-
Notifications
You must be signed in to change notification settings - Fork 907
/
mnist_3.1_convolutional_bigger_dropout.py
159 lines (133 loc) · 8.4 KB
/
mnist_3.1_convolutional_bigger_dropout.py
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
# encoding: UTF-8
# Copyright 2016 Google.com
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import tensorflow as tf
import tensorflowvisu
import math
import mnistdata
print("Tensorflow version " + tf.__version__)
tf.set_random_seed(0)
# Download images and labels into mnist.test (10K images+labels) and mnist.train (60K images+labels)
mnist = mnistdata.read_data_sets("data", one_hot=True, reshape=False)
# neural network structure for this sample:
#
# · · · · · · · · · · (input data, 1-deep) X [batch, 28, 28, 1]
# @ @ @ @ @ @ @ @ @ @ -- conv. layer 6x6x1=>6 stride 1 W1 [5, 5, 1, 6] B1 [6]
# ∶∶∶∶∶∶∶∶∶∶∶∶∶∶∶∶∶∶∶ Y1 [batch, 28, 28, 6]
# @ @ @ @ @ @ @ @ -- conv. layer 5x5x6=>12 stride 2 W2 [5, 5, 6, 12] B2 [12]
# ∶∶∶∶∶∶∶∶∶∶∶∶∶∶∶ Y2 [batch, 14, 14, 12]
# @ @ @ @ @ @ -- conv. layer 4x4x12=>24 stride 2 W3 [4, 4, 12, 24] B3 [24]
# ∶∶∶∶∶∶∶∶∶∶∶ Y3 [batch, 7, 7, 24] => reshaped to YY [batch, 7*7*24]
# \x/x\x\x/ ✞ -- fully connected layer (relu+dropout) W4 [7*7*24, 200] B4 [200]
# · · · · Y4 [batch, 200]
# \x/x\x/ -- fully connected layer (softmax) W5 [200, 10] B5 [10]
# · · · Y [batch, 10]
# input X: 28x28 grayscale images, the first dimension (None) will index the images in the mini-batch
X = tf.placeholder(tf.float32, [None, 28, 28, 1])
# correct answers will go here
Y_ = tf.placeholder(tf.float32, [None, 10])
# variable learning rate
lr = tf.placeholder(tf.float32)
# Probability of keeping a node during dropout = 1.0 at test time (no dropout) and 0.75 at training time
pkeep = tf.placeholder(tf.float32)
# step for variable learning rate
step = tf.placeholder(tf.int32)
# three convolutional layers with their channel counts, and a
# fully connected layer (the last layer has 10 softmax neurons)
K = 6 # first convolutional layer output depth
L = 12 # second convolutional layer output depth
M = 24 # third convolutional layer
N = 200 # fully connected layer
W1 = tf.Variable(tf.truncated_normal([6, 6, 1, K], stddev=0.1)) # 6x6 patch, 1 input channel, K output channels
B1 = tf.Variable(tf.constant(0.1, tf.float32, [K]))
W2 = tf.Variable(tf.truncated_normal([5, 5, K, L], stddev=0.1))
B2 = tf.Variable(tf.constant(0.1, tf.float32, [L]))
W3 = tf.Variable(tf.truncated_normal([4, 4, L, M], stddev=0.1))
B3 = tf.Variable(tf.constant(0.1, tf.float32, [M]))
W4 = tf.Variable(tf.truncated_normal([7 * 7 * M, N], stddev=0.1))
B4 = tf.Variable(tf.constant(0.1, tf.float32, [N]))
W5 = tf.Variable(tf.truncated_normal([N, 10], stddev=0.1))
B5 = tf.Variable(tf.constant(0.1, tf.float32, [10]))
# The model
stride = 1 # output is 28x28
Y1 = tf.nn.relu(tf.nn.conv2d(X, W1, strides=[1, stride, stride, 1], padding='SAME') + B1)
stride = 2 # output is 14x14
Y2 = tf.nn.relu(tf.nn.conv2d(Y1, W2, strides=[1, stride, stride, 1], padding='SAME') + B2)
stride = 2 # output is 7x7
Y3 = tf.nn.relu(tf.nn.conv2d(Y2, W3, strides=[1, stride, stride, 1], padding='SAME') + B3)
# reshape the output from the third convolution for the fully connected layer
YY = tf.reshape(Y3, shape=[-1, 7 * 7 * M])
Y4 = tf.nn.relu(tf.matmul(YY, W4) + B4)
YY4 = tf.nn.dropout(Y4, pkeep)
Ylogits = tf.matmul(YY4, W5) + B5
Y = tf.nn.softmax(Ylogits)
# cross-entropy loss function (= -sum(Y_i * log(Yi)) ), normalised for batches of 100 images
# TensorFlow provides the softmax_cross_entropy_with_logits function to avoid numerical stability
# problems with log(0) which is NaN
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=Ylogits, labels=Y_)
cross_entropy = tf.reduce_mean(cross_entropy)*100
