Serialisation of tensorflow models without using collections

tensorflow_serialisation.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Definition and serialisation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Define coefficients\n",
    "p = 5\n",
    "coefficient_values = np.random.normal(0, 1, p)\n",
    "\n",
    "# Create a graph\n",
    "with tf.Graph().as_default() as original_graph:\n",
    "    placeholder = tf.placeholder(tf.float32, [None, p], name='placeholder')\n",
    "    coefficients = tf.Variable(coefficient_values, name='coefficients', dtype=tf.float32)\n",
    "    predictor = tf.reduce_sum(coefficients * placeholder, 1, name='predictor')\n",
    "    init_op = tf.initialize_all_variables()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Create a session and initialise\n",
    "session = tf.Session(graph=original_graph)\n",
    "session.run(init_op)\n",
    "\n",
    "# Make sure the graph performs as expected\n",
    "X = np.random.normal(0, 1, (100, p))\n",
    "actual = session.run(predictor, {placeholder: X})\n",
    "desired = np.dot(X, coefficient_values)\n",
    "np.testing.assert_allclose(actual, desired, 1e-3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Serialise the graph\n",
    "with original_graph.as_default():\n",
    "    saver = tf.train.Saver()\n",
    "    saver.save(session, 'linear_model.tf')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Deserialisation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Deserialise the graph\n",
    "with tf.Graph().as_default() as restored_graph:\n",
    "    restored_session = tf.Session(graph=restored_graph)\n",
    "    saver = tf.train.import_meta_graph('linear_model.tf.meta')\n",
    "    saver.restore(session, 'linear_model.tf')\n",
    "    restored_session.run(tf.initialize_all_variables())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Option 1 gets the tensors (see https://github.com/tensorflow/tensorflow/issues/3378 and \n",
    "# http://stackoverflow.com/a/37870634/419116 for information on the :0 at the end)\n",
    "restored_placeholder = restored_graph.get_tensor_by_name('placeholder:0')\n",
    "restored_predictor = restored_graph.get_tensor_by_name('predictor:0')\n",
    "actual = restored_session.run(restored_predictor, {restored_placeholder: X})\n",
    "np.testing.assert_allclose(actual, desired, 1e-3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# Option 2 simply uses their names\n",
    "actual = restored_session.run('predictor:0', {'placeholder:0': X})\n",
    "np.testing.assert_allclose(actual, desired, 1e-3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<tf.Tensor 'placeholder:0' shape=(?, 5) dtype=float32>"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# The shape is persisted\n",
    "restored_graph.get_tensor_by_name('placeholder:0')"
   ]
  }
 ],
 "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": 0
}