MNIST-Handwritten-Digit-Recognition-python/launch.py at master · danif93/MNIST-Handwritten-Digit-Recognition-python · GitHub

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

# coding: utf-8

# In[ ]:


import mnist_loader as load
import neural_network as neuronet
import activation_functions as af
import optimisation_functions as of
import cost_functions as cf
import learningStep_generators as lg

import numpy as np
import matplotlib.pyplot as plt
get_ipython().magic('matplotlib inline')


# In[ ]:


# parameters initialization

maxEpochs = 50              # the neural net will not train over maxEpochs
earlyStopParameter = 5      # number of last error entries the net will take in consideration for early stopping
earlyStopThreshold = 0.14   # errors-vector norm threshold for deciding early stop

batchSize = 10              # stochastic gradient descent batch size
regularParam = 3.0          # lambda
hiddenUnit = 20             # neurons number for the single hidden unit layer

decayStep = lg.harmonicSeries(3) # at each epoch the step will shrink over the harmonic series
constStep = lg.constant(3)       # at each epoch the step will remain constant

weightInitErr = 0.3         # if requested, a low delta for weights initialization

tr, va, te = load.loadMnist() # train, validation and test datasets


# In[ ]:


# Simple execution

sigmoidNetCross = [neuronet.NeuralNetwork([784, hiddenUnit, 10], unit=af.Sigmoid, cost=cf.MeanSquareError)]
sgdC = of.StochasticGradientDescent(sigmoidNetCross)
validAccuracy, trainAccuracy, testAccuracy = sgdC.SGD(tr, va, te, maxEpochs, batchSize, decayStep, earlyStopParameter, earlyStopThreshold)


# In[ ]:


# Constant - Decaying step comparison
# considerations about oscillating, saturation, early stopping and learning break

sigmoidNetCross = [neuronet.NeuralNetwork([784, hiddenUnit, 10], unit=af.Sigmoid, cost=cf.CrossEntropy)]
sgdC = of.StochasticGradientDescent(sigmoidNetCross)
validAccuracyC, trainAccuracyC, testAccuracyC = sgdC.SGD(tr, va, te, maxEpochs, batchSize, constStep, earlyStopParameter, earlyStopThreshold)
sigmoidNetCross = [neuronet.NeuralNetwork(net=[784, hiddenUnit, 10], unit=af.Sigmoid, cost=cf.CrossEntropy)]
sgdC = of.StochasticGradientDescent(sigmoidNetCross)
validAccuracyD, trainAccuracyD, testAccuracyD = sgdC.SGD(tr, va, te, maxEpochs, batchSize, decayStep, earlyStopParameter, earlyStopThreshold)

nonZeroConstant = len(np.argwhere(validAccuracyC[0]))
nonZeroDecaying = len(np.argwhere(validAccuracyD[0]))

plt.title('Constant/Decaying learning step')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.ylim([80,100])
plt.plot(np.arange(nonZeroConstant),validAccuracyC[0][:nonZeroConstant], label='Costant')
plt.plot(nonZeroConstant, testAccuracyC[0], 'x', label='Costant - Test')
plt.plot(np.arange(nonZeroDecaying),validAccuracyD[0][:nonZeroDecaying], label='Harmonic Decaying')
plt.plot(nonZeroDecaying, testAccuracyD[0], 'x', label='Harmonic Decaying - Test')
plt.grid()
plt.legend()


# In[ ]:


# Cross Entropy - MSE comparison

sigmoidNetCross = [neuronet.NeuralNetwork([784, hiddenUnit, 10], unit=af.Sigmoid, cost=cf.CrossEntropy, weightError=weightInitErr)]
sigmoidNetMSE = [neuronet.NeuralNetwork(net=[784, hiddenUnit, 10], unit=af.Sigmoid, cost=cf.MeanSquareError, weightError=weightInitErr)]
sgdC = of.StochasticGradientDescent(sigmoidNetCross)
sgdM = of.StochasticGradientDescent(sigmoidNetMSE)

validAccuracyM, trainAccuracyM, testAccuracyM = sgdM.SGD(tr, va, te, maxEpochs, batchSize, decayStep, earlyStopParameter, earlyStopThreshold)
validAccuracyC, trainAccuracyC, testAccuracyC = sgdC.SGD(tr, va, te, maxEpochs, batchSize, decayStep, earlyStopParameter, earlyStopThreshold)

nonZeroCross = len(np.argwhere(validAccuracyC[0]))
nonZeroMSE = len(np.argwhere(validAccuracyM[0]))

plt.title('Cross-Entropy / MSE training result')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.ylim([0,100])
plt.plot(np.arange(nonZeroCross), validAccuracyC[0][:nonZeroCross], label='Cross-Entropy Validation')
plt.plot(nonZeroCross, testAccuracyC[0], 'x', label='Cross-Entropy - Test')
plt.plot(np.arange(nonZeroMSE), validAccuracyM[0][:nonZeroMSE], label='MSE')
plt.plot(nonZeroMSE, testAccuracyM[0], 'x', label='MSE - Test')

plt.grid()
plt.legend()


