LectureMultiTask.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Multi-modal and Multi-Task Learning (MTL) Overview  "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## To do\n",
    "Explain evaluation process, Bland Altman (Mostly done)\n",
    "\n",
    "Section of visualization and the history of applications (In progress)\n",
    "\n",
    "Theory, very technical, why does it prevent overfitting & regularization (Very sparse)\n",
    "\n",
    "Separate Loss Function, talk about it?\n",
    "\n",
    "Transfer Learning - Relate History"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Why Multi Task\n",
    "Multi task learning helps in scenarios where there is a lot of general data, but very limited data for each class.\n",
    "The example used in this notebook is predicting student test scores.\n",
    "There are many students across the country and the data from each of them can be used to train a shared model.\n",
    "\n",
    "However, only a handful of samples (~150) are available for each school.\n",
    "Each school has distinct behavior in students that influences which factors are strong predictors of test scores.\n",
    "By treating all students as part of the sam data set, we make it difficult for the model to adjust itself to the nuances of each school.\n",
    "\n",
    "We can combat this by creating an artificial partition in the dataset on the school and giving each school a specific part of the model.\n",
    "Situations such as these are where multi task learning is useful."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Preface\n",
    "\n",
    "##### A Common Source of Confusion\n",
    "\n",
    "To aid in processing mutli-task learning for the first time it is helpful to reiterate the distinction between targets and tasks.\n",
    "These two are *not* the same.\n",
    "When switching from a traditional ML model to a multi-task model, it is still important to note that the target of the model is still to predict test scores.\n",
    "The tasks are not pieces of output, but categories of data.\n",
    "\n",
    "In summary: \n",
    "- Task: a breakpoint in the dataset that could be helpful to impart on the model before beginning training\n",
    "- Target: the attribute the model is attempting to predict (same as traditional ML)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# MTL Summary\n",
    "In Machine Learning, our traditional goal is to optimize for a specific target. \n",
    "We do this by training a model and refining it until the performance improvement flatlines. \n",
    "All of the data is processed in the same way.\n",
    "By treating all the data as identical, we tend to leave out other pieces of potentially valuable information that could improve our score. \n",
    "In order to fix this, we present a Multi-Task Learning (MTL) approach.\n",
    "\n",
    "MTL has been succesfully implemented and achieved increasingly better results than its single-task rival. Ranging from applications such as natural language processing (Collobert and Watson, 2008) to drug discovery (Ramsundar, et al., 2015), it has been shown to optimize most architectures if the dataset fits the situation. Although it is also known as joint learning, typically as soon as you start optimizing more than one loss function concurrently, you are considering to be multi-tasking. To give a more concise definition:\n",
    "\n",
    "“MTL improves generalization by leveraging the domain-specific information contained in the training signals of related tasks\". - [Caruana, 1998]\n",
    "\n",
    "<img src=\"images/multi-task-graph.png\" width=\"400\">\n",
    "\n",
    "An advantage to this type of model, is that while some of the hidden layers are shared among the data, they each have their own individual output layers, which greatly reduces the risk of overfitting. This is due to the model having a lot more data to learn from than it normally would.\n",
    "\n",
    "By presenting a model with similar data but different goals, such as predicting scores of different schools around a country, we can make our overall performance better for our original task. Due to the split of data needed to train a multi-task model, one typically needs a very large, complete, and even dataset to train/test with."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Multi Task Applications\n",
    "\n",
    "In terms of specific applications, we will go down a history and talk about the ones we mentioned before:\n",
    "\n",
    "NLP - Collobert and Watson 2008\n",
    "Speech Recognition - Deng et al., 2013\n",
    "Computer Vision - Girshick, 2015\n",
    "Drug Discovery - Ramsundar et al., 2015 \n",
    "\n",
    "https://arxiv.org/pdf/1707.08114.pdf\n",
    "https://arxiv.org/pdf/1707.08114.pdf\n",
    "\n",
    "\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## So why does MTL work?\n",
    "\n",
    "### Implicit data augmentation\n",
    "By having more than one task to train a model with, we passively acquire more data that has different noise patterns. For example, instead of overfitting on a single task A, we are able to learn with similar yet fundamentaly different task B, thus creating a more diverse and general representation.\n",
    "\n",
    "### Attention focusing\n",
    "If a task has a high amount of noise or the data is insufficient, a model can have issues solving which features are more relevent that others. However, with an MTL approach, features become easier to differentiate in terms of importance as the multitude of tasks will increase the amount of data available to test on.\n",
    "\n",
    "### Eavesdropping\n",
    "Similar to how some people are better at different individual tasks, so are models. Due to MTL, different models are able to \"eavesdrop\" on each other with shared layer, and spread the benefits of each to one another.\n",
    "\n",
    "### Representation bias\n",
    "MTL tends to prefer representations that the other tasks also focus on. This way, the model is better at generalizing new added tasks in the future and perform well on them if they are part of the same enviroment.\n",
    "\n",
    "### Regularization\n",
