Reward-augmented maximum likelihood learning for autoencoders.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import tensorflow as tf\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def build_model(num_timesteps, num_samples, vocab_size, embedding_size,\n",
    "                keep_prob=1.0, cell_size=256):\n",
    "    \"\"\"\n",
    "    Build an RML-enabled seq2seq model which accepts `num_samples` different\n",
    "    target output sequences for a given input sequence.\n",
    "    \n",
    "    Returns:\n",
    "        inputs: List of `batch_size` tensors; list of input ID tensors\n",
    "        targets: List of lists of `batch_size` tensors; each sublist is a\n",
    "            sequence of target ID tensors\n",
    "        loss: Sequence loss averaged over the `num_samples`-many target\n",
    "            sequences\n",
    "    \"\"\"\n",
    "    cell = tf.nn.rnn_cell.GRUCell(cell_size)\n",
    "    if keep_prob < 1.0:\n",
    "        cell = tf.nn.rnn_cell.DropoutWrapper(cell, output_keep_prob=keep_prob)\n",
    "        \n",
    "    inputs = [tf.placeholder(tf.int32, shape=(None,), name=\"inputs/%i\" % t)\n",
    "              for t in range(num_timesteps)]\n",
    "    targets = [[tf.placeholder(tf.int32, shape=(None,), name=\"targets/%i/%i\" % (i, t))\n",
    "                for t in range(num_timesteps)]\n",
    "               for i in range(num_samples)]\n",
    "    real_targets = [tf.placeholder(tf.int32, shape=(None,), name=\"targets/real/%i\" % t)\n",
    "                    for t in range(num_timesteps)]\n",
    "    \n",
    "    # Decoder inputs: prepend GO token to target sequence; remove last value from targets\n",
    "    decoder_inputs = [tf.pack([tf.zeros_like(targets_i[0])] + targets_i[:num_timesteps - 1])\n",
    "                      for targets_i in targets]\n",
    "    real_decoder_inputs = [tf.zeros_like(real_targets[0])] + real_targets[:-1]\n",
    "    \n",
    "    weights = [tf.ones_like(targets[0][0], dtype=tf.float32)\n",
    "               for t in range(num_timesteps)]\n",
    "    \n",
    "    def seq2seq_loss(dec_inp_i, targets_i):\n",
    "        outputs, state = tf.nn.seq2seq.embedding_rnn_seq2seq(\n",
    "            inputs, targets_i, cell, vocab_size, vocab_size, embedding_size)\n",
    "        \n",
    "        loss = tf.nn.seq2seq.sequence_loss(outputs, targets_i, weights)\n",
    "        return loss\n",
    "    \n",
    "    # Build a separate seq2seq graph for each of the `num_samples` feedforwards\n",
    "    # we perform for each batch.\n",
    "    seq2seq_loss = tf.make_template(\"seq2seq_loss\", seq2seq_loss)\n",
    "    losses = [seq2seq_loss(dec_inp_i, targets_i)\n",
    "              for dec_inp_i, targets_i in zip(decoder_inputs, targets)]\n",
    "    # Hill-climb with mean gradients = hill-climb with mean loss.\n",
    "    loss = tf.add_n(losses) / float(len(losses))\n",
    "    \n",
    "    real_loss = seq2seq_loss(real_decoder_inputs, real_targets)\n",
    "    \n",
    "    return inputs, targets, loss, real_targets, real_loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def modify_with_hamming_distance(y, distance, vocab_size):\n",
    "    assert distance >= 0\n",
    "    assert distance <= len(y)\n",
    "    if distance == 0:\n",
    "        return np.copy(y)\n",
    "    \n",
    "    idxs = np.random.choice(len(y), replace=False, size=distance)\n",
    "    replacements = np.random.choice(vocab_size, size=distance)\n",
    "    # TODO: account for case where replacement is same as existing char\n",
    "    \n",
    "    ret = np.copy(y)\n",
    "    ret[idxs] = replacements\n",
    "    return ret\n",
    "\n",
    "def sample_targets_hamming_distance(y, n, vocab_size, p=0.8):\n",
    "    \"\"\"\n",
    "    Sample `n` partially-correct targets by modifying the correct target\n",
    "    sequence `y` under a Hamming distance reward function.\n",
    "    \"\"\"\n",
    "    distances = np.minimum(len(y), np.random.geometric(p, n) - 1)\n",
    "    return [modify_with_hamming_distance(y, distance, vocab_size)\n",
    "            for distance in distances]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def train_batch(model, train_op, sess, batch_size, num_timesteps,\n",
    "                num_samples, vocab_size):\n",
    "    inputs, targets, loss, real_targets, real_loss = model\n",
    "    \n",
    "    b_inputs = np.random.choice(vocab_size, size=(num_timesteps, batch_size),\n",
    "                                replace=True)\n",
