1 # Copyright 2016 The TensorFlow Authors. All Rights Reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 """Convolutional Neural Network Estimator for MNIST, built with tf.layers.""" 15 16 from __future__ import absolute_import 17 from __future__ import division 18 from __future__ import print_function 19 20 import numpy as np 21 import tensorflow as tf 22 23 tf.logging.set_verbosity(tf.logging.INFO) 24 25 26 def cnn_model_fn(features, labels, mode): 27 """Model function for CNN.""" 28 # Input Layer 29 # Reshape X to 4-D tensor: [batch_size, width, height, channels] 30 # MNIST images are 28x28 pixels, and have one color channel 31 input_layer = tf.reshape(features["x"], [-1, 28, 28, 1]) 32 33 # Convolutional Layer #1 34 # Computes 32 features using a 5x5 filter with ReLU activation. 35 # Padding is added to preserve width and height. 36 # Input Tensor Shape: [batch_size, 28, 28, 1] 37 # Output Tensor Shape: [batch_size, 28, 28, 32] 38 conv1 = tf.layers.conv2d( 39 inputs=input_layer, 40 filters=32, 41 kernel_size=[5, 5], 42 padding="same", 43 activation=tf.nn.relu) 44 45 # Pooling Layer #1 46 # First max pooling layer with a 2x2 filter and stride of 2 47 # Input Tensor Shape: [batch_size, 28, 28, 32] 48 # Output Tensor Shape: [batch_size, 14, 14, 32] 49 pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2) 50 51 # Convolutional Layer #2 52 # Computes 64 features using a 5x5 filter. 53 # Padding is added to preserve width and height. 54 # Input Tensor Shape: [batch_size, 14, 14, 32] 55 # Output Tensor Shape: [batch_size, 14, 14, 64] 56 conv2 = tf.layers.conv2d( 57 inputs=pool1, 58 filters=64, 59 kernel_size=[5, 5], 60 padding="same", 61 activation=tf.nn.relu) 62 63 # Pooling Layer #2 64 # Second max pooling layer with a 2x2 filter and stride of 2 65 # Input Tensor Shape: [batch_size, 14, 14, 64] 66 # Output Tensor Shape: [batch_size, 7, 7, 64] 67 pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2) 68 69 # Flatten tensor into a batch of vectors 70 # Input Tensor Shape: [batch_size, 7, 7, 64] 71 # Output Tensor Shape: [batch_size, 7 * 7 * 64] 72 pool2_flat = tf.reshape(pool2, [-1, 7 * 7 * 64]) 73 74 # Dense Layer 75 # Densely connected layer with 1024 neurons 76 # Input Tensor Shape: [batch_size, 7 * 7 * 64] 77 # Output Tensor Shape: [batch_size, 1024] 78 dense = tf.layers.dense(inputs=pool2_flat, units=1024, activation=tf.nn.relu) 79 80 # Add dropout operation; 0.6 probability that element will be kept 81 dropout = tf.layers.dropout( 82 inputs=dense, rate=0.4, training=mode == tf.estimator.ModeKeys.TRAIN) 83 84 # Logits layer 85 # Input Tensor Shape: [batch_size, 1024] 86 # Output Tensor Shape: [batch_size, 10] 87 logits = tf.layers.dense(inputs=dropout, units=10) 88 89 predictions = { 90 # Generate predictions (for PREDICT and EVAL mode) 91 "classes": tf.argmax(input=logits, axis=1), 92 # Add `softmax_tensor` to the graph. It is used for PREDICT and by the 93 # `logging_hook`. 94 "probabilities": tf.nn.softmax(logits, name="softmax_tensor") 95 } 96 if mode == tf.estimator.ModeKeys.PREDICT: 97 return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions) 98 99 # Calculate Loss (for both TRAIN and EVAL modes) 100 loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits) 101 102 # Configure the Training Op (for TRAIN mode) 103 if mode == tf.estimator.ModeKeys.TRAIN: 104 optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001) 105 train_op = optimizer.minimize( 106 loss=loss, 107 global_step=tf.train.get_global_step()) 108 return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op) 109 110 # Add evaluation metrics (for EVAL mode) 111 eval_metric_ops = { 112 "accuracy": tf.metrics.accuracy( 113 labels=labels, predictions=predictions["classes"])} 114 return tf.estimator.EstimatorSpec( 115 mode=mode, loss=loss, eval_metric_ops=eval_metric_ops) 116 117 118 def main(unused_argv): 119 # Load training and eval data 120 mnist = tf.contrib.learn.datasets.load_dataset("mnist") 121 train_data = mnist.train.images # Returns np.array 122 train_labels = np.asarray(mnist.train.labels, dtype=np.int32) 123 eval_data = mnist.test.images # Returns np.array 124 eval_labels = np.asarray(mnist.test.labels, dtype=np.int32) 125 126 # Create the Estimator 127 mnist_classifier = tf.estimator.Estimator( 128 model_fn=cnn_model_fn, model_dir="/tmp/mnist_convnet_model") 129 130 # Set up logging for predictions 131 # Log the values in the "Softmax" tensor with label "probabilities" 132 tensors_to_log = {"probabilities": "softmax_tensor"} 133 logging_hook = tf.train.LoggingTensorHook( 134 tensors=tensors_to_log, every_n_iter=50) 135 136 # Train the model 137 train_input_fn = tf.estimator.inputs.numpy_input_fn( 138 x={"x": train_data}, 139 y=train_labels, 140 batch_size=100, 141 num_epochs=None, 142 shuffle=True) 143 mnist_classifier.train( 144 input_fn=train_input_fn, 145 steps=20000, 146 hooks=[logging_hook]) 147 148 # Evaluate the model and print results 149 eval_input_fn = tf.estimator.inputs.numpy_input_fn( 150 x={"x": eval_data}, 151 y=eval_labels, 152 num_epochs=1, 153 shuffle=False) 154 eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn) 155 print(eval_results) 156 157 158 if __name__ == "__main__": 159 tf.app.run() 160