    "MTL acts as a passive regularizer by having an inductive bias (set of assumptions the learner uses to predict outputs), thus reducing the potential for overfitting and the overall Rademacher complexity of the model (the ability of it to fit random noise).\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Evaluation Success: Bland-Altman plots\n",
    "In order to evauluate our results, we will be using a Bland-Altman (Tukey mean-difference plot) graph to measure the agreeableness between the actual target values and our predicted target values. \n",
    "\n",
    "<img src=\"images/bland-altman.png\" width=\"600\">\n",
    "\n",
    "Given two sets of data of n samples each, you plot them on a graph with the following equation:\n",
    "\n",
    "$$S(x, y) = ( \\dfrac{S_1 + S_2}{2} , S_1 - S_2 )$$\n",
    "\n",
    "Your x-axis will be the mean of the numbers, and the y-axis will be the difference. The three lines on the graph represent the the mean +/- 2 Standard Deviations, creating a 95% confidence interval centered around our mean error.\n",
    "\n",
    "The reason this graph is useful is because it helps mitigate errors from concluding that two pieces of data correlate with each other based solely on the correlation coefficient (least square fitting). Back in 1986, Bland and Altman showed that two sets of data can have a high correlation but not be in agreement at all. Plotting the data as a function of the mean vs the difference has proven a more accurate way to determine if two sets of data agree with each other, not just correlate. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# School Data\n",
    "\n",
    "## School Data Understanding\n",
    "\n",
    "We will be using a UK school dataset comprised of 140 different school locations with a total of 15,000 students. Each student has 28 different features, each representing a demographic aspect (gender, age, etc).\n",
    "\n",
    "To give a more detailed explanation of the data, this was taken from the data source itself:\n",
    "\n",
    "Column    Description              Coding\n",
    "\n",
    "     1         Year                     1985=1; 1986=2; 1987=3\n",
    "\n",
    "     2-4       School                   Codes 1-139\n",
    "\n",
    "     5-6       Exam Score               Numeric score\n",
    "\n",
    "     7-8       % FSM                    Percent. students eligible for free\n",
    "                                        school meals\n",
    "\n",
    "     9-10      % VR1 band               Percent. students in school in VR band\n",
    "                                        1\n",
    "\n",
    "     11        Gender                   Male=0; Female=1\n",
    "\n",
    "     12        VR band of student       VR1=2; VR2=3; VR3=1\n",
    "\n",
    "     13-14     Ethnic group of          ESWI=1*; African=2; Arab=3;\n",
    "               student                  Bangladeshi=4; Caribbean=5;\n",
    "                                        Greek=6;Indian=7;Pakistani=8;\n",
    "                                        S.E.Asian=9;Turkish=10; Other=11\n",
    "\n",
    "                                        Mixed=1; Male=2; Female=3\n",
    "     15        School gender\n",
    "                                        Maintained=1; Church of\n",
    "     16        School denomination      England=2; Roman Catholic=3\n",
    "\n",
    "Originally there were only 16 features, but the categorical ones were replace with a binary feature for each possible value. In total, there were 27 plus the bias term. Each school is treated as an individual task and the exam score is the target.\n",
    "\n",
    "In this notebooke will be exploring three different ways to create a classifier and compare their performances.\n",
    "In the next few sections, we will be showing three different methods of completing the task on this data: 1) running each school through its own model, 2) running all the schools together into one model, 3) using a Multi-Task approach and having some of the layers shared by each school, and some individual.\n",
    "\n",
    "## School Data Processing\n",
    "\n",
    "Unfortunately this is an extremely messy dataset so we have quite a bit of preprocessing that we will need to perform."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "import scipy.io\n",
    "from sklearn.datasets.base import Bunch\n",
    "from matplotlib import pyplot as plt\n",
    "import numpy as np\n",
    "from sklearn.model_selection import train_test_split\n",
    "\n",
    "# load the data for each individual school\n",
    "school_indices = [scipy.io.loadmat('data/school_splits/school_%d_indexes.mat' % (x)) for x in range(1, 11)]\n",
    "raw_data = scipy.io.loadmat('data/school_splits/school_b.mat')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "deletable": false,
    "editable": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "We have an input array of shape: (15362, 28)\n",
      "We have an output array of shape: (15362, 1)\n"
     ]
    }
   ],
   "source": [
    "# changed data to float because yhat will be floats\n",
    "X = np.array(raw_data['x'].T)\n",
    "y = np.array(raw_data['y'])\n",
    "print(\"We have an input array of shape:\", X.shape)\n",
    "print(\"We have an output array of shape:\", y.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "# Matlab - arrays are indexed at 1, needs adapting\n",
    "start_index = np.ravel(raw_data[\"task_indexes\"] - 1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "tasks = []\n",
    "for i in range(len(start_index)):\n",
    "    start = start_index[i]\n",
    "    if i == len(start_index) - 1:\n",
    "        end = -1\n",
    "    else:\n",
    "        end = start_index[i + 1]\n",
    "    descr = \"School data: school {}\".format(i + 1)\n",
    "    id = \"School {}\".format(i + 1)\n",
    "    tasks.append(Bunch(data=X[start:end],\n",
    "                       target=y[start:end],\n",
    "                       DESCR=descr,\n",
    "                       ID=id))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "\n",
    "def bland_altman_plot(data1, data2, std1, std2, *args, **kwargs):\n",
    "    data1     = np.asarray(data1)\n",
    "    data2     = np.asarray(data2)\n",
