{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "Use shift-enter to execute a code block and move to the next one." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": true }, "outputs": [], "source": [ "# 2.1 Import libraries.\n", "import math\n", "import os\n", "from six.moves import xrange # pylint: disable=redefined-builtin\n", "import numpy as np\n", "import tensorflow as tf\n", "from tensorflow.contrib.learn.python.learn.datasets.mnist import read_data_sets\n", "\n", "%matplotlib inline\n", "import matplotlib.pyplot as plt" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "# 2.2 Define some constants.\n", "# The MNIST dataset has 10 classes, representing the digits 0 through 9.\n", "NUM_CLASSES = 10\n", "\n", "# The MNIST images are always 28x28 pixels.\n", "IMAGE_SIZE = 28\n", "IMAGE_PIXELS = IMAGE_SIZE * IMAGE_SIZE\n", "\n", "# Batch size. Must be evenly dividable by dataset sizes.\n", "BATCH_SIZE = 100\n", "EVAL_BATCH_SIZE = 1\n", "\n", "# Number of units in hidden layers.\n", "HIDDEN1_UNITS = 128\n", "HIDDEN2_UNITS = 32\n", "\n", "# Maximum number of training steps.\n", "MAX_STEPS = 2000\n", "\n", "# Directory to put the training data.\n", "TRAIN_DIR=\"/tmp/mnist\"" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "# 2.3 Get input data: get the sets of images and labels for training, validation, and\n", "# test on MNIST.\n", "data_sets = read_data_sets(TRAIN_DIR, False)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": true }, "outputs": [], "source": [ "# 2.4 Build inference graph.\n", "def mnist_inference(images, hidden1_units, hidden2_units):\n", " \"\"\"Build the MNIST model up to where it may be used for inference.\n", " Args:\n", " images: Images placeholder.\n", " hidden1_units: Size of the first hidden layer.\n", " hidden2_units: Size of the second hidden layer.\n", " Returns:\n", " logits: Output tensor with the computed logits.\n", " \"\"\"\n", " # Hidden 1\n", " with tf.name_scope('hidden1'):\n", " weights = tf.Variable(\n", " tf.truncated_normal([IMAGE_PIXELS, hidden1_units],\n", " stddev=1.0 / math.sqrt(float(IMAGE_PIXELS))),\n", " name='weights')\n", " biases = tf.Variable(tf.zeros([hidden1_units]),\n", " name='biases')\n", " hidden1 = tf.nn.relu(tf.matmul(images, weights) + biases)\n", " # Hidden 2\n", " with tf.name_scope('hidden2'):\n", " weights = tf.Variable(\n", " tf.truncated_normal([hidden1_units, hidden2_units],\n", " stddev=1.0 / math.sqrt(float(hidden1_units))),\n", " name='weights')\n", " biases = tf.Variable(tf.zeros([hidden2_units]),\n", " name='biases')\n", " hidden2 = tf.nn.relu(tf.matmul(hidden1, weights) + biases)\n", " # Linear\n", " with tf.name_scope('softmax_linear'):\n", " weights = tf.Variable(\n", " tf.truncated_normal([hidden2_units, NUM_CLASSES],\n", " stddev=1.0 / math.sqrt(float(hidden2_units))),\n", " name='weights')\n", " biases = tf.Variable(tf.zeros([NUM_CLASSES]),\n", " name='biases')\n", " logits = tf.matmul(hidden2, weights) + biases\n", "\n", " # Uncomment the following line to see what we have constructed.\n", " # tf.train.write_graph(tf.get_default_graph().as_graph_def(),\n", " # \"/tmp\", \"inference.pbtxt\", as_text=True)\n", " return logits" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": true }, "outputs": [], "source": [ "# 2.5 Build training graph.\n", "def mnist_training(logits, labels, learning_rate):\n", " \"\"\"Build the training graph.\n", "\n", " Args:\n", " logits: Logits tensor, float - [BATCH_SIZE, NUM_CLASSES].\n", " labels: Labels tensor, int32 - [BATCH_SIZE], with values in the\n", " range [0, NUM_CLASSES).\n", " learning_rate: The learning rate to use for gradient descent.\n", " Returns:\n", " train_op: The Op for training.\n", " loss: The Op for calculating loss.\n", " \"\"\"\n", " # Create an operation that calculates loss.\n", " labels = tf.to_int64(labels)\n", " cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(\n", " logits, labels, name='xentropy')\n", " loss = tf.reduce_mean(cross_entropy, name='xentropy_mean')\n", " # Create the gradient descent optimizer with the given learning rate.\n", " optimizer = tf.train.GradientDescentOptimizer(learning_rate)\n", " # Create a variable to track the global step.