Plot inter-class and intra-class distances from files.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Plot Cityscapes inter-class : intra-class distance ratio histogram"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import pathlib\n",
    "import tqdm\n",
    "import numpy as np\n",
    "import pickle\n",
    "from scipy.ndimage.filters import uniform_filter\n",
    "from matplotlib import pyplot as plt\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Change `OUT_PATH` to a directory in which you want to save the plots."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "IN_PATH = pathlib.Path(r'./results/distances')\n",
    "# OUT_PATH = pathlib.Path(r'....')\n",
    "OUT_PATH = None"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We originally ran the `intra_inter_class_patch_dist.py` program on a cluster, using `--batch_size=10`, with the job number determining the batch `--batch=${JOB_INDEX}`. Please pardon our use of the term batch here. In this case we are dividing the 1000 patches (the default, set using `--n_patches`) into batches of 10, therefore dividing the work between 100 processes that can be processed in parallel.\n",
    "\n",
    "The code below loads all the output generated and joins it together."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "same_image_intra_class_dists: (1000, 1000)\n",
      "same_image_intra_class_coords: (1000, 1000, 3)\n",
      "same_image_inter_class_dists: (1000, 1000)\n",
      "same_image_inter_class_coords: (1000, 1000, 3)\n",
      "other_image_intra_class_dists: (1000, 1000)\n",
      "other_image_intra_class_coords: (1000, 1000, 3)\n",
      "other_image_inter_class_dists: (1000, 1000)\n",
      "other_image_inter_class_coords: (1000, 1000, 3)\n",
      "anchor_negative_img_dir_y_x_cls: (1000, 5)\n",
      "boundary_dists: (1000,)\n"
     ]
    }
   ],
   "source": [
    "def dict_concat(x, y):\n",
    "    j = dict()\n",
    "    j.update(x)\n",
    "    for k in y.keys():\n",
    "        if k in j:\n",
    "            if isinstance(y[k], list):\n",
    "                j[k] = j[k] + y[k]\n",
    "            elif isinstance(y[k], np.ndarray):\n",
    "                j[k] = np.append(j[k], y[k], axis=0)\n",
    "            else:\n",
    "                pass\n",
    "        else:\n",
    "            j[k] = y[k]\n",
    "    return j\n",
    "\n",
    "def join(data):\n",
    "    data_out = {}\n",
    "    for k, v in data.items():\n",
    "        if v[0].ndim > 0:\n",
    "            lengths = np.array([len(x) for x in v])\n",
    "            max_len = lengths.max()\n",
    "            if (lengths != max_len).any():\n",
    "                max_val = np.array([x.max() for x in v]).max()\n",
    "                v = [np.pad(x, [[0, max_len-len(x)]] + [[0,0]] * (x.ndim-1), mode='constant',\n",
    "                            constant_values=max_val) for x in v]\n",
    "        data_out[k] = np.array(v)\n",
    "    return data_out\n",
    "\n",
    "data = {}\n",
    "for i in range(1, 101):\n",
    "    path = IN_PATH / 'intra_inter_class_distances_{}.pkl'.format(i)\n",
    "    with path.open('rb') as f_in:\n",
    "        batch = pickle.load(f_in)\n",
    "        data = dict_concat(data, batch)\n",
    "        \n",
    "data = join(data)\n",
    "\n",
    "for k in data.keys():\n",
    "    print('{}: {}'.format(k, data[k].shape))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [],
   "source": [
    "anchor_negative_img_dir_y_x_cls = data['anchor_negative_img_dir_y_x_cls']\n",
    "same_image_intra_class_dists = data['same_image_intra_class_dists']\n",
    "same_image_intra_class_coords = data['same_image_intra_class_coords']\n",
    "same_image_inter_class_dists = data['same_image_inter_class_dists']\n",
    "same_image_inter_class_coords = data['same_image_inter_class_coords']\n",
    "other_image_intra_class_dists = data['other_image_intra_class_dists']\n",
    "other_image_intra_class_coords = data['other_image_intra_class_coords']\n",
    "other_image_inter_class_dists = data['other_image_inter_class_dists']\n",
    "other_image_inter_class_coords = data['other_image_inter_class_coords']\n",
    "boundary_dists = data['boundary_dists']\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "True"
      ]
     },
     "execution_count": 34,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "((anchor_negative_img_dir_y_x_cls[:, None, :] == anchor_negative_img_dir_y_x_cls[None, :, :]).all(axis=2) == np.eye(1000)).all()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We want to look at the input data distribution, considering inter-class and intra-class variation.\n",
    "Although we use inter-class and intra-class distances here.\n",
    "\n",
    "We randomly chose 1,000 anchor patches from the Cityscapes dataset that are centred on pixels that are adjacent to a class boundary.\n",
    "\n",
    "For the inter-class distances, we compute the $L^2$ distance between the anchor patch $A_i$ and a negative patch $N_i$ where the central pixels of $N_i$ is chosen such that the central pixels of $A_i$ and $N_i$ are immediate neighbours that lie either side of a class boundary.\n",
    "\n",
    "For the intra-class distance, we consider the distance over which the model must generalise within the same class. For a given anchor patch $A_i$ we search for a positive patch $P_i$ from the set of patches $S_i$, where $S_i$ is:\n",
    "\n",
    "- all patches that can be extracted from the entire dataset in a sliding window fashion\n",
    "- where the patches in $S_i$ do not come from the same image as $A_i$\n",
    "- where the central pixel of the patches in $S_i$ has the same class as that of $A_i$\n",
    "\n",
    "We choose the patch $P_i$ from $S_i$ such that has the smallest $L^2$ distance to $A_i$. Thus thus distance represents the best case distance over which the model must generalise within a class.\n",
    "\n",
    "\n",
    "#### The figures\n",
    "Below, we plot two figures.\n",
    "\n",
    "First we plot a histogram of the distances $|N_i - A_i|$ and $|P_i - A_i|$.\n",
    "\n",
    "Secondly we plot  histogram of the ratio $\\frac{|N_i - A_i|}{|P_i - A_i|}$."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x432 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x216 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(figsize=(12,6))\n",
    "plt.hist(boundary_dists, bins=50)\n",
    "plt.hist(other_image_intra_class_dists[:, 0], bins=100, alpha=0.5)\n",
    "plt.show()\n",
    "\n",
    "plt.figure(figsize=(6,3))\n",
    "# plt.hist(other_image_intra_class_dists[:, 0] / boundary_dists, bins=50)\n",
    "plt.hist(boundary_dists / other_image_intra_class_dists[:, 0], bins=50)\n",
    "plt.gcf().subplots_adjust(bottom=0.15)\n",
    "plt.title('Cityscapes training set')\n",
    "plt.ylabel('Frequency')\n",
    "plt.xlabel('Distance ratio - inter-class dist : intra-class dist')\n",
    "if OUT_PATH is not None:\n",
    "    plt.savefig(str(OUT_PATH / 'cityscapes_dist_ratio_hist.pdf'))\n",
    "    plt.savefig(str(OUT_PATH / 'cityscapes_dist_ratio_hist_png.png'))\n",
    "\n",
    "plt.show()\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x432 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "closer = (same_image_dists <= other_image_dists[:, 0:1])\n",
    "offset_dists = np.sqrt(((same_image_coords[:, :, 1:3] - img_dir_ij[:, None, 2:4])**2).sum(axis=2))\n",
    "f = (closer * offset_dists).max(axis=1)\n",
    "plt.figure(figsize=(12,6))\n",
    "plt.hist(f, bins=200)\n",
    "plt.show()\n",
    "\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.7.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}