    "    mean      = np.mean([data1, data2], axis=0)\n",
    "    diff      = data1 - data2                   # Difference between data1 and data2\n",
    "    md        = np.mean(diff)                   # Mean of the difference\n",
    "    sd        = np.std(diff, axis=0)            # Standard deviation of the difference\n",
    "    std1      = np.asarray(std1)\n",
    "    std2      = np.asarray(std2)\n",
    "\n",
    "    plt.scatter(mean, diff, s=10+(std1 * std2), *args, **kwargs)\n",
    "    plt.axhline(md,           color='gray', linestyle='--')\n",
    "    plt.axhline(md + 1.96*sd, color='gray', linestyle='--')\n",
    "    plt.axhline(md - 1.96*sd, color='gray', linestyle='--')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": false,
    "editable": false
   },
   "source": [
    "## Independent School Models\n",
    "\n",
    "First, let's train each school on its own individual model.\n",
    "Each of the 141 models will only have an average of 107.4 data points between both train and test data."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "deletable": false,
    "editable": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/yeti/anaconda3/lib/python3.6/site-packages/h5py/__init__.py:34: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n",
      "  from ._conv import register_converters as _register_converters\n",
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "import keras\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Reshape\n",
    "from keras.layers import Dense, Dropout, Activation, Flatten"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "models_ind = []\n",
    "test_daters_ind = []\n",
    "for x in tasks:\n",
    "    #X_train = x.data\n",
    "    #y_train = x.target\n",
    "    X_train, X_test, y_train, y_test = train_test_split(x.data, x.target, test_size=0.2, random_state=42)\n",
    "\n",
    "    mlp = Sequential()\n",
    "    mlp.add(Dense(input_dim=X_train.shape[1], units=30, activation='relu'))\n",
    "    mlp.add( Dense(units=15, activation='relu') )\n",
    "    mlp.add(Dense(1))\n",
    "\n",
    "    mlp.compile(loss='mean_squared_error',\n",
    "                optimizer='rmsprop',\n",
    "                metrics=[])\n",
    "    \n",
    "    mlp.fit(X_train, y_train, epochs=20, shuffle=True, verbose=0)\n",
    "    models_ind.append(mlp)\n",
    "    test_daters_ind.append([X_test, y_test])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "import seaborn as sns\n",
    "from sklearn.metrics import mean_squared_error, accuracy_score, precision_score, mean_squared_log_error"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false,
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "y_ind=[]\n",
    "y_ind_std=[]\n",
    "yhat_ind=[]\n",
    "yhat_ind_std=[]\n",
    "mses=[]\n",
    "\n",
    "for t, mlp in zip(test_daters_ind, models_ind):\n",
    "    X_test, y_test = t\n",
    "\n",
    "    yhat_mlp = mlp.predict(X_test)\n",
    "    yhat_rounded = np.around(yhat_mlp)\n",
    "    \n",
    "    y_test_plot = y_test.ravel()\n",
    "    yhat_mlp_plot = yhat_mlp.ravel()\n",
    "    \n",
    "    y_ind.append(np.mean(y_test_plot))\n",
    "    y_ind_std.append(np.std(y_test_plot))\n",
    "    yhat_ind.append(np.mean(yhat_mlp_plot))\n",
    "    yhat_ind_std.append(np.std(yhat_mlp_plot))\n",
    "\n",
    "    mse = mean_squared_error(y_test, yhat_mlp)\n",
    "    mses.append(mse)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "deletable": false,
    "editable": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAA4AAAAJcCAYAAACov8q3AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x13ccb8e10>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(figsize=(15,10))\n",
    "bland_altman_plot(y_ind, yhat_ind, y_ind_std, yhat_ind_std)\n",
    "plt.title('Bland-Altman Test Scores')\n",
    "plt.xlabel('Mean Score', fontsize=8)\n",
    "plt.ylabel('Diff Score', fontsize=8)\n",
    "plt.show()\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This graph is a Bland Altman plot, indended to show the relationship between two sets of values by plotting the difference between them on the y-axis and the average of them on the x-axis. This way, you can see the residual difference between our predictions and the actual numbers. \n",
    "   \n",
    "Each dot size represents its group standard deviation, the larger it is, the larger the STD is. In addition, each dot represents a different school, with each point being the average of all the points for that category.  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "deletable": false,
    "editable": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Average of Mean Squared Error:  202.08546447753906\n"
     ]
    }
   ],
   "source": [
    "avg = float(np.mean(mses))\n",
    "print (\"Average of Mean Squared Error: \", avg)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "By using Avg Mean Square Error, we can see how close we are to the line of best fit. Based on the graph, we can already tell we are not close."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## General School Model\n",
    "\n",
    "Let's see if using a single model for all the schools increases our accuracy.  We will have one model with all 15000 training instances."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "all_data = {\"data\": [], \"target\": [], \"data_test\": [], \"target_test\": []}\n",
    "for bunch in tasks:\n",
    "    x_temp, y_temp = [bunch.data, bunch.target]\n",
    "    X_train, X_test, y_train, y_test = train_test_split(x_temp, y_temp, test_size=0.2, random_state=42)\n",
    "    all_data[\"data\"].append(X_train)\n",
    "    all_data[\"target\"].append(y_train)        \n",
    "    all_data[\"data_test\"].append(X_test)\n",
    "    all_data[\"target_test\"].append(y_test)\n",
    "\n",
    "X_train = np.concatenate(all_data[\"data\"])\n",
    "y_train = np.concatenate(all_data[\"target\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "deletable": false,
    "editable": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x13d50cc50>"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "generalized_mlp = Sequential()\n",
    "generalized_mlp.add(Dense(input_dim=X_train.shape[1], units=30, activation='relu'))\n",
