KDD_Train_NLI_Using_Attention.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Train a NLI Model Using Attention\n",
    "\n",
    "In the lecture, we talked about in details how to train a natural language inference method using attention mechanisms. At a high level, it consists of three jointly trained steps: attending, comparing, and aggregating. In this lab, we will illustrate the three steps line by line in code.\n",
    "\n",
    "![Using Attention and MLP to train a Natural Language Inference (NLI) Model from scratch](https://d2l.ai/_images/nli_attention.svg)\n",
    "\n",
    "Let's first import all the packages for this notebook."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Looking in links: https://d2l.ai/whl.html\n",
      "Collecting d2l==0.13.2\n",
      "  Downloading d2l-0.13.2-py3-none-any.whl (39 kB)\n",
      "Requirement already satisfied: jupyter in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from d2l==0.13.2) (1.0.0)\n",
      "Requirement already satisfied: numpy in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from d2l==0.13.2) (1.18.1)\n",
      "Requirement already satisfied: matplotlib in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from d2l==0.13.2) (3.1.3)\n",
      "Requirement already satisfied: pandas in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from d2l==0.13.2) (1.0.3)\n",
      "Requirement already satisfied: notebook in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jupyter->d2l==0.13.2) (6.0.3)\n",
      "Requirement already satisfied: ipykernel in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jupyter->d2l==0.13.2) (5.1.4)\n",
      "Requirement already satisfied: nbconvert in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jupyter->d2l==0.13.2) (5.6.1)\n",
      "Requirement already satisfied: jupyter-console in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jupyter->d2l==0.13.2) (6.1.0)\n",
      "Requirement already satisfied: ipywidgets in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jupyter->d2l==0.13.2) (7.5.1)\n",
      "Requirement already satisfied: qtconsole in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jupyter->d2l==0.13.2) (4.7.2)\n",
      "Requirement already satisfied: kiwisolver>=1.0.1 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from matplotlib->d2l==0.13.2) (1.1.0)\n",
      "Requirement already satisfied: cycler>=0.10 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from matplotlib->d2l==0.13.2) (0.10.0)\n",
      "Requirement already satisfied: pyparsing!=2.0.4,!=2.1.2,!=2.1.6,>=2.0.1 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from matplotlib->d2l==0.13.2) (2.4.6)\n",
      "Requirement already satisfied: python-dateutil>=2.1 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from matplotlib->d2l==0.13.2) (2.8.1)\n",
      "Requirement already satisfied: pytz>=2017.2 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from pandas->d2l==0.13.2) (2019.3)\n",
      "Requirement already satisfied: jupyter-core>=4.6.1 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (4.6.3)\n",
      "Requirement already satisfied: prometheus-client in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (0.7.1)\n",
      "Requirement already satisfied: jupyter-client>=5.3.4 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (6.1.2)\n",
      "Requirement already satisfied: jinja2 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (2.11.1)\n",
      "Requirement already satisfied: traitlets>=4.2.1 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (4.3.3)\n",
      "Requirement already satisfied: nbformat in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (5.0.4)\n",
      "Requirement already satisfied: tornado>=5.0 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (6.0.4)\n",
      "Requirement already satisfied: ipython-genutils in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (0.2.0)\n",
      "Requirement already satisfied: Send2Trash in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (1.5.0)\n",
      "Requirement already satisfied: pyzmq>=17 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (18.1.1)\n",
      "Requirement already satisfied: terminado>=0.8.1 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from notebook->jupyter->d2l==0.13.2) (0.8.3)\n",
      "Requirement already satisfied: ipython>=5.0.0 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from ipykernel->jupyter->d2l==0.13.2) (7.13.0)\n",
      "Requirement already satisfied: defusedxml in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from nbconvert->jupyter->d2l==0.13.2) (0.6.0)\n",
      "Requirement already satisfied: bleach in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from nbconvert->jupyter->d2l==0.13.2) (3.1.4)\n",
      "Requirement already satisfied: mistune<2,>=0.8.1 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from nbconvert->jupyter->d2l==0.13.2) (0.8.4)\n",
      "Requirement already satisfied: testpath in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from nbconvert->jupyter->d2l==0.13.2) (0.4.4)\n",
      "Requirement already satisfied: pandocfilters>=1.4.1 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from nbconvert->jupyter->d2l==0.13.2) (1.4.2)\n",
      "Requirement already satisfied: pygments in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from nbconvert->jupyter->d2l==0.13.2) (2.6.1)\n",
      "Requirement already satisfied: entrypoints>=0.2.2 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from nbconvert->jupyter->d2l==0.13.2) (0.3)\n",
      "Requirement already satisfied: prompt-toolkit!=3.0.0,!=3.0.1,<3.1.0,>=2.0.0 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jupyter-console->jupyter->d2l==0.13.2) (3.0.4)\n",
      "Requirement already satisfied: widgetsnbextension~=3.5.0 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from ipywidgets->jupyter->d2l==0.13.2) (3.5.1)\n",
      "Requirement already satisfied: qtpy in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from qtconsole->jupyter->d2l==0.13.2) (1.9.0)\n",
      "Requirement already satisfied: setuptools in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from kiwisolver>=1.0.1->matplotlib->d2l==0.13.2) (46.1.3.post20200330)\n",
      "Requirement already satisfied: six in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from cycler>=0.10->matplotlib->d2l==0.13.2) (1.14.0)\n",
      "Requirement already satisfied: MarkupSafe>=0.23 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jinja2->notebook->jupyter->d2l==0.13.2) (1.1.1)\n",
      "Requirement already satisfied: decorator in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from traitlets>=4.2.1->notebook->jupyter->d2l==0.13.2) (4.4.2)\n",
      "Requirement already satisfied: jsonschema!=2.5.0,>=2.4 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from nbformat->notebook->jupyter->d2l==0.13.2) (3.2.0)\n",
      "Requirement already satisfied: pickleshare in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from ipython>=5.0.0->ipykernel->jupyter->d2l==0.13.2) (0.7.5)\n",
      "Requirement already satisfied: backcall in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from ipython>=5.0.0->ipykernel->jupyter->d2l==0.13.2) (0.1.0)\n",
      "Requirement already satisfied: pexpect; sys_platform != \"win32\" in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from ipython>=5.0.0->ipykernel->jupyter->d2l==0.13.2) (4.8.0)\n",
      "Requirement already satisfied: jedi>=0.10 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from ipython>=5.0.0->ipykernel->jupyter->d2l==0.13.2) (0.15.2)\n",
      "Requirement already satisfied: webencodings in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from bleach->nbconvert->jupyter->d2l==0.13.2) (0.5.1)\n",
      "Requirement already satisfied: wcwidth in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from prompt-toolkit!=3.0.0,!=3.0.1,<3.1.0,>=2.0.0->jupyter-console->jupyter->d2l==0.13.2) (0.1.9)\n",