    "    b_targets = np.copy(b_inputs)\n",
    "    b_targets_real = np.copy(b_targets)\n",
    "    b_targets = [sample_targets_hamming_distance(b_targets_i, num_samples, vocab_size)\n",
    "                 for b_targets_i in b_targets]\n",
    "    b_targets = np.array(b_targets).transpose(1, 0, 2)\n",
    "    \n",
    "    feed = {inputs[t]: b_inputs_t for t, b_inputs_t in enumerate(b_inputs)}\n",
    "    feed.update({real_targets[t]: b_targets_real_t\n",
    "                 for t, b_targets_real_t in enumerate(b_targets_real)})\n",
    "    feed.update({targets[j][t]: b_targets_samp_t\n",
    "                 for j, b_targets_samp in enumerate(b_targets)\n",
    "                 for t, b_targets_samp_t in enumerate(b_targets_samp)})\n",
    "                 \n",
    "    loss_b, real_loss_b, _ = sess.run((loss, real_loss, train_op), feed)\n",
    "    return loss_b, real_loss_b"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def train_model(num_samples, lr=0.1):\n",
    "    NUM_TIMESTEPS = 10\n",
    "    VOCAB_SIZE = 300\n",
    "    EMBEDDING_SIZE = 128\n",
    "\n",
    "    g = tf.Graph()\n",
    "    with g.as_default():\n",
    "        model = build_model(NUM_TIMESTEPS, num_samples, VOCAB_SIZE,\n",
    "                            EMBEDDING_SIZE)\n",
    "        optimizer = tf.train.MomentumOptimizer(lr, 0.9)\n",
    "        train_op = optimizer.minimize(model[2])\n",
    "\n",
    "    sess = tf.InteractiveSession(graph=g)\n",
    "    sess.run(tf.initialize_all_variables())\n",
    "    \n",
    "    losses, real_losses = [], []\n",
    "    for i in range(500):\n",
    "        loss, real_loss = train_batch(model, train_op, sess, 64, NUM_TIMESTEPS,\n",
    "                                      num_samples, VOCAB_SIZE)\n",
    "        losses.append(loss)\n",
    "        real_losses.append(real_loss)\n",
    "        if i % 100 == 0:\n",
    "            print(\"\\t\", i)\n",
    "                \n",
    "    return losses, real_losses"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.5 1\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.5 2\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.5 4\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.5 8\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.1 1\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.1 2\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.1 4\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.1 8\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.05 1\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.05 2\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.05 4\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n",
      "0.05 8\n",
      "\t 0\n",
      "\t 100\n",
      "\t 200\n",
      "\t 300\n",
      "\t 400\n"
     ]
    }
   ],
   "source": [
    "losses = {}\n",
    "\n",
    "for lr in [0.5, 0.1, 0.05]:\n",
    "    for num_samples in [1, 2, 4, 8]:\n",
    "        print(lr, num_samples)\n",
    "        losses[(lr, num_samples)] = train_model(num_samples, lr=lr)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAgAAAAF6CAYAAABvKeDAAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nIiIi45ASABERkXHo/wN18YhLySlVoQAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f7805ff1ba8>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "learning_rates = set(lr for lr, _ in losses.keys())\n",
    "fig, axes = plt.subplots(len(learning_rates), 1, sharex=True, sharey=True)\n",
    "\n",
    "for learning_rate, axis in zip(learning_rates, axes):\n",
    "    lines = []\n",
    "    axis.set_title(\"lr=%.02f\"%learning_rate)\n",
    "    for (lr_j, num_samples), (_, real_losses_j) in losses.items():\n",
    "        if lr_j != learning_rate: continue\n",
    "        line, = axis.plot(real_losses_j, label=str(num_samples))\n",
    "        lines.append(line)\n",
    "    axis.legend(handles=lines)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7f780647e6d8>"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAgAAAAFkCAYAAABW9YMrAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f780647e8d0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plots = []\n",
    "for (lr, num_samples), (_, real_losses_j) in losses.items():\n",
    "    if lr != 0.5: continue\n",
    "    plot, = plt.plot(real_losses_j, label=str(num_samples))\n",
    "    plots.append(plot)\n",
    "plt.legend(handles=plots)"
   ]
  },
  {
   "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": "3.5.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}