{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Dynamic Recurrent Neural Network.\n",
"\n",
"TensorFlow implementation of a Recurrent Neural Network (LSTM) that performs dynamic computation over sequences with variable length. This example is using a toy dataset to classify linear sequences. The generated sequences have variable length.\n",
"\n",
"- Author: Aymeric Damien\n",
"- Project: https://github.com/aymericdamien/TensorFlow-Examples/"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## RNN Overview\n",
"\n",
"<img src=\"http://colah.github.io/posts/2015-08-Understanding-LSTMs/img/RNN-unrolled.png\" alt=\"nn\" style=\"width: 600px;\"/>\n",
"\n",
"References:\n",
"- [Long Short Term Memory](http://deeplearning.cs.cmu.edu/pdfs/Hochreiter97_lstm.pdf), Sepp Hochreiter & Jurgen Schmidhuber, Neural Computation 9(8): 1735-1780, 1997."
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"from __future__ import print_function\n",
"\n",
"import tensorflow as tf\n",
"import random"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"# ====================\n",
"# TOY DATA GENERATOR\n",
"# ====================\n",
"\n",
"class ToySequenceData(object):\n",
" \"\"\" Generate sequence of data with dynamic length.\n",
" This class generate samples for training:\n",
" - Class 0: linear sequences (i.e. [0, 1, 2, 3,...])\n",
" - Class 1: random sequences (i.e. [1, 3, 10, 7,...])\n",
"\n",
" NOTICE:\n",
" We have to pad each sequence to reach 'max_seq_len' for TensorFlow\n",
" consistency (we cannot feed a numpy array with inconsistent\n",
" dimensions). The dynamic calculation will then be perform thanks to\n",
" 'seqlen' attribute that records every actual sequence length.\n",
" \"\"\"\n",
" def __init__(self, n_samples=1000, max_seq_len=20, min_seq_len=3,\n",
" max_value=1000):\n",
" self.data = []\n",
" self.labels = []\n",
" self.seqlen = []\n",
" for i in range(n_samples):\n",
" # Random sequence length\n",
" len = random.randint(min_seq_len, max_seq_len)\n",
" # Monitor sequence length for TensorFlow dynamic calculation\n",
" self.seqlen.append(len)\n",
" # Add a random or linear int sequence (50% prob)\n",
" if random.random() < .5:\n",
" # Generate a linear sequence\n",
" rand_start = random.randint(0, max_value - len)\n",
" s = [[float(i)/max_value] for i in\n",
" range(rand_start, rand_start + len)]\n",
" # Pad sequence for dimension consistency\n",
" s += [[0.] for i in range(max_seq_len - len)]\n",
" self.data.append(s)\n",
" self.labels.append([1., 0.])\n",
" else:\n",
" # Generate a random sequence\n",
" s = [[float(random.randint(0, max_value))/max_value]\n",
" for i in range(len)]\n",
" # Pad sequence for dimension consistency\n",
" s += [[0.] for i in range(max_seq_len - len)]\n",
" self.data.append(s)\n",
" self.labels.append([0., 1.])\n",
" self.batch_id = 0\n",
"\n",
" def next(self, batch_size):\n",
" \"\"\" Return a batch of data. When dataset end is reached, start over.\n",
" \"\"\"\n",
" if self.batch_id == len(self.data):\n",
" self.batch_id = 0\n",
" batch_data = (self.data[self.batch_id:min(self.batch_id +\n",
" batch_size, len(self.data))])\n",
" batch_labels = (self.labels[self.batch_id:min(self.batch_id +\n",
" batch_size, len(self.data))])\n",
" batch_seqlen = (self.seqlen[self.batch_id:min(self.batch_id +\n",
" batch_size, len(self.data))])\n",
" self.batch_id = min(self.batch_id + batch_size, len(self.data))\n",
" return batch_data, batch_labels, batch_seqlen"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"# ==========\n",
"# MODEL\n",
"# ==========\n",
"\n",
"# Parameters\n",
"learning_rate = 0.01\n",
"training_steps = 10000\n",
"batch_size = 128\n",
"display_step = 200\n",
"\n",
"# Network Parameters\n",
"seq_max_len = 20 # Sequence max length\n",
"n_hidden = 64 # hidden layer num of features\n",
"n_classes = 2 # linear sequence or not\n",
"\n",
"trainset = ToySequenceData(n_samples=1000, max_seq_len=seq_max_len)\n",
"testset = ToySequenceData(n_samples=500, max_seq_len=seq_max_len)\n",
"\n",
"# tf Graph input\n",
