEoaBJAnDQJKEYSBJwjCQJGEYSJIwDCRJwNxBN6BpbutVg+5AUg94ZiBJMgwkSYaBJAnDQJLEOMIgyfoke5I81FU7Lslwksfb87xWT5LrkowkeSDJO7u2Wd3WfzzJ6q76u5I82La5Lkl6PUhJ0qGN58zgRmDlQbXLgC1VtRTY0l4DnA0sbY81wPXQCQ/gCuA04FTgigMB0tb5aNd2Bx9LkjTFDhsGVfUDYN9B5VXAhra8ATi3q35TddwFHJvkBOAsYLiq9lXV88AwsLK998aququqCripa1+SpD6Z6OcMFlTV7rb8DLCgLS8Enu5ab2erHaq+c5T6qJKsoXPGwYknnjjB1jWd/WjHc4NuQZqRJn0Duf1GXz3oZTzHWldVQ1U1NH/+/H4cUpJmhYmGwbPtEg/teU+r7wIWd623qNUOVV80Sl2S1EcTDYONwIEZQauBO7vqF7ZZRSuAF9rlpM3AmUnmtRvHZwKb23svJlnRZhFd2LUvSVKfHPaeQZJbgPcAxyfZSWdW0NXAbUkuAp4EzmurbwLOAUaAl4GPAFTVviSfA+5p6322qg7clP44nRlLrwG+2x6SpD46bBhU1QVjvHX6KOsWsHaM/awH1o9S3wacfLg+JElTx08gS5IMA0mSYSBJwjCQJGEYSJIwDCRJGAaSJAwDSRKGgSQJw0CShGEgScIwkCRhGEiSmPifvZQ0za14al0fj3ZNH4+lifDMQJJkGEiSDANJEoaBJAnDQJKEYSBJwqmlkvrg2uHH+nq8S894a1+PNxN4ZiBJMgwkSYaBJAnDQJKEYSBJwtlEM8/WqwbdgaRpaFJhkOQJ4CXgV8D+qhpKchzwTWAJ8ARwXlU9nyTAl4FzgJeBD1fVfW0/q4F/bLv9fFVtmExfktRP/Zw6O1XTZntxmejPqmp5VQ2115cBW6pqKbClvQY4G1jaHmuA6wFaeFwBnAacClyRZF4P+pIkjdNU3DNYBRz4zX4DcG5X/abquAs4NskJwFnAcFXtq6rngWFg5RT0JUkaw2TDoIDvJ7k3yZpWW1BVu9vyM8CCtrwQeLpr252tNlb9NyRZk2Rbkm179+6dZOuSpAMmewP5T6pqV5LfBYaT/LT7zaqqJDXJY3Tvbx2wDmBoaKhn+5Wk2W5SZwZVtas97wG+Teea/7Pt8g/teU9bfRewuGvzRa02Vl2S1CcTDoMkr0vyhgPLwJnAQ8BGYHVbbTVwZ1veCFyYjhXAC+1y0mbgzCTz2o3jM1tNktQnk7lMtAD4dmfGKHOB/6iq7yW5B7gtyUXAk8B5bf1NdKaVjtCZWvoRgKral+RzwD1tvc9W1b5J9KU++tGO5wbdgqQemHAYVNUO4B2j1J8DTh+lXsDaMfa1Hlg/0V4kSZPj11FIkgwDSZJhIEnCMJAkYRhIkjAMJEn49wwk9cGKp9b1+YjX9Pl4059nBpIkw0CSZBhIkjAMJEkYBpIknE3UH1uvGnQHknRIhsEM41dKS5oILxNJkgwDSZJhIEnCewaSNGn9/bqNqfmqjdkZBs7ukaRfMzvDQNLM5i98R8wwkDTjOMX6yHkDWZJkGEiSDANJEoaBJAnDQJKEYSBJwqmlfeE0N0lHu6PmzCDJyiSPJhlJctmg+5Gk2eSoCIMkc4CvAGcDy4ALkiwbbFeSNHscFWEAnAqMVNWOqnoFuBVYNeCeJGnWOFruGSwEnu56vRM47eCVkqwB1rSX/53k0Qke73jg5xPc9mg3k8cGM3t8jm366t/4/uaLk93D749WPFrCYFyqah0w6e+KTbKtqoZ60NJRZyaPDWb2+Bzb9DUTxne0XCbaBSzuer2o1SRJfXC0hME9wNIkJyU5Bjgf2DjgniRp1jgqLhNV1f4kFwObgTnA+qraPoWH7OefJeq3mTw2mNnjc2zT17QfX6pq0D1IkgbsaLlMJEkaIMNAkjR7wiDJ4iRbkzycZHuSSwbd01RIMifJT5J8Z9C99FKSY5PcnuSnSR5J8keD7qlXklza/k0+lOSWJL896J4mI8n6JHuSPNRVOy7JcJLH2/O8QfY4UWOM7Z/bv8sHknw7ybEDbHHCZk0YAPuBT1bVMmAFsHaGfuXFJcAjg25iCnwZ+F5V/QHwDmbIGJMsBP4OGKqqk+lMoDh/sF1N2o3AyoNqlwFbqmopsKW9no5u5DfHNgycXFV/CDwGXN7vpnph1oRBVe2uqvva8kt0fpgsHGxXvZVkEfAXwDcG3UsvJfkd4E+BGwCq6pWq+sVAm+qtucBrkswFXgv814D7mZSq+gGw76DyKmBDW94AnNvPnnpltLFV1feran97eRedz0lNO7MmDLolWQKcAtw94FZ67V+Avwf+b8B99NpJwF7g39olsG8ked2gm+qFqtoFXAM8BewGXqiq7w+2qymxoKp2t+VngAWDbGYK/TXw3UE3MRGzLgySvB74FvCJqnpx0P30SpL3AXuq6t5B9zIF5gLvBK6vqlOA/2H6Xmb4Ne3a+So6gfd7wOuS/NVgu5pa1ZnPPuPmtCf5BzqXo28edC8TMavCIMmr6ATBzVV1x6D76bF3A+9P8gSdb319b5J/H2xLPbMT2FlVB87kbqcTDjPBnwM/q6q9VfW/wB3AHw+4p6nwbJITANrzngH301NJPgy8D/jLmqYf3po1YZAkdK45P1JVXxp0P71WVZdX1aKqWkLnBuR/VtWM+A2zqp4Bnk7ytlY6HXh4gC310lPAiiSvbf9GT2eG3Bw/yEZgdVteDdw5wF56KslKOpdn319VLw+6n4maNWFA5zfnD9H5jfn+9jhn0E1p3P4WuDnJA8By4J8G205vtLOd24H7gAfp/J+c1l9tkOQW4EfA25LsTHIRcDVwRpLH6ZwNXT3IHidqjLH9K/AGYLj9XPnaQJucIL+OQpI0q84MJEljMAwkSYaBJMkwkCRhGEiSMAwkSRgGkiTg/wFE2qwNbU3cPAAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "write: resources/override_labels/tiny_imagenet-train-beta-2-1-seed-0.csv.gz\n",
