RBM.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Initialization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Extracting MNIST_data/train-images-idx3-ubyte.gz\n",
      "Extracting MNIST_data/train-labels-idx1-ubyte.gz\n",
      "Extracting MNIST_data/t10k-images-idx3-ubyte.gz\n",
      "Extracting MNIST_data/t10k-labels-idx1-ubyte.gz\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fb21c361bd0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "%matplotlib inline\n",
    "import matplotlib.pyplot as plt\n",
    "import matplotlib.image as mpimg\n",
    "import numpy as np\n",
    "import matplotlib.cm as cm\n",
    "plt.gray();\n",
    "\n",
    "import tensorflow as tf\n",
    "tf.set_random_seed(0)\n",
    "import time\n",
    "\n",
    "#load up MNIST data\n",
    "import input_data\n",
    "mnist_data = input_data.read_data_sets('MNIST_data', one_hot=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "IMAGE_HEIGHT = 28\n",
    "IMAGE_WIDTH = 28\n",
    "IMAGE_CHANNELS = 1\n",
    "IMAGE_SIZE = IMAGE_HEIGHT*IMAGE_WIDTH\n",
    "NUM_V = IMAGE_SIZE\n",
    "NUM_H = NUM_V\n",
    "LOGDIR = 'RBM'\n",
    "BATCH_SIZE = 20 #10\n",
    "LEARN_RATE = 0.1 #0.01 #0.25 #0.1 #10.0\n",
    "W_STDDEV = 0.01 #2.0\n",
    "WEIGHT_COST = 0.01"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/coolguy/anaconda/lib/python2.7/site-packages/ipykernel/__main__.py:4: RuntimeWarning: divide by zero encountered in log\n"
     ]
    }
   ],
   "source": [
    "## use full dataset to help with initalizing biases\n",
    "init_batch = mnist_data.train.next_batch(mnist_data.train.num_examples)[0][:,0:IMAGE_SIZE]\n",
    "prob_data = np.mean(init_batch, axis=0)\n",
    "vis_init = np.nan_to_num(np.log(prob_data/(1-prob_data)))\n",
    "vis_init[np.abs(vis_init)>1.0e38]=-1000 #rescale gigantic biases"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "#define placeholders, functions, and variables for visible/hidden units\n",
    "\n",
    "#joint global variables\n",
    "with tf.name_scope('joint') as scope:\n",
    "    joint_scope = scope\n",
    "    #weights between hidden and visible layers\n",
    "    W = tf.Variable(tf.truncated_normal(shape=[NUM_V,NUM_H], stddev=W_STDDEV), name='weights') \n",
    "\n",
    "#placeholders\n",
    "with tf.name_scope('visible') as scope:\n",
    "    vis_scope = scope\n",
    "    v_probabilities = tf.placeholder(\"float\", shape=[None, NUM_V], name='vis_probs')\n",
    "    v_state = tf.placeholder(\"float\", shape=[None, NUM_V], name='vis_state')\n",
    "    \n",
    "with tf.name_scope('hidden') as scope:\n",
    "    hid_scope = scope    \n",
    "    h_probabilities = tf.placeholder(\"float\", shape=[None, NUM_H], name='hid_probs')\n",
    "    h_state = tf.placeholder(\"float\", shape=[None, NUM_H], name='hid_state')\n",
    "\n",
    "with tf.name_scope(vis_scope) as scope:\n",
    "    #bias variable\n",
    "    a = tf.Variable(vis_init, name='bias')\n",
    "    #activation function\n",
    "    v_prob = tf.nn.sigmoid(tf.matmul(h_state, W, transpose_b=True)+a, name=\"visible_probs\")\n",
    "    #conversion function for probabilities to binary values\n",
    "    v_prob_to_state = tf.to_float(tf.less_equal(tf.random_uniform(shape=tf.TensorShape([BATCH_SIZE,NUM_V])), v_probabilities), name='prob_to_state')\n",
    "    #convert to form more conducive to tensor product\n",
    "    v_reshaped = tf.reshape(v_state, [BATCH_SIZE, NUM_V, 1], name='vis_reshaped')\n",
    "    #average over training set\n",
    "    v_mean = tf.reduce_mean(v_state,0, name='vis_batch_mean')\n",
    "    \n",
    "with tf.name_scope(hid_scope) as scope:\n",
    "    #bias variable\n",
    "    b = tf.Variable(tf.zeros(shape=[NUM_H]), name='bias')\n",
    "    #activation function\n",
    "    h_prob = tf.nn.sigmoid(tf.matmul(v_state, W,)+b, name=\"hidden_probs\")\n",
    "    #conversion function for probabilities to binary values\n",
    "    h_prob_to_state = tf.to_float(tf.less_equal(tf.random_uniform(shape=tf.TensorShape([BATCH_SIZE,NUM_H])), h_probabilities), name='prob_to_state')\n",
    "    #convert to form more conducive to tensor product\n",
    "    h_reshaped = tf.reshape(h_state, [BATCH_SIZE, 1, NUM_H], name='hid_reshaped')\n",
    "    h_mean = tf.reduce_mean(h_state,0, name='hid_batch_mean')\n",
    "\n",
    "with tf.name_scope(joint_scope) as scope: #to do: fix this\n",
    "    v_times_h = tf.batch_matmul(v_reshaped,h_reshaped, name='vh_outer') \n",
    "    vh_expval = tf.reduce_mean(v_times_h,0, name='vh_batch_mean')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "#cost function \n",
    "with tf.name_scope('cost') as scope:\n",
    "    #placeholders to hold the expectation values for feeding into cost function\n",
    "    v_expval_data = tf.placeholder(\"float\", shape=[NUM_V,], name='vis_expval_data')\n",
    "    h_expval_data = tf.placeholder(\"float\", shape=[NUM_H,], name='hid_expval_data')\n",
    "    vh_expval_data = tf.placeholder(\"float\", shape=[NUM_V, NUM_H], name='vh_expval_data')\n",
    "    v_expval_model = tf.placeholder(\"float\", shape=[NUM_V,], name='vis_expval_model')\n",
    "    h_expval_model = tf.placeholder(\"float\", shape=[NUM_H,], name='hid_expval_model')\n",
    "    vh_expval_model = tf.placeholder(\"float\", shape=[NUM_V, NUM_H], name='vh_expval_model')\n",
    "\n",
    "    a_term = tf.reduce_sum(tf.mul(a, v_expval_data - v_expval_model), name = 'a_term')\n",
    "    b_term = tf.reduce_sum(tf.mul(b, h_expval_data - h_expval_model), name = 'b_term')\n",
    "    w_term = tf.reduce_sum(tf.reshape(tf.mul(W, vh_expval_data - vh_expval_model), [-1]), name = 'w_term')\n",
    "    E = a_term + b_term + w_term\n",
    "    log_prob = -E #tf.log(-E)\n",
    "    \n",
    "    weight_cost = WEIGHT_COST*0.5*(tf.reduce_sum(tf.square(W)) )# + \\\n",
    "#                                   tf.reduce_sum(tf.square(a)) + \\\n",
    "#                                   tf.reduce_sum(tf.square(b)) + \\\n",
    "#                                   ) \n",
    "    \n",
    "    cost = log_prob #+ weight_cost"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "#training ops\n",
    "\n",
    "#summarizers\n",
    "tf.scalar_summary('cost', cost)\n",
    "\n",
    "#generated_images = tf.placeholder(\"float\", shape=[None, IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNELS], name='gen_image')\n",
    "#tf.image_summary('sample_images', generated_images)\n",
    "\n",
    "#instantiate gradient optimizer\n",
    "optimizer = tf.train.GradientDescentOptimizer(LEARN_RATE)\n",
    "\n",
    "#global training step counter\n",