"x = tf.placeholder(\"float\", [None, seq_max_len, 1])\n",
"y = tf.placeholder(\"float\", [None, n_classes])\n",
"# A placeholder for indicating each sequence length\n",
"seqlen = tf.placeholder(tf.int32, [None])\n",
"\n",
"# Define weights\n",
"weights = {\n",
" 'out': tf.Variable(tf.random_normal([n_hidden, n_classes]))\n",
"}\n",
"biases = {\n",
" 'out': tf.Variable(tf.random_normal([n_classes]))\n",
"}"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"def dynamicRNN(x, seqlen, weights, biases):\n",
"\n",
" # Prepare data shape to match `rnn` function requirements\n",
" # Current data input shape: (batch_size, n_steps, n_input)\n",
" # Required shape: 'n_steps' tensors list of shape (batch_size, n_input)\n",
" \n",
" # Unstack to get a list of 'n_steps' tensors of shape (batch_size, n_input)\n",
" x = tf.unstack(x, seq_max_len, 1)\n",
"\n",
" # Define a lstm cell with tensorflow\n",
" lstm_cell = tf.contrib.rnn.BasicLSTMCell(n_hidden)\n",
"\n",
" # Get lstm cell output, providing 'sequence_length' will perform dynamic\n",
" # calculation.\n",
" outputs, states = tf.contrib.rnn.static_rnn(lstm_cell, x, dtype=tf.float32,\n",
" sequence_length=seqlen)\n",
"\n",
" # When performing dynamic calculation, we must retrieve the last\n",
" # dynamically computed output, i.e., if a sequence length is 10, we need\n",
" # to retrieve the 10th output.\n",
" # However TensorFlow doesn't support advanced indexing yet, so we build\n",
" # a custom op that for each sample in batch size, get its length and\n",
" # get the corresponding relevant output.\n",
"\n",
" # 'outputs' is a list of output at every timestep, we pack them in a Tensor\n",
" # and change back dimension to [batch_size, n_step, n_input]\n",
" outputs = tf.stack(outputs)\n",
" outputs = tf.transpose(outputs, [1, 0, 2])\n",
"\n",
" # Hack to build the indexing and retrieve the right output.\n",
" batch_size = tf.shape(outputs)[0]\n",
" # Start indices for each sample\n",
" index = tf.range(0, batch_size) * seq_max_len + (seqlen - 1)\n",
" # Indexing\n",
" outputs = tf.gather(tf.reshape(outputs, [-1, n_hidden]), index)\n",
"\n",
" # Linear activation, using outputs computed above\n",
" return tf.matmul(outputs, weights['out']) + biases['out']"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {
"collapsed": false
},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"/Users/aymeric.damien/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/gradients_impl.py:93: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.\n",
" \"Converting sparse IndexedSlices to a dense Tensor of unknown shape. \"\n"
]
}
],
"source": [
"pred = dynamicRNN(x, seqlen, weights, biases)\n",
"\n",
"# Define loss and optimizer\n",
"cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))\n",
"optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cost)\n",
"\n",
"# Evaluate model\n",
"correct_pred = tf.equal(tf.argmax(pred,1), tf.argmax(y,1))\n",
"accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n",
"\n",
"# Initialize the variables (i.e. assign their default value)\n",
"init = tf.global_variables_initializer()"
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {
"collapsed": false,
"scrolled": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Step 1, Minibatch Loss= 0.864517, Training Accuracy= 0.42188\n",
"Step 200, Minibatch Loss= 0.686012, Training Accuracy= 0.43269\n",
"Step 400, Minibatch Loss= 0.682970, Training Accuracy= 0.48077\n",
"Step 600, Minibatch Loss= 0.679640, Training Accuracy= 0.50962\n",
"Step 800, Minibatch Loss= 0.675208, Training Accuracy= 0.53846\n",
"Step 1000, Minibatch Loss= 0.668636, Training Accuracy= 0.56731\n",
"Step 1200, Minibatch Loss= 0.657525, Training Accuracy= 0.62500\n",
"Step 1400, Minibatch Loss= 0.635423, Training Accuracy= 0.67308\n",
"Step 1600, Minibatch Loss= 0.580433, Training Accuracy= 0.75962\n",
"Step 1800, Minibatch Loss= 0.475599, Training Accuracy= 0.81731\n",
"Step 2000, Minibatch Loss= 0.434865, Training Accuracy= 0.83654\n",