      "Requirement already satisfied: pyrsistent>=0.14.0 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jsonschema!=2.5.0,>=2.4->nbformat->notebook->jupyter->d2l==0.13.2) (0.16.0)\n",
      "Requirement already satisfied: attrs>=17.4.0 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jsonschema!=2.5.0,>=2.4->nbformat->notebook->jupyter->d2l==0.13.2) (19.3.0)\n",
      "Requirement already satisfied: importlib-metadata; python_version < \"3.8\" in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jsonschema!=2.5.0,>=2.4->nbformat->notebook->jupyter->d2l==0.13.2) (1.5.0)\n",
      "Requirement already satisfied: ptyprocess>=0.5 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from pexpect; sys_platform != \"win32\"->ipython>=5.0.0->ipykernel->jupyter->d2l==0.13.2) (0.6.0)\n",
      "Requirement already satisfied: parso>=0.5.2 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from jedi>=0.10->ipython>=5.0.0->ipykernel->jupyter->d2l==0.13.2) (0.5.2)\n",
      "Requirement already satisfied: zipp>=0.5 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from importlib-metadata; python_version < \"3.8\"->jsonschema!=2.5.0,>=2.4->nbformat->notebook->jupyter->d2l==0.13.2) (2.2.0)\n",
      "Installing collected packages: d2l\n",
      "Successfully installed d2l-0.13.2\n",
      "\u001b[33mWARNING: You are using pip version 20.0.2; however, version 20.1.1 is available.\n",
      "You should consider upgrading via the '/home/ec2-user/anaconda3/envs/mxnet_p36/bin/python -m pip install --upgrade pip' command.\u001b[0m\n",
      "Collecting mxnet-cu101mkl==1.6.0\n",
      "  Downloading mxnet_cu101mkl-1.6.0-py2.py3-none-manylinux1_x86_64.whl (711.0 MB)\n",
      "\u001b[K     |████████████████████████████████| 711.0 MB 2.7 kB/s  eta 0:00:01    |█▉                              | 40.6 MB 10.5 MB/s eta 0:01:04     |███████▍                        | 164.9 MB 57.4 MB/s eta 0:00:10     |█████████████████████▏          | 470.3 MB 64.6 MB/s eta 0:00:04     |█████████████████████▊          | 482.5 MB 64.6 MB/s eta 0:00:04     |█████████████████████████████   | 645.8 MB 66.0 MB/s eta 0:00:01     |███████████████████████████████ | 690.9 MB 66.0 MB/s eta 0:00:01\n",
      "\u001b[?25hRequirement already satisfied, skipping upgrade: requests<3,>=2.20.0 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from mxnet-cu101mkl==1.6.0) (2.23.0)\n",
      "Requirement already satisfied, skipping upgrade: numpy<2.0.0,>1.16.0 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from mxnet-cu101mkl==1.6.0) (1.18.1)\n",
      "Requirement already satisfied, skipping upgrade: graphviz<0.9.0,>=0.8.1 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from mxnet-cu101mkl==1.6.0) (0.8.4)\n",
      "Requirement already satisfied, skipping upgrade: idna<3,>=2.5 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from requests<3,>=2.20.0->mxnet-cu101mkl==1.6.0) (2.9)\n",
      "Requirement already satisfied, skipping upgrade: chardet<4,>=3.0.2 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from requests<3,>=2.20.0->mxnet-cu101mkl==1.6.0) (3.0.4)\n",
      "Requirement already satisfied, skipping upgrade: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from requests<3,>=2.20.0->mxnet-cu101mkl==1.6.0) (1.25.8)\n",
      "Requirement already satisfied, skipping upgrade: certifi>=2017.4.17 in /home/ec2-user/anaconda3/envs/mxnet_p36/lib/python3.6/site-packages (from requests<3,>=2.20.0->mxnet-cu101mkl==1.6.0) (2020.4.5.2)\n",
      "Installing collected packages: mxnet-cu101mkl\n",
      "Successfully installed mxnet-cu101mkl-1.6.0\n",
      "\u001b[33mWARNING: You are using pip version 20.0.2; however, version 20.1.1 is available.\n",
      "You should consider upgrading via the '/home/ec2-user/anaconda3/envs/mxnet_p36/bin/python -m pip install --upgrade pip' command.\u001b[0m\n",
      "aws-mxnet-cu101mkl                 1.6.0              \n",
      "keras-mxnet                        2.2.4.2            \n",
      "mxnet-cu101mkl                     1.6.0              \n",
      "mxnet-model-server                 1.0.8              \n",
      "\u001b[33mWARNING: You are using pip version 20.0.2; however, version 20.1.1 is available.\n",
      "You should consider upgrading via the '/home/ec2-user/anaconda3/envs/mxnet_p36/bin/python -m pip install --upgrade pip' command.\u001b[0m\n"
     ]
    }
   ],
   "source": [
    "!pip install d2l==0.13.2 -f https://d2l.ai/whl.html \n",
    "!pip install -U mxnet-cu101mkl==1.6.0\n",
    "!pip list | egrep mxnet"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "from d2l import mxnet as d2l\n",
    "import mxnet as mx\n",
    "from mxnet import autograd, gluon, init, np, npx\n",
    "from mxnet.gluon import nn\n",
    "\n",
    "npx.set_np()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Part I. Model Designing\n",
    "### Step 1: Attending\n",
    "\n",
    "The first step is to align words in one text sequence to each word in the other sequence. Suppose that the premise is \"i do need sleep\" and the hypothesis is \"i am tired\". Due to semantical similarity, we may wish to align \"i\" in the hypothesis with \"i\" in the premise, and align \"tired\" in the hypothesis with \"sleep\" in the premise. Likewise, we may wish to align \"i\" in the premise with \"i\" in the hypothesis, and align \"need\" and \"sleep\" in the premise with \"tired\" in the hypothesis. Note that such alignment is soft using weighted average, where ideally large weights are associated with the words to be aligned. For ease of demonstration, the above plot shows such alignment in a hard way.\n",
    "\n",
    "Now we describe the soft alignment using attention mechanisms in more detail. Denote by $\\mathbf{A} = (\\mathbf{a}_1, \\ldots, \\mathbf{a}_m)$ and $\\mathbf{B} = (\\mathbf{b}_1, \\ldots, \\mathbf{b}_n)$ the premise and hypothesis, whose number of words are $m$ and $n$, respectively, where $\\mathbf{a}_i, \\mathbf{b}j \\in \\mathbb{R}^{d}$ ($i = 1, \\ldots, m, j = 1, \\ldots, n$) is a $d$-dimensional word embedding vector. For soft alignment, we compute the attention weights $e_{ij} \\in \\mathbb{R}$ as\n",
    "\n",
    "$$e_{ij} = f(\\mathbf{a}_i)^\\top f(\\mathbf{b}_j),$$\n",
    "\n",
    "where the function $f$ is a multilayer perceptron defined in the following mlp function. The output dimension of $f$ is specified by the num_hiddens argument of mlp."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "def mlp(num_hiddens, flatten):\n",
    "    net = nn.Sequential()\n",
    "    net.add(nn.Dropout(0.2))\n",
    "    net.add(nn.Dense(num_hiddens, activation='relu', flatten=flatten))\n",
    "    net.add(nn.Dropout(0.2))\n",
    "    net.add(nn.Dense(num_hiddens, activation='relu', flatten=flatten))\n",
    "    return net"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "It should be highlighted that, in in the above equation $f$ takes inputs $\\mathbf{a}_i$ and $\\mathbf{b}_j$ separately rather than takes a pair of them together as the input. This decomposition trick leads to only $m + n$ applications (linear complexity) of $f$ rather than $mn$ applications (quadratic complexity).\n",
    "\n",
    "Normalizing the attention weights in the above equation, we compute the weighted average of all the word embeddings in the hypothesis to obtain representation of the hypothesis that is softly aligned with the word indexed by $i$ in the premise:\n",
    "\n",
    "$$ \\boldsymbol{\\beta}_i = \\sum_{j=1}^{n}\\frac{\\exp(e_{ij})}{ \\sum_{k=1}^{n} \\exp(e_{ik})} \\mathbf{b}_j. $$\n",
    "\n",
    "Likewise, we compute soft alignment of premise words for each word indexed by $j$ in the hypothesis:\n",
    "\n",
    "$$ \\boldsymbol{\\alpha}_j = \\sum_{i=1}^{m}\\frac{\\exp(e_{ij})}{ \\sum_{k=1}^{m} \\exp(e_{kj})} \\mathbf{a}_i. $$\n",
    "\n",
    "Below we define the Attend class to compute the soft alignment of hypotheses (beta) with input premises A and soft alignment of premises (alpha) with input hypotheses B."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Attend(nn.Block):\n",
