{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Jensen-Shannon Divergence & Cross-Entropy Loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import timm\n",
    "import torch\n",
    "import torch.nn.functional as F\n",
    "from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy\n",
    "from timm.loss import JsdCrossEntropy\n",
    "from timm.data.mixup import mixup_target\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's create a example of the `output` of a model, and our `labels`. Note we have 3 output predictions, but only 1 label. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "output = F.one_hot(torch.tensor([0,9,0])).float()\n",
    "labels=torch.tensor([0])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "If we set label `smoothing` and `alpha` to 0, then we will have the regular `cross_entropy loss`, if we look only at the first element of our output and labels. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "jsd = JsdCrossEntropy(smoothing=0,alpha=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor(1.4612)"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "jsd(output,labels)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor(1.4612)"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "base_loss = F.cross_entropy(output[0,None],labels[0,None])\n",
    "base_loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "jsd = JsdCrossEntropy(num_splits=1,smoothing=0,alpha=0)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We can also change the number of splits,changing the size of each group. In `Augmix` this would equate to the number of transformation mixtures. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "jsd = JsdCrossEntropy(num_splits=2,smoothing=0,alpha=0)\n",
    "output = F.one_hot(torch.tensor([0,9,1,0])).float()\n",
    "labels=torch.tensor([0,9])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(tensor(1.4612), tensor(1.4612))"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "jsd(output,labels),F.cross_entropy(output[[0,1]],labels)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "By default we have 1 label for 3 predictions, this is a two part loss, and measures both cross entropy and jason-shannon divergence. Jason-shannon entropy does not need a label, instead measuring the how significantly different the 3 predictions are."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "jsd = JsdCrossEntropy(smoothing=0)\n",
    "output = F.one_hot(torch.tensor([0,0,0]),num_classes=10).float()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([-0.1000,  0.1000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,\n",
       "         0.0000,  0.0000])"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "deltas = torch.cat((torch.zeros([2,10]),torch.tensor([[-1,1,0,0,0,0,0,0,0,0]])))*0.1\n",
    "deltas[2]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "deltas=(torch.arange(-10,11))[...,None,None]*deltas"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "losses = [jsd((output+delta),labels)-base_loss for delta in deltas]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The below graph shows how changes in one of the model's outputs(prediction), in a group, effects the Jason-Shannon Divergence. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "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.plot([ .1*i-1 for i in range(len(losses))],[loss for loss in losses])\n",
    "plt.ylabel('JS Divergence')\n",
    "plt.xlabel('Change in output')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Converted 00_model_architectures.ipynb.\n",
      "Converted 01_training_scripts.ipynb.\n",
      "Converted 02_dataset.ipynb.\n",
      "Converted 03_loss.cross_entropy.ipynb.\n",
      "Converted 04_models.ipynb.\n",
      "Converted 05_loss.jsd_cross_entropy.ipynb.\n",
      "Converted index.ipynb.\n"
     ]
    }
   ],
   "source": [
    "#hide\n",
    "from nbdev.export import notebook2script\n",
    "notebook2script()"
   ]
  }
 ],
 "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": "3.8.5"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {},
   "toc_section_display": true,
   "toc_window_display": true
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}