1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
|
{
"worksheets": [
{
"cells": [
{
"metadata": {
"id": "4embtkV0pNxM",
"colab_type": "text"
},
"cell_type": "markdown",
"source": "Deep Learning\n=============\n\nAssignment 4\n------------\n\nPreviously in `2_fullyconnected.ipynb` and `3_regularization.ipynb`, we trained fully connected networks to classify [notMNIST](http://yaroslavvb.blogspot.com/2011/09/notmnist-dataset.html) characters.\n\nThe goal of this assignment is make the neural network convolutional."
},
{
"metadata": {
"id": "tm2CQN_Cpwj0",
"colab_type": "code",
"colab": {
"autoexec": {
"startup": false,
"wait_interval": 0
}
},
"cellView": "both"
},
"cell_type": "code",
"input": "# These are all the modules we'll be using later. Make sure you can import them\n# before proceeding further.\nimport cPickle as pickle\nimport numpy as np\nimport tensorflow as tf",
"language": "python",
"outputs": []
},
{
"metadata": {
"id": "y3-cj1bpmuxc",
"colab_type": "code",
"colab": {
"autoexec": {
"startup": false,
"wait_interval": 0
},
"output_extras": [
{
"item_id": 1
}
]
},
"cellView": "both",
"executionInfo": {
"elapsed": 11948,
"status": "ok",
"timestamp": 1446658914837,
"user": {
"color": "",
"displayName": "",
"isAnonymous": false,
"isMe": true,
"permissionId": "",
"photoUrl": "",
"sessionId": "0",
"userId": ""
},
"user_tz": 480
},
"outputId": "016b1a51-0290-4b08-efdb-8c95ffc3cd01"
},
"cell_type": "code",
"input": "pickle_file = 'notMNIST.pickle'\n\nwith open(pickle_file, 'rb') as f:\n save = pickle.load(f)\n train_dataset = save['train_dataset']\n train_labels = save['train_labels']\n valid_dataset = save['valid_dataset']\n valid_labels = save['valid_labels']\n test_dataset = save['test_dataset']\n test_labels = save['test_labels']\n del save # hint to help gc free up memory\n print 'Training set', train_dataset.shape, train_labels.shape\n print 'Validation set', valid_dataset.shape, valid_labels.shape\n print 'Test set', test_dataset.shape, test_labels.shape",
"language": "python",
"outputs": [
{
"output_type": "stream",
"stream": "stdout",
"text": "Training set (200000, 28, 28) (200000,)\nValidation set (10000, 28, 28) (10000,)\nTest set (18724, 28, 28) (18724,)\n"
}
]
},
{
"metadata": {
"id": "L7aHrm6nGDMB",
"colab_type": "text"
},
"cell_type": "markdown",
"source": "Reformat into a TensorFlow-friendly shape:\n- convolutions need the image data formatted as a cube (width by height by #channels)\n- labels as float 1-hot encodings."
},
{
"metadata": {
"id": "IRSyYiIIGIzS",
"colab_type": "code",
"colab": {
"autoexec": {
"startup": false,
"wait_interval": 0
},
"output_extras": [
{
"item_id": 1
}
]
},
"cellView": "both",
"executionInfo": {
"elapsed": 11952,
"status": "ok",
"timestamp": 1446658914857,
"user": {
"color": "",
"displayName": "",
"isAnonymous": false,
"isMe": true,
"permissionId": "",
"photoUrl": "",
"sessionId": "0",
"userId": ""
},
"user_tz": 480
},
"outputId": "650a208c-8359-4852-f4f5-8bf10e80ef6c"
},
"cell_type": "code",
"input": "image_size = 28\nnum_labels = 10\nnum_channels = 1 # grayscale\n\nimport numpy as np\n\ndef reformat(dataset, labels):\n dataset = dataset.reshape(\n (-1, image_size, image_size, num_channels)).astype(np.float32)\n labels = (np.arange(num_labels) == labels[:,None]).astype(np.float32)\n return dataset, labels\ntrain_dataset, train_labels = reformat(train_dataset, train_labels)\nvalid_dataset, valid_labels = reformat(valid_dataset, valid_labels)\ntest_dataset, test_labels = reformat(test_dataset, test_labels)\nprint 'Training set', train_dataset.shape, train_labels.shape\nprint 'Validation set', valid_dataset.shape, valid_labels.shape\nprint 'Test set', test_dataset.shape, test_labels.shape",
"language": "python",
"outputs": [
{
"output_type": "stream",
"stream": "stdout",
"text": "Training set (200000, 28, 28, 1) (200000, 10)\nValidation set (10000, 28, 28, 1) (10000, 10)\nTest set (18724, 28, 28, 1) (18724, 10)\n"
}
]
},
{
"metadata": {
"id": "AgQDIREv02p1",
"colab_type": "code",
"colab": {
"autoexec": {
"startup": false,
"wait_interval": 0
}
},
"cellView": "both"
},
"cell_type": "code",
"input": "def accuracy(predictions, labels):\n return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))\n / predictions.shape[0])",
"language": "python",
"outputs": []
},
{
"metadata": {
"id": "5rhgjmROXu2O",
"colab_type": "text"
},
"cell_type": "markdown",
"source": "Let's build a small network with two convolutional layers, followed by one fully connected layer. Convolutional networks are more expensive computationally, so we'll limit its depth and number of fully connected nodes."