    "generalized_mlp.add(Dense(units=15, activation='relu'))\n",
    "generalized_mlp.add(Dense(1))\n",
    "\n",
    "generalized_mlp.compile(loss='mean_squared_error',\n",
    "            optimizer='rmsprop',\n",
    "            metrics=[])\n",
    "\n",
    "generalized_mlp.fit(X_train, y_train, epochs=20, shuffle=True, verbose=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "y_general=[]\n",
    "y_general_std=[]\n",
    "yhat_general=[]\n",
    "yhat_general_std=[]\n",
    "mses_general=[]\n",
    "\n",
    "for X_test, y_test in zip(all_data[\"data_test\"], all_data[\"target_test\"]):\n",
    "    yhat_mlp = generalized_mlp.predict(X_test)\n",
    "    yhat_rounded = np.around(yhat_mlp)\n",
    "    \n",
    "    y_test_plot = y_test.ravel()\n",
    "    yhat_mlp_plot = yhat_mlp.ravel()\n",
    "    \n",
    "    y_general.append(np.mean(y_test_plot))\n",
    "    y_general_std.append(np.std(y_test_plot))\n",
    "    yhat_general.append(np.mean(yhat_mlp_plot))\n",
    "    yhat_general_std.append(np.std(y_test_plot))\n",
    "    \n",
    "    mse = mean_squared_error(y_test, yhat_mlp)\n",
    "    mses_general.append(mse)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "deletable": false,
    "editable": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAA4kAAAJcCAYAAABUquF9AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x1406d3198>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(figsize=(15,10))\n",
    "bland_altman_plot(y_general, yhat_general, y_general_std, yhat_general_std)\n",
    "plt.title('Bland-Altman Test Scores')\n",
    "plt.xlabel('Mean Score', fontsize=8)\n",
    "plt.ylabel('Diff Score', fontsize=8)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "deletable": false,
    "editable": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Average of Mean Squared Error:  107.19197845458984\n"
     ]
    }
   ],
   "source": [
    "avg = float(np.mean(mses_general))\n",
    "print (\"Average of Mean Squared Error: \", avg)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Our Avg Mean Squared Error is significantly better, clearly showing better results, although we still have an issue of underpredicting at higher test scores."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## School Multi-Task Model\n",
    "\n",
    "### mean_squared_error"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "from keras.layers import Input, Dense\n",
    "from keras.models import Model\n",
    "\n",
    "inputs = Input(shape=(X[0].shape[0],))\n",
    "shared_mlp = Dense(64, activation='relu')(inputs)\n",
    "shared_mlp = Dense(64, activation='relu')(shared_mlp)\n",
    "\n",
    "models_mtl = []\n",
    "\n",
    "for _ in range(len(tasks)):\n",
    "    personal_layer = Dense(10, activation='relu')(shared_mlp)\n",
    "    output_layer = Dense(1)(personal_layer)\n",
    "    model = Model(inputs=inputs, outputs=output_layer)\n",
    "    model.compile(optimizer='rmsprop', loss='mean_squared_error')\n",
    "    models_mtl.append(model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "all_data = []\n",
    "all_test = []\n",
    "\n",
    "weights = {}\n",
    "for bunch in tasks:\n",
    "    ID = int(bunch.ID.split(\" \")[-1]) - 1\n",
    "    X_train, X_test, y_train, y_test = train_test_split(bunch.data, bunch.target, test_size=0.4, random_state=42)\n",
    "    for row, target in zip(X_train, y_train):\n",
    "        instance = {}\n",
    "        instance[\"data\"] = row.reshape((1, 28))\n",
    "        instance[\"school\"] = ID\n",
    "        instance[\"target\"] = target\n",
    "        all_data.append(instance)\n",
    "        \n",
    "    for row, target in zip(X_test, y_test):\n",
    "        instance = {}\n",
    "        instance[\"data\"] = row.reshape((1, 28))\n",
    "        instance[\"school\"] = ID\n",
    "        instance[\"target\"] = target\n",
    "        all_test.append(instance)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "After separating the data correctly, shuffle it all back together and run it through 20 epochs on the shared model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "import random\n",
    "random.shuffle(all_data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "EPOCHS = 10\n",
    "for _underscore in range(10):\n",
    "    for instance in all_data:\n",
    "        models_mtl[instance[\"school\"]].train_on_batch(instance[\"data\"], instance[\"target\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "school_y = [ [] for x in range(139)]\n",
    "school_yhat = [ [] for x in range(139)]\n",
    "\n",
    "for test in all_test:\n",
    "    X_test = test[\"data\"]\n",
    "    y_test = test[\"target\"]\n",
    "    \n",
    "    mlp = models_mtl[test[\"school\"]]\n",
    "    \n",
    "    yhat_mlp = mlp.predict(X_test)\n",
    "    yhat_rounded = np.around(yhat_mlp)\n",
    "    \n",
    "    y_test_plot = y_test.ravel()\n",
    "    yhat_mlp_plot = yhat_mlp.ravel()\n",
    "\n",
    "    school_y[test[\"school\"]].append(y_test_plot)\n",
    "    school_yhat[test[\"school\"]].append(yhat_mlp_plot)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "y_mtl=[]\n",
    "y_mtl_std=[]\n",
    "yhat_mtl=[]\n",
    "yhat_mtl_std=[]\n",
    "mses_mtl=[]\n",
    "\n",
    "plt.figure(figsize=(15,10))\n",
    "\n",
    "for x, y in zip(school_y, school_yhat):\n",
    "    y_mtl.append(np.mean(x))\n",
    "    y_mtl_std.append(np.std(x))\n",
    "    yhat_mtl.append(np.mean(y))\n",
    "    yhat_mtl_std.append(np.std(y))\n",
    "    \n",
    "    mse = mean_squared_error(x, y)\n",
    "    mses_mtl.append(mse)\n",
    "    \n",
    "bland_altman_plot(y_mtl, yhat_mtl, y_mtl_std, yhat_mtl_std)\n",
    "plt.title('Bland-Altman Test Scores')\n",
    "plt.xlabel('Mean Score', fontsize=8)\n",
    "plt.ylabel('Diff Score', fontsize=8)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "avg = float(np.mean(mses_mtl))\n",
    "print (\"Average of Mean Squared Error: \", avg)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### mean_absolute_error"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "from sklearn.metrics import mean_absolute_error\n",
    "\n",
    "inputs = Input(shape=(X[0].shape[0],))\n",