    "global_step = tf.Variable(0, name='global_step', trainable=False)\n",
    "\n",
    "#train_op updates the trainable weights and the global step\n",
    "train_op = optimizer.minimize(cost, global_step = global_step)\n",
    "\n",
    "# collect all summaries together into one op\n",
    "summary_op = tf.merge_all_summaries()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Run things in session"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "#start a tensorflow interactive session\n",
    "sess = tf.InteractiveSession()\n",
    "\n",
    "#initialize variables in session\n",
    "init = tf.initialize_all_variables()\n",
    "sess.run(init)\n",
    "\n",
    "#instantiate a SummaryWriter to save the graph and summaries for inspection\n",
    "summary_writer = tf.train.SummaryWriter(LOGDIR,sess.graph_def)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def update_vhv(start_state, num_iterations):\n",
    "    \n",
    "    #update hidden layer using start_state clamped to visible layer\n",
    "    hidden_probs = sess.run(h_prob, feed_dict={v_state: start_state})\n",
    "    hidden_state = sess.run(h_prob_to_state, feed_dict={h_probabilities: hidden_probs})\n",
    "    \n",
    "    #update visible layer\n",
    "    visible_probs = sess.run(v_prob, feed_dict={h_state: hidden_state})\n",
    "    visible_state = sess.run(v_prob_to_state, feed_dict={v_probabilities: visible_probs})\n",
    "    \n",
    "    #repeat until desired number of iterations\n",
    "    if num_iterations > 1:\n",
    "        return update_vhv(visible_state, num_iterations-1)\n",
    "    else:\n",
    "        #use probabilities in final output\n",
    "        return visible_probs\n",
    "    \n",
    "def update_hvh(start_state, num_iterations):\n",
    "\n",
    "    #update visible layer\n",
    "    visible_probs = sess.run(v_prob, feed_dict={h_state: start_state})\n",
    "    visible_state = sess.run(v_prob_to_state, feed_dict={v_probabilities: visible_probs})\n",
    "    \n",
    "    #update hidden layer using start_state clamped to visible layer\n",
    "    hidden_probs = sess.run(h_prob, feed_dict={v_state: visible_state})\n",
    "    hidden_state = sess.run(h_prob_to_state, feed_dict={h_probabilities: hidden_probs})\n",
    "        \n",
    "    #repeat until desired number of iterations\n",
    "    if num_iterations > 1:\n",
    "        return update_hvh(hidden_state, num_iterations-1)\n",
    "    else:\n",
    "        #use probabilities in final output\n",
    "        return hidden_probs\n",
    "    \n",
    "def build_feed_dict(batch, v_expval_data_pl, h_expval_data_pl, vh_expval_data_pl, v_expval_model_pl, h_expval_model_pl, vh_expval_model_pl):\n",
    "    \n",
    "    ##calculate <v/h/v*h>_data\n",
    "    v_values_data = batch\n",
    "    h_probs_data = sess.run(h_prob, {v_state: batch}) #do we need to randomly select subsets of training images?\n",
    "    h_values_data = sess.run(h_prob_to_state,{h_probabilities: h_probs_data})\n",
    "    vh_values_data = sess.run(v_times_h, {v_state: v_values_data, h_state: h_values_data})\n",
    "    \n",
    "    v_expval_data_curr = sess.run(v_mean, {v_state: v_values_data}) #simply average over the training data here\n",
    "    h_expval_data_curr = sess.run(h_mean, {h_state: h_values_data})\n",
    "    vh_expval_data_curr = sess.run(vh_expval, {v_state: v_values_data, h_state: h_values_data})\n",
    "    \n",
    "    ##calculate <v/h/v*h>_model\n",
    "    visible_probs_cd = update_vhv(batch, cd_steps) #start with training vector (randomly selected?)\n",
    "    visible_values_cd = sess.run(v_prob_to_state, {v_probabilities: visible_probs_cd})\n",
    "\n",
    "    hidden_start_probs = sess.run(h_prob, {v_state: batch}) #do we need to randomly select a training image?\n",
    "    hidden_start_state = sess.run(h_prob_to_state,{h_probabilities: hidden_start_probs})\n",
    "    hidden_probs_cd = update_hvh(hidden_start_state, cd_steps)\n",
    "    hidden_values_cd = sess.run(h_prob_to_state, {h_probabilities: hidden_probs_cd})\n",
    "    \n",
    "    #should we use visible probs for these instead?\n",
    "    vh_values_model = sess.run(v_times_h, {v_state: visible_values_cd, h_state: hidden_values_cd})\n",
    "  \n",
    "    v_expval_model_curr = sess.run(v_mean, feed_dict={v_state: visible_values_cd}) #or use probs?\n",
    "    h_expval_model_curr = sess.run(h_mean, {h_state: hidden_values_cd})\n",
    "    vh_expval_model_curr = sess.run(vh_expval, {v_state: visible_values_cd, h_state: hidden_values_cd})     \n",
    "    \n",
    "    feed_dict = {\n",
    "            v_expval_data_pl: v_expval_data_curr,\n",
    "            h_expval_data_pl:  h_expval_data_curr,\n",
    "            vh_expval_data_pl: vh_expval_data_curr,\n",
    "            v_expval_model_pl: v_expval_model_curr,\n",
    "            h_expval_model_pl: h_expval_model_curr,\n",
    "            vh_expval_model_pl: vh_expval_model_curr\n",
    "    }\n",
    "    \n",
    "    return feed_dict"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 120,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Step 41300 (epoch 16): cost = -3.58881187439 (17.5565178394 sec)\n",
      "Step 41400 (epoch 16): cost = -6.87475013733 (36.0822207928 sec)\n",
      "Step 41500 (epoch 16): cost = -21.8811149597 (48.4088590145 sec)\n",
      "Step 41600 (epoch 16): cost = -36.1162719727 (57.1798160076 sec)\n",
      "Step 41700 (epoch 16): cost = -7.48731994629 (56.7106370926 sec)\n",
      "Step 41800 (epoch 16): cost = -14.8740091324 (56.9105870724 sec)\n",
      "Step 41900 (epoch 16): cost = -4.21817588806 (58.2222390175 sec)\n",
      "Step 42000 (epoch 16): cost = -17.8524017334 (57.5009598732 sec)\n",
      "Step 42100 (epoch 16): cost = -13.4559440613 (57.7034208775 sec)\n",
      "Step 42200 (epoch 16): cost = -12.6276226044 (57.980645895 sec)\n",
      "Step 42300 (epoch 16): cost = -8.82434463501 (59.9017159939 sec)\n",
      "Step 42400 (epoch 16): cost = -15.7976360321 (57.7374539375 sec)\n",
      "Step 42500 (epoch 16): cost = -14.7830543518 (59.6955561638 sec)\n",
      "Step 42600 (epoch 16): cost = -5.53691005707 (63.8802371025 sec)\n",
      "Step 42700 (epoch 16): cost = -14.7918539047 (68.3269188404 sec)\n",
      "Step 42800 (epoch 16): cost = -24.3357086182 (67.0303530693 sec)\n",
      "Step 42900 (epoch 16): cost = -16.2083530426 (61.3408010006 sec)\n",
      "Step 43000 (epoch 16): cost = -74.742149353 (62.6882050037 sec)\n",
      "Step 43100 (epoch 16): cost = -37.8027954102 (64.4831390381 sec)\n",
      "Step 43200 (epoch 16): cost = -9.89011192322 (77.6715099812 sec)\n",
