example_sum.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "np.random.seed(1337)\n",
    "\n",
    "from keras import models\n",
    "from keras import optimizers\n",
    "from keras.layers import Input\n",
    "from keras.layers import Dense\n",
    "from keras.layers import LSTM\n",
    "from keras.layers import GlobalAveragePooling1D\n",
    "\n",
    "from attention.layers import AttentionLayer"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Generating the dataset\n",
    "\n",
    "This taks is a bit more challenging than the copy task on `example.ipynb`. In this case we want to sum\n",
    "both numbers marked with $1$ in the second line of the tape.\n",
    "\n",
    "\n",
    "Each sample is a \"tape\" formed by two rows: the top row $y[0]$ contains random numbers sampled from the interval $[0, 1)$, while the second, $y[1]$, is formed by a string of zeros, except for two positions, which have a one. The model is trained as a regressor to produce on the output the sum of the marked values from the first row. For example, given:\n",
    "\n",
    "$$ X = \n",
    "\\begin{bmatrix}\n",
    "    0.13 & 0.01 & 0.11 & 0.32 & 0.24 & 0.01 \\\\\n",
    "    0    & 1    & 0    & 0    &   1  & 0    \\\\\n",
    "\\end{bmatrix},\n",
    "$$\n",
    "\n",
    "the model output should be $\\mathcal{M}_\\theta(X) = 0.25$. We train two models, one using the attention mechanism and another using an average pooling. This models should be a bit more powerful than the one seen in `example.ipynb`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[0.26202468 1.        ]\n",
      " [0.15868397 0.        ]\n",
      " [0.27812652 0.        ]\n",
      " [0.45931689 0.        ]\n",
      " [0.32100054 0.        ]\n",
      " [0.51839282 0.        ]\n",
      " [0.26194293 0.        ]\n",
      " [0.97608528 0.        ]\n",
      " [0.73281455 0.        ]\n",
      " [0.11527423 0.        ]\n",
      " [0.38627507 0.        ]\n",
      " [0.62850118 1.        ]\n",
      " [0.12505793 0.        ]\n",
      " [0.98354861 0.        ]\n",
      " [0.44322487 0.        ]\n",
      " [0.78955834 0.        ]\n",
      " [0.79411858 0.        ]\n",
      " [0.36126157 0.        ]\n",
      " [0.41610394 0.        ]\n",
      " [0.58425813 0.        ]\n",
      " [0.76017177 0.        ]\n",
      " [0.18780841 0.        ]\n",
      " [0.28816715 0.        ]\n",
      " [0.67021886 0.        ]\n",
      " [0.49964826 0.        ]\n",
      " [0.17856868 0.        ]\n",
      " [0.4131413  0.        ]\n",
      " [0.19919524 0.        ]\n",
      " [0.5316994  0.        ]\n",
      " [0.8323707  0.        ]\n",
      " [0.18525095 0.        ]\n",
      " [0.95735922 0.        ]]\n",
      "sum = 0.8905258545552933\n",
      "\n",
      "[[0.42541467 0.        ]\n",
      " [0.50400704 0.        ]\n",
      " [0.51047095 0.        ]\n",
      " [0.01579145 0.        ]\n",
      " [0.73169007 0.        ]\n",
      " [0.99330504 1.        ]\n",
      " [0.16287753 0.        ]\n",
      " [0.12663478 0.        ]\n",
      " [0.37483418 0.        ]\n",
      " [0.69321944 0.        ]\n",
      " [0.00290103 0.        ]\n",
      " [0.36922906 0.        ]\n",
      " [0.05867933 0.        ]\n",
      " [0.78933609 0.        ]\n",
      " [0.34976921 0.        ]\n",
      " [0.70252372 0.        ]\n",
      " [0.49122908 0.        ]\n",
      " [0.97287233 1.        ]\n",
      " [0.83596788 0.        ]\n",
      " [0.61023733 0.        ]\n",
      " [0.56451898 0.        ]\n",
      " [0.99738914 0.        ]\n",
      " [0.25472414 0.        ]\n",
      " [0.01437786 0.        ]\n",
      " [0.08904507 0.        ]\n",
      " [0.9389826  0.        ]\n",
      " [0.9734594  0.        ]\n",
      " [0.49148153 0.        ]\n",
      " [0.34093597 0.        ]\n",
      " [0.72285762 0.        ]\n",
      " [0.01089676 0.        ]\n",
      " [0.75995318 0.        ]]\n",
      "sum = 1.9661773697596354\n",
      "\n",
      "[[0.67140279 0.        ]\n",
      " [0.19044553 0.        ]\n",
      " [0.66646465 0.        ]\n",
      " [0.9111224  0.        ]\n",
      " [0.16211743 0.        ]\n",
      " [0.91057832 0.        ]\n",
      " [0.3243579  0.        ]\n",
      " [0.70029842 0.        ]\n",
      " [0.26545034 0.        ]\n",
      " [0.51895865 0.        ]\n",
      " [0.17746798 0.        ]\n",
      " [0.46784597 0.        ]\n",
      " [0.44896452 1.        ]\n",
      " [0.39606462 0.        ]\n",
      " [0.79192543 0.        ]\n",
      " [0.49339824 0.        ]\n",
      " [0.7239479  0.        ]\n",
      " [0.79302303 0.        ]\n",
      " [0.3515241  0.        ]\n",
      " [0.90965352 1.        ]\n",
      " [0.7128875  0.        ]\n",
      " [0.89724763 0.        ]\n",
      " [0.42020833 0.        ]\n",
      " [0.4675161  0.        ]\n",
      " [0.9039272  0.        ]\n",
      " [0.63354861 0.        ]\n",
      " [0.53282946 0.        ]\n",
      " [0.23751077 0.        ]\n",
      " [0.94650392 0.        ]\n",
      " [0.55943012 0.        ]\n",
      " [0.75735153 0.        ]\n",
      " [0.23614921 0.        ]]\n",
      "sum = 1.3586180436732245\n",
      "\n"
     ]
    }
   ],
   "source": [
    "N_SAMPLES = 1000\n",
    "N_TIMESTEPS = 32\n",
    "\n",
    "# X = (b, t, d)\n",
    "X = np.random.rand(N_SAMPLES, N_TIMESTEPS, 1)\n",
    "F = np.zeros_like(X)\n",
    "\n",
    "# X[b, t, 0] = number\n",