    "shared_mlp = Dense(64, activation='relu')(inputs)\n",
    "shared_mlp = Dense(64, activation='relu')(shared_mlp)\n",
    "\n",
    "models_mtl = []\n",
    "\n",
    "for _ in range(len(tasks)):\n",
    "    personal_layer = Dense(10, activation='relu')(shared_mlp)\n",
    "    output_layer = Dense(1)(personal_layer)\n",
    "    model = Model(inputs=inputs, outputs=output_layer)\n",
    "    model.compile(optimizer='rmsprop', loss='mean_absolute_error')\n",
    "    models_mtl.append(model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "all_data = []\n",
    "all_test = []\n",
    "\n",
    "weights = {}\n",
    "for bunch in tasks:\n",
    "    ID = int(bunch.ID.split(\" \")[-1]) - 1\n",
    "    X_train, X_test, y_train, y_test = train_test_split(bunch.data, bunch.target, test_size=0.4, random_state=42)\n",
    "    for row, target in zip(X_train, y_train):\n",
    "        instance = {}\n",
    "        instance[\"data\"] = row.reshape((1, 28))\n",
    "        instance[\"school\"] = ID\n",
    "        instance[\"target\"] = target\n",
    "        all_data.append(instance)\n",
    "        \n",
    "    for row, target in zip(X_test, y_test):\n",
    "        instance = {}\n",
    "        instance[\"data\"] = row.reshape((1, 28))\n",
    "        instance[\"school\"] = ID\n",
    "        instance[\"target\"] = target\n",
    "        all_test.append(instance)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "import random\n",
    "random.shuffle(all_data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "EPOCHS = 10\n",
    "for _underscore in range(10):\n",
    "    for instance in all_data:\n",
    "        models_mtl[instance[\"school\"]].train_on_batch(instance[\"data\"], instance[\"target\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "school_ya = [ [] for x in range(139)]\n",
    "school_yhata = [ [] for x in range(139)]\n",
    "\n",
    "for test in all_test:\n",
    "    X_test = test[\"data\"]\n",
    "    y_test = test[\"target\"]\n",
    "    \n",
    "    mlp = models_mtl[test[\"school\"]]\n",
    "    \n",
    "    yhat_mlp = mlp.predict(X_test)\n",
    "    yhat_rounded = np.around(yhat_mlp)\n",
    "    \n",
    "    y_test_plot = y_test.ravel()\n",
    "    yhat_mlp_plot = yhat_mlp.ravel()\n",
    "\n",
    "    school_ya[test[\"school\"]].append(y_test_plot)\n",
    "    school_yhata[test[\"school\"]].append(yhat_mlp_plot)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "y_mtla=[]\n",
    "y_mtl_stda=[]\n",
    "yhat_mtla=[]\n",
    "yhat_mtl_stda=[]\n",
    "maes_mtl=[]\n",
    "\n",
    "plt.figure(figsize=(15,10))\n",
    "\n",
    "for x, y in zip(school_ya, school_yhata):\n",
    "    y_mtla.append(np.mean(x))\n",
    "    y_mtl_stda.append(np.std(x))\n",
    "    yhat_mtla.append(np.mean(y))\n",
    "    yhat_mtl_stda.append(np.std(y))\n",
    "    \n",
    "    mae = mean_absolute_error(x, y)\n",
    "    maes_mtl.append(mse)\n",
    "    \n",
    "bland_altman_plot(y_mtla, yhat_mtla, y_mtl_stda, yhat_mtl_stda)\n",
    "plt.title('Bland-Altman Test Scores')\n",
    "plt.xlabel('Mean Score', fontsize=8)\n",
    "plt.ylabel('Diff Score', fontsize=8)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "avg = float(np.mean(maes_mtl))\n",
    "print (\"Average of Mean Absolute Error: \", avg)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "So the results were a bit dissapointing. It turns out the General Model proved best for this problem, but that could be due to a variety of reasons, such as lack of data or a poorly built Multi-Task model. However, it should be noted that while the Avg Mean Squared Error is higher, the graph shows that the problem of overfitting to low test scores is less in the Multi-Task Model than the other previous two. So while it did score a higher error value, it did prove better on solving the overfitting problem.\n",
    "\n",
    "Now we will move on to an example that best shows how a Multi Task model can be extremely effective."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Computer Survey Data\n",
    "\n",
    "### Computer Survey Data Understanding\n",
    "This data set was obtained on pmtkdata project's web site:\n",
    "    https://code.google.com/p/pmtkdata/.\n",
    "It was used by \"Argyriou, Evgeniou, Pontil - Convex multi-task feature learning - ML 2008\" and \"Kumar, Daume III - Learning Task Grouping and Overlap in Multi-Task Learning - ICML 2012\".\n",
    "\n",
    "The data set's properties are:\n",
    "   \n",
    "20 samples (all subjects have the same samples)\n",
    "13 features (describing computer's intrinsic (e.g. CPU speed, RAM) and extrinsic (e.g. Hot line support, Color of unit) features)\n",
    "regression task of predicting subject's likelihood of purchasing a personal computer\n",
    "   \n",
    "Each data point is 20 x 14, which is each of the computers for that respective student. The computers are the students and the students are the school while the ratings are the test scores and the comp specs are the demographs."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "import os.path\n",
    "import scipy.io\n",
    "from sklearn.datasets.base import Bunch\n",
    "from matplotlib import pyplot as plt\n",
    "import numpy as np\n",
    "from sklearn.model_selection import train_test_split\n",
    "\n",
    "def load_computer_survey_data():\n",
    "    \"\"\"\n",
    "    Returns\n",
    "    -------\n",
    "    tasks -- list\n",
    "        A list of Bunch objects that correspond to regression tasks, each task\n",
    "        corresponding to one subject.\n",
    "    \"\"\"\n",
    "    \n",
    "    #matlab_file = os.path.join(\"data/data/computer/conjointAnalysisComputerBuyers.mat\")\n",
    "    mat = scipy.io.loadmat(\"data/computer_survey/conjointAnalysisComputerBuyers.mat\")\n",
    "    # feature names are taken from the ComputerSurveyFull.xls file, Sheet named\n",
    "    # Design Matrix\n",
    "    feature_names = [\"cnst\", \"Hot line\", \"RAM\", \"Screen size\", \"CPU speed\",\n",
    "                     \"Hard disk\", \"CD ROM\", \"Cache\", \"Color of unit\",\n",
    "                     \"Availability\", \"Warranty\", \"Software\", \"Guarantee\",\n",
    "                     \"Price\"]\n",