    "    def __init__(self, num_hiddens, **kwargs):\n",
    "        super(Attend, self).__init__(**kwargs)\n",
    "        self.f = mlp(num_hiddens=num_hiddens, flatten=False)\n",
    "\n",
    "    def forward(self, A, B):\n",
    "        # Shape of A/B: (batch_size, #words in sequence A/B, embed_size)\n",
    "        # Shape of f_A/f_B: (batch_size, #words in sequence A/B, num_hiddens)\n",
    "        f_A = self.f(A)\n",
    "        f_B = self.f(B)\n",
    "        # Shape of e: (batch_size, #words in sequence A, #words in sequence B)\n",
    "        e = npx.batch_dot(f_A, f_B, transpose_b=True)\n",
    "        # Shape of beta: (batch_size, #words in sequence A, embed_size), where\n",
    "        # sequence B is softly aligned with each word (axis 1 of beta) in\n",
    "        # sequence A\n",
    "        beta = npx.batch_dot(npx.softmax(e), B)\n",
    "        # Shape of alpha: (batch_size, #words in sequence B, embed_size),\n",
    "        # where sequence A is softly aligned with each word (axis 1 of alpha)\n",
    "        # in sequence B\n",
    "        alpha = npx.batch_dot(npx.softmax(e.transpose(0, 2, 1)), A)\n",
    "        return beta, alpha"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Step 2: Comparing\n",
    "\n",
    "In the next step, we compare a word in one sequence with the other sequence that is softly aligned with that word. Note that in soft alignment, all the words from one sequence, though with probably different attention weights, will be compared with a word in the other sequence. \n",
    "\n",
    "In the comparing step, we feed the concatenation (operator $[\\cdot, \\cdot]$) of words from one sequence and aligned words from the other sequence into a function $g$ (a multilayer perceptron):\n",
    "\n",
    "$$\\mathbf{v}_{A,i} = g([\\mathbf{a}_i, \\boldsymbol{\\beta}_i]), i = 1, \\ldots, n;$$\n",
    "$$\\mathbf{v}_{B,j} = g([\\mathbf{b}_j, \\boldsymbol{\\alpha}_j]), j = 1, \\ldots, n.$$\n",
    "\n",
    "\n",
    "In the above equations, $\\mathbf{v}{A,i}$ is the comparison between word $i$ in the premise and all the hypothesis words that are softly aligned with word $i$; while $\\mathbf{v}{B,j}$ is the comparison between word $j$ in the hypothesis and all the premise words that are softly aligned with word $j$. The following Compare class defines such as comparing step."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Compare(nn.Block):\n",
    "    def __init__(self, num_hiddens, **kwargs):\n",
    "        super(Compare, self).__init__(**kwargs)\n",
    "        self.g = mlp(num_hiddens=num_hiddens, flatten=False)\n",
    "\n",
    "    def forward(self, A, B, beta, alpha):\n",
    "        V_A = self.g(np.concatenate([A, beta], axis=2))\n",
    "        V_B = self.g(np.concatenate([B, alpha], axis=2))\n",
    "        return V_A, V_B"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Step 3. Aggregating\n",
    "With two sets of comparison vectors $\\mathbf{v}_{A,i}$ ($i = 1, \\ldots, m$) and $\\mathbf{v}_{B,j}$ ($j = 1, \\ldots, n$) on hand, in the last step we will aggregate such information to infer the logical relationship. We begin by summing up both sets:\n",
    "\n",
    "$$ \\mathbf{v}_A = \\sum_{i=1}^{m}{\\mathbf{v}_{A,i}}, \\quad \\mathbf{v}_B = \\sum_{j=1}^{n}\\mathbf{v}_{B,j}. $$\n",
    "\n",
    "Next we feed the concatenation of both summarization results into function $h$ (a multilayer perceptron) to obtain the classification result of the logical relationship:\n",
    "\n",
    "$$ \\hat{\\mathbf{y}} = h([\\mathbf{v}_A, \\mathbf{v}_B]). $$\n",
    "\n",
    "The aggregation step is defined in the following Aggregate class."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Aggregate(nn.Block):\n",
    "    def __init__(self, num_hiddens, num_outputs, **kwargs):\n",
    "        super(Aggregate, self).__init__(**kwargs)\n",
    "        self.h = mlp(num_hiddens=num_hiddens, flatten=True)\n",
    "        self.h.add(nn.Dense(num_outputs))\n",
    "\n",
    "    def forward(self, V_A, V_B):\n",
    "        # Sum up both sets of comparison vectors\n",
    "        V_A = V_A.sum(axis=1)\n",
    "        V_B = V_B.sum(axis=1)\n",
    "        # Feed the concatenation of both summarization results into an MLP\n",
    "        Y_hat = self.h(np.concatenate([V_A, V_B], axis=1))\n",
    "        return Y_hat"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Putting All Things Together\n",
    "By putting the attending, comparing, and aggregating steps together, we define the `Attention` model to jointly train these three steps."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Attention(nn.Block):\n",
    "    def __init__(self, vocab, embed_size, num_hiddens, **kwargs):\n",
    "        super(Attention, self).__init__(**kwargs)\n",
    "        self.embedding = nn.Embedding(len(vocab), embed_size)\n",
    "        self.attend = Attend(num_hiddens)\n",
    "        self.compare = Compare(num_hiddens)\n",
    "        # There are 3 possible outputs: entailment, contradiction, and neutral\n",
    "        self.aggregate = Aggregate(num_hiddens, 3)\n",
    "\n",
    "    def forward(self, X):\n",
    "        premises, hypotheses = X\n",
    "        A = self.embedding(premises)\n",
    "        B = self.embedding(hypotheses)\n",
    "        beta, alpha = self.attend(A, B)\n",
    "        V_A, V_B = self.compare(A, B, beta, alpha)\n",
    "        Y_hat = self.aggregate(V_A, V_B)\n",
    "        return Y_hat"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Part II. Training and Evaluating the Model\n",
    "Now we will train and evaluate the defined `Attention` model on the Stanford Natural Language Inference (SNLI) dataset. We begin by [reading the SNLI dataset using D2L](https://d2l.ai/chapter_natural-language-processing-applications/natural-language-inference-and-dataset.html)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Downloading ../data/snli_1.0.zip from https://nlp.stanford.edu/projects/snli/snli_1.0.zip...\n",
      "read 549367 examples\n",
      "read 9824 examples\n"
     ]
    }
   ],
   "source": [
    "# Define the batch size and sequence length\n",
    "batch_size, num_steps = 256, 50\n",
    "train_iter, test_iter, vocab = d2l.load_data_snli(batch_size, num_steps)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We use the pretrained $100$-dimensional GloVe embedding to represent the input tokens. Thus, we predefine the dimension of word vectors $\\mathbf{a}_i$ and $\\mathbf{b}_j$ as $100$. The output dimension of functions $f$ in step 1 and $g$ in step 2 is set to $200$. Then we create a model instance, initialize its parameters, and load the GloVe embedding to initialize vectors of input tokens."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Downloading ../data/glove.6B.100d.zip from http://d2l-data.s3-accelerate.amazonaws.com/glove.6B.100d.zip...\n"
     ]
    }
   ],
   "source": [
    "embed_size, num_hiddens, ctx = 100, 200, d2l.try_all_gpus()\n",
    "att = Attention(vocab, embed_size, num_hiddens)\n",
    "att.initialize(init.Xavier(), ctx=ctx)\n",
    "glove_embedding = d2l.TokenEmbedding('glove.6b.100d') ## Load GloVe pretrained model\n",
    "embeds = glove_embedding[vocab.idx_to_token]\n",
    "att.embedding.weight.set_data(embeds)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Next, we define a function to take multiple inputs (i.e., premises and hypotheses) in minibatches."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "def split_batch_multi_inputs(X, y, ctx_list):\n",
    "    \"\"\"Split multi-input `X` and `y` into multiple devices.\"\"\"\n",
    "    X = list(zip(*[gluon.utils.split_and_load(\n",
    "        feature, ctx_list, even_split=False) for feature in X]))\n",
    "    return (X, gluon.utils.split_and_load(y, ctx_list, even_split=False))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Then, we define the hyperparameters for the training as below:\n",
    "1. The learning rate `lr`;\n",
