Elman network

Untitled.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Elman network\n",
    "\n",
    "**Elman, J. L. (1990). Finding Structure in Time. Cognitive Science, 14(2), 179–211.**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import tensorflow as tf\n",
    "\n",
    "np.random.seed(0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "m = 2\n",
    "n_0 = 3\n",
    "n_1 = 4\n",
    "n_2 = 5\n",
    "\n",
    "parameters = {\n",
    "    \"w_x\": np.random.randn(n_1, n_0),\n",
    "    \"w_c\": np.random.randn(n_1, n_1),\n",
    "    \"a_1\": np.random.randn(n_1, m),\n",
    "    \"w_1\": np.random.randn(n_1, n_0 + n_1),\n",
    "    \"b_1\": np.random.randn(n_1, 1),\n",
    "    \"w_2\": np.random.randn(n_2, n_1),\n",
    "    \"b_2\": np.random.randn(n_2, 1),\n",
    "}\n",
    "\n",
    "for key in parameters.keys():\n",
    "    parameters[key] = tf.constant(parameters[key], dtype=np.float)\n",
    "\n",
    "x_val = np.random.randn(n_0, m)\n",
    "y_val = np.random.randn(n_2, m)\n",
    "\n",
    "# Input units\n",
    "\n",
    "x = tf.placeholder(tf.float32, shape=(n_0, m), name=\"x\")\n",
    "y = tf.placeholder(tf.float32, shape=(n_2, m), name=\"y\")\n",
    "w_x = tf.Variable(parameters[\"w_x\"], name=\"w_x\")\n",
    "\n",
    "# Context units\n",
    "\n",
    "w_c = tf.Variable(parameters[\"w_c\"], trainable=False, name=\"w_c\")\n",
    "a_1 = tf.Variable(parameters[\"a_1\"], name=\"a_1\")\n",
    "\n",
    "# First layer\n",
    "\n",
    "a_0 = tf.concat([x, a_1], 0)\n",
    "\n",
    "# Hidden units\n",
    "\n",
    "w_1 = tf.concat([w_x, w_c], 1, name=\"w_1\")\n",
    "b_1 = tf.Variable(parameters[\"b_1\"], name=\"b_1\")\n",
    "z_1 = tf.add(tf.matmul(w_1, a_0), b_1, name=\"z_1\")\n",
    "a_1 = tf.assign(a_1, tf.sigmoid(z_1))\n",
    "\n",
    "# Output units\n",
    "\n",
    "w_2 = tf.Variable(parameters[\"w_2\"], name=\"w_2\")\n",
    "b_2 = tf.Variable(parameters[\"b_2\"], name=\"b_2\")\n",
    "z_2 = tf.add(tf.matmul(w_2, a_1), b_2, name=\"z_2\")\n",
    "a_2 = tf.sigmoid(z_2, name=\"a_2\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "WARNING: Logging before flag parsing goes to stderr.\n",
      "W0707 08:26:40.687984 140685639608128 deprecation.py:323] From /usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/losses/losses_impl.py:121: add_dispatch_support.<locals>.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "Use tf.where in 2.0, which has the same broadcast rule as np.where\n"
     ]
    }
   ],
   "source": [
    "learning_rate = 1e-3\n",
    "\n",
    "# Loss & Training\n",
    "\n",
    "prediction = tf.cast(tf.math.round(a_2), tf.float32)\n",
    "loss = tf.losses.mean_squared_error(y, a_2)\n",
    "train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "with tf.Session() as sess:\n",
    "    sess.run(tf.global_variables_initializer())\n",
    "    \n",
    "    writer = tf.summary.FileWriter(\"graph\", sess.graph)\n",
    "    sess.run(train_step, {x: x_val, y: y_val})\n",
    "    writer.close()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "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": "3.6.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}

After changes suggested by jdehesa:

# Input units

x = tf.placeholder(tf.float32, shape=(n_0, m), name="x")
y = tf.placeholder(tf.float32, shape=(n_2, m), name="y")
w_x = tf.Variable(parameters["w_x"], name="w_x")

# Context units

w_c = tf.Variable(parameters["w_c"], name="w_c")
a_c = tf.Variable(parameters["a_c"], trainable=False, name="a_c")

# First layer

a_0 = tf.concat([x, a_c], 0, name="a_0")

# Hidden units

w_1 = tf.concat([w_x, w_c], 1, name="w_1")
b_1 = tf.Variable(parameters["b_1"], name="b_1")
z_1 = tf.add(tf.matmul(w_1, a_0), b_1, name="z_1")
a_1 = tf.sigmoid(z_1)
with tf.control_dependencies([tf.assign(a_c, a_1)]):
    a_1 = tf.identity(a_1)

# Output units

w_2 = tf.Variable(parameters["w_2"], name="w_2")
b_2 = tf.Variable(parameters["b_2"], name="b_2")
z_2 = tf.add(tf.matmul(w_2, a_1), b_2, name="z_2")
a_2 = tf.sigmoid(z_2, name="a_2")