    "    # extract X and combined y data\n",
    "    X = mat[\"designMarix\"].astype(\"float\")\n",
    "    y_combined = mat[\"likeBuy\"].astype(\"float\")\n",
    "    # create a task for each subject\n",
    "    tasks = []\n",
    "    for subject in range(len(y_combined)):\n",
    "        descr = \"Computer Survey data: Subject {}\".format(subject)\n",
    "        id = \"Subject {}\".format(subject)\n",
    "        y = np.array(y_combined[subject], copy=True)\n",
    "        tasks.append(Bunch(data=X,\n",
    "                           target=y.reshape(20, 1),\n",
    "                           feature_names=feature_names,\n",
    "                           DESCR=descr,\n",
    "                           ID=id))\n",
    "    return tasks"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "comp_tasks = load_computer_survey_data()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": false,
    "editable": false
   },
   "source": [
    "## Individual Computer Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "import keras\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Reshape\n",
    "from keras.layers import Dense, Dropout, Activation, Flatten\n",
    "from sklearn.metrics import mean_squared_error, accuracy_score, precision_score, mean_squared_log_error"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "modelsc_ind = []\n",
    "testc_daters_ind = []\n",
    "for x in comp_tasks:\n",
    "    X_train, X_test, y_train, y_test = train_test_split(x.data, x.target, test_size=0.25, random_state=42)\n",
    "\n",
    "    mlp = Sequential()\n",
    "    mlp.add(Dense(input_dim=X_train.shape[1], units=30, activation='relu'))\n",
    "    mlp.add( Dense(units=15, activation='relu') )\n",
    "    mlp.add(Dense(1))\n",
    "\n",
    "    mlp.compile(loss='mean_squared_error',\n",
    "                optimizer='rmsprop',\n",
    "                metrics=[])\n",
    "    \n",
    "    mlp.fit(X_train, y_train, epochs=10, shuffle=True, verbose=0)\n",
    "    modelsc_ind.append(mlp)\n",
    "    testc_daters_ind.append([X_test, y_test])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "yc_ind=[]\n",
    "yc_ind_std=[]\n",
    "yhatc_ind=[]\n",
    "yhatc_ind_std=[]\n",
    "msesc=[]\n",
    "\n",
    "for t, mlp in zip(testc_daters_ind, modelsc_ind):\n",
    "    X_test, y_test = t\n",
    "\n",
    "    yhat_mlp = mlp.predict(X_test)\n",
    "    yhat_rounded = np.around(yhat_mlp)\n",
    "    \n",
    "    y_test_plot = y_test.ravel()\n",
    "    yhat_mlp_plot = yhat_mlp.ravel()\n",
    "    \n",
    "    yc_ind.append(np.mean(y_test_plot))\n",
    "    yc_ind_std.append(np.std(y_test_plot))\n",
    "    yhatc_ind.append(np.mean(yhat_mlp_plot))\n",
    "    yhatc_ind_std.append(np.std(yhat_mlp_plot))\n",
    "\n",
    "    mse = mean_squared_error(y_test, yhat_mlp)\n",
    "    msesc.append(mse)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "plt.figure(figsize=(15,10))\n",
    "bland_altman_plot(yc_ind, yhatc_ind, yc_ind_std, yhatc_ind_std)\n",
    "plt.title('Bland-Altman Test Scores')\n",
    "plt.xlabel('Mean Score', fontsize=8)\n",
    "plt.ylabel('Diff Score', fontsize=8)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "avg = float(np.mean(msesc))\n",
    "print (\"Average of Mean Squared Error: \", avg)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## General Computer Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "all_comp_data = {\"data\": [], \"target\": [], \"data_test\": [], \"target_test\": []}\n",
    "for bunch in comp_tasks:\n",
    "    x_temp, y_temp = [bunch.data, bunch.target]\n",
    "    X_train, X_test, y_train, y_test = train_test_split(x_temp, y_temp, test_size=0.25, random_state=42)\n",
    "    all_comp_data[\"data\"].append(X_train)\n",
    "    all_comp_data[\"target\"].append(y_train)        \n",
    "    all_comp_data[\"data_test\"].append(X_test)\n",
    "    all_comp_data[\"target_test\"].append(y_test)\n",
    "\n",
    "X_train_comp = np.concatenate(all_comp_data[\"data\"])\n",
    "y_train_comp = np.concatenate(all_comp_data[\"target\"]).reshape(2850, 1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "generalizedc_mlp = Sequential()\n",
    "generalizedc_mlp.add(Dense(input_dim=X_train_comp.shape[1], units=20, activation='relu'))\n",
    "generalizedc_mlp.add(Dense(units=10, activation='relu'))\n",
    "generalizedc_mlp.add(Dense(1))\n",
    "\n",
    "generalizedc_mlp.compile(loss='mean_squared_error',\n",
    "            optimizer='rmsprop',\n",
    "            metrics=[])\n",
    "\n",
    "generalizedc_mlp.fit(X_train_comp, y_train_comp, epochs=20, shuffle=True, verbose=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "yc_general=[]\n",
    "yc_general_std=[]\n",
    "yhatc_general=[]\n",
    "yhatc_general_std=[]\n",
    "msesc_general=[]\n",
    "\n",
    "for X_test, y_test in zip(all_comp_data[\"data_test\"], all_comp_data[\"target_test\"]):\n",
    "    yhat_mlp = generalizedc_mlp.predict(X_test)\n",
    "\n",
    "    y_test_plot = y_test.ravel()\n",
    "    yhat_mlp_plot = yhat_mlp.ravel()\n",
    "    \n",
    "    yc_general.append(np.mean(y_test_plot))\n",
    "    yc_general_std.append(np.std(y_test_plot))\n",
    "    yhatc_general.append(np.mean(yhat_mlp_plot))\n",
    "    yhatc_general_std.append(np.std(yhat_mlp_plot))\n",
    "    \n",
    "    mse = mean_squared_error(y_test, yhat_mlp)\n",
    "    msesc_general.append(mse)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false,
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "plt.figure(figsize=(15,10))\n",
    "bland_altman_plot(yc_general, yhatc_general, yc_general_std, yhatc_general_std)\n",
    "plt.title('Bland-Altman Test Scores')\n",
    "plt.xlabel('Mean Score', fontsize=8)\n",
    "plt.ylabel('Diff Score', fontsize=8)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "avg = float(np.mean(msesc_general))\n",
    "print (\"Average Mean Squared Error: \", avg)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Computer Survey Multi-Task Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "from keras.layers import Input, Dense\n",
    "from keras.models import Model\n",
    "\n",
    "inputs = Input(shape=(X_train_comp.shape[1],))\n",
    "shared_mlp = Dense(64, activation='relu')(inputs)\n",
