example/recommenders/demo1-MF2-fancy.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Matrix Factorization (MF) part 2: Getting Fancy\n",
    "Demonstrates matrix factorization with MXNet on the [MovieLens 100k](http://grouplens.org/datasets/movielens/100k/) dataset.  This is an extension of [part 1](demo1-MF.ipynb) where we try fancy optimizers and network structures.\n",
    "\n",
    "You need to have python package pandas and bokeh installed (pip install pandas bokeh)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import mxnet as mx\n",
    "from movielens_data import get_data_iter, max_id\n",
    "from matrix_fact import train"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# If MXNet is not compiled with GPU support (e.g. on OSX), set to [mx.cpu(0)]\n",
    "# Can be changed to [mx.gpu(0), mx.gpu(1), ..., mx.gpu(N-1)] if there are N GPUs\n",
    "ctx = [mx.gpu(0)]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "train_test_data = get_data_iter(batch_size=100)\n",
    "max_user, max_item = max_id('./ml-100k/u.data')\n",
    "(max_user, max_item)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Linear MF\n",
    "Same as before, but this time with the [Adam optimizer](https://arxiv.org/abs/1412.6980) which will often converge much faster than SGD w/ momentum as we used before.  You should see this model over-fitting quickly. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def plain_net(k):\n",
    "    # input\n",
    "    user = mx.symbol.Variable('user')\n",
    "    item = mx.symbol.Variable('item')\n",
    "    score = mx.symbol.Variable('score')\n",
    "    # user feature lookup\n",
    "    user = mx.symbol.Embedding(data = user, input_dim = max_user, output_dim = k) \n",
    "    # item feature lookup\n",
    "    item = mx.symbol.Embedding(data = item, input_dim = max_item, output_dim = k)\n",
    "    # predict by the inner product, which is elementwise product and then sum\n",
    "    pred = user * item\n",
    "    pred = mx.symbol.sum_axis(data = pred, axis = 1)\n",
    "    pred = mx.symbol.Flatten(data = pred)\n",
    "    # loss layer\n",
    "    pred = mx.symbol.LinearRegressionOutput(data = pred, label = score)\n",
    "    return pred\n",
    "\n",
    "net1 = plain_net(64)\n",
    "mx.viz.plot_network(net1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "results1 = train(net1, train_test_data, num_epoch=25, learning_rate=0.001, optimizer='adam', ctx=ctx)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Neural Network (non-linear) MF\n",
    "The non-linear model converges strangely with Adam."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "def get_one_layer_mlp(hidden, k):\n",
    "    # input\n",
    "    user = mx.symbol.Variable('user')\n",
    "    item = mx.symbol.Variable('item')\n",
    "    score = mx.symbol.Variable('score')\n",
    "    # user latent features\n",
    "    user = mx.symbol.Embedding(data = user, input_dim = max_user, output_dim = k)\n",
    "    user = mx.symbol.Activation(data = user, act_type='relu')\n",
    "    user = mx.symbol.FullyConnected(data = user, num_hidden = hidden)\n",
    "    # item latent features\n",
    "    item = mx.symbol.Embedding(data = item, input_dim = max_item, output_dim = k)\n",
    "    item = mx.symbol.Activation(data = item, act_type='relu')\n",
    "    item = mx.symbol.FullyConnected(data = item, num_hidden = hidden)\n",
    "    # predict by the inner product\n",
    "    pred = user * item\n",
    "    pred = mx.symbol.sum_axis(data = pred, axis = 1)\n",
    "    pred = mx.symbol.Flatten(data = pred)\n",
    "    # loss layer\n",
    "    pred = mx.symbol.LinearRegressionOutput(data = pred, label = score)\n",
    "    return pred\n",
    "\n",
    "net2 = get_one_layer_mlp(64, 64)\n",
    "mx.viz.plot_network(net2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "results2 = train(net2, train_test_data, num_epoch=20, learning_rate=0.001, optimizer='adam', ctx=ctx)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Deep Neural Network (Residual Network / ResNet)\n",