    "1. The number of epochs to train `num_epochs`;\n",
    "1. A Gluon Trainer `trainer`: to apply an Optimizer on the model parameters. we will use \"[Adam](https://d2l.ai/chapter_optimization/adam.html)\" to optinize the model;\n",
    "1. A Gluon Loss function `loss`: we will use softmax cross entropy loss, since this is a multiclass classification;\n",
    "1. A D2L Timer `timer`: to measure training time;\n",
    "1. A D2L Animator `animator` to plot the training loss and accuracy."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Finally, we can train and evaluate the `Attention` model on the SNLI dataset. The model will update its weights in each minibatch using [`train_batch_ch13`](https://d2l.ai/chapter_computer-vision/image-augmentation.html#using-a-multi-gpu-training-model) function in D2L. The function will simply train each minibatch and record the loss and accuracy of the minibatch."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "loss 0.523, train acc 0.792, test acc 0.817\n",
      "5732.8 examples/sec on [gpu(0)]\n"
     ]
    },
    {
     "data": {
      "image/svg+xml": [
       "<?xml version=\"1.0\" encoding=\"utf-8\" standalone=\"no\"?>\n",
       "<!DOCTYPE svg PUBLIC \"-//W3C//DTD SVG 1.1//EN\"\n",
       "  \"http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd\">\n",
       "<!-- Created with matplotlib (https://matplotlib.org/) -->\n",
       "<svg height=\"184.455469pt\" version=\"1.1\" viewBox=\"0 0 235.784375 184.455469\" width=\"235.784375pt\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\">\n",
       " <defs>\n",
       "  <style type=\"text/css\">\n",
       "*{stroke-linecap:butt;stroke-linejoin:round;}\n",
       "  </style>\n",
       " </defs>\n",
       " <g id=\"figure_1\">\n",
       "  <g id=\"patch_1\">\n",
       "   <path d=\"M 0 184.455469 \n",
       "L 235.784375 184.455469 \n",
       "L 235.784375 -0 \n",
       "L 0 -0 \n",
       "z\n",
       "\" style=\"fill:none;\"/>\n",
       "  </g>\n",
       "  <g id=\"axes_1\">\n",
       "   <g id=\"patch_2\">\n",
       "    <path d=\"M 30.103125 146.899219 \n",
       "L 225.403125 146.899219 \n",
       "L 225.403125 10.999219 \n",
       "L 30.103125 10.999219 \n",
       "z\n",
       "\" style=\"fill:#ffffff;\"/>\n",
       "   </g>\n",
       "   <g id=\"matplotlib.axis_1\">\n",
       "    <g id=\"xtick_1\">\n",
       "     <g id=\"line2d_1\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 30.103125 146.899219 \n",
       "L 30.103125 10.999219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_2\">\n",
       "      <defs>\n",
       "       <path d=\"M 0 0 \n",
       "L 0 3.5 \n",
       "\" id=\"m990c732ede\" style=\"stroke:#000000;stroke-width:0.8;\"/>\n",
       "      </defs>\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"30.103125\" xlink:href=\"#m990c732ede\" y=\"146.899219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_1\">\n",
       "      <!-- 0 -->\n",
       "      <defs>\n",
       "       <path d=\"M 31.78125 66.40625 \n",
       "Q 24.171875 66.40625 20.328125 58.90625 \n",
       "Q 16.5 51.421875 16.5 36.375 \n",
       "Q 16.5 21.390625 20.328125 13.890625 \n",
       "Q 24.171875 6.390625 31.78125 6.390625 \n",
       "Q 39.453125 6.390625 43.28125 13.890625 \n",
       "Q 47.125 21.390625 47.125 36.375 \n",
       "Q 47.125 51.421875 43.28125 58.90625 \n",
       "Q 39.453125 66.40625 31.78125 66.40625 \n",
       "z\n",
       "M 31.78125 74.21875 \n",
       "Q 44.046875 74.21875 50.515625 64.515625 \n",
       "Q 56.984375 54.828125 56.984375 36.375 \n",
       "Q 56.984375 17.96875 50.515625 8.265625 \n",
       "Q 44.046875 -1.421875 31.78125 -1.421875 \n",
       "Q 19.53125 -1.421875 13.0625 8.265625 \n",
       "Q 6.59375 17.96875 6.59375 36.375 \n",
       "Q 6.59375 54.828125 13.0625 64.515625 \n",
       "Q 19.53125 74.21875 31.78125 74.21875 \n",
       "z\n",
       "\" id=\"DejaVuSans-48\"/>\n",
       "      </defs>\n",
       "      <g transform=\"translate(26.921875 161.497656)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-48\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"xtick_2\">\n",
       "     <g id=\"line2d_3\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 78.928125 146.899219 \n",
       "L 78.928125 10.999219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_4\">\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"78.928125\" xlink:href=\"#m990c732ede\" y=\"146.899219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_2\">\n",
       "      <!-- 1 -->\n",
       "      <defs>\n",
       "       <path d=\"M 12.40625 8.296875 \n",
       "L 28.515625 8.296875 \n",
       "L 28.515625 63.921875 \n",
       "L 10.984375 60.40625 \n",
       "L 10.984375 69.390625 \n",
       "L 28.421875 72.90625 \n",
       "L 38.28125 72.90625 \n",
       "L 38.28125 8.296875 \n",
       "L 54.390625 8.296875 \n",
       "L 54.390625 0 \n",
       "L 12.40625 0 \n",
       "z\n",
       "\" id=\"DejaVuSans-49\"/>\n",
       "      </defs>\n",
       "      <g transform=\"translate(75.746875 161.497656)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-49\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"xtick_3\">\n",
       "     <g id=\"line2d_5\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 127.753125 146.899219 \n",
       "L 127.753125 10.999219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_6\">\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"127.753125\" xlink:href=\"#m990c732ede\" y=\"146.899219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_3\">\n",
       "      <!-- 2 -->\n",
       "      <defs>\n",
       "       <path d=\"M 19.1875 8.296875 \n",
       "L 53.609375 8.296875 \n",
       "L 53.609375 0 \n",
       "L 7.328125 0 \n",
       "L 7.328125 8.296875 \n",
       "Q 12.9375 14.109375 22.625 23.890625 \n",
       "Q 32.328125 33.6875 34.8125 36.53125 \n",
       "Q 39.546875 41.84375 41.421875 45.53125 \n",
       "Q 43.3125 49.21875 43.3125 52.78125 \n",
       "Q 43.3125 58.59375 39.234375 62.25 \n",
       "Q 35.15625 65.921875 28.609375 65.921875 \n",
       "Q 23.96875 65.921875 18.8125 64.3125 \n",
       "Q 13.671875 62.703125 7.8125 59.421875 \n",
       "L 7.8125 69.390625 \n",
       "Q 13.765625 71.78125 18.9375 73 \n",
       "Q 24.125 74.21875 28.421875 74.21875 \n",
       "Q 39.75 74.21875 46.484375 68.546875 \n",
       "Q 53.21875 62.890625 53.21875 53.421875 \n",
       "Q 53.21875 48.921875 51.53125 44.890625 \n",
       "Q 49.859375 40.875 45.40625 35.40625 \n",
       "Q 44.1875 33.984375 37.640625 27.21875 \n",
       "Q 31.109375 20.453125 19.1875 8.296875 \n",
       "z\n",
       "\" id=\"DejaVuSans-50\"/>\n",
       "      </defs>\n",
       "      <g transform=\"translate(124.571875 161.497656)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-50\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"xtick_4\">\n",
       "     <g id=\"line2d_7\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 176.578125 146.899219 \n",
       "L 176.578125 10.999219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_8\">\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"176.578125\" xlink:href=\"#m990c732ede\" y=\"146.899219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_4\">\n",
       "      <!-- 3 -->\n",
       "      <defs>\n",
       "       <path d=\"M 40.578125 39.3125 \n",
       "Q 47.65625 37.796875 51.625 33 \n",
       "Q 55.609375 28.21875 55.609375 21.1875 \n",
       "Q 55.609375 10.40625 48.1875 4.484375 \n",
       "Q 40.765625 -1.421875 27.09375 -1.421875 \n",
       "Q 22.515625 -1.421875 17.65625 -0.515625 \n",
       "Q 12.796875 0.390625 7.625 2.203125 \n",
       "L 7.625 11.71875 \n",
       "Q 11.71875 9.328125 16.59375 8.109375 \n",
       "Q 21.484375 6.890625 26.8125 6.890625 \n",
       "Q 36.078125 6.890625 40.9375 10.546875 \n",
       "Q 45.796875 14.203125 45.796875 21.1875 \n",
       "Q 45.796875 27.640625 41.28125 31.265625 \n",