    "shared_mlp = Dense(64, activation='relu')(shared_mlp)\n",
    "\n",
    "models_mtl = []\n",
    "\n",
    "for _ in range(len(comp_tasks)):\n",
    "    personal_layer = Dense(10, activation='relu')(shared_mlp)\n",
    "    output_layer = Dense(1)(personal_layer)\n",
    "    model = Model(inputs=inputs, outputs=output_layer)\n",
    "    model.compile(optimizer='rmsprop', loss='mean_squared_error')\n",
    "    models_mtl.append(model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "all_comp_data = []\n",
    "all_comp_test = []\n",
    "\n",
    "weights = {}\n",
    "for bunch in comp_tasks:\n",
    "    ID = int(bunch.ID.split(\" \")[-1])\n",
    "    X_train, X_test, y_train, y_test = train_test_split(bunch.data, bunch.target, test_size=0.25, random_state=42)\n",
    "    for row, target in zip(X_train, y_train):\n",
    "        instance = {}\n",
    "        instance[\"data\"] = row.reshape((1, 14))\n",
    "        instance[\"subject\"] = ID\n",
    "        instance[\"target\"] = target\n",
    "        all_comp_data.append(instance)\n",
    "        \n",
    "    for row, target in zip(X_test, y_test):\n",
    "        instance = {}\n",
    "        instance[\"data\"] = row.reshape((1, 14))\n",
    "        instance[\"subject\"] = ID\n",
    "        instance[\"target\"] = target\n",
    "        all_comp_test.append(instance)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "import random\n",
    "random.shuffle(all_comp_data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "EPOCHS = 10\n",
    "for _underscore in range(10):\n",
    "    for instance in all_comp_data:\n",
    "        models_mtl[instance[\"subject\"]].train_on_batch(instance[\"data\"], instance[\"target\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "computer_ya = [ [] for x in range(190)]\n",
    "computer_yhata = [ [] for x in range(190)]\n",
    "\n",
    "for test in all_comp_test:\n",
    "    X_test = test[\"data\"]\n",
    "    y_test = test[\"target\"]\n",
    "    \n",
    "    mlp = models_mtl[test[\"subject\"]]\n",
    "    \n",
    "    yhat_mlp = mlp.predict(X_test)\n",
    "    \n",
    "    y_test_plot = y_test.ravel()\n",
    "    yhat_mlp_plot = yhat_mlp.ravel()\n",
    "\n",
    "    computer_ya[test[\"subject\"]].append(y_test_plot)\n",
    "    computer_yhata[test[\"subject\"]].append(yhat_mlp_plot)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "yc_mtl=[]\n",
    "yc_mtl_std=[]\n",
    "yhatc_mtl=[]\n",
    "yhatc_mtl_std=[]\n",
    "msesc_mtl=[]\n",
    "\n",
    "plt.figure(figsize=(15,10))\n",
    "\n",
    "for x, y in zip(computer_ya, computer_yhata):\n",
    "    yc_mtl.append(np.mean(x))\n",
    "    yc_mtl_std.append(np.std(x))\n",
    "    yhatc_mtl.append(np.mean(y))\n",
    "    yhatc_mtl_std.append(np.std(y))\n",
    "    \n",
    "    mse = mean_squared_error(x, y)\n",
    "    msesc_mtl.append(mse)\n",
    "    \n",
    "bland_altman_plot(yc_mtl, yhatc_mtl, yc_mtl_std, yhatc_mtl_std)\n",
    "plt.title('Bland-Altman Test Scores')\n",
    "plt.xlabel('Mean Score', fontsize=8)\n",
    "plt.ylabel('Diff Score', fontsize=8)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "avg = float(np.mean(msesc_mtl))\n",
    "print (\"Average of Mean Squared Error: \", avg)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# ChemBL\n",
    "\n",
    "ChEMBL is an actively maintained and used chemical database of bioactive molecules with drug-like properties. The European Bioinformatics Institute (EBI) is in charge of keeping up to date with it. \n",
    "\n",
    "\n",
    "## ChemBL Data\n",
    "The first part of the code is meant to take in the ChemBL molecules and transform them into a readable data set in order for us to use it with rdkit. \n",
    "\n",
    "TODO: Pray to Pavel that he will help us again."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "import pandas, numpy\n",
    "\n",
    "def cleanUp(filename, BA='IC50', UNITS='nM'):\n",
    "   # Read in the file.\n",
    "   d = pandas.read_table(filename, sep=r',(?=(?:[^\"]*\"[^\"]*\")*[^\"]*$)', engine='python')\n",
    "\n",
    "   # Eliminate duplicates.\n",
    "   #d =d.drop_duplicates(['ACTIVITY_ID'])\n",
    "\n",
    "   # Retain only IC50 entries\n",
    "   # (do not mess with the conversion -- like, from logIC50 or pIC50 to IC50 -- as the fraction of\n",
    "   # those entries is negligible).\n",
    "   d = d[d.STANDARD_TYPE==BA]\n",
    "\n",
    "   # Restrict to exclusively human targets\n",
    "   # (this will take care of unwanted 'Starlite ...' entries as well).\n",
    "   #d = d[d.NAME.str.startswith('h')]\n",
    "\n",
    "   # Retain only the nM entries.\n",
    "   d = d[d.STANDARD_UNITS==UNITS]\n",
    "\n",
    "   # Integerize the DOM_ID column (the COMPOUND_ID column type is already 'int').\n",
    "   d.DOC_ID = d.DOC_ID.astype(int)\n",
    "\n",
    "   # Delete idle columns.\n",
    "   #d.drop([Column Name or list],inplace=True,axis=1)\n",
    "\n",
    "   return d\n",
    "\n",
    "def binarizeInteractions(d, MIN=100, MAX=10000):\n",
    "   # Apply the interaction criterion.\n",
    "   ci = d[((d.STANDARD_RELATION=='=') | (d.STANDARD_RELATION=='~') | (d.STANDARD_RELATION=='<') | (d.STANDARD_RELATION=='<=')) & (d.STANDARD_VALUE<=MIN)]\n",
    "   # Apply the non-interaction criterion.\n",
    "   no = d[((d.STANDARD_RELATION=='=') | (d.STANDARD_RELATION=='~') | (d.STANDARD_RELATION=='>') | (d.STANDARD_RELATION=='>>') | (d.STANDARD_RELATION=='>=')) & (d.STANDARD_VALUE>=MAX)]\n",
    "   print(\"Look up how to suppress the warning message below.\")\n",
    "   ci.STANDARD_VALUE = 1\n",
    "   no.STANDARD_VALUE = 0\n",
    "   return pandas.concat([ci, no])\n",
    "\n",
    "d = cleanUp(\"generated_data/molecules.csv\")\n",
    "cino = binarizeInteractions(d)\n",
    "# Remove entries with less than 50 ligands per target.\n",
    "cino = cino.groupby('MOLGREGNO').filter(lambda x: len(x) > 49)\n",
    "\n",
    "dom_ID = numpy.unique( cino.DOC_ID.value_counts().index )\n",
    "com_ID = numpy.unique( cino.RECORD_ID.value_counts().index )\n",
    "print (com_ID.size, dom_ID.size)\n",
    "\n",
    "# Initialize the target-compound matrix.\n",