      "Step 43300 (epoch 16): cost = -28.2562637329 (72.8230280876 sec)\n",
      "Step 43400 (epoch 16): cost = 2.43044376373 (53.5844311714 sec)\n",
      "Step 43500 (epoch 16): cost = -3.07593536377 (52.4429500103 sec)\n",
      "Step 43600 (epoch 16): cost = -31.3041610718 (53.593572855 sec)\n",
      "Step 43700 (epoch 16): cost = -3.42064666748 (53.8062028885 sec)\n",
      "Step 43800 (epoch 16): cost = -42.1017112732 (51.2785210609 sec)\n",
      "Step 43900 (epoch 16): cost = -22.0262145996 (56.8974950314 sec)\n",
      "Step 44000 (epoch 17): cost = -5.04652404785 (53.6566848755 sec)\n",
      "Step 44100 (epoch 17): cost = -23.4507026672 (61.4576430321 sec)\n",
      "Step 44200 (epoch 17): cost = -31.8945159912 (67.6274988651 sec)\n",
      "Step 44300 (epoch 17): cost = 11.8982982635 (83.9159669876 sec)\n",
      "Step 44400 (epoch 17): cost = -20.1353054047 (76.0482490063 sec)\n",
      "Step 44500 (epoch 17): cost = -11.7569198608 (75.710463047 sec)\n",
      "Step 44600 (epoch 17): cost = -16.8159980774 (77.3503940105 sec)\n",
      "Step 44700 (epoch 17): cost = 5.91460561752 (103.400488138 sec)\n",
      "Step 44800 (epoch 17): cost = 0.00745677947998 (76.4677100182 sec)\n",
      "Step 44900 (epoch 17): cost = -17.8680725098 (106.251493216 sec)\n",
      "Step 45000 (epoch 17): cost = -18.8347930908 (87.7529320717 sec)\n",
      "Step 45100 (epoch 17): cost = -5.27837562561 (120.272450924 sec)\n",
      "Step 45200 (epoch 17): cost = -10.396276474 (71.9492180347 sec)\n",
      "Step 45300 (epoch 17): cost = -12.5195789337 (73.0653760433 sec)\n",
      "Step 45400 (epoch 17): cost = -4.89378070831 (78.6125969887 sec)\n",
      "Step 45500 (epoch 17): cost = -18.8473434448 (79.145195961 sec)\n",
      "Step 45600 (epoch 17): cost = -34.9617500305 (92.5878901482 sec)\n",
      "Step 45700 (epoch 17): cost = -61.4136428833 (85.5298321247 sec)\n",
      "Step 45800 (epoch 17): cost = -16.7882785797 (91.1526529789 sec)\n",
      "Step 45900 (epoch 17): cost = -23.1754283905 (83.0632989407 sec)\n",
      "Step 46000 (epoch 17): cost = -24.1523361206 (79.8390440941 sec)\n",
      "Step 46100 (epoch 17): cost = -1.25892925262 (68.724378109 sec)\n",
      "Step 46200 (epoch 17): cost = -10.4862785339 (69.2375750542 sec)\n",
      "Step 46300 (epoch 17): cost = -19.0916023254 (68.2954189777 sec)\n",
      "Step 46400 (epoch 17): cost = -20.82213974 (65.6271028519 sec)\n",
      "Step 46500 (epoch 17): cost = -16.4332618713 (79.0264649391 sec)\n",
      "Step 46600 (epoch 17): cost = 0.606267929077 (68.8205981255 sec)\n",
      "Step 46700 (epoch 17): cost = -15.8211402893 (68.7681407928 sec)\n",
      "Step 46800 (epoch 18): cost = -10.4616451263 (88.0419290066 sec)\n",
      "Step 46900 (epoch 18): cost = 5.69402122498 (114.219311953 sec)\n",
      "Step 47000 (epoch 18): cost = -19.0538406372 (64.9753420353 sec)\n",
      "Step 47100 (epoch 18): cost = -19.1687469482 (73.8139169216 sec)\n",
      "Step 47200 (epoch 18): cost = -7.68465995789 (70.6223618984 sec)\n",
      "Step 47300 (epoch 18): cost = -47.328453064 (75.0093312263 sec)\n",
      "Step 47400 (epoch 18): cost = -8.02725601196 (72.3034229279 sec)\n",
      "Step 47500 (epoch 18): cost = -1.44691848755 (70.4880120754 sec)\n",
      "Step 47600 (epoch 18): cost = -14.2449808121 (104.452627897 sec)\n",
      "Step 47700 (epoch 18): cost = -28.2056999207 (76.986716032 sec)\n",
      "Step 47800 (epoch 18): cost = -7.49584627151 (67.8092541695 sec)\n",
      "Step 47900 (epoch 18): cost = -12.3945579529 (74.4944450855 sec)\n",
      "Step 48000 (epoch 18): cost = 2.81218719482 (72.574406147 sec)\n",
      "Step 48100 (epoch 18): cost = -19.8547229767 (72.8855509758 sec)\n",
      "Step 48200 (epoch 18): cost = -7.33100128174 (72.0085430145 sec)\n",
      "Step 48300 (epoch 18): cost = -6.18923377991 (88.4286179543 sec)\n",
      "Step 48400 (epoch 18): cost = -10.6890268326 (75.1397600174 sec)\n",
      "Step 48500 (epoch 18): cost = -26.9362716675 (78.6081438065 sec)\n",
      "Step 48600 (epoch 18): cost = -14.8190994263 (67.0898149014 sec)\n",
      "Step 48700 (epoch 18): cost = -15.2442960739 (72.966301918 sec)\n",
      "Step 48800 (epoch 18): cost = -20.8708152771 (81.6115019321 sec)\n",
      "Step 48900 (epoch 18): cost = -8.34210205078 (68.2312488556 sec)\n",
      "Step 49000 (epoch 18): cost = -22.0488414764 (75.9776310921 sec)\n",
      "Step 49100 (epoch 18): cost = -15.197013855 (66.5725839138 sec)\n",
      "Step 49200 (epoch 18): cost = -30.6386241913 (67.9348509312 sec)\n",
      "Step 49300 (epoch 18): cost = -40.0578155518 (67.4293489456 sec)\n",
      "Step 49400 (epoch 18): cost = 0.499853134155 (65.002051115 sec)\n",
      "Step 49500 (epoch 19): cost = -14.1930885315 (65.3365569115 sec)\n",
      "Step 49600 (epoch 19): cost = -17.5314292908 (77.9629089832 sec)\n",
      "Step 49700 (epoch 19): cost = -10.9905071259 (70.4654459953 sec)\n",
      "Step 49800 (epoch 19): cost = 2.7689666748 (74.391493082 sec)\n",
      "Step 49900 (epoch 19): cost = -16.4721183777 (87.1134989262 sec)\n",
      "Step 50000 (epoch 19): cost = -26.0241260529 (66.7045378685 sec)\n",
      "Step 50100 (epoch 19): cost = -0.811944961548 (71.243997097 sec)\n",
      "Step 50200 (epoch 19): cost = -10.1623306274 (85.8297710419 sec)\n",
      "Step 50300 (epoch 19): cost = -37.5479850769 (88.1633028984 sec)\n",
      "Step 50400 (epoch 19): cost = -13.4895763397 (86.4009890556 sec)\n",
      "Step 50500 (epoch 19): cost = -1.72083282471 (78.3383228779 sec)\n",
      "Step 50600 (epoch 19): cost = 5.89131498337 (77.1449358463 sec)\n",
      "Step 50700 (epoch 19): cost = -15.4944190979 (70.9178040028 sec)\n",
      "Step 50800 (epoch 19): cost = 7.8623342514 (66.2590949535 sec)\n",
      "Step 50900 (epoch 19): cost = -10.5547866821 (67.1036250591 sec)\n",
      "Step 51000 (epoch 19): cost = -5.71610355377 (68.4853820801 sec)\n",
      "Step 51100 (epoch 19): cost = -19.6188259125 (74.288823843 sec)\n",
      "Step 51200 (epoch 19): cost = -4.12885665894 (70.0933029652 sec)\n",
      "Step 51300 (epoch 19): cost = -10.3088779449 (3148.12154484 sec)\n",
      "Step 51400 (epoch 19): cost = -9.21058082581 (32.4235990047 sec)\n",
      "Step 51500 (epoch 19): cost = -17.7934875488 (31.8011069298 sec)\n",
      "Step 51600 (epoch 19): cost = -6.23769187927 (31.3497650623 sec)\n",
      "Step 51700 (epoch 19): cost = -62.2499160767 (34.6676502228 sec)\n",
      "Step 51800 (epoch 19): cost = -14.3364610672 (39.0088069439 sec)\n",
      "Step 51900 (epoch 19): cost = -12.3203678131 (38.7767879963 sec)\n",
      "Step 52000 (epoch 19): cost = -3.43066620827 (40.0338799953 sec)\n",
