Wing IDE编译TesorFlow中Mnist convolutional 实例
1 # 2 # http://www.cnblogs.com/mydebug/ 3 # 4 from __future__ import absolute_import 5 from __future__ import division 6 from __future__ import print_function 7 8 import gzip 9 import os 10 import sys 11 sys.path.append("这里是numpy的路径")//提示：如果没有这句import numpy会报错 12 import tensorflow.python.platform 13 14 import numpy 15 from six.moves import urllib 16 from six.moves import xrange # pylint: disable=redefined-builtin 17 import tensorflow as tf 18 19 SOURCE_URL = 'http://yann.lecun.com/exdb/mnist/' 20 WORK_DIRECTORY = '/TensorFlow/data' 21 IMAGE_SIZE = 28 22 NUM_CHANNELS = 1 23 PIXEL_DEPTH = 255 24 NUM_LABELS = 10 25 VALIDATION_SIZE = 5000 # Size of the validation set. 26 SEED = 66478 # Set to None for random seed. 27 BATCH_SIZE = 64 28 NUM_EPOCHS = 10 29 30 31 tf.app.flags.DEFINE_boolean("self_test", False, "True if running a self test.") 32 FLAGS = tf.app.flags.FLAGS 33 34 35 def maybe_download(filename): 36 """Download the data from Yann's website, unless it's already here.""" 37 if not os.path.exists(WORK_DIRECTORY): 38 os.mkdir(WORK_DIRECTORY) 39 filepath = os.path.join(WORK_DIRECTORY, filename) 40 if not os.path.exists(filepath): 41 filepath, _ = urllib.request.urlretrieve(SOURCE_URL + filename, filepath) 42 statinfo = os.stat(filepath) 43 print('Succesfully downloaded', filename, statinfo.st_size, 'bytes.') 44 return filepath 45 46 47 def extract_data(filename, num_images): 48 """Extract the images into a 4D tensor [image index, y, x, channels]. 49 50 Values are rescaled from [0, 255] down to [-0.5, 0.5]. 51 """ 52 print('Extracting', filename) 53 with gzip.open(filename) as bytestream: 54 bytestream.read(16) 55 buf = bytestream.read(IMAGE_SIZE * IMAGE_SIZE * num_images) 56 data = numpy.frombuffer(buf, dtype=numpy.uint8).astype(numpy.float32) 57 data = (data - (PIXEL_DEPTH / 2.0)) / PIXEL_DEPTH 58 data = data.reshape(num_images, IMAGE_SIZE, IMAGE_SIZE, 1) 59 return data 60 61 62 def extract_labels(filename, num_images): 63 """Extract the labels into a 1-hot matrix [image index, label index].""" 64 print('Extracting', filename) 65 with gzip.open(filename) as bytestream: 66 bytestream.read(8) 67 buf = bytestream.read(1 * num_images) 68 labels = numpy.frombuffer(buf, dtype=numpy.uint8) 69 # Convert to dense 1-hot representation. 70 return (numpy.arange(NUM_LABELS) == labels[:, None]).astype(numpy.float32) 71 72 73 def fake_data(num_images): 74 """Generate a fake dataset that matches the dimensions of MNIST.""" 75 data = numpy.ndarray( 76 shape=(num_images, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS), 77 dtype=numpy.float32) 78 labels = numpy.zeros(shape=(num_images, NUM_LABELS), dtype=numpy.float32) 79 for image in xrange(num_images): 80 label = image % 2 81 data[image, :, :, 0] = label - 0.5 82 labels[image, label] = 1.0 83 return data, labels 84 85 86 def error_rate(predictions, labels): 87 """Return the error rate based on dense predictions and 1-hot labels.""" 88 return 100.0 - ( 89 100.0 * 90 numpy.sum(numpy.argmax(predictions, 1) == numpy.argmax(labels, 1)) / 91 predictions.shape[0]) 92 93 94 def main(argv=None): # pylint: disable=unused-argument 95 if FLAGS.self_test: 96 print('Running self-test.') 97 train_data, train_labels = fake_data(256) 98 validation_data, validation_labels = fake_data(16) 99 test_data, test_labels = fake_data(256) 100 num_epochs = 1 101 else: 102 # Get the data. 103 train_data_filename = maybe_download('train-images-idx3-ubyte.gz') 104 train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz') 105 test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz') 106 test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz') 107 108 # Extract it into numpy arrays. 