"Step 2200, Minibatch Loss= 0.423690, Training Accuracy= 0.85577\n",
"Step 2400, Minibatch Loss= 0.417472, Training Accuracy= 0.85577\n",
"Step 2600, Minibatch Loss= 0.412906, Training Accuracy= 0.85577\n",
"Step 2800, Minibatch Loss= 0.409193, Training Accuracy= 0.85577\n",
"Step 3000, Minibatch Loss= 0.406035, Training Accuracy= 0.86538\n",
"Step 3200, Minibatch Loss= 0.403287, Training Accuracy= 0.87500\n",
"Step 3400, Minibatch Loss= 0.400862, Training Accuracy= 0.87500\n",
"Step 3600, Minibatch Loss= 0.398704, Training Accuracy= 0.86538\n",
"Step 3800, Minibatch Loss= 0.396768, Training Accuracy= 0.86538\n",
"Step 4000, Minibatch Loss= 0.395017, Training Accuracy= 0.86538\n",
"Step 4200, Minibatch Loss= 0.393422, Training Accuracy= 0.86538\n",
"Step 4400, Minibatch Loss= 0.391957, Training Accuracy= 0.85577\n",
"Step 4600, Minibatch Loss= 0.390600, Training Accuracy= 0.85577\n",
"Step 4800, Minibatch Loss= 0.389334, Training Accuracy= 0.86538\n",
"Step 5000, Minibatch Loss= 0.388143, Training Accuracy= 0.86538\n",
"Step 5200, Minibatch Loss= 0.387015, Training Accuracy= 0.86538\n",
"Step 5400, Minibatch Loss= 0.385940, Training Accuracy= 0.86538\n",
"Step 5600, Minibatch Loss= 0.384907, Training Accuracy= 0.86538\n",
"Step 5800, Minibatch Loss= 0.383904, Training Accuracy= 0.85577\n",
"Step 6000, Minibatch Loss= 0.382921, Training Accuracy= 0.86538\n",
"Step 6200, Minibatch Loss= 0.381941, Training Accuracy= 0.86538\n",
"Step 6400, Minibatch Loss= 0.380947, Training Accuracy= 0.86538\n",
"Step 6600, Minibatch Loss= 0.379912, Training Accuracy= 0.86538\n",
"Step 6800, Minibatch Loss= 0.378796, Training Accuracy= 0.86538\n",
"Step 7000, Minibatch Loss= 0.377540, Training Accuracy= 0.86538\n",
"Step 7200, Minibatch Loss= 0.376041, Training Accuracy= 0.86538\n",
"Step 7400, Minibatch Loss= 0.374130, Training Accuracy= 0.85577\n",
"Step 7600, Minibatch Loss= 0.371514, Training Accuracy= 0.85577\n",
"Step 7800, Minibatch Loss= 0.367723, Training Accuracy= 0.85577\n",
"Step 8000, Minibatch Loss= 0.362049, Training Accuracy= 0.85577\n",
"Step 8200, Minibatch Loss= 0.353558, Training Accuracy= 0.85577\n",
"Step 8400, Minibatch Loss= 0.341072, Training Accuracy= 0.86538\n",
"Step 8600, Minibatch Loss= 0.323062, Training Accuracy= 0.87500\n",
"Step 8800, Minibatch Loss= 0.299278, Training Accuracy= 0.89423\n",
"Step 9000, Minibatch Loss= 0.273857, Training Accuracy= 0.90385\n",
"Step 9200, Minibatch Loss= 0.248392, Training Accuracy= 0.91346\n",
"Step 9400, Minibatch Loss= 0.221348, Training Accuracy= 0.92308\n",
"Step 9600, Minibatch Loss= 0.191947, Training Accuracy= 0.92308\n",
"Step 9800, Minibatch Loss= 0.159308, Training Accuracy= 0.93269\n",
"Step 10000, Minibatch Loss= 0.136938, Training Accuracy= 0.96154\n",
"Optimization Finished!\n",
"Testing Accuracy: 0.952\n"
]
}
],
"source": [
"# Start training\n",
"with tf.Session() as sess:\n",
"\n",
" # Run the initializer\n",
" sess.run(init)\n",
"\n",
" for step in range(1, training_steps+1):\n",
" batch_x, batch_y, batch_seqlen = trainset.next(batch_size)\n",
" # Run optimization op (backprop)\n",
" sess.run(optimizer, feed_dict={x: batch_x, y: batch_y,\n",
" seqlen: batch_seqlen})\n",
" if step % display_step == 0 or step == 1:\n",
" # Calculate batch accuracy & loss\n",
" acc, loss = sess.run([accuracy, cost], feed_dict={x: batch_x, y: batch_y,\n",
" seqlen: batch_seqlen})\n",
" print(\"Step \" + str(step) + \", Minibatch Loss= \" + \\\n",
" \"{:.6f}\".format(loss) + \", Training Accuracy= \" + \\\n",
" \"{:.5f}\".format(acc))\n",
"\n",
" print(\"Optimization Finished!\")\n",
"\n",
" # Calculate accuracy\n",
" test_data = testset.data\n",
" test_label = testset.labels\n",
" test_seqlen = testset.seqlen\n",
" print(\"Testing Accuracy:\", \\\n",
" sess.run(accuracy, feed_dict={x: test_data, y: test_label,\n",
" seqlen: test_seqlen}))"
]
}
],
"metadata": {
"anaconda-cloud": {},
"kernelspec": {
"display_name": "Python [default]",
"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": "2.7.12"
}
},
"nbformat": 4,
"nbformat_minor": 1
}