},
{
"metadata": {
"id": "IZYv70SvvOan",
"colab_type": "code",
"colab": {
"autoexec": {
"startup": false,
"wait_interval": 0
}
},
"cellView": "both"
},
"cell_type": "code",
"input": "batch_size = 16\npatch_size = 5\ndepth = 16\nnum_hidden = 64\n\ngraph = tf.Graph()\n\nwith graph.as_default():\n\n # Input data.\n tf_train_dataset = tf.placeholder(\n tf.float32, shape=(batch_size, image_size, image_size, num_channels))\n tf_train_labels = tf.placeholder(tf.float32, shape=(batch_size, num_labels))\n tf_valid_dataset = tf.constant(valid_dataset)\n tf_test_dataset = tf.constant(test_dataset)\n \n # Variables.\n layer1_weights = tf.Variable(tf.truncated_normal(\n [patch_size, patch_size, num_channels, depth], stddev=0.1))\n layer1_biases = tf.Variable(tf.zeros([depth]))\n layer2_weights = tf.Variable(tf.truncated_normal(\n [patch_size, patch_size, depth, depth], stddev=0.1))\n layer2_biases = tf.Variable(tf.constant(1.0, shape=[depth]))\n layer3_weights = tf.Variable(tf.truncated_normal(\n [image_size / 4 * image_size / 4 * depth, num_hidden], stddev=0.1))\n layer3_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))\n layer4_weights = tf.Variable(tf.truncated_normal(\n [num_hidden, num_labels], stddev=0.1))\n layer4_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))\n \n # Model.\n def model(data):\n conv = tf.nn.conv2d(data, layer1_weights, [1, 2, 2, 1], padding='SAME')\n hidden = tf.nn.relu(conv + layer1_biases)\n conv = tf.nn.conv2d(hidden, layer2_weights, [1, 2, 2, 1], padding='SAME')\n hidden = tf.nn.relu(conv + layer2_biases)\n shape = hidden.get_shape().as_list()\n reshape = tf.reshape(hidden, [shape[0], shape[1] * shape[2] * shape[3]])\n hidden = tf.nn.relu(tf.matmul(reshape, layer3_weights) + layer3_biases)\n return tf.matmul(hidden, layer4_weights) + layer4_biases\n \n # Training computation.\n logits = model(tf_train_dataset)\n loss = tf.reduce_mean(\n tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))\n \n # Optimizer.\n optimizer = tf.train.GradientDescentOptimizer(0.05).minimize(loss)\n \n # Predictions for the training, validation, and test data.\n train_prediction = tf.nn.softmax(logits)\n valid_prediction = tf.nn.softmax(model(tf_valid_dataset))\n test_prediction = tf.nn.softmax(model(tf_test_dataset))",
"language": "python",
"outputs": []
},
{
"metadata": {
"id": "noKFb2UovVFR",
"colab_type": "code",
"colab": {
"autoexec": {
"startup": false,
"wait_interval": 0
},
"output_extras": [
{
"item_id": 37
}
]
},
"cellView": "both",
"executionInfo": {
"elapsed": 63292,
"status": "ok",
"timestamp": 1446658966251,
"user": {
"color": "",
"displayName": "",
"isAnonymous": false,
"isMe": true,
"permissionId": "",
"photoUrl": "",
"sessionId": "0",
"userId": ""
},
"user_tz": 480
},
"outputId": "28941338-2ef9-4088-8bd1-44295661e628"
},
"cell_type": "code",
"input": "num_steps = 1001\n\nwith tf.Session(graph=graph) as session:\n tf.initialize_all_variables().run()\n print \"Initialized\"\n for step in xrange(num_steps):\n offset = (step * batch_size) % (train_labels.shape[0] - batch_size)\n batch_data = train_dataset[offset:(offset + batch_size), :, :, :]\n batch_labels = train_labels[offset:(offset + batch_size), :]\n feed_dict = {tf_train_dataset : batch_data, tf_train_labels : batch_labels}\n _, l, predictions = session.run(\n [optimizer, loss, train_prediction], feed_dict=feed_dict)\n if (step % 50 == 0):\n print \"Minibatch loss at step\", step, \":\", l\n print \"Minibatch accuracy: %.1f%%\" % accuracy(predictions, batch_labels)\n print \"Validation accuracy: %.1f%%\" % accuracy(\n valid_prediction.eval(), valid_labels)\n print \"Test accuracy: %.1f%%\" % accuracy(test_prediction.eval(), test_labels)",