      "Step 52100 (epoch 19): cost = -29.3558082581 (44.3625361919 sec)\n",
      "Step 52200 (epoch 19): cost = -8.76265525818 (46.5113320351 sec)\n",
      "Step 52300 (epoch 20): cost = -4.07636642456 (39.5479240417 sec)\n",
      "Step 52400 (epoch 20): cost = -35.2089767456 (47.4816529751 sec)\n",
      "Step 52500 (epoch 20): cost = -8.97724437714 (47.536921978 sec)\n",
      "Step 52600 (epoch 20): cost = -3.65865898132 (47.4938678741 sec)\n",
      "Step 52700 (epoch 20): cost = -20.1201591492 (47.652061224 sec)\n",
      "Step 52800 (epoch 20): cost = -12.5881347656 (47.4887208939 sec)\n",
      "Step 52900 (epoch 20): cost = -6.48471927643 (47.6449670792 sec)\n",
      "Step 53000 (epoch 20): cost = 12.9016113281 (47.5152771473 sec)\n",
      "Step 53100 (epoch 20): cost = -36.3945808411 (47.5860290527 sec)\n",
      "Step 53200 (epoch 20): cost = -21.8353977203 (47.4144568443 sec)\n",
      "Step 53300 (epoch 20): cost = -29.4082069397 (47.6048769951 sec)\n",
      "Step 53400 (epoch 20): cost = -26.3429393768 (47.5246250629 sec)\n",
      "Step 53500 (epoch 20): cost = -26.1090316772 (47.8892400265 sec)\n",
      "Step 53600 (epoch 20): cost = -23.7721633911 (47.4226779938 sec)\n",
      "Step 53700 (epoch 20): cost = -2.94399452209 (48.0742909908 sec)\n",
      "Step 53800 (epoch 20): cost = -11.781879425 (64.0440218449 sec)\n",
      "Step 53900 (epoch 20): cost = -9.24857711792 (45.7090411186 sec)\n",
      "Step 54000 (epoch 20): cost = -59.634727478 (49.1796748638 sec)\n",
      "Step 54100 (epoch 20): cost = -7.57622337341 (47.3764140606 sec)\n",
      "Step 54200 (epoch 20): cost = -11.0804719925 (46.3801109791 sec)\n",
      "Step 54300 (epoch 20): cost = -14.3688316345 (53.1999590397 sec)\n",
      "Step 54400 (epoch 20): cost = -14.2554473877 (43.5611407757 sec)\n",
      "Step 54500 (epoch 20): cost = -17.6048622131 (46.9061830044 sec)\n",
      "Step 54600 (epoch 20): cost = -26.8916969299 (48.5768640041 sec)\n",
      "Step 54700 (epoch 20): cost = -17.4938545227 (53.20322299 sec)\n",
      "Step 54800 (epoch 20): cost = -15.4226531982 (46.7162330151 sec)\n",
      "Step 54900 (epoch 20): cost = -13.8368396759 (44.7665281296 sec)\n"
     ]
    }
   ],
   "source": [
    "# train model\n",
    "num_epochs = 20\n",
    "cd_steps = 25 #15\n",
    "\n",
    "#LEARN_RATE = 0.3  #1.0 #good starting rate\n",
    "tf.set_random_seed(0)\n",
    "\n",
    "start_time = time.time()\n",
    "while mnist_data.train.epochs_completed <= num_epochs:\n",
    "\n",
    "    #Note: the next_batch method gives both the image and its label.\n",
    "    #We only take the image (slot '0')\n",
    "    #Also allow image to be cropped to IMAGE_SIZE pixels\n",
    "    batch = mnist_data.train.next_batch(BATCH_SIZE)[0][:,0:IMAGE_SIZE]\n",
    "    \n",
    "    feed_dict = build_feed_dict(batch, v_expval_data, h_expval_data, vh_expval_data, v_expval_model, h_expval_model, vh_expval_model)\n",
    "\n",
    "    _, cost_value = sess.run([train_op, cost],feed_dict)\n",
    "    \n",
    "    if global_step.eval() % 100 == 0:\n",
    "        duration = time.time() - start_time\n",
    "        start_time = time.time()\n",
    "        \n",
    "        print 'Step {0} (epoch {1}): cost = {2} ({3} sec)'.format(global_step.eval(), mnist_data.train.epochs_completed, cost_value, duration)\n",
    "        \n",
    "        #saver.save(sess, LOGDIR, global_step)\n",
    "        \n",
    "        #update event log\n",
    "        \n",
    "        #sample_images = generate_image(batch, 15)\n",
    "        #sess.run(gen_images, {gen_images: sample_images})\n",
    "        #tf.image_summary('sample_images', sample_images)\n",
    "        \n",
    "        summary_str = sess.run(summary_op, feed_dict=feed_dict)\n",
    "        summary_writer.add_summary(summary_str, step)\n",
    "        \n",
    "        "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 121,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'20epoch-RBM-54997'"
      ]
     },
     "execution_count": 121,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "#save model\n",
    "saver = tf.train.Saver()\n",
    "saver.save(sess, '20epoch-RBM', global_step = global_step)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 85,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "#reload model\n",
    "chkpt = tf.train.get_checkpoint_state('/home/coolguy/code/tensorflow/tensorflow-training')\n",
    "saver.restore(sess, chkpt.model_checkpoint_path)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Workspace"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 122,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def generate_state(start_state, num_iterations):\n",
    "    #starting with any visible state, perform gibbs sampling to generate new visible state\n",
    "    vis_probs = update_vhv(start_state, num_iterations)\n",
    "    #gen_state = sess.run(v_prob_to_state, {v_probabilities: vis_probs})\n",
    "    gen_state = vis_probs #gives greyscale image based on probs\n",
    "    \n",
    "    #convert this state to an image for display\n",
    "    #gen_image = tf.reshape(gen_state, [BATCH_SIZE, IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNELS])\n",
    "    return gen_state\n",
    "\n",
    "def to_image(flat_tensor):\n",
    "    if type(flat_tensor) == tf.python.framework.ops.Tensor:\n",
    "        flat_array = tf.squeeze(flat_tensor).eval()\n",
    "    else:\n",
    "        flat_array = flat_tensor\n",
    "        \n",
    "    if len(flat_array.shape) == 1: #no explicit batches\n",
    "        flat_length = flat_array.shape\n",
    "        square_length = int(np.sqrt(flat_length))\n",
    "        square_array = flat_array.reshape([square_length, square_length])\n",
    "    else:\n",
    "        batch_size, flat_length = flat_array.shape\n",
    "        square_length = int(np.sqrt(flat_length))\n",
    "        square_array = flat_array.reshape([batch_size, square_length, square_length])\n",
    "    return square_array"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 123,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "init_bias = to_image(vis_init)\n",
    "init_bias_remove_giant = init_bias*(np.abs(init_bias)<1.0e38).astype(float) #remove gigantic biases\n",
    "init_bias_scaled = init_bias_remove_giant + np.min(init_bias_remove_giant)*(np.abs(init_bias)>1.0e38).astype(float) #rescale giantic biases and recombine\n",
    "#plt.imshow(init_bias_scaled); #initial biases\n",
    "\n",
    "tf.set_random_seed(0)\n",