109 train_data = extract_data(train_data_filename, 60000) 110 train_labels = extract_labels(train_labels_filename, 60000) 111 test_data = extract_data(test_data_filename, 10000) 112 test_labels = extract_labels(test_labels_filename, 10000) 113 114 # Generate a validation set. 115 validation_data = train_data[:VALIDATION_SIZE, :, :, :] 116 validation_labels = train_labels[:VALIDATION_SIZE] 117 train_data = train_data[VALIDATION_SIZE:, :, :, :] 118 train_labels = train_labels[VALIDATION_SIZE:] 119 num_epochs = NUM_EPOCHS 120 train_size = train_labels.shape[0] 121 122 # This is where training samples and labels are fed to the graph. 123 # These placeholder nodes will be fed a batch of training data at each 124 # training step using the {feed_dict} argument to the Run() call below. 125 train_data_node = tf.placeholder( 126 tf.float32, 127 shape=(BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS)) 128 train_labels_node = tf.placeholder(tf.float32, 129 shape=(BATCH_SIZE, NUM_LABELS)) 130 # For the validation and test data, we'll just hold the entire dataset in 131 # one constant node. 132 validation_data_node = tf.constant(validation_data) 133 test_data_node = tf.constant(test_data) 134 135 # The variables below hold all the trainable weights. They are passed an 136 # initial value which will be assigned when when we call: 137 # {tf.initialize_all_variables().run()} 138 conv1_weights = tf.Variable( 139 tf.truncated_normal([5, 5, NUM_CHANNELS, 32], # 5x5 filter, depth 32. 140 stddev=0.1, 141 seed=SEED)) 142 conv1_biases = tf.Variable(tf.zeros([32])) 143 conv2_weights = tf.Variable( 144 tf.truncated_normal([5, 5, 32, 64], 145 stddev=0.1, 146 seed=SEED)) 147 conv2_biases = tf.Variable(tf.constant(0.1, shape=[64])) 148 fc1_weights = tf.Variable( # fully connected, depth 512. 149 tf.truncated_normal( 150 [IMAGE_SIZE // 4 * IMAGE_SIZE // 4 * 64, 512], 151 stddev=0.1, 152 seed=SEED)) 153 fc1_biases = tf.Variable(tf.constant(0.1, shape=[512])) 154 fc2_weights = tf.Variable( 155 tf.truncated_normal([512, NUM_LABELS], 156 stddev=0.1, 157 seed=SEED)) 158 fc2_biases = tf.Variable(tf.constant(0.1, shape=[NUM_LABELS])) 159 160 # We will replicate the model structure for the training subgraph, as well 161 # as the evaluation subgraphs, while sharing the trainable parameters. 162 def model(data, train=False): 163 """The Model definition.""" 164 # 2D convolution, with 'SAME' padding (i.e. the output feature map has 165 # the same size as the input). Note that {strides} is a 4D array whose 166 # shape matches the data layout: [image index, y, x, depth]. 167 conv = tf.nn.conv2d(data, 168 conv1_weights, 169 strides=[1, 1, 1, 1], 170 padding='SAME') 171 # Bias and rectified linear non-linearity. 172 relu = tf.nn.relu(tf.nn.bias_add(conv, conv1_biases)) 173 # Max pooling. The kernel size spec {ksize} also follows the layout of 174 # the data. Here we have a pooling window of 2, and a stride of 2. 175 pool = tf.nn.max_pool(relu, 176 ksize=[1, 2, 2, 1], 177 strides=[1, 2, 2, 1], 178 padding='SAME') 179 conv = tf.nn.conv2d(pool, 180 conv2_weights, 181 strides=[1, 1, 1, 1], 182 padding='SAME') 183 relu = tf.nn.relu(tf.nn.bias_add(conv, conv2_biases)) 184 pool = tf.nn.max_pool(relu, 185 ksize=[1, 2, 2, 1], 186 strides=[1, 2, 2, 1], 187 padding='SAME') 188 # Reshape the feature map cuboid into a 2D matrix to feed it to the 189 # fully connected layers. 