"language": "python",
"outputs": [
{
"output_type": "stream",
"stream": "stdout",
"text": "Initialized\nMinibatch loss at step 0 : 3.51275\nMinibatch accuracy: 6.2%\nValidation accuracy: 12.8%\nMinibatch loss at step 50 : 1.48703\nMinibatch accuracy: 43.8%\nValidation accuracy: 50.4%\nMinibatch loss at step 100 : 1.04377\nMinibatch accuracy: 68.8%\nValidation accuracy: 67.4%\nMinibatch loss at step 150 : 0.601682\nMinibatch accuracy: 68.8%\nValidation accuracy: 73.0%\nMinibatch loss at step 200 : 0.898649\nMinibatch accuracy: 75.0%\nValidation accuracy: 77.8%\nMinibatch loss at step 250 : 1.3637\nMinibatch accuracy: 56.2%\nValidation accuracy: 75.4%\nMinibatch loss at step 300 : 1.41968\nMinibatch accuracy: 62.5%\nValidation accuracy: 76.0%\nMinibatch loss at step 350 : 0.300648\nMinibatch accuracy: 81.2%\nValidation accuracy: 80.2%\nMinibatch loss at step 400 : 1.32092\nMinibatch accuracy: 56.2%\nValidation accuracy: 80.4%\nMinibatch loss at step 450 : 0.556701\nMinibatch accuracy: 81.2%\nValidation accuracy: 79.4%\nMinibatch loss at step 500 : 1.65595\nMinibatch accuracy: 43.8%\nValidation accuracy: 79.6%\nMinibatch loss at step 550 : 1.06995\nMinibatch accuracy: 75.0%\nValidation accuracy: 81.2%\nMinibatch loss at step 600 : 0.223684\nMinibatch accuracy: 100.0%\nValidation accuracy: 82.3%\nMinibatch loss at step 650 : 0.619602\nMinibatch accuracy: 87.5%\nValidation accuracy: 81.8%\nMinibatch loss at step 700 : 0.812091\nMinibatch accuracy: 75.0%\nValidation accuracy: 82.4%\nMinibatch loss at step 750 : 0.276302\nMinibatch accuracy: 87.5%\nValidation accuracy: 82.3%\nMinibatch loss at step 800 : 0.450241\nMinibatch accuracy: 81.2%\nValidation accuracy: 82.3%\nMinibatch loss at step 850 : 0.137139\nMinibatch accuracy: 93.8%\nValidation accuracy: 82.3%\nMinibatch loss at step 900 : 0.52664\nMinibatch accuracy: 75.0%\nValidation accuracy: 82.2%\nMinibatch loss at step 950 : 0.623835\nMinibatch accuracy: 87.5%\nValidation accuracy: 82.1%\nMinibatch loss at step 1000 : 0.243114\nMinibatch accuracy: 93.8%\nValidation accuracy: 82.9%\nTest accuracy: 90.0%\n"
}
]
},
{
"metadata": {
"id": "KedKkn4EutIK",
"colab_type": "text"
},
"cell_type": "markdown",
"source": "---\nProblem 1\n---------\n\nThe convolutional model above uses convolutions with stride 2 to reduce the dimensionality. Replace the strides a max pooling operation (`nn.max_pool()`) of stride 2 and kernel size 2.\n\n---"
},
{
"metadata": {
"id": "klf21gpbAgb-",
"colab_type": "text"
},
"cell_type": "markdown",
"source": "---\nProblem 2\n---------\n\nTry to get the best performance you can using a convolutional net. Look for example at the classic [LeNet5](http://yann.lecun.com/exdb/lenet/) architecture, adding Dropout, and/or adding learning rate decay.\n\n---"
}
]
}
],
"metadata": {
"name": "4_convolutions.ipynb",
"colabVersion": "0.3.2",
"colab_views": {},
"colab_default_view": {}
},
"nbformat": 3,
"nbformat_minor": 0
}
|
