forked from martin-gorner/tensorflow-mnist-tutorial
-
Notifications
You must be signed in to change notification settings - Fork 0
/
mnist_2.1_five_layers_relu_lrdecay.py
150 lines (125 loc) · 7.12 KB
/
mnist_2.1_five_layers_relu_lrdecay.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
# 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
from tensorflow.contrib.learn.python.learn.datasets.mnist import read_data_sets
tf.set_random_seed(0)
# neural network with 5 layers
#
# · · · · · · · · · · (input data, flattened pixels) X [batch, 784] # 784 = 28*28
# \x/x\x/x\x/x\x/x\x/ -- fully connected layer (relu) W1 [784, 200] B1[200]
# · · · · · · · · · Y1 [batch, 200]
# \x/x\x/x\x/x\x/ -- fully connected layer (relu) W2 [200, 100] B2[100]
# · · · · · · · Y2 [batch, 100]
# \x/x\x/x\x/ -- fully connected layer (relu) W3 [100, 60] B3[60]
# · · · · · Y3 [batch, 60]
# \x/x\x/ -- fully connected layer (relu) W4 [60, 30] B4[30]
# · · · Y4 [batch, 30]
# \x/ -- fully connected layer (softmax) W5 [30, 10] B5[10]
# · Y5 [batch, 10]
# Download images and labels into mnist.test (10K images+labels) and mnist.train (60K images+labels)
mnist = read_data_sets("data", one_hot=True, reshape=False, validation_size=0)
# 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)
# five layers and their number of neurons (tha last layer has 10 softmax neurons)
L = 200
M = 100
N = 60
O = 30
# Weights initialised with small random values between -0.2 and +0.2
# When using RELUs, make sure biases are initialised with small *positive* values for example 0.1 = tf.ones([K])/10
W1 = tf.Variable(tf.truncated_normal([784, L], stddev=0.1)) # 784 = 28 * 28
B1 = tf.Variable(tf.ones([L])/10)
W2 = tf.Variable(tf.truncated_normal([L, M], stddev=0.1))
B2 = tf.Variable(tf.ones([M])/10)
W3 = tf.Variable(tf.truncated_normal([M, N], stddev=0.1))
B3 = tf.Variable(tf.ones([N])/10)
W4 = tf.Variable(tf.truncated_normal([N, O], stddev=0.1))
B4 = tf.Variable(tf.ones([O])/10)
W5 = tf.Variable(tf.truncated_normal([O, 10], stddev=0.1))
B5 = tf.Variable(tf.zeros([10]))
# The model
XX = tf.reshape(X, [-1, 784])
Y1 = tf.nn.relu(tf.matmul(XX, W1) + B1)
Y2 = tf.nn.relu(tf.matmul(Y1, W2) + B2)
Y3 = tf.nn.relu(tf.matmul(Y2, W3) + B3)
Y4 = tf.nn.relu(tf.matmul(Y3, W4) + B4)
Ylogits = tf.matmul(Y4, 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
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)
# learning rate decay
max_learning_rate = 0.003
min_learning_rate = 0.0001
decay_speed = 2000.0 # 0.003-0.0001-2000=>0.9826 done in 5000 iterations
learning_rate = min_learning_rate + (max_learning_rate - min_learning_rate) * math.exp(-i/decay_speed)
# compute training values for visualisation
if update_train_data:
a, c, im, w, b = sess.run([accuracy, cross_entropy, I, allweights, allbiases], {X: batch_X, Y_: batch_Y})
print(str(i) + ": accuracy:" + str(a) + " loss: " + str(c) + " (lr:" + str(learning_rate) + ")")
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})
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, lr: learning_rate})
datavis.animate(training_step, iterations=10000+1, train_data_update_freq=20, test_data_update_freq=100, more_tests_at_start=True)
# 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()))
# Some results to expect:
# (In all runs, if sigmoids are used, all biases are initialised at 0, if RELUs are used,
# all biases are initialised at 0.1 apart from the last one which is initialised at 0.)
## learning rate = 0.003, 10K iterations
# final test accuracy = 0.9788 (sigmoid - slow start, training cross-entropy not stabilised in the end)
# final test accuracy = 0.9825 (relu - above 0.97 in the first 1500 iterations but noisy curves)
## now with learning rate = 0.0001, 10K iterations
# final test accuracy = 0.9722 (relu - slow but smooth curve, would have gone higher in 20K iterations)
## decaying learning rate from 0.003 to 0.0001 decay_speed 2000, 10K iterations
# final test accuracy = 0.9746 (sigmoid - training cross-entropy not stabilised)
# final test accuracy = 0.9824 (relu - training set fully learned, test accuracy stable)