       "Q 36.765625 34.90625 28.71875 34.90625 \n",
       "L 20.21875 34.90625 \n",
       "L 20.21875 43.015625 \n",
       "L 29.109375 43.015625 \n",
       "Q 36.375 43.015625 40.234375 45.921875 \n",
       "Q 44.09375 48.828125 44.09375 54.296875 \n",
       "Q 44.09375 59.90625 40.109375 62.90625 \n",
       "Q 36.140625 65.921875 28.71875 65.921875 \n",
       "Q 24.65625 65.921875 20.015625 65.03125 \n",
       "Q 15.375 64.15625 9.8125 62.3125 \n",
       "L 9.8125 71.09375 \n",
       "Q 15.4375 72.65625 20.34375 73.4375 \n",
       "Q 25.25 74.21875 29.59375 74.21875 \n",
       "Q 40.828125 74.21875 47.359375 69.109375 \n",
       "Q 53.90625 64.015625 53.90625 55.328125 \n",
       "Q 53.90625 49.265625 50.4375 45.09375 \n",
       "Q 46.96875 40.921875 40.578125 39.3125 \n",
       "z\n",
       "\" id=\"DejaVuSans-51\"/>\n",
       "      </defs>\n",
       "      <g transform=\"translate(173.396875 161.497656)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-51\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"xtick_5\">\n",
       "     <g id=\"line2d_9\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 225.403125 146.899219 \n",
       "L 225.403125 10.999219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_10\">\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"225.403125\" xlink:href=\"#m990c732ede\" y=\"146.899219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_5\">\n",
       "      <!-- 4 -->\n",
       "      <defs>\n",
       "       <path d=\"M 37.796875 64.3125 \n",
       "L 12.890625 25.390625 \n",
       "L 37.796875 25.390625 \n",
       "z\n",
       "M 35.203125 72.90625 \n",
       "L 47.609375 72.90625 \n",
       "L 47.609375 25.390625 \n",
       "L 58.015625 25.390625 \n",
       "L 58.015625 17.1875 \n",
       "L 47.609375 17.1875 \n",
       "L 47.609375 0 \n",
       "L 37.796875 0 \n",
       "L 37.796875 17.1875 \n",
       "L 4.890625 17.1875 \n",
       "L 4.890625 26.703125 \n",
       "z\n",
       "\" id=\"DejaVuSans-52\"/>\n",
       "      </defs>\n",
       "      <g transform=\"translate(222.221875 161.497656)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-52\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"text_6\">\n",
       "     <!-- epoch -->\n",
       "     <defs>\n",
       "      <path d=\"M 56.203125 29.59375 \n",
       "L 56.203125 25.203125 \n",
       "L 14.890625 25.203125 \n",
       "Q 15.484375 15.921875 20.484375 11.0625 \n",
       "Q 25.484375 6.203125 34.421875 6.203125 \n",
       "Q 39.59375 6.203125 44.453125 7.46875 \n",
       "Q 49.3125 8.734375 54.109375 11.28125 \n",
       "L 54.109375 2.78125 \n",
       "Q 49.265625 0.734375 44.1875 -0.34375 \n",
       "Q 39.109375 -1.421875 33.890625 -1.421875 \n",
       "Q 20.796875 -1.421875 13.15625 6.1875 \n",
       "Q 5.515625 13.8125 5.515625 26.8125 \n",
       "Q 5.515625 40.234375 12.765625 48.109375 \n",
       "Q 20.015625 56 32.328125 56 \n",
       "Q 43.359375 56 49.78125 48.890625 \n",
       "Q 56.203125 41.796875 56.203125 29.59375 \n",
       "z\n",
       "M 47.21875 32.234375 \n",
       "Q 47.125 39.59375 43.09375 43.984375 \n",
       "Q 39.0625 48.390625 32.421875 48.390625 \n",
       "Q 24.90625 48.390625 20.390625 44.140625 \n",
       "Q 15.875 39.890625 15.1875 32.171875 \n",
       "z\n",
       "\" id=\"DejaVuSans-101\"/>\n",
       "      <path d=\"M 18.109375 8.203125 \n",
       "L 18.109375 -20.796875 \n",
       "L 9.078125 -20.796875 \n",
       "L 9.078125 54.6875 \n",
       "L 18.109375 54.6875 \n",
       "L 18.109375 46.390625 \n",
       "Q 20.953125 51.265625 25.265625 53.625 \n",
       "Q 29.59375 56 35.59375 56 \n",
       "Q 45.5625 56 51.78125 48.09375 \n",
       "Q 58.015625 40.1875 58.015625 27.296875 \n",
       "Q 58.015625 14.40625 51.78125 6.484375 \n",
       "Q 45.5625 -1.421875 35.59375 -1.421875 \n",
       "Q 29.59375 -1.421875 25.265625 0.953125 \n",
       "Q 20.953125 3.328125 18.109375 8.203125 \n",
       "z\n",
       "M 48.6875 27.296875 \n",
       "Q 48.6875 37.203125 44.609375 42.84375 \n",
       "Q 40.53125 48.484375 33.40625 48.484375 \n",
       "Q 26.265625 48.484375 22.1875 42.84375 \n",
       "Q 18.109375 37.203125 18.109375 27.296875 \n",
       "Q 18.109375 17.390625 22.1875 11.75 \n",
       "Q 26.265625 6.109375 33.40625 6.109375 \n",
       "Q 40.53125 6.109375 44.609375 11.75 \n",
       "Q 48.6875 17.390625 48.6875 27.296875 \n",
       "z\n",
       "\" id=\"DejaVuSans-112\"/>\n",
       "      <path d=\"M 30.609375 48.390625 \n",
       "Q 23.390625 48.390625 19.1875 42.75 \n",
       "Q 14.984375 37.109375 14.984375 27.296875 \n",
       "Q 14.984375 17.484375 19.15625 11.84375 \n",
       "Q 23.34375 6.203125 30.609375 6.203125 \n",
       "Q 37.796875 6.203125 41.984375 11.859375 \n",
       "Q 46.1875 17.53125 46.1875 27.296875 \n",
       "Q 46.1875 37.015625 41.984375 42.703125 \n",
       "Q 37.796875 48.390625 30.609375 48.390625 \n",
       "z\n",
       "M 30.609375 56 \n",
       "Q 42.328125 56 49.015625 48.375 \n",
       "Q 55.71875 40.765625 55.71875 27.296875 \n",
       "Q 55.71875 13.875 49.015625 6.21875 \n",
       "Q 42.328125 -1.421875 30.609375 -1.421875 \n",
       "Q 18.84375 -1.421875 12.171875 6.21875 \n",
       "Q 5.515625 13.875 5.515625 27.296875 \n",
       "Q 5.515625 40.765625 12.171875 48.375 \n",
       "Q 18.84375 56 30.609375 56 \n",
       "z\n",
       "\" id=\"DejaVuSans-111\"/>\n",
       "      <path d=\"M 48.78125 52.59375 \n",
       "L 48.78125 44.1875 \n",
       "Q 44.96875 46.296875 41.140625 47.34375 \n",
       "Q 37.3125 48.390625 33.40625 48.390625 \n",
       "Q 24.65625 48.390625 19.8125 42.84375 \n",
       "Q 14.984375 37.3125 14.984375 27.296875 \n",
       "Q 14.984375 17.28125 19.8125 11.734375 \n",
       "Q 24.65625 6.203125 33.40625 6.203125 \n",
       "Q 37.3125 6.203125 41.140625 7.25 \n",
       "Q 44.96875 8.296875 48.78125 10.40625 \n",
       "L 48.78125 2.09375 \n",
       "Q 45.015625 0.34375 40.984375 -0.53125 \n",
       "Q 36.96875 -1.421875 32.421875 -1.421875 \n",
       "Q 20.0625 -1.421875 12.78125 6.34375 \n",
       "Q 5.515625 14.109375 5.515625 27.296875 \n",
       "Q 5.515625 40.671875 12.859375 48.328125 \n",
       "Q 20.21875 56 33.015625 56 \n",
       "Q 37.15625 56 41.109375 55.140625 \n",
       "Q 45.0625 54.296875 48.78125 52.59375 \n",
       "z\n",
       "\" id=\"DejaVuSans-99\"/>\n",
       "      <path d=\"M 54.890625 33.015625 \n",
       "L 54.890625 0 \n",
       "L 45.90625 0 \n",
       "L 45.90625 32.71875 \n",
       "Q 45.90625 40.484375 42.875 44.328125 \n",
       "Q 39.84375 48.1875 33.796875 48.1875 \n",
       "Q 26.515625 48.1875 22.3125 43.546875 \n",
       "Q 18.109375 38.921875 18.109375 30.90625 \n",
       "L 18.109375 0 \n",
       "L 9.078125 0 \n",
       "L 9.078125 75.984375 \n",
       "L 18.109375 75.984375 \n",
       "L 18.109375 46.1875 \n",
       "Q 21.34375 51.125 25.703125 53.5625 \n",
       "Q 30.078125 56 35.796875 56 \n",
       "Q 45.21875 56 50.046875 50.171875 \n",
       "Q 54.890625 44.34375 54.890625 33.015625 \n",
       "z\n",
       "\" id=\"DejaVuSans-104\"/>\n",
       "     </defs>\n",
       "     <g transform=\"translate(112.525 175.175781)scale(0.1 -0.1)\">\n",
       "      <use xlink:href=\"#DejaVuSans-101\"/>\n",
       "      <use x=\"61.523438\" xlink:href=\"#DejaVuSans-112\"/>\n",
       "      <use x=\"125\" xlink:href=\"#DejaVuSans-111\"/>\n",
       "      <use x=\"186.181641\" xlink:href=\"#DejaVuSans-99\"/>\n",
       "      <use x=\"241.162109\" xlink:href=\"#DejaVuSans-104\"/>\n",
       "     </g>\n",
       "    </g>\n",
       "   </g>\n",
       "   <g id=\"matplotlib.axis_2\">\n",
       "    <g id=\"ytick_1\">\n",
       "     <g id=\"line2d_11\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 30.103125 146.899219 \n",
       "L 225.403125 146.899219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_12\">\n",
       "      <defs>\n",