    "batch = mnist_data.train.next_batch(BATCH_SIZE)[0][:,0:IMAGE_SIZE]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 124,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD8CAYAAABTq8lnAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nqREUPCE1goInpEZQ8ITUiH8BQ51oOsnSvfMAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fb1f1286cd0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(to_image(batch)[0]);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 125,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD8CAYAAABTq8lnAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fb1f0ea65d0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "sample_images = generate_state(batch, 500) #15\n",
    "plt.imshow(to_image(sample_images)[0]);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 126,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD8CAYAAABTq8lnAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n4B2nRrjgHadGuOAdp0a44B2nRvx/hGBEGoqM4BoAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fb1ee58ef10>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(to_image(batch)[2]);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 127,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD8CAYAAABTq8lnAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAIABJREFUeJztfVvIddtZ3jO+9Z3+f/9/aCV1J4RdtxcKpQgJhVBIS7wI\nEhG03kQCYiixeGFTES9MvDCpXtQKBrEXUmsiiRUPVEzjRa1RKo0XHlJiY2u0CtmQhGTvFLv3f/hO\na61v9OL73vW/61nv+44x5zrOb40HBnPM85hjjGe8h3GYKeeMhoaG/cDBthPQ0NCwOTTCNzTsERrh\nGxr2CI3wDQ17hEb4hoY9QiN8Q8MeoTfhU0rvTCn9ZUrpr1NKP7rKRDU0NKwHqU8/fEppBOCvALwD\nwJcB/CmAd+ecP6+uaR38DQ1bRM458bG+Ev6tAP4m5/xSznkM4NcAfNcyiWtoaFg/+hL+TQC+qPa/\ndHusoaFhh9GX8E1db2gYIPoS/ssAXlD7L+BGyjc0NOww+hL+MwC+KaX0YkrpGMD3APjk6pLV0NCw\nDhz2uSnnPEkp/UsA/xXACMBHtIe+oaFhN9GrW67qwa1brqFhq1hlt1xDQ8MA0Qjf0LBHaIRvaNgj\nNMI3NOwRGuEbGvYIjfANDXuERviGhj1CI3xDwx6hEb6hYY/QCN/QsEdohG9o2CM0wjc07BEa4Rsa\n9giN8A0Ne4RG+IaGPUIjfEPDHqERvqFhj9AI39CwR2iEb2jYIzTCNzTsERrhGxr2CI3wDQ17hEb4\nhoY9QiN8Q8MeoRG+oWGP0Ajf0LBHaIRvaNgj9PqZZENDF6S08Iuzpe/t+8zoX4ql/yyu6z+Mm0Qj\nfMNKwUSMiGmd08eseErJjQPzpGSCyn5p6x3rg11rJPaa8F5l3JVCWkYyeljnt3kEtc57xywSy/bg\n4GB23gqCnHNIaj7vHeO4h+g8f9+269ZeEb6WQNsspHWQ3Hr+Gn8THm457l1rkVkIf3BwMBeXbS2p\n+waGHOtSX9ad/yUsRfiU0ksAHgGYAhjnnN+6ikStGsuSaNlCWjeJ+yCltNJKV6N+W1vrei8I0b0A\nlMl9fX09t43ism8h57ygVVh54eXxqvO/FstK+AzgW3POf7uKxKwDqyRbn0LaRbILVlHpIpubJXSJ\n9JrYXnw0Grlbi9h8TBPeikueyLGDgwOX9AJNfs7PSFhsg/SrUOl3tkZ3VeFrMr9LIa2K7Kt6zqor\nl2eTRzZ2tGUVnbej0WiO4Bwiic3k1mE6nc7KNaU0R3DJMz6m64yuExHxt22/A6uR8L+XUpoC+Pc5\n5/+wgjStBDUk8WzJUsFsqvBWrR1sstKVSG8Rnm103hdiHx4emnEAC5LbIvl0Ol0gu+xznssxqyw0\n2bnuWHXJyv9NNwTLEv5tOeevpJT+HoBPpZT+Muf86VUkbBlEROnaEHS1wbq+27Nlo22Erl1OtY0c\n26wePHKUbHe+37PrrUZBGgO5Xhx4lkovZOetjosaz+dFykdOQa47NQ3BJrEU4XPOX7ndfi2l9FsA\n3gpgq4T3pHbpOo1Sq1zzDC8tnpTzVFod18+0iGNJNs+ele/0vNBdG7PSec/LzQTRjjIhruyLJE4p\nzSTzdDpdyBdLg5D88aS8t+VjkfMv8uZbxLfi60ZvwqeU7gMY5Zwfp5SeA/BtAP71ylLWL01m3Nr3\njpUKp2s69L5FbB3EPtXOKB3Xz+DnAggrbuSJtiqrpRl0Re29bDsz0UXiasJrsgssU0BrAx7hLZLr\n4DUQOk26UfJIz+e3YdcvI+GfB/Bbt5XtEMCv5Jx/dyWp6gGP7CXiRyrYMkS33hM5pDwbVfa9QSf6\nXdPpFJPJZLbV8el0Gkp/Vle1bdqnUnYxF/g+JrrE9TkP8k7LscdqvWXbW0HyT66RdMl5/Z2WplQj\nQDZF/t6Ezzl/AcCbV5iW3iiRvQvpAZvsNfZrKQ2W7anD0dERDg8PF8LR0RFGo5FJeB3G4zEmkwnG\n4/FcfDQaYTwem04sLTU9lb8Etk9Leek9Q4ikya1Jz/laalRHo9FcnloOPI/wQvLpdIrRaDR3/ODg\nYI7okn5Pykek37TEH/xIu5KNXkN6S8r3KYSo4WHCW/3IR0dHYfDMAVH3r66uMB6P57ZaxbXUWHlG\njS3KiIhem5cWGZjsWqXX+cmQfJC4EP34+BjHx8dzhLfI72lIo9FoFj84OMBkMlmQ6rqhqlXpLayb\n9IMmfGQrW9sorvdXRXZLulsqvKjtR0dHs8rJ4ejoKBxaCtwQ/vLyEoeHhwtkB56pyFoVtVRSK7Aq\nbTncrDztUnktkmuyc76yn0HUdpbwx8fHODk5mb2jRPjJZDIrE4kL6bmx0flTMjd0vkVm0zpJP1jC\nl1TsruQH4kpsFYCVhtL7PHtdJNHJycks6P3j42Ozb1pL+4uLi5lU0yaAVEqtigrZNVjK6wpsaUGc\nRx5q1Hm53yN7NOJN0itqu0d40WSY8BIXsuugG2XdeDLR9Vj+6Ns9kjPWRfrBEt5DDQmjuFWBuSGw\n3lOyr7WKqW1zHddk5yCEt9R6UWVZokulZOeShia4JpXlwNTxkppuoVZzknNCJInrey3ycwMlpJf8\n1Y2fdlhKYLKLH0TiOm913ol2YDWWnlDZpN2uMUjCl6S7B4u43vP1tdY+p0NXMMshp6WN2OQSl62W\n6rw9Ojoy36m/jQmupZRIIKtvWz+D84DjVp7yvd51rBXo93vv4ngU+Dm6QZRgpV3AZD86OpqRXrQm\ni/Ri5wOYk/jsdOT3bprswEAJb6FrI+BJMCa3fj7b5DqeUlroBuIhoBaR+ZhuBNiG54rNTi4tbaTS\n6jRoKc62aE3+dq2cnuSvaWSs/RLhLfVZk16bOXw+pWSSXex43TWq06IbV6uh0j0gOl2lb/eOLYs7\nQ3gLfTUB776S2s7daqyyn56ezoWTk5O5ffbKa23A6laKHE/SNXd1dTWrrGzz61BbsUoqfM1zrMpv\nHbfeU0t4S8Jr7Yslv3jfdd5JuUlc7tP5LvktDapAN6rcuJbIv07cacJ3gSV9eN+yy9lG1yRltf3e\nvXuzcP/+/bl9IbxUMo5LXzD3F7N3WSrs1dXV7Bki4UXF1OmOUGtrczy6zrJpa9/bRbrLc9isimbc\nsXTXxD88PHTJLnkcmT1eXjeVvgAr40oOtb7P9shuOc90P7pWzyV+enqK+/fvz8Jzzz03F7937545\n6EZ73TWpmeQ6XF1d4fj4eNZFx5KNv6GUJ31szpLNH+V9dG9NsDQYq1eEA0t3DpIG1qR0/nJ6u2hR\nm1DrB0f4CMvY8dZzLKect9VOOSG4qOwnJye4d+8ennvuOTx48GBh++DBgxnhuVJqiaQrJI+mk8E2\nEr+8vFyQ8DJwpJbsnB+R1PXO16LWFOhjw2sJr00sHQ4PDzGdThfyV0wibVJpTUo3yDw8WRykllrf\n5btXicETvq80Lz3Pcux4A194VJe2zUVdv3//Ph4+fIgHDx7g4cOHc0HIH6mbKSVcXV3NjaLT+zJ8\nVsgupoSW8EJ6dlytAl0let93dCE7AFOl95yimvBCdN1IMOHFbNImE6fTsuMjG34djjqNQRO+JJ09\nSe09o+SUi5ZXOjw8nLPJ2V73CC/7zz33nOlc0l1BOs6qZQ1x+1SkyEZe9bu6pqfktLPIztJdk140\nICaobLUTVjfOniPUk+ybUN09DJbwFomtgorUqZKNrgtSVxhu+Uej0YJTznLQiSQX+120AJEwXEmt\n/mMmukj5i4sLXF5ezkl9Ufm9aZ7yrFrbWN5fcpatCpGX3oOVh17PCYeUbkbieeZPSTXvkn/bwqAI\nXyPRu7SyvB912fDkFktSWN53fUw77eSYOPV0t4+1FWi1Ukh9eXk5I7wmPTv5rGmhtWS3CK/TpLFs\nhdbqbXSNJdVlqxtpz0nHDTeABU++1Z3JaattDGvzZZ3SflCEF3j2pyWlPfXMeqanTusuNw564IzX\n7SZxbdPLOblfz4az0i/2oCXhNeEjCc+E98hbK6FWTXR5BpdRSTrq45ajLiK8Jj2A2ag5q9++huxe\n2rYp1TUGSXgGF4YnKUvPYEcZVxjtdddx2bJ0t/ra9fh42deEt75HVE3g2VBOcTAx4YX0mvC8gEMk\n4eUdEeGj7arhNQAeiSzCR2TXpM85mxK+Vp3ntJUa021gsIT3HHKWHV4ivKcC6n09Uo6ltI5bkl1C\naegsf5u1r1V6UeuF6KzSRxK+iyrvSfV1V1pN9khN5mPWQJuS/a4lvLbfoxGKVnojolvXbxqDI3zJ\nWcektwjv+QIix5wMqCnZ5BHhuQ+YK2FUATThRFKzhL+4uFggu+W446DfUZJUm6q4THZu9Dxzoq+E\nl/y3tLxliV6rBWnbfV12/OAIz/BIH3m6veewR1cHGUyjic5ed4vs2n73tAepWECsMns2vPbSd1Hp\nu6jz61bd9Tdax2olJ9eBGhteE16WsGKnXV+1Xp+30lyTn6sk/yAJ76m8Ftm9YY98f0rJlLx6uKwm\nvBBdj5rzuuWE8F66JFgDN4SYQm5Lqp+fn8+CRXxW62u89ZIOnZ51o8bsskw43rfUee4/j7zxtV1x\ngE1gb39bdrvGYAjveeR53xpsoSUoP0viBwcHCzPUeK66luSWCq+dcNZIt6hCScWwJsaI+v706VM8\nefJkFp4+fTp3TBoATXrPW2/Z9JbnfluV1HJcchlH2hyXv6VdRWWiUdJ6+Bq9r89b201jMIT3EJGe\nC1tfb2kFliNNj41nm51tdO1519qCVDLtTLQqFs980yr51dXVjOBMdNn3+uNLA3AiR96m4TXk7Ixl\n4kdEtyR8iezafyDbGqLrY1Z82xgU4T3VKrLddavO1+rCldFy3qozQniR8jxsVrzwWsLrSsZdhFyx\nAMw543j12cvLywXCc9Akt/rked48S3dgN7qOgPJAKia+R3prIlLklJN3a0RmTo1U9+LbyN9BEL6k\nzvNx7XHXhe1VHgCzfnYrSHccq/SszrN2wCp9ZBdqlV4PqJFwcXHhEl0kvdVQaC3Bkugs3XVaJL4K\nWA65CBHJrS7XEulZykd2u5f+GpU+isu+9exNYecJX+vI8Wx43cJHUkJLeG9FmpJKb60lryWLZ04I\nrO42ccSdn5+bJNcqvTYDeCtLMEWrtgKrl0TW82rKlMvXG0zFxywpzl2tllpfaog5PzyP/DLE3wR2\nmvB9JYLXwnuVJ6U0J+E1yfXAmpLTzpukoSW89W0p2d1t4oA7OzvD2dmZSXIhvxDe+82UJjz3AGjC\nM1ZFdj5eKluL9FHjHjnsIg+9p3VZpC857zzNaBeILthZwnchu75Ht/RM+kgdlL52TXpriqs32Eb6\n2XXF4kpWgqXSn5+fz8hesuEt777er6m020SJ5CXpXiJ75LQrqfV9CB5JeL52U9hZwnuocebwcMqa\n6aeHh4dzUj0K1prxsm58pDJaqqFsrZFzosY/efIEjx8/niO9SH1R+S8uLsxRdDVeeKsy9mlwV41I\nqkdquyfNSx56DZ1Hnu+D87I0noGfvw0MhvCRQ0XOa7td/3FEuslYCjDhvQEz3L9e08cuadKIKshk\nMpmT6E+fPsXjx4/x+PFjPHr0aLZ98uTJjOjiibccchHZOU3rgm7kutzD+WaR3XLE6YlIVvnwczw1\n3iK692fZmoZAPzPCJsyAwRA+Akt3TXhNeovsEpfJMdY0ViZ915VPALsS6TAej2dSXQj/5MkTPHr0\naBZE0j99+nRGeN3H7nne+1S+VZZNH+3B0+A02a3RkNwgW42y92wNyaMaotdMSiphU+UxaMJz66xV\nPS3hhaxs11uEZ5JrsnsVqtTlJtBk5AokhBcnnVblX3vtNTx69GhOnWcJLza655SzVMxI5Vx1OdU4\n6rx7PQlvjYaMyF4yBwFbleffR1vkt+6LSB/l+TrLo0j4lNJHAXwHgFdyzt9ye+zrAPw6gG8A8BKA\nd+WcX11bKu10LcQ9lV5I69l+THivW85TGSMJoqGlBnvRpb+dJbyo8q+++urcmHmx3Vmlj8wGScO6\nyiN6dheyW/mnJXykxelRjnqBSc9ur9HG9N9lolmH2zSfahFPIbvBLwF4Jx17P4BP5Zy/GcDv3+5v\nDBbZWaXXXnfPGWetEW8tUcUSvqQuWna8rkT87zc9441t+EePHuG1117Da6+9NnPcaQkvg228Srhu\nld7rylrls3WDLmWspbvlQC1J+ajvHfBJz70gJbJ3zet1NxJFCZ9z/nRK6UU6/J0A3n4b/xiAP8CG\nSQ/Y4+hZ3dOEt7pnJM5/b+XVaTTZ2WnnDZ21KhGriXpdOk/CC+F5aWq9dp2o9PKemkq3bGWyCN7H\nUWc9U28tHw2r9FxGntPOaphL5VRjx9c47fgd20BfG/75nPPLt/GXATy/ovQUYUmTGqdd9KMHvcBF\n9H92axSd2PA1ziAt4fm3UCzhZYCNOOxee+21uVFzPIpuOp3O3hFtN4FlSa+fI1uP9OynseYx1Drt\nIpU+WhuwL+lLWEeZLe20yznnlNJGmyvLkWP1v3Prz904PDIumjxT652nvJmLa6nOY+WlX52DNABn\nZ2emLclOI+/d6yiDmvPeuz3tgMs0ctRpU00vNKJt+chxJ/Acddwwc4PrOe0sG34X7HegP+FfTim9\nIef81ZTSGwG8sspEeeDWOOqX1ZWDJbK1thmvMxeNifcmwwi4sFNKcwNr9MIVEn/69OlcP7tehbZ2\nXTr9/mXzeVXwnsXHuZfFGq2opyqLf0WvOCRxvRqwXgLcapjZt8L/6OOGmSckWSMYNfr6S9bVQPQl\n/CcBvAfAv73dfmJlKQpgEd0bSsmE12qeRX5tm1sSnaU6k17g2WrWpBgtyXUfu/SzixdenHLRghUe\nulacGrL3bRA8Z6vAGg6ttTGtvvPKQw8ePJhzuup1Cbz5DJI319fXs0aZfStMel5fYNn+dwvb7pb7\nVdw46F6fUvoigB8H8FMAfiOl9F7cdsutLYWL6VkgPUsDa1BGtHyV1afLzrmamVaWk0x3xemfRmjH\nnB5V542k86a3blKy9yG6Jcm9rUV2bsC1E1avLfjw4UNzdCQTniF5GBGeFxSxSL8Kokt61okaL/27\nnVPvWHFaQnjqvFbpIwnvhS7X8Si9qDuH7UJtu1vj5C3Cs4TnkXRRn6+kpUv+LnO+5nruwuO45Yux\nfCws4fUfeK31/vXgKA2dP5EzlX/04a0EzM/eRTt+cCPtPEedJRWstelqGwDL1recSTV9uNphZxFe\nPPF6gkxpJN2qyL4OFb5Gqlv+DyZ75IBllf7hw4cLi5BoG15mK1q9F1oLE03MIr1nw0fDmPW7dgGD\nIrwl5SP7vUR8bgQ8h16pP1er9Ex27fW1VHo9uMYbOisS3rMT+3b7bJLset8LXSW8OO1EpY/KU9R5\nKRuJ68bT8sp7TjshfaRp7SLpB0l4y2lXo9Lz4pRMeq/bLvLMW6oiYHt/WaXXzrpHjx4tdMfpobOa\n8Pr5fclem9+ruj8ieakB5/Kz/g/w8OHDhd4UHm8h5WH5W7grTkt3TXrvr7xW2Uh8lzAowgP1Tjur\nAajptuOuIB08G1Tg9eWWnEF6aWkePWc56LjLT5xONXm3zNZClwptjXDTcf5BJ/+0s7REuIydt3pw\nDg4O5hx0Ov0s5a0yi37BvQpn3aYwGMJHqp/VmnuTJrgSWPueyq7TItAktNR4HsBhzaHmyhJJQamw\n1pbTxvCIZsW94aeeDRwd02mLhrl6i4hKeN3rXoeHDx/O+t3FG8/OOXauShDpzg23LktL4tfMkNPY\nVekODIjwAov0Ftm5QfCIbx33hl8yNNks6c6eX32s1IdukV7eyVtLwlvprc0Hbgwlrt/n9UZ4+9fX\n125eS4gWED09PcWDBw/wute9bu5PP7rrzXOsWumv0dCsobXRHHhdLruKQRG+5KFnFTyS6tYzSg2A\nBUvC84yqaNKFRXo2GyQNJeluPUPvW2aKF6z89IjsEYVDqaGJ/swrhJcg6ryo/5Fj1WrAo3KMzLKo\n/HbdYQcMkPAW6a3RcCXilwjuqfMaXKheRbHGXnP3WvStsrWkOm/5GTruOTflmNdDIUF/HxNc27/8\n/bLPtjVveeoy2+k8jVn3t8s3RGaJzqNIwteSnq8fAgZJeMuj60mmkhrpXcv2K9uwGjWqoCcluMKU\nPNks1Vml93wNWsJHow6tIISKpDc3cByfTCYLec37mszRTz94rQImvJd/UkaRlK9ptEsa2i6TfzCE\n1xWaJYMn4SNVPmoAPNILLKcUSz6rskSed3mW/l7WaPT7IpXeI7/klTcmIQpHR0cL5ornq/BCyYzg\nv/Jy0La99R8/ySPP/+JpbFF5djHLPGGwSxgM4YFYwteo831tdy1dNTzHldclF0l4/k7reyPpbkl5\nj/AyEYW7vax1APSxkm9Cj1Dj+Hg8XignjuuJMDrIMZ6qzKPpLO87a2eRM7ZWU4vKjuvHrmFQhAfi\nMfWWRLfUR895Z1WGSCowvJZe71u+B5FQnoov1/OzahoKHbx5/hbJLdJHJooQPgpeAy1xITZLdjnm\nDY9mdZ7zQ+Kizuv8K0nyyO8yNPsdGCDhgbqx2R7ZI0keeXVZwnPF0enwGiGpoKKO8rBZIZn1F9jx\neOx+O7/b21rS0YuzOi8S3graTvek+2QyMcvCU+mjBS28GYzcKDPhtaS3BtZEU1+5rKO4tb8rGCTh\nNaxCjZxyJfJbjjLrXZG97aVBq9IyAUQqFoBZxTs5OTHXrfOcURIszUbvd5kpyOcODw+LUlBPKrF+\nammlScf599ua8NFUZXmOV1Z8TPtZPMJ7E2Os8mfsKtmBgRE+Utdq1HxP3Y8886zSa2mu93U69LNG\no5v+a02gk5OThQqVUnJVYSF8lx4HrwfD6npjInlzCiLpHjnttIS3GkQ5xl1yvDw4p4lNA6tuWPXH\nI7w3dFZL+ZLJtstkBwZGeIblgeUKxdLdInok3T0pz8SvaXREwjPZgZtRcJ7Ti73cnubiTfwpEcWT\nmLxfkvBMdD5m5a+Ol/744zlmLQnv1Rcpu0jCW4tb6DLXGBLZgQESPlLVukh4i/Q1DrtIwltpEAl/\nfX2No6MjTCYTHB8fz1WklG5+Vx11aU2n0yJpuV+dj1lk5rhOO+dl5LCz7Hkdn06nc/lj5a83tFZs\neK/hZglfqjs1Kr0n4XU9GALBGYMjPGCPkrIqkSetupBdP193fWmycnfPwcFNF5pstVpvkV3sayaM\nJtH19bWpgnsz/6z5/l0clZzPKaVqsnvqvqUp6WNed6BsI+3s4CD+pwqX3fX19cLqNrwyrWXDD5Xo\ngkESXuBVUE/Cl2x6i+wcB+Ynzchxq49XV0IhvCa7vFcIq6UKkynnXFytp+Rpt8wVdk5GXugS4UvB\ngn5HlH4hvKeFyDd4ajdraTUS3pPuQ8agCa/BpIxs6kh1Zymkn2cRHFgc0CGSXSBSXsaj6+drD7qu\nZDouobRMV9Sldnx8bOYLfx87p3SIGqSo/1o3WjpPOB6tRCQDazz735PwHtmZ8H3XnR+axN9ZwnNB\n9X2GjtcUWO17PfVUKp5VCQ4P57NbN0SitlsjveQ4d5dFY+G9lXrkvV7Dxmm3rrOCNHJW4MbFaiwB\nLDTIVrqj/NewGhRroA33JHgTnYY2hNbDzhIe6EZ69qR6kqp0f/Q+S+3Vx1mF994h92jv+mQymVVI\naysmQSlYS3VFhLe+l8nvaU1M9C6ak5UnJb8Kp91LPz/bIrxF+i5kH5pkF+w04YF+kr7UjdJFJbPe\n76XHs42tCsvddaLyWkEG5lhdaZanno97ktLbcuPnkZmdk120KCsvuzQUUTlYmh2AcBYcO+tYnfeG\n0w6N9DtPeMCX1lHmR/d0eWdEbuuYSPmU0oyoTDbp0x6NRnN2rpVufazUV14aTchpZzJ5RNdxj+wW\n6fVU1BqpXNNbYO1b8CSypMsjve5dYAnPE2aGRnZgIITvAo/0fE2ppS6p9laDcHBws26adtoJ8TVR\nrq+vMRqNFux07zv086PhqVGotX/l2yzSRxIewALBuREowfK6R8T3yoTBZGennafSs6e+RmPZdQyG\n8F0ytSTpl3l2RAgACzatkECIoSuP5wiKtAdtDlhdVJFZET2fvy2S7vwOuaYk5a33aXhdbVGavXOe\naeSp9Z7TrkT6IZEdGBDhu8JS5/hc38JiQnjnNXkiVZ3JbtnWJTKX7HJOnwUmu/WNEfEtkstxbd54\naal12Fnn+Dt03CJ9NP6/i3QfGgZJ+FLrzS25jmtVmu/TEkuD1cXIhvNUSzYDPHLVjBOorfge2Hyw\nvslqoKxxAvqYNfRWH+P8sPLHavxEM+Jj1rM8mz3nvJAm3QdvzZTzSM/5MiQMivAlkluOGF2oLJH0\nHG0hPLBYqfQ5SYfeWmnzGqVoPxpFFo2Qq1V9dTo5zTXnaobQRtdE0h3Awtz86XSKo6OjWTp0j4Pk\nl2gQusHwxjOwg46nIJd+B83lz/EhYFCEB2zSe5WNR1FpskiQe7T6KefEAcfv57jeRqT2tBC5xnO4\njUaLq7FK2vRxgafGW+pp1N3EIZrcY9m+XB5e+mRfxs2Px2OcnJwsODPF2cnSnUnHqrvWQLhOeGTX\n3+LlzxAxKMKXpHtkm43H4wUS6Uqpu9MEFuk96Q4gHDRjpVHHhfDWjD5JA2sAbA54pJdrLFOG4x7p\nhTDRElYlqa9hkV7Ifnp6Opce9hHw3HfWvriO6LxmCW/9/12TnvNm6CgSPqX0UQDfAeCVnPO33B77\nEIDvB/C128s+kHP+nXUlkhGRvmYBBk0sXRmEdBq6i03erbd8LCIUE4FtRW+ut5Zs2gmm97nBsshv\nST3WMjypn3OeI4T+0aI1rdSS8pwu3j89PZ0bYiz5KuWl81w3ArpO1NaNSMLzt1gaBNeBoaBGwv8S\ngH8H4OPqWAbw4Zzzh9eSKgOeKmqp9NySS9BEn0wmc6S3nHUaJeJHtiOrlFYQUltz3uV+7nPnPCl5\n6j3tQjd4kVkipNBbHS/Z8QKP9JrsOt8lH/genQ9WXbF8O9Pp1FXpSzb8XbDji4TPOX86pfSicWq5\nmS09UWo7kG9mAAAdJklEQVTFIylvSXcJvICCdgKxqivp4G1EdqsR0nFNeBlqK2TXf33R6j6nySK6\ntnE96auJb5kispU/3eq/3uqtp8pbKr1Op4AJn9L8Wvr6u3hMA9ePGkcuaynemnaR025oWMaGf19K\n6fsAfAbAj+ScX11RmkJ4hPekvLVMskh3XqdNQ1euyGFTIrzlW+CuIKlkPEWWK7NIMyG7VnO1Hc+E\nlzinQ5PAIjxrLPJba/m/Pccjh53kb6TSi4NOCK3Jfnx8POd01d2rVt3gsvAcuePx2LXhLaedVR+H\nhL6E/3kAP3Eb/0kAPwPgvStJUYAae82SpFdXV7MVX7yuLmB+rLrYxZpYFun1cUtd1kHbv9aiC9ay\nVDqutRNveSrAV+ktj7ret3wPHuGZ9B7hdX4IonRqovNfcrhbUve06AbAatBYs/IckCVV3qqPQ0Iv\nwuecX5F4SukXAfz2ylK0BHSBa7JfXl7ObEBP5ZtMJu6acUI2y67V+x6RrAbIqmi1C0x6pI+Q82K3\nGqezRqWPgiY3Owen06mrgUjQEt2a729JWt1IeAT3Ane/RY5Mz3E3NPQifErpjTnnr9zufjeAP19d\nkrqBJS2rziJFWUpbDjWeSz6dPhv4oW1GyxOvG5lSZbPOi5TS3XHeYo0e4S3tQ+eT5zCU93tEl2OW\nzauDRxiJRwOH5DustfFlq8GSXghv5benUVljByyye3VuiKjplvtVAG8H8PqU0hcBfBDAt6aU3owb\nb/0XAPzAWlNpwGpx2RuuJTyP9WYJf3x8jPF4PFfhdKEL4SMPfNRtJWvLa8Iz4Uoj7UpS3nIy6hB1\nWYpKH2kwXmMlxzwfgMS975O4R3it0gM22VnCR2Rn9d0iPeedV++Ghhov/buNwx9dQ1p6Q0s0S8LL\nemgscdi5xAtJau