\n", " global_step = tf.Variable(0, name='global_step', trainable=False)\n", " # Use the optimizer to apply the gradients that minimize the loss\n", " # (and also increment the global step counter) as a single training step.\n", " train_op = optimizer.minimize(loss, global_step=global_step)\n", "\n", " # Uncomment the following line to see what we have constructed.\n", " # tf.train.write_graph(tf.get_default_graph().as_graph_def(),\n", " # \"/tmp\", \"train.pbtxt\", as_text=True)\n", "\n", " return train_op, loss" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "# 2.6 Build the complete graph for feeding inputs, training, and saving checkpoints.\n", "mnist_graph = tf.Graph()\n", "with mnist_graph.as_default():\n", " # Generate placeholders for the images and labels.\n", " images_placeholder = tf.placeholder(tf.float32) \n", " labels_placeholder = tf.placeholder(tf.int32)\n", " tf.add_to_collection(\"images\", images_placeholder) # Remember this Op.\n", " tf.add_to_collection(\"labels\", labels_placeholder) # Remember this Op.\n", "\n", " # Build a Graph that computes predictions from the inference model.\n", " logits = mnist_inference(images_placeholder,\n", " HIDDEN1_UNITS,\n", " HIDDEN2_UNITS)\n", " tf.add_to_collection(\"logits\", logits) # Remember this Op.\n", "\n", " # Add to the Graph the Ops that calculate and apply gradients.\n", " train_op, loss = mnist_training(logits, labels_placeholder, 0.01)\n", "\n", " # Add the variable initializer Op.\n", " init = tf.initialize_all_variables()\n", "\n", " # Create a saver for writing training checkpoints.\n", " saver = tf.train.Saver()\n", " \n", " # Uncomment the following line to see what we have constructed.\n", " # tf.train.write_graph(tf.get_default_graph().as_graph_def(),\n", " # \"/tmp\", \"complete.pbtxt\", as_text=True)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "# 2.7 Run training for MAX_STEPS and save checkpoint at the end.\n", "with tf.Session(graph=mnist_graph) as sess:\n", " # Run the Op to initialize the variables.\n", " sess.run(init)\n", "\n", " # Start the training loop.\n", " for step in xrange(MAX_STEPS):\n", " # Read a batch of images and labels.\n", " images_feed, labels_feed = data_sets.train.next_batch(BATCH_SIZE)\n", "\n", " # Run one step of the model. The return values are the activations\n", " # from the `train_op` (which is discarded) and the `loss` Op. To\n", " # inspect the values of your Ops or variables, you may include them\n", " # in the list passed to sess.run() and the value tensors will be\n", " # returned in the tuple from the call.\n", " _, loss_value = sess.run([train_op, loss],\n", " feed_dict={images_placeholder: images_feed,\n", " labels_placeholder: labels_feed})\n", "\n", " # Print out loss value.\n", " if step % 1000 == 0:\n", " print('Step %d: loss = %.2f' % (step, loss_value))\n", "\n", " # Write a checkpoint.\n", " checkpoint_file = os.path.join(TRAIN_DIR, 'checkpoint')\n", " saver.save(sess, checkpoint_file, global_step=step)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "# 2.8 Run evaluation based on the saved checkpoint.\n", "with tf.Session(graph=tf.Graph()) as sess:\n", " saver = tf.train.import_meta_graph(\n", " os.path.join(TRAIN_DIR, \"checkpoint-1999.meta\"))\n", " saver.restore(\n", " sess, os.path.join(TRAIN_DIR, \"checkpoint-1999\"))\n", "\n", " # Retrieve the Ops we 'remembered'.\n", " logits = tf.get_collection(\"logits\")[0]\n", " images_placeholder = tf.get_collection(\"images\")[0]\n", " labels_placeholder = tf.get_collection(\"labels\")[0]\n", " \n", " # Add an Op that chooses the top k predictions.\n", " eval_op = tf.nn.top_k(logits)\n", " \n", " # Run evaluation.\n", " images_feed, labels_feed = data_sets.validation.next_batch(EVAL_BATCH_SIZE)\n", " imgplot = plt.imshow(np.reshape(images_feed, (28, 28)))\n", " prediction = sess.run(eval_op,\n", " feed_dict={images_placeholder: images_feed,\n", " labels_placeholder: labels_feed})\n", " print(\"Ground truth: %d\\nPrediction: %d\" % (labels_feed, prediction.indices[0][0]))" ] }, { "cell_type": "markdown", "metadata": { "collapsed": true }, "source": [ "

Back to [1_linear_regression_model.ipynb](1_linear_regression_model.ipynb).

\n", "

Next to [Bonus Lab](extras/extras_0_deepdream.ipynb).

" ] } ], "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": "2.7.12" } }, "nbformat": 4, "nbformat_minor": 1 }