       "       <path d=\"M 0 0 \n",
       "L -3.5 0 \n",
       "\" id=\"mb2e8a9c9c7\" style=\"stroke:#000000;stroke-width:0.8;\"/>\n",
       "      </defs>\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"30.103125\" xlink:href=\"#mb2e8a9c9c7\" y=\"146.899219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_7\">\n",
       "      <!-- 0.0 -->\n",
       "      <defs>\n",
       "       <path d=\"M 10.6875 12.40625 \n",
       "L 21 12.40625 \n",
       "L 21 0 \n",
       "L 10.6875 0 \n",
       "z\n",
       "\" id=\"DejaVuSans-46\"/>\n",
       "      </defs>\n",
       "      <g transform=\"translate(7.2 150.698437)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-48\"/>\n",
       "       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n",
       "       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-48\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"ytick_2\">\n",
       "     <g id=\"line2d_13\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 30.103125 119.719219 \n",
       "L 225.403125 119.719219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_14\">\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"30.103125\" xlink:href=\"#mb2e8a9c9c7\" y=\"119.719219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_8\">\n",
       "      <!-- 0.2 -->\n",
       "      <g transform=\"translate(7.2 123.518437)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-48\"/>\n",
       "       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n",
       "       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-50\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"ytick_3\">\n",
       "     <g id=\"line2d_15\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 30.103125 92.539219 \n",
       "L 225.403125 92.539219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_16\">\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"30.103125\" xlink:href=\"#mb2e8a9c9c7\" y=\"92.539219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_9\">\n",
       "      <!-- 0.4 -->\n",
       "      <g transform=\"translate(7.2 96.338437)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-48\"/>\n",
       "       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n",
       "       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-52\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"ytick_4\">\n",
       "     <g id=\"line2d_17\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 30.103125 65.359219 \n",
       "L 225.403125 65.359219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_18\">\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"30.103125\" xlink:href=\"#mb2e8a9c9c7\" y=\"65.359219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_10\">\n",
       "      <!-- 0.6 -->\n",
       "      <defs>\n",
       "       <path d=\"M 33.015625 40.375 \n",
       "Q 26.375 40.375 22.484375 35.828125 \n",
       "Q 18.609375 31.296875 18.609375 23.390625 \n",
       "Q 18.609375 15.53125 22.484375 10.953125 \n",
       "Q 26.375 6.390625 33.015625 6.390625 \n",
       "Q 39.65625 6.390625 43.53125 10.953125 \n",
       "Q 47.40625 15.53125 47.40625 23.390625 \n",
       "Q 47.40625 31.296875 43.53125 35.828125 \n",
       "Q 39.65625 40.375 33.015625 40.375 \n",
       "z\n",
       "M 52.59375 71.296875 \n",
       "L 52.59375 62.3125 \n",
       "Q 48.875 64.0625 45.09375 64.984375 \n",
       "Q 41.3125 65.921875 37.59375 65.921875 \n",
       "Q 27.828125 65.921875 22.671875 59.328125 \n",
       "Q 17.53125 52.734375 16.796875 39.40625 \n",
       "Q 19.671875 43.65625 24.015625 45.921875 \n",
       "Q 28.375 48.1875 33.59375 48.1875 \n",
       "Q 44.578125 48.1875 50.953125 41.515625 \n",
       "Q 57.328125 34.859375 57.328125 23.390625 \n",
       "Q 57.328125 12.15625 50.6875 5.359375 \n",
       "Q 44.046875 -1.421875 33.015625 -1.421875 \n",
       "Q 20.359375 -1.421875 13.671875 8.265625 \n",
       "Q 6.984375 17.96875 6.984375 36.375 \n",
       "Q 6.984375 53.65625 15.1875 63.9375 \n",
       "Q 23.390625 74.21875 37.203125 74.21875 \n",
       "Q 40.921875 74.21875 44.703125 73.484375 \n",
       "Q 48.484375 72.75 52.59375 71.296875 \n",
       "z\n",
       "\" id=\"DejaVuSans-54\"/>\n",
       "      </defs>\n",
       "      <g transform=\"translate(7.2 69.158437)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-48\"/>\n",
       "       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n",
       "       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-54\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"ytick_5\">\n",
       "     <g id=\"line2d_19\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 30.103125 38.179219 \n",
       "L 225.403125 38.179219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_20\">\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"30.103125\" xlink:href=\"#mb2e8a9c9c7\" y=\"38.179219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_11\">\n",
       "      <!-- 0.8 -->\n",
       "      <defs>\n",
       "       <path d=\"M 31.78125 34.625 \n",
       "Q 24.75 34.625 20.71875 30.859375 \n",
       "Q 16.703125 27.09375 16.703125 20.515625 \n",
       "Q 16.703125 13.921875 20.71875 10.15625 \n",
       "Q 24.75 6.390625 31.78125 6.390625 \n",
       "Q 38.8125 6.390625 42.859375 10.171875 \n",
       "Q 46.921875 13.96875 46.921875 20.515625 \n",
       "Q 46.921875 27.09375 42.890625 30.859375 \n",
       "Q 38.875 34.625 31.78125 34.625 \n",
       "z\n",
       "M 21.921875 38.8125 \n",
       "Q 15.578125 40.375 12.03125 44.71875 \n",
       "Q 8.5 49.078125 8.5 55.328125 \n",
       "Q 8.5 64.0625 14.71875 69.140625 \n",
       "Q 20.953125 74.21875 31.78125 74.21875 \n",
       "Q 42.671875 74.21875 48.875 69.140625 \n",
       "Q 55.078125 64.0625 55.078125 55.328125 \n",
       "Q 55.078125 49.078125 51.53125 44.71875 \n",
       "Q 48 40.375 41.703125 38.8125 \n",
       "Q 48.828125 37.15625 52.796875 32.3125 \n",
       "Q 56.78125 27.484375 56.78125 20.515625 \n",
       "Q 56.78125 9.90625 50.3125 4.234375 \n",
       "Q 43.84375 -1.421875 31.78125 -1.421875 \n",
       "Q 19.734375 -1.421875 13.25 4.234375 \n",
       "Q 6.78125 9.90625 6.78125 20.515625 \n",
       "Q 6.78125 27.484375 10.78125 32.3125 \n",
       "Q 14.796875 37.15625 21.921875 38.8125 \n",
       "z\n",
       "M 18.3125 54.390625 \n",
       "Q 18.3125 48.734375 21.84375 45.5625 \n",
       "Q 25.390625 42.390625 31.78125 42.390625 \n",
       "Q 38.140625 42.390625 41.71875 45.5625 \n",
       "Q 45.3125 48.734375 45.3125 54.390625 \n",
       "Q 45.3125 60.0625 41.71875 63.234375 \n",
       "Q 38.140625 66.40625 31.78125 66.40625 \n",
       "Q 25.390625 66.40625 21.84375 63.234375 \n",
       "Q 18.3125 60.0625 18.3125 54.390625 \n",
       "z\n",
       "\" id=\"DejaVuSans-56\"/>\n",
       "      </defs>\n",
       "      <g transform=\"translate(7.2 41.978437)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-48\"/>\n",
       "       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n",
       "       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-56\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"ytick_6\">\n",
       "     <g id=\"line2d_21\">\n",
       "      <path clip-path=\"url(#p4d7681706a)\" d=\"M 30.103125 10.999219 \n",
       "L 225.403125 10.999219 \n",
       "\" style=\"fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;\"/>\n",
       "     </g>\n",
       "     <g id=\"line2d_22\">\n",
       "      <g>\n",
       "       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"30.103125\" xlink:href=\"#mb2e8a9c9c7\" y=\"10.999219\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "     <g id=\"text_12\">\n",
       "      <!-- 1.0 -->\n",
       "      <g transform=\"translate(7.2 14.798437)scale(0.1 -0.1)\">\n",
       "       <use xlink:href=\"#DejaVuSans-49\"/>\n",
       "       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-46\"/>\n",
       "       <use x=\"95.410156\" xlink:href=\"#DejaVuSans-48\"/>\n",