+s9up7zjjt/NHebDkWdctplRdYdLgJISLSR2TLOS9ItojwFuktM0DvWyTR+95I\nQgnW3270vkATXdv8nirvkb2k0nt+myFjUCPtAL//k4ksBX94eIjLy8u50WZeV43+Qyq37Nqjb3U9\nTSaTmRPr8vJywaGlu61YssjW+i6NEuEtNZq1EK7sEeE5eA2ebK2y0duSuRL9+kqvussDqCJThbWt\nyIbnhp674+4CBkf4CJrQUuAyDh3AQgUVkuuCt1p2aSwsokp8PB7j7OwM5+fnZmAvttVoRBIWqCc8\nO8us3os+hLc0iJInW6PGLInW2bdGSsr1nrPOMrEss6ok4e8K8QdPeC1BLLLrEXSsgkuQP8KwGg88\nUx8jp5fY7+fn53j69CnOzs4Wwvn5udtlxyT1nF8R4XkgkdVF6PkPZCv56KnlOr+teKlb0Op54Li1\nVLUcZyelbhw86e6p9NzX7kn5u0J0wSAJb0kdq5vt6upqYWCHpYpLsCSi7rKL7N+rqyucnZ2ZhH/6\n9OlMwnvqsCWR9TGP8JbTyiJ8lHaL8FZcEA2eibreJJ3X19dmXEvhy8vLufX1xRejHXuR/V0TLD/F\nXbPZGYMifERwqdTa68tTR637NJG1diAV7vz8fPZfM8tRpe/RKr3Ez87OZnY8VzirokaVl/0Umugy\nAMXSIizS83ssm9ur+NrUKe2zFiDfIOWjnZXcCLOE1pLccqx6/fuek5AbJn2vPnaXMBjCc0Vkkmiy\nWwXqeedFauhKZtmPco1nh4/HY3O4qXbkdZE6nj2pp8sK6eVbI+2B06zPWeS2tkyAWqILNNF1EIel\np4no0ZL677va7yHwuu6sbkBLG+Hn3DUMhvDA4jh6kWwShPRWIbKHXtR03aUjaqMVeCy6ZRpYE0pk\nqwemMKn1fhS3Kq5ISIl7moIl/fV5ySNrq/Nf56l1PFKDufyYcDo/LRvccrBK2jWsRkU3jNbAH6sB\n0N9a+rahYJCEt9RbTXqP8CIRZTuZTIp2scRZglrSk6eL6mmX4/HY9W7zMc++FJJ4FbekPXgOQssT\n7cV5X5NDtABrK+flG+ReLp+oW42701il18+MVHlrLkWNlL8LpB8U4QF7LHyplZbrhcQW0T2Vr5ZQ\n0+l0oQtIx6Xby3MSWcf5WCSRvPR5JoJnMnBe15SHoFbCC+lZOls2vKjypZlsGiXSe3XFIzvHh0z6\nQRHes98jostW7F5v4IfuvvMK25LAEtfOO2vMvO7n1t/SJZScTFH6arWIKO+tPLGuKdnw8l6+1xor\noEnPZGcJz+VWkvQe6Uu2/JBJPyjCA/OkF7KL08civL6+pOLxPV3C9fXiijbszWeye3FvXxARrmQ2\neEHury0DCzoPZZ+vtWxuSYPnpecZhtYgKZ2GiOyRKm81GHcNgyF8RDQAc8S3IITnSqArQBcpaKXF\nG8ghQe6r3XYhopUvJXL3IXspDUBMfKvs9CjBPqaHRXDp49d5Yf1pl/9BzwLCI/5QpfxgCC/gCiOF\noVV7S9UHnqmUkY0eSUh+P1diq+9b7+tv0FvrmLetyRd+rnVsk8h5vktP0iOk0XGBRWD2ubBzVT9P\nP0e2sgz2yckJTk9PZ+HevXu4f//+XH3gMrUIzg3bEDAowuvKz44fi+gSl3sih5dcU6sK6+sBLEgi\nzxPO3+LFa67lvPHS2eU564ImNZeJp5ZbE2Usosu+9U36WZrwEu7duzcbO6Hvl3LjXh8LQ5L2gyI8\nYKuEAOYIronOBRXZbTWE19dZ6ZGtpYrK9dY3dYl7+6XgvX/VYIlunbc0GI/0VnepFfiZrOqfnJzM\nSM8SXhNeS3Y9H2OIEp0xaMKzas+kB/zhnxzXz/YIr6/Rz7RIzw1AiejesVrSR5I9ep/Oi1VWZIv0\nnnT31HmL9NZ4Ce5l4WfIeZbwTHqtnYnjUJsLUV4NpTEYFOFZBWSnjya9oLYgPNIyYaNtjYSt+cbS\nMasRqJH40TuA1Vdaq5Fl0kffYKn0rMbr4/Iuy/6fTqeziTia7Pfu3ZsbCal7Cq6urhbIrvPKMyF2\nmfSDIjywWDm0My26RyRApDJ7hPfU4q7bLt8XnYsI34fo6wYT39sXeKQtqfV8r3jaZdCVJeEvLy9x\n7969hRWJ9Oy8u0T6wRA+knJaurMjTl9vEbeklteSPUprCVZaPXTRNPjba9O5rgpbIjqr9SWys0rP\nDQX7UVjC83r0PNBHE75rvuwq6QdDeIau4LqAmey6S06uj8hfIjy/uws8R1YfDcCT9l6Dtsw7Vw1L\novN+1C1nee2Z8Fwvcs5zv6MW0vMqtnqgjyzC4TkF5R1efdhFu37whOcuNSEsnzs4OAgJ4anwnoT3\n4iXvdOma2u/2GiAvbqV5V+CpxZ79zlJdb7UaL2Uu79B/sRHC83RhrdKLBnBycjJbU69UP6zv2iVp\nP0jCS8sqcX3cG69dIrqO9yU7p60m/X2+3Xt/zbku79gULMmoSewNutHXsSqvG3ipD7JajhCeh+jq\nabn6/3miFYhfQGuBesvSnnsktpG3jMER3spIvWWiW/fw1iM8H9PP4+d7+5Ea34f0XQnf59mbgJVH\nkifWwBtW36P5EKzNSZDhs3oRDd0DI5JdpPvFxcWs715+SCL+ACE/8MyH5H2nbgi2TfzBEV5gkV2O\na9KLOm/ZWh75lyG7lU5B1C/d5bv5uSWJ3iWNmwB/t6fSawnv2fGWs9ZruLWEZ7IfHBzMEV5WK7II\nr00A6zu8usnftw3SD5bwwGLrKcd0i+tlvEeSUpzf3ye9pWPR/d67a8wN7/pNwtK2PKedZcNbkl5f\nr5/J5Sc2PJNd7h2Px7Nhtvfu3cPp6emM8MfHxzO1n8d6WPXDkuyW6bLpchgk4T21HpifIGN57eUe\na98je3TPMmnvc6/3/q7724ZFfMEqPPT6ubIVCc+SXZY504Q/OzubEV6cdlZ9Yo2y9hu3hUER3iK6\njutM1Z76Gs+5jkdkX0XBWSptTSPQx7TYlYrmwbNxI9JHKr1+Bn/74eGhSXZpCGQh0rOzswUJL5qB\nTreYi7L1SC/YBSk/KMIzSqT3roueF235ur7gdHYp9CET3bJtLZPLsss9e90ivIDLXM+q4/wfjUZz\nQ271iDzx6ntz5IXwgL3AB3+/TtOmMTjCR9LRawAiQtWo9961q/wGjcjje9chZNSBj3eFlvy6O08a\nAF4gwxqgc3p6ipxvfshpEZ6nabOqH9n0m5TygyO8Ba8RKJHdeo7eWudWhUiN7yvt7yqY6JaE9kw3\n6zprUg4TnkkvEp7fJaTWKxvr4yXn8aYxSMJbZCmpi95zomNRA7AKsN3asIhlNSydt6z+a/VeCC9j\n6Jnsp6enc4TXEly66DRE/Zd4V81zXQgJn1J6AcDHAXw9gAzgF3LOP5dS+joAvw7gGwC8BOBdOedX\n15zWOXQlfe0zOV5zb+3zaxuerg68uwBLgrM6z9cB/voGXv5oCS+2txBfvPWWhNeE12kSsuvZegLW\nCLZNdgA4KJwfA/jhnPM/BPCPAfxgSukfAHg/gE/lnL8ZwO/f7m8ctQ44rzJFtmLNu7valH2ebfkY\n7hLZPR8K70ekB+IJOJ4TkL3/lkrPS2Jpr71eENMbCWj1HniN1CYQSvic81cBfPU2/iSl9HkAbwLw\nnQDefnvZxwD8AbZMevZ+9nGKROrjKknWVaLfJYJHiMjuXdMFTHhW5y3pzhIewJwar/9PyOkVSW45\n+SQ9O6XSa6SUXgTwFgB/DOD5nPPLt6deBvD8ylPWEZ7jTuJdnxXtrxp9VPq7DEuaR5qNJTm964Ts\n+tlyzlLptZfe+jGn/ouRNYdDwjJdsatESaUHAKSUHgD4TQA/lHN+rM/lm1TvhAhaBVE3TXYvDXdd\nqkffF6ngpT73Eix12nuXVvlZfS8N87XSE53bFIqETykd4Ybsv5xz/sTt4ZdTSm+4Pf9GAK+sL4nd\nsAuE3TaG+s0lG9v6D2BX8mipq6e26vH11krFy37PrmhtIeHTTSo/AuAvcs4/q059EsB7buPvAfAJ\nvneXEDnB+Lpdwq6lZxPQRGYJa5G+C5ks5y0vJR4tS34XNK+SDf82AN8L4HMppc/eHvsAgJ8C8Bsp\npffitltubSnsAbbnrfMa2x4MEaH0LXztkMES3vKgW+Pou8DqnfH+I1Aiedf8trqRN42Sl/4P4WsB\n71h9claHVRBlVwi0K+nYBCyVPrKbeSBMCRbRS5J+Vd/kOWc3Wb5VTruhYp+IclfATjMmvZ4a21XC\nWyq6JdlLi5jqZ5W+pSZ9m7Tv7zThgeUmXTRsFtpr7nnJvamxtahR6Vehynte+m3jzhNeowvxWwOx\nHXg2fKTS94Flt5ds+Nr6E5F92x77vSK8oG+hNawONaMLu4QaRN53/umnpNEyMaL19LwhvruCQc6W\n2xS24VTZd+h+cf4xhPwcQka46b7ziPiWd9wivVzrOQ554I2O6wk58g0W0bdtXjbCV2AXulP2AZrs\nmvCybPTJycmsARDCM+kF1th1HS9pCtFIO6vXQH5YKd/AS15tm+iCRvgeaA3A6sDEtCS8SHf9Dzgt\n4ZnwPHFKP5+3FtnZj8Bkj0if87P17WrV+dYtNzDsmp02REj+CYH1n1zlTzDyHziW8JFzjclcI+Et\nx6FH9JpFNT1sQ+o3wq8QjfTLQQhgqfQSLHVeO9uY7NF7PDsemP83PZOdie8tmw3YGsw21ftG+Ap0\nkeCN9P2hveWWDc+kn0wmbt95rfpeo9LrabOelC/NmNsVs68RPgAX2i6OnLprKHnpNdkjld7a8nsi\nlb52emxJrd817KXTrlQQXc5HXUG70qoPBUx2lvAs5YXwotpHfeF9+vS5e67U567v875v29g7wkdk\nLp2zCiwif+vHLyPnZ6vBROr80dHRguNO2/NMeB7ZViI9ryVfM/quZqSe9b3bxF4RvmuLzOdKkt0j\n+C7actuGEF3HvS45Jryn2pckcA3ZS42CNdEmahQ8v8G2sDeE70r2ZdT6SLI3qf8MTHrPQ390dISL\ni4s5tV5LeCG9VsEt8kfE7RJK4/BXtWLOOrAXhO9C6tKxrpK92fjz0CTnfc9hd3l5idFoZEp47p5j\nVV4WrATQi+AemUtSnrsKrXzYBu484VdJdt6vkexNoi/CIj0wv/yzlvDiNBMJb9nw0+l0of9bx3lZ\n6lqiRyZAqSFoEn7DWIbsXbvfmOQ10n5fpTzwLL84D5nwursrsuHlPv1HGRni2oXsNQ66SK0v2e/b\nxp0l/CrJbjl/rGeUJPs+E9wDO7P05JnxeDwnta2uOU16Ud91eQnZa1T5kkS3GgWtlUQ2/K6U+50l\nfC26kt065kkrfSza35XKsG1YBBJpPxqN5sbXW0H3mYtk18G6h82D0nHuGWCi14z609+7adxZwrOd\nWDpuoe91kWOK9zneYENLV/bki62vVXlNfk14udbaXlxc4Pz8HBcXF27gXgJrTD9rBZJ+vd0W7izh\ngW7kLmFVz5E0WRLea6D2FfztWo22RuNpwlukn0wmc+YAT8oRsnukj7oES7a8/h42YzaJO014oEx6\n77w34Ma7xnueJ835vFfw+y79LXu6RsKXCC9B+wQiyW6NA4hW3vEcd9uW9Hee8EC9pO9zTR/Jr0mu\nt3LOe+86K0nNd2zq/d572MYXwglhmfBM/MlkMkdwicu2K+Etlb5Gwm8Te0F4oCxpPXiS3oqXbHd9\nzNrys6y0rLrSdGmw1qFteO/3pKImu3jyPQnPwSK6JvH5+XnRhrdW3okG3kRqfHPa7QhqVfwatV/Q\nVZ3verwP+volNtmzoMnhSfguhNck5yCkjmx4Vuktp50n4fX3WN+4CTTCB4iIXSJ9pEVYKnyXfvu7\n0pXXtcER8lg2/Hg8rpbwTHwhc0mlZycfz9YbwgSavSJ8rQMvsrE9lAbeeJ55a78Gy5C+5lsieNrK\nqmA56jypLuQ9Pz/H0dGR2R1nqfSszpeC7r7z+uEjgq8jn/pgrwiv0cV7Xru1UCL3Kshfi9rehmi/\nxsGo0fVbLI/8wcGBOWVW1O/j4+PZ8lOa8Bz3uuWsqbZR0AODPBW+JOH75s+y2FvCA/6cbAC9ye7Z\n/961pQLv0phY50vPtvY9cyXKL4anzfA13JiwVPcku0hnkeyy/JSQPKX5qbJ64I211WaC3vJxyytf\nS+5dQEj4lNILAD4O4OsBZAC/kHP+uZTShwB8P4Cv3V76gZzz76wzoeuCJ+lrJbjAs8F1fB2VoKbB\nse7x9oUoVlw70KK8qc0zuVanwVLjLQkv3Wh6ccnRaDRLMxM+pWQS2CO4d6zLHPgS+bfRKJQk/BjA\nD+ec/yyl9ADA/0gpfQo35P9wzvnDa0/hihHZnyUpViPtS+8SLCO5rev7XmuR3SI9E76G7F0aTJbu\nTHpRyfUceS3ZZcFJ/gZusJi0er9Wla+x25noVj5tAyHhc85fBfDV2/iTlNLnAbzp9vRqxppuAR4R\nPYLreI1qHxHQs3/Xabvzuzku+6XARC+p6n2+x5Lw8n5W6fXvo1NKM8Lr79HxyLa+vr42pX2tlI+I\nb33jtlBtw6eUXgTwFgB/BOBtAN6XUvo+AJ8B8CM551fXkcB1wVODWcLL+Vp7np+hsQlSWyhJdR2P\nVOIu9mmNDe/dowkvqjgT3vpXPE+P5a2XH4Jahx1rBlFDwt+1bVQR/lad/08AfuhW0v88gJ+4Pf2T\nAH4GwHvXk8T1w7Ijrf0abEKFXyYtEdlly2SXbVSxvYag63ex1NVp1IRnsuec55azqtFgeO07z6Yv\nSflatX4XUCR8SukIwG8C+I85508AQM75FXX+FwH89tpSuEGwZPZU9BqV3joXqcCrQo05oeORClx6\n3qoRkV0kvHSt6Z8/yHWa8Po+Hfe67Hhgjjcf3htZF42uk2/bBZS89AnARwD8Rc75Z9XxN+acv3K7\n+90A/nx9SdwsLHU8Uts90nvPrnHq1FSOWiJ66qxHeOvbOf21oQ8s0l9f3/x2mbvnrL+9WN/G3+dJ\nevH869F33g8srd9VD6GLriTh3wbgewF8LqX02dtjPwbg3SmlN+PGW/8FAD+wviTuBkqkL92jj8nW\nsu1KlSMyGUookd1KP+93IfmqSC/gIbVM9JINH2kxQnieWMODcyxPvdUlt6soeen/EPb/5/7LepKz\nG7CkPJ8H+tneniMneg43NPp4BM+Gjwih06m/z7NPS3brsja8PiZkt37cqL8tMk2iPOPps6zi8yi7\n0uy4XST/Xo+0i+Cp9oA/yMYqYNYI+Fm1lSIiexf1vgsZPNLreF9NxYJ+n6xJp215Jjw3UNYy1dwI\nRGaU+Aes5a+E8Npm91R5/exdI30jfAAmOB/X50q2e3TMu7dGfe8i9UuqrQWrAtfEo++KoO8Rsksj\nIMNjWbLrBiKy0UtaCg+55WARvOtU2G2jEb4CHvH1ub7PLF1TI9lrJbxcW+uo48pbIr/3XV3yyPID\nSBplmO14PJ6d19KWCc/7bC5w3FsNVxoBy4ypIfouEb8RvgNWWXDWs7oS19rW3CNxi/BWZfVI7anz\n1n4NLBNCp9tKK4+59ybOaFPBUsl5JR1rG+XDst++KTTCrwHLSH3Pjl4nmGhyrGSX1tjuXfPCSouA\nh9tyWi2SlwjP4+pLU2GX1WC2jUb4HUMt6TUx+vQWWM/2bFyOy77ecnyV0N/Iw22n0+nsOk+Vl2NM\neKsBKC1XZaVtSGiEHwiY2Ksku6dKl2z0TTqoOE1CTn3McthZNjyT2SI/k/2uoBF+B+Gp8SyN1012\na6vvt+LeNX3SZh0XIlrH2SfhEd6S3BbhI8m+zPdtE43wO4pa0gtq7Pyo4lo2PN/TRaVdlgyeGWOR\nXnvogbjbkYlskZ7j1vcMkewAkNaV8JTSMHNkB1HTt973eaWuxq526zrqUzSYhu116x7ZlnwTpcZg\nXd+3LuScFwq3SfgBIJLkq5Cky5xfRRq6Pl+r79qnobcct54XaTQl/8VQ0Qg/MCxb6frY/btS0S2C\neqQuaT5dTJFd+f5VoBF+zzD0yntXbOltwZoJ19DQcEfRCN/QsEdohG9o2CM0wjc07BHW1g/f0NCw\ne2gSvqFhj9AI39CwR9gI4VNK70wp/WVK6a9TSj+6iXd2QUrppZTS51JKn00p/ckOpOejKaWXU0p/\nro59XUrpUyml/5NS+t2U0t/ZsfR9KKX0pds8/GxK6Z1bStsLKaX/llL63yml/5VS+le3x3ci/4L0\nbST/1m7Dp5RGAP4KwDsAfBnAnwJ4d87582t9cQeklL4A4B/lnP9222kBgJTSPwXwBMDHc87fcnvs\npwH835zzT982mn835/z+HUrfBwE8zlv+wWhK6Q0A3pDVD1AB/DMA/xw7kH9B+t6FDeTfJiT8WwH8\nTc75pZzzGMCvAfiuDby3Kzb3i5UCcs6fBvD/6PB3AvjYbfxjuKkkW4GTPmAH8jDn/NWc85/dxp8A\nkB+g7kT+BekDNpB/myD8mwB8Ue1/Cc8+cFeQAfxeSukzKaV/se3EOHg+5/zybfxlAM9vMzEO3pdS\n+p8ppY9s0+QQpJRexM0PUP8YO5h/Kn1/dHto7fm3CcIPod/vbTnntwD4dgA/eKuy7izyjR22a/n6\n8wC+EcCbAXwFNz8Y3Rpu1eXfxM0PUB/rc7uQf4l+0IoN5d8mCP9lAC+o/RdwI+V3Bvn2P3k5568B\n+C3cmCG7hpdv7T+klN4I4JXC9RtFzvmVfAsAv4gt5mF69gPUX863P0DFDuVfcn7Quon82wThPwPg\nm1JKL6aUjgF8D4BPbuC9VUgp3U8pPbyNPwfg27CbP8f8JID33MbfA+ATwbUbxy2JBFv7wWi6mRe7\n8ANU7Ej+eenbVP5tZKRdSunbAfwsgBGAj+Sc/83aX1qJlNI34kaqAzfThX9l2+lLKf0qgLcDeD1u\n7M0fB/CfAfwGgL8P4CUA78o5v7oj6fsggG/FjTqacfuDUWUzbzJt/wTAfwfwOTxT2z8A4E+wA/nn\npO/HALwbG8i/NrS2oWGP0EbaNTTsERrhGxr2CI3wDQ17hEb4hoY9QiN8Q8MeoRG+oWGP0Ajf0LBH\naIRvaNgj/H/8GCqJnNQQWgAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fb1f0dd3810>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(to_image(sample_images)[2]);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 128,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD8CAYAAABTq8lnAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n20mlTGTdUZcAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fb1f0dc1110>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "#run gibbs sampling from random inputs\n",
    "random_gibbs_images = generate_state(np.random.random(size=[BATCH_SIZE,IMAGE_HEIGHT*IMAGE_WIDTH]), 200)\n",
    "plt.imshow(to_image(random_gibbs_images)[0]);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[-0.01409554, -0.00536683, -0.00565238, ...,  0.01831558,\n",
       "        -0.00584126, -0.00820349],\n",
       "       [ 0.00997176,  0.01120596, -0.00595941, ...,  0.01863942,\n",
       "        -0.00650183,  0.00103016],\n",
       "       [ 0.01185286, -0.00741214,  0.00017319, ..., -0.01414159,\n",
       "         0.01406672, -0.00414365],\n",
       "       ..., \n",
       "       [-0.00112381,  0.00889921, -0.00948953, ..., -0.01556171,\n",
       "         0.00801014,  0.0082342 ],\n",
       "       [-0.00777897, -0.00974959, -0.01762774, ..., -0.00301312,\n",
       "         0.00079663, -0.01516621],\n",
       "       [ 0.01714878,  0.00978682, -0.01005803, ..., -0.01142717,\n",
       "        -0.00302883, -0.01486796]], dtype=float32)"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "W.eval()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2.4584517"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.max(np.abs(W.eval()))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "a_vals = a.eval()\n",
    "a_vals = a_vals*(np.abs(a_vals)<1.0e38).astype(float) #kill off ridiculously high biases"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "1000.0"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.max(np.abs(a_vals))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.0012470795772969723"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.min(np.abs(a_vals))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "1.6849988"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.max(np.abs(b.eval()))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "1.5832484e-08"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.min(np.abs(b.eval()))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "7.4040145e-08"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.min(np.abs(W.eval()))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "start_state = batch\n",
    "hidden_probs = sess.run(h_prob, feed_dict={v_state: start_state})\n",
    "hidden_state = sess.run(h_prob_to_state, feed_dict={h_probabilities: hidden_probs})\n",
    "\n",
    "visible_probs = sess.run(v_prob, feed_dict={h_state: hidden_state})\n",
    "visible_state = sess.run(v_prob_to_state, feed_dict={v_probabilities: visible_probs})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "1.0"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.max(hidden_probs[0]-hidden_probs[7])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD8CAYAAABTq8lnAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nZkwAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fb1f1253dd0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(to_image(hidden_probs)[7]);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD8CAYAAABTq8lnAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fb1f11af250>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(to_image(hidden_probs)[0]);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x7fb1f1056910>"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAPwAAAD8CAYAAABTq8lnAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nmru2Q0NVNKwAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fb1f113cd50>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(to_image(hidden_probs[0]-hidden_probs[7]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}