# accuracy of the trained model, between 0 (worst) and 1 (best)
correct_prediction = tf.equal(tf.argmax(Y, 1), tf.argmax(Y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# matplotlib visualisation
allweights = tf.concat([tf.reshape(W1, [-1]), tf.reshape(W2, [-1]), tf.reshape(W3, [-1]), tf.reshape(W4, [-1]), tf.reshape(W5, [-1])], 0)
allbiases = tf.concat([tf.reshape(B1, [-1]), tf.reshape(B2, [-1]), tf.reshape(B3, [-1]), tf.reshape(B4, [-1]), tf.reshape(B5, [-1])], 0)
I = tensorflowvisu.tf_format_mnist_images(X, Y, Y_)
It = tensorflowvisu.tf_format_mnist_images(X, Y, Y_, 1000, lines=25)
datavis = tensorflowvisu.MnistDataVis()
# training step, the learning rate is a placeholder
# the learning rate is: # 0.0001 + 0.003 * (1/e)^(step/2000)), i.e. exponential decay from 0.003->0.0001
lr = 0.0001 + tf.train.exponential_decay(0.003, step, 2000, 1/math.e)
train_step = tf.train.AdamOptimizer(lr).minimize(cross_entropy)
# init
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
# You can call this function in a loop to train the model, 100 images at a time
def training_step(i, update_test_data, update_train_data):
# training on batches of 100 images with 100 labels
batch_X, batch_Y = mnist.train.next_batch(100)
# compute training values for visualisation
if update_train_data:
a, c, im, w, b, l = sess.run([accuracy, cross_entropy, I, allweights, allbiases, lr],
feed_dict={X: batch_X, Y_: batch_Y, pkeep: 1.0, step: i})
print(str(i) + ": accuracy:" + str(a) + " loss: " + str(c) + " (lr:" + str(l) + ")")
datavis.append_training_curves_data(i, a, c)
datavis.update_image1(im)
datavis.append_data_histograms(i, w, b)
# compute test values for visualisation
if update_test_data:
a, c, im = sess.run([accuracy, cross_entropy, It], {X: mnist.test.images, Y_: mnist.test.labels, pkeep: 1.0})
print(str(i) + ": ********* epoch " + str(i*100//mnist.train.images.shape[0]+1) + " ********* test accuracy:" + str(a) + " test loss: " + str(c))
datavis.append_test_curves_data(i, a, c)
datavis.update_image2(im)
# the backpropagation training step
sess.run(train_step, {X: batch_X, Y_: batch_Y, step: i, pkeep: 0.75})
datavis.animate(training_step, 10001, train_data_update_freq=20, test_data_update_freq=100)
# to save the animation as a movie, add save_movie=True as an argument to datavis.animate
# to disable the visualisation use the following line instead of the datavis.animate line
# for i in range(10000+1): training_step(i, i % 100 == 0, i % 20 == 0)
print("max test accuracy: " + str(datavis.get_max_test_accuracy()))
## All runs 10K iterations:
# layers 4 8 12 200, patches 5x5str1 5x5str2 4x4str2 best 0.989
# layers 4 8 12 200, patches 5x5str1 4x4str2 4x4str2 best 0.9892
# layers 6 12 24 200, patches 5x5str1 4x4str2 4x4str2 best 0.9908 after 10000 iterations but going downhill from 5000 on
# layers 6 12 24 200, patches 5x5str1 4x4str2 4x4str2 dropout=0.75 best 0.9922 (but above 0.99 after 1400 iterations only)
# layers 4 8 12 200, patches 5x5str1 4x4str2 4x4str2 dropout=0.75, best 0.9914 at 13700 iterations
# layers 9 16 25 200, patches 5x5str1 4x4str2 4x4str2 dropout=0.75, best 0.9918 at 10500 (but 0.99 at 1500 iterations already, 0.9915 at 5800)
# layers 9 16 25 300, patches 5x5str1 4x4str2 4x4str2 dropout=0.75, best 0.9916 at 5500 iterations (but 0.9903 at 1200 iterations already)
# attempts with 2 fully-connected layers: no better 300 and 100 neurons, dropout 0.75 and 0.5, 6x6 5x5 4x4 patches no better
# layers 6 12 24 200, patches 6x6str1 5x5str2 4x4str2 no dropout best 0.9906 after 3100 iterations (avove 0.99 from iteration 1400)
#*layers 6 12 24 200, patches 6x6str1 5x5str2 4x4str2 dropout=0.75 best 0.9928 after 12800 iterations (but consistently above 0.99 after 1300 iterations only, 0.9916 at 2300 iterations, 0.9921 at 5600, 0.9925 at 20000)
#*same with dacaying learning rate 0.003-0.0001-2000: best 0.9931 (on other runs max accuracy 0.9921, 0.9927, 0.9935, 0.9929, 0.9933)