       "      </g>\n",
       "     </g>\n",
       "    </g>\n",
       "   </g>\n",
       "   <g id=\"line2d_23\">\n",
       "    <path clip-path=\"url(#p4d7681706a)\" d=\"M 39.840823 1.977042 \n",
       "L 49.601273 17.90498 \n",
       "L 59.361722 27.300269 \n",
       "L 69.122172 33.804384 \n",
       "L 78.882622 38.459678 \n",
       "L 88.665823 61.842488 \n",
       "L 98.426273 62.592635 \n",
       "L 108.186722 63.321372 \n",
       "L 117.947172 64.08175 \n",
       "L 127.707622 64.782206 \n",
       "L 137.490823 70.882929 \n",
       "L 147.251273 70.909761 \n",
       "L 157.011722 71.269234 \n",
       "L 166.772172 71.459017 \n",
       "L 176.532622 71.705859 \n",
       "L 186.315823 75.542988 \n",
       "L 196.076273 75.710255 \n",
       "L 205.836722 75.812363 \n",
       "L 215.597172 75.649117 \n",
       "L 225.357622 75.78832 \n",
       "\" style=\"fill:none;stroke:#1f77b4;stroke-linecap:square;stroke-width:1.5;\"/>\n",
       "   </g>\n",
       "   <g id=\"line2d_24\">\n",
       "    <path clip-path=\"url(#p4d7681706a)\" d=\"M 39.840823 84.927265 \n",
       "L 49.601273 73.571947 \n",
       "L 59.361722 67.144919 \n",
       "L 69.122172 62.892579 \n",
       "L 78.882622 59.947423 \n",
       "L 88.665823 45.450631 \n",
       "L 98.426273 45.140035 \n",
       "L 108.186722 44.824902 \n",
       "L 117.947172 44.497497 \n",
       "L 127.707622 44.14489 \n",
       "L 137.490823 41.389371 \n",
       "L 147.251273 41.425875 \n",
       "L 157.011722 41.266453 \n",
       "L 166.772172 41.176842 \n",
       "L 176.532622 41.052047 \n",
       "L 186.315823 39.337705 \n",
       "L 196.076273 39.232523 \n",
       "L 205.836722 39.198287 \n",
       "L 215.597172 39.289136 \n",
       "L 225.357622 39.219159 \n",
       "\" style=\"fill:none;stroke:#ff7f0e;stroke-linecap:square;stroke-width:1.5;\"/>\n",
       "   </g>\n",
       "   <g id=\"line2d_25\">\n",
       "    <path clip-path=\"url(#p4d7681706a)\" d=\"M 78.928125 41.917022 \n",
       "L 127.753125 38.721491 \n",
       "L 176.578125 36.591137 \n",
       "L 225.403125 35.871796 \n",
       "\" style=\"fill:none;stroke:#2ca02c;stroke-linecap:square;stroke-width:1.5;\"/>\n",
       "   </g>\n",
       "   <g id=\"patch_3\">\n",
       "    <path d=\"M 30.103125 146.899219 \n",
       "L 30.103125 10.999219 \n",
       "\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n",
       "   </g>\n",
       "   <g id=\"patch_4\">\n",
       "    <path d=\"M 225.403125 146.899219 \n",
       "L 225.403125 10.999219 \n",
       "\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n",
       "   </g>\n",
       "   <g id=\"patch_5\">\n",
       "    <path d=\"M 30.103125 146.899219 \n",
       "L 225.403125 146.899219 \n",
       "\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n",
       "   </g>\n",
       "   <g id=\"patch_6\">\n",
       "    <path d=\"M 30.103125 10.999219 \n",
       "L 225.403125 10.999219 \n",
       "\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n",
       "   </g>\n",
       "   <g id=\"legend_1\">\n",
       "    <g id=\"patch_7\">\n",
       "     <path d=\"M 37.103125 141.899219 \n",
       "L 114.871875 141.899219 \n",
       "Q 116.871875 141.899219 116.871875 139.899219 \n",
       "L 116.871875 96.864844 \n",
       "Q 116.871875 94.864844 114.871875 94.864844 \n",
       "L 37.103125 94.864844 \n",
       "Q 35.103125 94.864844 35.103125 96.864844 \n",
       "L 35.103125 139.899219 \n",
       "Q 35.103125 141.899219 37.103125 141.899219 \n",
       "z\n",
       "\" style=\"fill:#ffffff;opacity:0.8;stroke:#cccccc;stroke-linejoin:miter;\"/>\n",
       "    </g>\n",
       "    <g id=\"line2d_26\">\n",
       "     <path d=\"M 39.103125 102.963281 \n",
       "L 59.103125 102.963281 \n",
       "\" style=\"fill:none;stroke:#1f77b4;stroke-linecap:square;stroke-width:1.5;\"/>\n",
       "    </g>\n",
       "    <g id=\"line2d_27\"/>\n",
       "    <g id=\"text_13\">\n",
       "     <!-- train loss -->\n",
       "     <defs>\n",
       "      <path d=\"M 18.3125 70.21875 \n",
       "L 18.3125 54.6875 \n",
       "L 36.8125 54.6875 \n",
       "L 36.8125 47.703125 \n",
       "L 18.3125 47.703125 \n",
       "L 18.3125 18.015625 \n",
       "Q 18.3125 11.328125 20.140625 9.421875 \n",
       "Q 21.96875 7.515625 27.59375 7.515625 \n",
       "L 36.8125 7.515625 \n",
       "L 36.8125 0 \n",
       "L 27.59375 0 \n",
       "Q 17.1875 0 13.234375 3.875 \n",
       "Q 9.28125 7.765625 9.28125 18.015625 \n",
       "L 9.28125 47.703125 \n",
       "L 2.6875 47.703125 \n",
       "L 2.6875 54.6875 \n",
       "L 9.28125 54.6875 \n",
       "L 9.28125 70.21875 \n",
       "z\n",
       "\" id=\"DejaVuSans-116\"/>\n",
       "      <path d=\"M 41.109375 46.296875 \n",
       "Q 39.59375 47.171875 37.8125 47.578125 \n",
       "Q 36.03125 48 33.890625 48 \n",
       "Q 26.265625 48 22.1875 43.046875 \n",
       "Q 18.109375 38.09375 18.109375 28.8125 \n",
       "L 18.109375 0 \n",
       "L 9.078125 0 \n",
       "L 9.078125 54.6875 \n",
       "L 18.109375 54.6875 \n",
       "L 18.109375 46.1875 \n",
       "Q 20.953125 51.171875 25.484375 53.578125 \n",
       "Q 30.03125 56 36.53125 56 \n",
       "Q 37.453125 56 38.578125 55.875 \n",
       "Q 39.703125 55.765625 41.0625 55.515625 \n",
       "z\n",
       "\" id=\"DejaVuSans-114\"/>\n",
       "      <path d=\"M 34.28125 27.484375 \n",
       "Q 23.390625 27.484375 19.1875 25 \n",
       "Q 14.984375 22.515625 14.984375 16.5 \n",
       "Q 14.984375 11.71875 18.140625 8.90625 \n",
       "Q 21.296875 6.109375 26.703125 6.109375 \n",
       "Q 34.1875 6.109375 38.703125 11.40625 \n",
       "Q 43.21875 16.703125 43.21875 25.484375 \n",
       "L 43.21875 27.484375 \n",
       "z\n",
       "M 52.203125 31.203125 \n",
       "L 52.203125 0 \n",
       "L 43.21875 0 \n",
       "L 43.21875 8.296875 \n",
       "Q 40.140625 3.328125 35.546875 0.953125 \n",
       "Q 30.953125 -1.421875 24.3125 -1.421875 \n",
       "Q 15.921875 -1.421875 10.953125 3.296875 \n",
       "Q 6 8.015625 6 15.921875 \n",
       "Q 6 25.140625 12.171875 29.828125 \n",
       "Q 18.359375 34.515625 30.609375 34.515625 \n",
       "L 43.21875 34.515625 \n",
       "L 43.21875 35.40625 \n",
       "Q 43.21875 41.609375 39.140625 45 \n",
       "Q 35.0625 48.390625 27.6875 48.390625 \n",
       "Q 23 48.390625 18.546875 47.265625 \n",
       "Q 14.109375 46.140625 10.015625 43.890625 \n",
       "L 10.015625 52.203125 \n",
       "Q 14.9375 54.109375 19.578125 55.046875 \n",
       "Q 24.21875 56 28.609375 56 \n",
       "Q 40.484375 56 46.34375 49.84375 \n",
       "Q 52.203125 43.703125 52.203125 31.203125 \n",
       "z\n",
       "\" id=\"DejaVuSans-97\"/>\n",
       "      <path d=\"M 9.421875 54.6875 \n",
       "L 18.40625 54.6875 \n",
       "L 18.40625 0 \n",
       "L 9.421875 0 \n",
       "z\n",
       "M 9.421875 75.984375 \n",
       "L 18.40625 75.984375 \n",
       "L 18.40625 64.59375 \n",
       "L 9.421875 64.59375 \n",
       "z\n",
       "\" id=\"DejaVuSans-105\"/>\n",
       "      <path d=\"M 54.890625 33.015625 \n",
       "L 54.890625 0 \n",
       "L 45.90625 0 \n",
       "L 45.90625 32.71875 \n",
       "Q 45.90625 40.484375 42.875 44.328125 \n",
       "Q 39.84375 48.1875 33.796875 48.1875 \n",
       "Q 26.515625 48.1875 22.3125 43.546875 \n",
       "Q 18.109375 38.921875 18.109375 30.90625 \n",
       "L 18.109375 0 \n",
       "L 9.078125 0 \n",
       "L 9.078125 54.6875 \n",
       "L 18.109375 54.6875 \n",
       "L 18.109375 46.1875 \n",
       "Q 21.34375 51.125 25.703125 53.5625 \n",
       "Q 30.078125 56 35.796875 56 \n",
       "Q 45.21875 56 50.046875 50.171875 \n",
       "Q 54.890625 44.34375 54.890625 33.015625 \n",
       "z\n",
       "\" id=\"DejaVuSans-110\"/>\n",
       "      <path id=\"DejaVuSans-32\"/>\n",
       "      <path d=\"M 9.421875 75.984375 \n",
       "L 18.40625 75.984375 \n",
       "L 18.40625 0 \n",