    "# X[b, t, 1] = flag\n",
    "X = np.concatenate([X, F], axis=2)\n",
    "\n",
    "Y = list()\n",
    "sampling_range = list(range(N_TIMESTEPS))\n",
    "for sample in range(N_SAMPLES):\n",
    "    a, b = np.random.choice(sampling_range, size=2, replace=False)\n",
    "    \n",
    "    X[sample][a][1] = 1\n",
    "    X[sample][b][1] = 1\n",
    "    \n",
    "    Y.append(X[sample][a][0] + X[sample][b][0])\n",
    "\n",
    "# Y = (b, sum, 1)\n",
    "Y = np.asarray(Y).reshape(-1, 1)\n",
    "\n",
    "for sample in range(3):\n",
    "    print(f'{X[sample]}\\nsum = {Y[sample][0]}\\n')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Training both models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "tape-in (InputLayer)         (None, 32, 2)             0         \n",
      "_________________________________________________________________\n",
      "lstm (LSTM)                  (None, 32, 4)             112       \n",
      "_________________________________________________________________\n",
      "eita_james (GlobalAveragePoo (None, 4)                 0         \n",
      "_________________________________________________________________\n",
      "linear (Dense)               (None, 1)                 5         \n",
      "=================================================================\n",
      "Total params: 117\n",
      "Trainable params: 117\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Epoch 1/20\n",
      "1000/1000 [==============================] - 2s 2ms/step - loss: 0.2385\n",
      "Epoch 2/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.1657\n",
      "Epoch 3/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.1626\n",
      "Epoch 4/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.1491\n",
      "Epoch 5/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.1206\n",
      "Epoch 6/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0980\n",
      "Epoch 7/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0633\n",
      "Epoch 8/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0539\n",
      "Epoch 9/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0484\n",
      "Epoch 10/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0505\n",
      "Epoch 11/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0360\n",
      "Epoch 12/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0300\n",
      "Epoch 13/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0292\n",
      "Epoch 14/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0277\n",
      "Epoch 15/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0216\n",
      "Epoch 16/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0178\n",
      "Epoch 17/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0153\n",
      "Epoch 18/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0143\n",
      "Epoch 19/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0124\n",
      "Epoch 20/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0148\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f67404ceeb8>"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "def build_average_model():\n",
    "    model_in = Input((N_TIMESTEPS, 2), name='tape-in')\n",
    "    vectors  = LSTM(units=4, return_sequences=True, name='lstm')(model_in)\n",
    "    avg_pool = GlobalAveragePooling1D(name='eita_james')(vectors)\n",
    "    output   = Dense(1, activation='linear', name='linear')(avg_pool)\n",
    "\n",
    "    model = models.Model(inputs=[model_in], outputs=[output])\n",
    "    model.summary()\n",
    "\n",
    "    # 0.05, 0.1\n",
    "    model.compile(optimizers.Adam(lr=1e-1), 'mse')\n",
    "    return model\n",
    "\n",
    "avg_model = build_average_model()\n",
    "avg_model.fit(X, Y, batch_size=32, epochs=20)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "tape-in (InputLayer)         (None, 32, 2)             0         \n",
      "_________________________________________________________________\n",
      "lstm (LSTM)                  (None, 32, 4)             112       \n",
      "_________________________________________________________________\n",
      "attention (AttentionLayer)   (None, 4)                 24        \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 1)                 5         \n",
      "=================================================================\n",
      "Total params: 141\n",
      "Trainable params: 141\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Epoch 1/20\n",
      "1000/1000 [==============================] - 2s 2ms/step - loss: 0.1960\n",
      "Epoch 2/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.1651\n",
      "Epoch 3/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.1626\n",
      "Epoch 4/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.1393\n",
      "Epoch 5/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0427\n",
      "Epoch 6/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0222\n",
      "Epoch 7/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0124\n",
      "Epoch 8/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0091\n",
      "Epoch 9/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0084\n",
      "Epoch 10/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0070\n",
      "Epoch 11/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0066\n",