190 pool_shape = pool.get_shape().as_list() 191 reshape = tf.reshape( 192 pool, 193 [pool_shape[0], pool_shape[1] * pool_shape[2] * pool_shape[3]]) 194 # Fully connected layer. Note that the '+' operation automatically 195 # broadcasts the biases. 196 hidden = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases) 197 # Add a 50% dropout during training only. Dropout also scales 198 # activations such that no rescaling is needed at evaluation time. 199 if train: 200 hidden = tf.nn.dropout(hidden, 0.5, seed=SEED) 201 return tf.matmul(hidden, fc2_weights) + fc2_biases 202 203 # Training computation: logits + cross-entropy loss. 204 logits = model(train_data_node, True) 205 loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits( 206 logits, train_labels_node)) 207 208 # L2 regularization for the fully connected parameters. 209 regularizers = (tf.nn.l2_loss(fc1_weights) + tf.nn.l2_loss(fc1_biases) + 210 tf.nn.l2_loss(fc2_weights) + tf.nn.l2_loss(fc2_biases)) 211 # Add the regularization term to the loss. 212 loss += 5e-4 * regularizers 213 214 # Optimizer: set up a variable that's incremented once per batch and 215 # controls the learning rate decay. 216 batch = tf.Variable(0) 217 # Decay once per epoch, using an exponential schedule starting at 0.01. 218 learning_rate = tf.train.exponential_decay( 219 0.01, # Base learning rate. 220 batch * BATCH_SIZE, # Current index into the dataset. 221 train_size, # Decay step. 222 0.95, # Decay rate. 223 staircase=True) 224 # Use simple momentum for the optimization. 225 optimizer = tf.train.MomentumOptimizer(learning_rate, 226 0.9).minimize(loss, 227 global_step=batch) 228 229 # Predictions for the minibatch, validation set and test set. 230 train_prediction = tf.nn.softmax(logits) 231 # We'll compute them only once in a while by calling their {eval()} method. 232 validation_prediction = tf.nn.softmax(model(validation_data_node)) 233 test_prediction = tf.nn.softmax(model(test_data_node)) 234 235 # Create a local session to run this computation. 236 with tf.Session() as s: 237 # Run all the initializers to prepare the trainable parameters. 238 tf.initialize_all_variables().run() 239 print('Initialized!') 240 # Loop through training steps. 241 for step in xrange(num_epochs * train_size // BATCH_SIZE): 242 # Compute the offset of the current minibatch in the data. 243 # Note that we could use better randomization across epochs. 244 offset = (step * BATCH_SIZE) % (train_size - BATCH_SIZE) 245 batch_data = train_data[offset:(offset + BATCH_SIZE), :, :, :] 246 batch_labels = train_labels[offset:(offset + BATCH_SIZE)] 247 # This dictionary maps the batch data (as a numpy array) to the 248 # node in the graph is should be fed to. 249 feed_dict = {train_data_node: batch_data, 250 train_labels_node: batch_labels} 251 # Run the graph and fetch some of the nodes. 252 _, l, lr, predictions = s.run( 253 [optimizer, loss, learning_rate, train_prediction], 254 feed_dict=feed_dict) 255 if step % 100 == 0: 256 print('Epoch %.2f' % (float(step) * BATCH_SIZE / train_size)) 257 print('Minibatch loss: %.3f, learning rate: %.6f' % (l, lr)) 258 print('Minibatch error: %.1f%%' % error_rate(predictions, batch_labels)) 259 print('Validation error: %.1f%%' % 260 error_rate(validation_prediction.eval(), validation_labels)) 261 sys.stdout.flush() 262 # Finally print the result! 263 test_error = error_rate(test_prediction.eval(), test_labels) 264 print('Test error: %.1f%%' % test_error) 265 if FLAGS.self_test: 266 print('test_error', test_error) 267 assert test_error == 0.0, 'expected 0.0 test_error, got %.2f' % ( 268 test_error,) 269 270 271 if __name__ == '__main__': 272 tf.app.run()
在Ming IDE中新建python文件，复制以上代码，点击Debug，运行成功。
注：注意源代码中的提示部分。