       "L 9.421875 0 \n",
       "z\n",
       "\" id=\"DejaVuSans-108\"/>\n",
       "      <path d=\"M 44.28125 53.078125 \n",
       "L 44.28125 44.578125 \n",
       "Q 40.484375 46.53125 36.375 47.5 \n",
       "Q 32.28125 48.484375 27.875 48.484375 \n",
       "Q 21.1875 48.484375 17.84375 46.4375 \n",
       "Q 14.5 44.390625 14.5 40.28125 \n",
       "Q 14.5 37.15625 16.890625 35.375 \n",
       "Q 19.28125 33.59375 26.515625 31.984375 \n",
       "L 29.59375 31.296875 \n",
       "Q 39.15625 29.25 43.1875 25.515625 \n",
       "Q 47.21875 21.78125 47.21875 15.09375 \n",
       "Q 47.21875 7.46875 41.1875 3.015625 \n",
       "Q 35.15625 -1.421875 24.609375 -1.421875 \n",
       "Q 20.21875 -1.421875 15.453125 -0.5625 \n",
       "Q 10.6875 0.296875 5.421875 2 \n",
       "L 5.421875 11.28125 \n",
       "Q 10.40625 8.6875 15.234375 7.390625 \n",
       "Q 20.0625 6.109375 24.8125 6.109375 \n",
       "Q 31.15625 6.109375 34.5625 8.28125 \n",
       "Q 37.984375 10.453125 37.984375 14.40625 \n",
       "Q 37.984375 18.0625 35.515625 20.015625 \n",
       "Q 33.0625 21.96875 24.703125 23.78125 \n",
       "L 21.578125 24.515625 \n",
       "Q 13.234375 26.265625 9.515625 29.90625 \n",
       "Q 5.8125 33.546875 5.8125 39.890625 \n",
       "Q 5.8125 47.609375 11.28125 51.796875 \n",
       "Q 16.75 56 26.8125 56 \n",
       "Q 31.78125 56 36.171875 55.265625 \n",
       "Q 40.578125 54.546875 44.28125 53.078125 \n",
       "z\n",
       "\" id=\"DejaVuSans-115\"/>\n",
       "     </defs>\n",
       "     <g transform=\"translate(67.103125 106.463281)scale(0.1 -0.1)\">\n",
       "      <use xlink:href=\"#DejaVuSans-116\"/>\n",
       "      <use x=\"39.208984\" xlink:href=\"#DejaVuSans-114\"/>\n",
       "      <use x=\"80.322266\" xlink:href=\"#DejaVuSans-97\"/>\n",
       "      <use x=\"141.601562\" xlink:href=\"#DejaVuSans-105\"/>\n",
       "      <use x=\"169.384766\" xlink:href=\"#DejaVuSans-110\"/>\n",
       "      <use x=\"232.763672\" xlink:href=\"#DejaVuSans-32\"/>\n",
       "      <use x=\"264.550781\" xlink:href=\"#DejaVuSans-108\"/>\n",
       "      <use x=\"292.333984\" xlink:href=\"#DejaVuSans-111\"/>\n",
       "      <use x=\"353.515625\" xlink:href=\"#DejaVuSans-115\"/>\n",
       "      <use x=\"405.615234\" xlink:href=\"#DejaVuSans-115\"/>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"line2d_28\">\n",
       "     <path d=\"M 39.103125 117.641406 \n",
       "L 59.103125 117.641406 \n",
       "\" style=\"fill:none;stroke:#ff7f0e;stroke-linecap:square;stroke-width:1.5;\"/>\n",
       "    </g>\n",
       "    <g id=\"line2d_29\"/>\n",
       "    <g id=\"text_14\">\n",
       "     <!-- train acc -->\n",
       "     <g transform=\"translate(67.103125 121.141406)scale(0.1 -0.1)\">\n",
       "      <use xlink:href=\"#DejaVuSans-116\"/>\n",
       "      <use x=\"39.208984\" xlink:href=\"#DejaVuSans-114\"/>\n",
       "      <use x=\"80.322266\" xlink:href=\"#DejaVuSans-97\"/>\n",
       "      <use x=\"141.601562\" xlink:href=\"#DejaVuSans-105\"/>\n",
       "      <use x=\"169.384766\" xlink:href=\"#DejaVuSans-110\"/>\n",
       "      <use x=\"232.763672\" xlink:href=\"#DejaVuSans-32\"/>\n",
       "      <use x=\"264.550781\" xlink:href=\"#DejaVuSans-97\"/>\n",
       "      <use x=\"325.830078\" xlink:href=\"#DejaVuSans-99\"/>\n",
       "      <use x=\"380.810547\" xlink:href=\"#DejaVuSans-99\"/>\n",
       "     </g>\n",
       "    </g>\n",
       "    <g id=\"line2d_30\">\n",
       "     <path d=\"M 39.103125 132.319531 \n",
       "L 59.103125 132.319531 \n",
       "\" style=\"fill:none;stroke:#2ca02c;stroke-linecap:square;stroke-width:1.5;\"/>\n",
       "    </g>\n",
       "    <g id=\"line2d_31\"/>\n",
       "    <g id=\"text_15\">\n",
       "     <!-- test acc -->\n",
       "     <g transform=\"translate(67.103125 135.819531)scale(0.1 -0.1)\">\n",
       "      <use xlink:href=\"#DejaVuSans-116\"/>\n",
       "      <use x=\"39.208984\" xlink:href=\"#DejaVuSans-101\"/>\n",
       "      <use x=\"100.732422\" xlink:href=\"#DejaVuSans-115\"/>\n",
       "      <use x=\"152.832031\" xlink:href=\"#DejaVuSans-116\"/>\n",
       "      <use x=\"192.041016\" xlink:href=\"#DejaVuSans-32\"/>\n",
       "      <use x=\"223.828125\" xlink:href=\"#DejaVuSans-97\"/>\n",
       "      <use x=\"285.107422\" xlink:href=\"#DejaVuSans-99\"/>\n",
       "      <use x=\"340.087891\" xlink:href=\"#DejaVuSans-99\"/>\n",
       "     </g>\n",
       "    </g>\n",
       "   </g>\n",
       "  </g>\n",
       " </g>\n",
       " <defs>\n",
       "  <clipPath id=\"p4d7681706a\">\n",
       "   <rect height=\"135.9\" width=\"195.3\" x=\"30.103125\" y=\"10.999219\"/>\n",
       "  </clipPath>\n",
       " </defs>\n",
       "</svg>\n"
      ],
      "text/plain": [
       "<Figure size 252x180 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Defined in file: ./chapter_computer-vision/image-augmentation.md\n",
    "lr, num_epochs = 0.001, 4\n",
    "trainer = gluon.Trainer(att.collect_params(), 'adam', {'learning_rate': lr})\n",
    "loss = gluon.loss.SoftmaxCrossEntropyLoss()\n",
    "num_batches  = len(train_iter)\n",
    "timer = d2l.Timer()\n",
    "animator = d2l.Animator(xlabel='epoch', xlim=[0, num_epochs], ylim=[0, 1],\n",
    "                        legend=['train loss', 'train acc', 'test acc'])\n",
    "\n",
    "\n",
    "for epoch in range(num_epochs):\n",
    "    # Store training_loss, training_accuracy, num_examples, num_features\n",
    "    metric = d2l.Accumulator(4)\n",
    "    for i, (features, labels) in enumerate(train_iter):\n",
    "        timer.start()\n",
    "        l, acc = d2l.train_batch_ch13(\n",
    "            att, features, labels, loss, trainer, ctx, split_batch_multi_inputs)\n",
    "        metric.add(l, acc, labels.shape[0], labels.size)\n",
    "        timer.stop()\n",
    "        if (i + 1) % (num_batches // 5) == 0:\n",
    "            animator.add(epoch + i / num_batches,\n",
    "                         (metric[0] / metric[2], metric[1] / metric[3],\n",
    "                          None))\n",
    "    test_acc = d2l.evaluate_accuracy_gpus(att, test_iter, split_batch_multi_inputs)\n",
    "    animator.add(epoch + 1, (None, None, test_acc))\n",
    "print(f'loss {metric[0] / metric[2]:.3f}, train acc '\n",
    "      f'{metric[1] / metric[3]:.3f}, test acc {test_acc:.3f}')\n",
    "print(f'{metric[2] * num_epochs / timer.sum():.1f} examples/sec on '\n",
    "      f'{str(ctx)}')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Finally, we can define the inference function to predict the logical relationship between a pair of premise and hypothesis."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "def predict_snli(net, vocab, premise, hypothesis):\n",
    "    premise = np.array(vocab[premise], ctx=d2l.try_gpu())\n",
    "    hypothesis = np.array(vocab[hypothesis], ctx=d2l.try_gpu())\n",
    "    label = np.argmax(net([premise.reshape((1, -1)),\n",
    "                           hypothesis.reshape((1, -1))]), axis=1)\n",
    "    return 'entailment' if label == 0 else 'contradiction' if label == 1 \\\n",
    "            else 'neutral'\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'contradiction'"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "predict_snli(att, vocab, ['that', 'is', 'good', '.'], ['that', 'is', 'bad', '.'])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Resources\n",
    "\n",
    "- Amazon SageMaker https://aws.amazon.com/sagemaker/\n",
    "- Amazon SageMaker Python SDK https://sagemaker.readthedocs.io/\n",
    "- GluonNLP http://gluon-nlp.mxnet.io/\n",
    "- GluonCV http://gluon-cv.mxnet.io/\n",
    "- GluonTS https://gluon-ts.mxnet.io/\n",
    "- Dive into Deep Learning http://d2l.ai/\n",
    "- MXNet Forum https://discuss.mxnet.io/"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "conda_mxnet_p36",
   "language": "python",
   "name": "conda_mxnet_p36"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}