    "Borrowing ideas from [Deep Residual Learning for Image Recognition (He, et al.)](https://arxiv.org/abs/1512.03385) to build a complex deep network that is aggressively regularized to avoid over-fitting, but still achieves good performance. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "def get_multi_layer_dropout_resnet(hidden, k):\n",
    "    # input\n",
    "    user = mx.symbol.Variable('user')\n",
    "    item = mx.symbol.Variable('item')\n",
    "    score = mx.symbol.Variable('score')\n",
    "    # user latent features\n",
    "    user1 = mx.symbol.Embedding(data = user, input_dim = max_user, output_dim = k)\n",
    "    user = mx.symbol.FullyConnected(data = user1, num_hidden = hidden)\n",
    "    user = mx.symbol.Activation(data = user, act_type='relu')\n",
    "    user = mx.symbol.Dropout(data=user, p=0.5)\n",
    "    user = mx.symbol.FullyConnected(data = user, num_hidden = hidden)\n",
    "    user2 = user + user1\n",
    "    user2 = mx.symbol.Dropout(data=user2, p=0.5)\n",
    "    user = mx.symbol.FullyConnected(data = user2, num_hidden = hidden)\n",
    "    user = mx.symbol.Activation(data = user, act_type='relu')\n",
    "    user = mx.symbol.Dropout(data=user, p=0.5)\n",
    "    user = mx.symbol.FullyConnected(data = user, num_hidden = hidden)\n",
    "    user = user + user2\n",
    "    # item latent features\n",
    "    item1 = mx.symbol.Embedding(data = item, input_dim = max_item, output_dim = k)\n",
    "    item = mx.symbol.FullyConnected(data = item1, num_hidden = hidden)\n",
    "    item = mx.symbol.Activation(data = item, act_type='relu')\n",
    "    item = mx.symbol.Dropout(data=item, p=0.5)    \n",
    "    item = mx.symbol.FullyConnected(data=item, num_hidden = hidden)\n",
    "    item2 = item + item1\n",
    "    item2 = mx.symbol.Dropout(data=item2, p=0.5)    \n",
    "    item = mx.symbol.FullyConnected(data = item2, num_hidden = hidden)\n",
    "    item = mx.symbol.Activation(data = item, act_type='relu')\n",
    "    item = mx.symbol.Dropout(data=item, p=0.5)    \n",
    "    item = mx.symbol.FullyConnected(data=item, num_hidden = hidden)\n",
    "    item = item + item2\n",
    "    # predict by the inner product\n",
    "    pred = user * item\n",
    "    pred = mx.symbol.sum_axis(data = pred, axis = 1)\n",
    "    pred = mx.symbol.Flatten(data = pred)\n",
    "    # loss layer\n",
    "    pred = mx.symbol.LinearRegressionOutput(data = pred, label = score)\n",
    "    return pred\n",
    "net3 = get_multi_layer_dropout_resnet(64, 64)\n",
    "mx.viz.plot_network(net3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Larger batch size makes GPU more efficient for this complex model\n",
    "train_test_data2 = get_data_iter(batch_size=200) \n",
    "results3 = train(net3, train_test_data2, num_epoch=25, learning_rate=0.001, optimizer='adam', ctx=ctx)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Visualizing results\n",
    "Compare accuracy and training time across the models."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import bokeh\n",
    "import bokeh.io\n",
    "import bokeh.plotting\n",
    "bokeh.io.output_notebook()\n",
    "import pandas as pd\n",
    "\n",
    "def viz_lines(fig, results, legend, color):\n",
    "    df = pd.DataFrame(results._data['eval'])\n",
    "    fig.line(df.elapsed,df.RMSE, color=color, legend=legend, line_width=2)\n",
    "    df = pd.DataFrame(results._data['train'])\n",
    "    fig.line(df.elapsed,df.RMSE, color=color, line_dash='dotted', alpha=0.1)\n",
    "\n",
    "fig = bokeh.plotting.Figure(x_axis_type='datetime', x_axis_label='Training time', y_axis_label='RMSE')\n",
    "viz_lines(fig, results1, \"Linear MF\", \"orange\")\n",
    "viz_lines(fig, results2, \"MLP\", \"blue\")\n",
    "viz_lines(fig, results3, \"ResNet\", \"red\")\n",
    "\n",
    "bokeh.io.show(fig)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "kernelspec": {
   "display_name": "Python [Root]",
   "language": "python",
   "name": "Python [Root]"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}