      "Epoch 12/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0051\n",
      "Epoch 13/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0047\n",
      "Epoch 14/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0047\n",
      "Epoch 15/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0044\n",
      "Epoch 16/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0040\n",
      "Epoch 17/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0033\n",
      "Epoch 18/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0028\n",
      "Epoch 19/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0029\n",
      "Epoch 20/20\n",
      "1000/1000 [==============================] - 1s 1ms/step - loss: 0.0035\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f673d7d75f8>"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "def build_attention_model():\n",
    "    model_in = Input((N_TIMESTEPS, 2), name='tape-in')\n",
    "    vectors  = LSTM(units=4, return_sequences=True, name='lstm')(model_in)\n",
    "    attended = AttentionLayer(name='attention')(vectors)\n",
    "    output   = Dense(1, activation='linear')(attended)\n",
    "\n",
    "    model = models.Model(inputs=[model_in], outputs=[output])\n",
    "    model.summary()\n",
    "\n",
    "    model.compile(optimizers.Adam(lr=5e-2), 'mse')\n",
    "    return model\n",
    "\n",
    "att_model = build_attention_model()\n",
    "att_model.fit(X, Y, batch_size=32, epochs=20)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Comparing the losses\n",
    "\n",
    "Interestingly on this particular case, the average model is almost as competitive as the attention model. Still, it's worhtwhile mentioning that the goal of these toy tasks are:\n",
    "\n",
    "1. To ensure that the attention layer works properly\n",
    "2. Visualize how it affects the model and how its weights are distributed across the timesteps\n",
    "3. The average pooling can be seen as a special case of the attention model is $\\alpha_i = \\alpha_j \\;\\; \\forall i, j$"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(avg_model.history.history['loss'], label='Avg model')\n",
    "plt.plot(att_model.history.history['loss'], label='Att model')\n",
    "plt.legend()\n",
    "plt.grid(alpha=0.3)\n",
    "plt.title('Loss vs epochs')\n",
    "plt.xlabel('Epochs')\n",
    "plt.ylabel('Loss');"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "tape-in (InputLayer)         (None, 32, 2)             0         \n",
      "_________________________________________________________________\n",
      "lstm (LSTM)                  (None, 32, 4)             112       \n",
      "_________________________________________________________________\n",
      "attention (AttentionLayer)   (None, 32)                24        \n",
      "=================================================================\n",
      "Total params: 136\n",
      "Trainable params: 136\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "def build_visualization_model(model):\n",
    "    model_in = Input((N_TIMESTEPS, 2), name='tape-in')\n",
    "    vectors  = LSTM(units=4, return_sequences=True, name='lstm',\n",
    "                    weights=model.layers[1].get_weights())(model_in)\n",
    "\n",
    "    weights = AttentionLayer(name='attention', return_attention=True,\n",
    "                             weights=model.layers[2].get_weights())(vectors)\n",
    "\n",
    "    model = models.Model(inputs=[model_in], outputs=[weights])\n",
    "    model.summary()\n",
    "\n",
    "    return model\n",
    "\n",
    "viz_model = build_visualization_model(att_model)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Visualizing the attention weights"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x216 with 3 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "def plot_attention(viz_model, sample):\n",
    "    # finding timesteps to be summed\n",
    "    sum_coordinates = np.where(sample[:, 1] > 0.90)[0]\n",
    "\n",
    "    # computing the attention coefficients and their maxima\n",
    "    alphas = viz_model.predict(sample[None, :])[0]\n",
    "    max_alpha = np.max(alphas)\n",
    "\n",
    "    # utilities to plot the correct timesteps in the plot\n",
    "    ys = np.linspace(0, max_alpha)\n",
    "    attention1 = np.ones_like(ys) * sum_coordinates[0]\n",
    "    attention2 = np.ones_like(ys) * sum_coordinates[1]\n",
    "\n",
    "    plt.plot(attention1, ys, 'r', alpha=0.5, label='Correct timestep')\n",
    "    plt.plot(attention2, ys, 'r', alpha=0.5)\n",
    "    plt.plot(alphas, label='Attention')\n",
    "\n",
    "    # making it pretty\n",
    "    plt.xlabel('Timesteps')\n",
    "    plt.ylabel('Attention coefficient')\n",
    "    plt.grid(alpha=0.4)\n",
    "    plt.legend()\n",
    "    plt.title(f'Correct timesteps @ {sum_coordinates[0]} and {sum_coordinates[1]}')\n",
    "\n",
    "\n",
    "picks = np.random.randint(0, N_SAMPLES, 3)\n",
    "    \n",
    "plt.figure(figsize=(20, 3))\n",
    "plt.subplot(131)\n",
    "plot_attention(viz_model, X[picks[0]])\n",
    "plt.subplot(132)\n",
    "plot_attention(viz_model, X[picks[1]])\n",
    "plt.subplot(133)\n",
    "plot_attention(viz_model, X[picks[2]])\n",
    "\n",
    "plt.show()"
   ]
  }
 ],
 "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": "3.6.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}