    "ma = pandas.DataFrame( numpy.nan*numpy.ones((com_ID.size, dom_ID.size)), columns=dom_ID, index=com_ID )\n",
    "# Fill in the target-compound matrix and save it as a file.\n",
    "for i, r in cino.iterrows():\n",
    "   ma[r.DOC_ID][r.RECORD_ID] = r.STANDARD_VALUE\n",
    "ma.to_csv(\"ChemBL/kinda_works.csv\")\n",
    "\n",
    "#d = pandas.read_csv('ChemBL/kinda_works.csv')\n",
    "#d.columns = d.columns.astype(int)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": false,
    "editable": false
   },
   "outputs": [],
   "source": [
    "from __future__ import print_function\n",
    "\n",
    "from rdkit import Chem\n",
    "from rdkit.Chem import MACCSkeys, AllChem\n",
    "from rdkit.Avalon import pyAvalonTools as fpAvalon\n",
    "from rdkit.Chem.AtomPairs import Pairs, Torsions\n",
    "from rdkit.Chem.Fingerprints import FingerprintMols\n",
    "from rdkit.Chem.ChemicalFeatures import BuildFeatureFactory\n",
    "from rdkit.Chem import rdMolDescriptors\n",
    "import pandas as pd\n",
    "import argparse\n",
    "import textwrap\n",
    "import getopt\n",
    "from argparse import ArgumentParser\n",
    "import time\n",
    "\n",
    "def CalculateFP_mod(smiles, N_BITS=512):\n",
    "    m = Chem.MolFromSmiles(smiles)\n",
    "    if m is None:\n",
    "        raise ValueError('SMILES cannot be converted to a RDKit molecules:', smiles)\n",
    "\n",
    "    return AllChem.GetMorganFingerprintAsBitVect(m, 2, nBits=N_BITS)\n",
    "\n",
    "\n",
    "if __name__ == \"__main__\":\n",
    "    parser = ArgumentParser()\n",
    "\n",
    "    parser.add_argument(\"-i\", \"--input\", dest=\"ligandsWithSmiles\", required=True,help=\"expected file containing 2 cols molregno,smiles\", metavar=\"FILE\")\n",
    "\n",
    "    parser.add_argument(\"-s\", \"--size\", dest=\"numBits\", default=512,help=\"Size of fingerprints\",metavar=\"NumBits\")\n",
    "\n",
    "\n",
    "\n",
    "    args = parser.parse_args()\n",
    "\n",
    "    print ( \"reading from file   %s\" %(args.ligandsWithSmiles))\n",
    "\n",
    "    ipFile = args.ligandsWithSmiles\n",
    "\n",
    "    df = pd.read_csv(ipFile)\n",
    "\n",
    "\n",
    "\n",
    "    numBits = int(args.numBits)\n",
    "    count = 0\n",
    "    errorCompounds = []\n",
    "\n",
    "\n",
    "\n",
    "    current_milli_time = lambda: int(round(time.time() * 1000))\n",
    "\n",
    "    opFile = \"ChemBL/CompoundFingerPrints_%d_%s.csv\"%(numBits,current_milli_time())\n",
    "\n",
    "    print(\"Generating output file with fingerprints %s\"%opFile)\n",
    "    with open(opFile,\"w\") as fpFile :\n",
    "\n",
    "        tempLine = \"\"\n",
    "\n",
    "        tempLine =\"molregno\"\n",
    "        for i in range(0,numBits):\n",
    "            tempLine=tempLine+\",\"+\"Compound_\"+str(i)\n",
    "        fpFile.writelines(tempLine+\"\\n\")\n",
    "\n",
    "        tempLine = \"\"\n",
    "\n",
    "        for row in df.iterrows():\n",
    "            mol= row[1][0]\n",
    "            smilesCode= row[1][1]\n",
    "\n",
    "            try:\n",
    "                count = count + 1\n",
    "                tempLine = str(mol)\n",
    "\n",
    "                temp = CalculateFP_mod(smilesCode)\n",
    "                strFP = temp.ToBitString()\n",
    "\n",
    "                for i in range(0,len(strFP)):\n",
    "                    tempLine=tempLine+\",\"+strFP[i]\n",
    "\n",
    "\n",
    "                fpFile.writelines(tempLine+\"\\n\")\n",
    "\n",
    "                if count%100000 ==0 :\n",
    "                    print(count)\n",
    "            except :\n",
    "                errorCompounds.append(mol)\n",
    "\n",
    "\n",
    "    print(\"Generating error.log for failed compounds\")\n",
    "    with open(\"ChemBL/error.log\",\"w\") as fp :\n",
    "        fp.write(\"Following compounds errored out \\n\")\n",
    "        fp.write(str(errorCompounds))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Multi-modal Summary\n",
    "A multi-modal design can be defined as a system designed to fufill one task based on a variety of similar input. The reason for this design is due to the many advantages it offers in the right situations. For example, in the medical field, a doctor uses multiple image data (such as a CT, MRI, ultrasound, etc) in order to be able to diagnose a patient more effectively. \n",
    "   \n",
    "There are a varierty of ways one can go about combining similar data together to be used. Early fusion is when the different data sources are combined together in the pre-processing phase through either PCA, ICA, or CCA and then presented to the machine learning algorithm to be processed. Later on, late fusion became the norm as ensemble classifiers became more and more popular. This method was typically easier and better to implement, especially if the different inputs varied significantly. However, recently, Deep Neural Network Architectures have paved a way for a third type, Indermediate fusion, a method of merging the modality throughout the many hidden layers and connecting the data on a very abstract, yet effective, level.\n",
    "<img src=\"images/multi-modal-graph.png\" width=\"700\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## References\n",
    "Image #1: http://stanford.edu/~eadeli/\n",
    "\n",
    "Image #2: http://ruder.io/multi-task/\n",
    "\n",
    "Bland-Altman picture & info : https://en.wikipedia.org/wiki/Bland%E2%80%93Altman_plot\n",
    "\n",
    "Source: https://keras.io/getting-started/functional-api-guide/#multi-input-and-multi-output-models\n",
    "\n",
    "Bland Altman: https://en.wikipedia.org/wiki/Bland%E2%80%93Altman_plot\n",
    "\n",
    "School Papers:\n",
    "\n",
    "Overview of Multi-Task Learning: https://arxiv.org/pdf/1706.05098.pdf\n",
    "\n",
    "Multi-modal Learning: https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8103116\n",
    "   \n",
    "2015 https://www.ijcai.org/Proceedings/15/Papers/512.pdf\n",
    "   \n",
    "2017 https://arxiv.org/pdf/1703.00977.pdf\n",
    "\n",
    "2017 https://arxiv.org/pdf/1611.03427.pdf\n",
    "\n",
    "data import: https://github.com/tjanez/PyMTL/blob/master/src/PyMTL/data.py"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}