# In[ ]:


# Training - Validation accuracy comparison

#SGD(trainingSet, validSet, testSet, numEpochs, batchSize, stepGen, earlyStopParam, earlyStopThrshld, _lambda=0.0, l2Regul=True, trainEval=False)
sigmoidNetCross = [neuronet.NeuralNetwork([784, hiddenUnit, 10], unit=af.Sigmoid, cost=cf.CrossEntropy)]
sgdC = of.StochasticGradientDescent(sigmoidNetCross)
validAccuracy, trainAccuracy, testAccuracy = sgdC.SGD(tr, va, te, maxEpochs, batchSize, decayStep, earlyStopParameter, earlyStopThreshold, trainEval=True)

nonZeroValues = len(np.argwhere(validAccuracy[0]))

plt.title('Training/Validation accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.ylim([80,100])
plt.plot(np.arange(nonZeroValues), validAccuracy[0][:nonZeroValues], label='Validation Accuracy')
plt.plot(np.arange(nonZeroValues), trainAccuracy[0][:nonZeroValues], label='Train Accuracy')
plt.plot(nonZeroValues, testAccuracy[0], 'x', label='Test Accuracy')
plt.legend()
plt.grid()


# In[ ]:


# Training - Validation accuracy comparison with l2 regularization

sigmoidNetCross = [neuronet.NeuralNetwork([784, hiddenUnit, 10], unit=af.Sigmoid, cost=cf.CrossEntropy)]
sgdC = of.StochasticGradientDescent(sigmoidNetCross)
validAccuracy, trainAccuracy, testAccuracy = sgdC.SGD(tr, va, te, maxEpochs, batchSize, decayStep, earlyStopParameter, earlyStopThreshold, _lambda=regularParam, trainEval=True)

nonZeroValues = len(np.argwhere(validAccuracy[0]))

plt.title('Training/Validation accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.ylim([80,100])
plt.plot(np.arange(nonZeroValues), validAccuracy[0][:nonZeroValues], label='Validation Accuracy')
plt.plot(np.arange(nonZeroValues), trainAccuracy[0][:nonZeroValues], label='Train Accuracy')
plt.plot(nonZeroValues, testAccuracy[0], 'x', label='Test Accuracy')
plt.legend()
plt.grid()


# In[ ]:


# Training - Test accuracy comparison with l1 regularization

sigmoidNetCross = [neuronet.NeuralNetwork([784, hiddenUnit, 10], unit=af.Sigmoid, cost=cf.CrossEntropy)]
sgdC = of.StochasticGradientDescent(sigmoidNetCross)
validAccuracy, trainAccuracy, testAccuracy = sgdC.SGD(tr, va, te, maxEpochs, batchSize, decayStep, earlyStopParameter, earlyStopThreshold, _lambda=regularParam, l2Regul=False, trainEval=True)

nonZeroValues = len(np.argwhere(validAccuracy[0]))

plt.title('Training/Validation accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.ylim([80,100])
plt.plot(np.arange(nonZeroValues), validAccuracy[0][:nonZeroValues], label='Validation Accuracy')
plt.plot(np.arange(nonZeroValues), trainAccuracy[0][:nonZeroValues], label='Train Accuracy')
plt.plot(nonZeroValues, testAccuracy[0], 'x', label='Test Accuracy')
plt.legend()
plt.grid()


# In[ ]:


# hold-out validation on hidden unit number

nets = [neuronet.NeuralNetwork([784, hiddUnitNumber, 10], unit=af.Sigmoid, cost=cf.CrossEntropy) for hiddUnitNumber in [50,100,150,200]]
sgdC = of.StochasticGradientDescent(nets)
validAccuracy, trainAccuracy, testAccuracy = sgdC.SGD(tr, va, te, maxEpochs, batchSize, decayStep, earlyStopParameter, earlyStopThreshold, _lambda=regularParam)

plt.title('Validation accuracy per hidden unit model')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.ylim([90,100])
for i in np.arange(len(nets)):
    nonZeroValues = len(np.argwhere(validAccuracy[i]))
    plt.plot(np.arange(nonZeroValues),validAccuracy[i][:nonZeroValues], label='{}'.format(nets[i].net[1:-1]))
    plt.plot(nonZeroValues, testAccuracy[i], 'x', label='{} - Test'.format(nets[i].net[1:-1]))
plt.legend()
plt.grid()


# In[ ]:


# AdaGrad algo - not working

sigmoidNetCross = neuronet.NeuralNetwork(net=[784, hiddenUnit, 10], unit=af.Sigmoid, cost=cf.CrossEntropy)
agC = of.AdaGrad(sigmoidNetCross)
testAccuracy, trainAccuracy = agC.AG(tr, va, te, maxEpochs, batchSize, constStep, _lambda=regularParam, trainEval=True)


plt.title('Training/Test accuracy ')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.ylim([80,100])
plt.plot(np.arange(epochs), testAccuracy, label='Test Accuracy')
plt.plot(np.arange(epochs), trainAccuracy, label='Train Accuracy')
plt.legend()
plt.grid()