      "exclude trivial nodes, mean len 9.0 -> 7.85\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "write: resources/override_labels/tiny_imagenet-train-beta-1.5-1-seed-0.csv.gz\n",
      "exclude trivial nodes, mean len 9.0 -> 7.85\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "write: resources/override_labels/tiny_imagenet-train-beta-1-1-seed-0.csv.gz\n",
      "exclude trivial nodes, mean len 9.0 -> 7.85\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "write: resources/override_labels/tiny_imagenet-train-beta-1-1.5-seed-0.csv.gz\n",
      "exclude trivial nodes, mean len 9.0 -> 7.85\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "write: resources/override_labels/tiny_imagenet-train-beta-1-2-seed-0.csv.gz\n"
     ]
    }
   ],
   "source": [
    "import csv\n",
    "import gzip\n",
    "import pathlib\n",
    "\n",
    "out_dir = pathlib.Path('resources/override_labels')\n",
    "out_dir.mkdir(parents=True, exist_ok=True)\n",
    "\n",
    "seed = 0\n",
    "\n",
    "beta_params = [\n",
    "    (3, 1),  # Quadratic bias toward leaves.\n",
    "    (2, 1),  # Linear bias toward leaves.\n",
    "    (1.5, 1),\n",
    "    (1, 1),  # Uniform.\n",
    "    (1, 1.5),\n",
    "    (1, 2),  # Linear bias toward root.\n",
    "]\n",
    "\n",
    "# with open('resources/hierarchy/imagenet_fiveai.csv') as f:\n",
    "#     tree, node_names = hier.make_hierarchy_from_edges(hier.load_edges(f))\n",
    "# dataset = torchvision.datasets.ImageNet('/home/jack/data/torchvision/imagenet/', 'train')\n",
    "# name = 'imagenet-train'\n",
    "\n",
    "with open('resources/hierarchy/tiny_imagenet_fiveai.csv') as f:\n",
    "    tree, node_names = hier.make_hierarchy_from_edges(hier.load_edges(f))\n",
    "dataset = datasets.TinyImageNet('/home/jack/data/manual/tiny_imagenet/', 'train')\n",
    "name = 'tiny_imagenet-train'\n",
    "\n",
    "# with open('resources/hierarchy/inat18_anon.csv') as f:\n",
    "#     tree, node_names = hier.make_hierarchy_from_edges(hier.load_edges(f))\n",
    "# dataset = datasets.INaturalist2018('/home/jack/data/manual/inaturalist2018/', 'train')\n",
    "# name = 'inat18-train'\n",
    "\n",
    "for a, b in beta_params:\n",
    "    label_order = tree.leaf_subset()\n",
    "    coarse_targets = degrade_labels(tree, label_order, dataset.targets, a, b, seed=seed)\n",
    "\n",
    "    fname = out_dir / f'{name}-beta-{a}-{b}-seed-{seed}.csv.gz'\n",
    "    print('write:', fname)\n",
    "    with gzip.open(fname, 'wt') as f:\n",
    "        w = csv.writer(f)\n",
    "        for i in range(len(dataset)):\n",
    "            coarse = coarse_targets[i]\n",
    "            original = label_order[dataset.targets[i]]\n",
    "            # Write coarse label and original label for sanity check.\n",
    "            w.writerow((node_names[coarse], node_names[original]))\n",
    "\n",
    "    # fname = out_dir / f'{name}-beta-{a}-{b}-seed-{seed}.npy'\n",
    "    # print('write:', fname)\n",
    "    # np.save(fname, coarse_targets)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "39c3e959",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}