-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbasic_nn.r
218 lines (163 loc) · 7.89 KB
/
basic_nn.r
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
## DEPDENCY #######################################################################################
source("https://raw.githubusercontent.com/oltkkol/vmod/master/basic_text.r", encoding="UTF-8")
source("https://raw.githubusercontent.com/oltkkol/vmod/master/basic_ml.r", encoding="UTF-8")
plotMetric <- function(a, b, env, ...){
if (is.null(env$train.acc.log)){
env$train.acc.log <- c()
env$test.acc.log <- c()
}
env$train.acc.log <- append(env$train.acc.log, env$metric$get(env$train.metric)$value)
env$test.acc.log <- append(env$test.acc.log, env$metric$get(env$eval.metric)$value)
layout(matrix(c(1,1,1,1,2,2,2,2), nrow=2, byrow=T))
plot(env$train.acc.log, type="o", col="blue") # :-/
points(env$test.acc.log, type="o", col="red")
hist(as.matrix( env$model$arg.params[[1]] ), breaks=30)
return (TRUE)
}
GetSoftmaxResult <- function(softmaxOutput){
return ( max.col( t(softmaxOutput) ) -1 )
}
PrepareDatasetsForNeuralNetwork <- function(datasets){
datasets$Train$X <- data.matrix(datasets$Train$X)
datasets$Test$X <- data.matrix(datasets$Test$X)
datasets$Train$Y <- as.numeric(datasets$Train$Y) - 1
datasets$Test$Y <- as.numeric(datasets$Test$Y) - 1
return (datasets)
}
stop_quietly();
####################################################################################################
## EXAMPLE 1: Sonar
####################################################################################################
# 1. Prepare data (see "PrepareDatasetsForNeuralNetwork" converting all to numeric and to data.matrix)
data(Sonar)
datasets <- PrepareTrainAndTest(Sonar, "Class", 2/3, scaleBy="z-score")
datasets <- PrepareDatasetsForNeuralNetwork(datasets)
# 2. Prepare NN Model: 20 -> 20 -> 20 -> 20 -> 2 Neural network with RELU activations
myNeuralNetwork <-(
mx.symbol.Variable("data") %>%
mx.symbol.FullyConnected(num_hidden = 20) %>%
mx.symbol.Activation(act_type = "relu") %>%
mx.symbol.FullyConnected(num_hidden = 20) %>%
mx.symbol.Activation(act_type = "relu") %>%
mx.symbol.FullyConnected(num_hidden = 20) %>%
mx.symbol.Activation(act_type = "relu") %>%
mx.symbol.FullyConnected(num_hidden = 20) %>%
mx.symbol.Activation(act_type = "relu") %>%
mx.symbol.FullyConnected(num_hidden = 2) %>%
mx.symbol.SoftmaxOutput()
)
# 3. Create & Train:
model <- mx.model.FeedForward.create( myNeuralNetwork,
X = datasets$Train$X,
y = datasets$Train$Y,
eval.data=list(data=datasets$Test$X, label=datasets$Test$Y),
optimizer="adam",
ctx=mx.cpu(),
num.round=200,
array.batch.size=10,
learning.rate=0.0001,
wd=0.000183,
eval.metric=mx.metric.accuracy,
epoch.end.callback=plotMetric )
# 4. Display our network
graph.viz(model$symbol)
# 5. Evaluate
EvaluateModelAndPlot(model, datasets$Train, datasets$Test, GetSoftmaxResult)
# 6. Benchmark with LDA, SVM & naiveBayes
datasets <- PrepareTrainAndTest(Sonar, "Class", 2/3, scaleBy="z-score")
#datasets <- PrepareTrainAndTest(Sonar, "Class", 2/3, scaleBy="minmax")
model <- lda(datasets$Train$X, datasets$Train$Y)
EvaluateModelAndPlot(model, datasets$Train, datasets$Test)
model <- svm(datasets$Train$X, datasets$Train$Y, kernel="linear")
EvaluateModelAndPlot(model, datasets$Train, datasets$Test)
model <- naiveBayes(datasets$Train$X, datasets$Train$Y)
EvaluateModelAndPlot(model, datasets$Train, datasets$Test)
####################################################################################################
## EXAMPLE 2: Easy authorship with BOW
####################################################################################################
## [1] Prepare BOW
# 1. Read files & Tokenize them (+ limit to random X words)
authorA_Files <- GetFilesContentsFromFolder("F:/VMOD/DATASETY/Asimov", "ASIMOV")
authorB_Files <- GetFilesContentsFromFolder("F:/VMOD/DATASETY/Foglar", "FOGLAR")
authorA_Tokens <- TokenizeTexts(authorA_Files)
authorB_Tokens <- TokenizeTexts(authorB_Files)
numberOfTokens <- 50
authorA_Tokens <- LimitTokensInTexts(authorA_Tokens, count=numberOfTokens, takeRandom=TRUE)
authorB_Tokens <- LimitTokensInTexts(authorB_Tokens, count=numberOfTokens, takeRandom=TRUE)
# 2. Bag of Words: per texts
allBOW <- MakeBOWModel( append( authorA_Tokens, authorB_Tokens) )
# 3. Bag Of Words: per author
authorA_Corpora <- MergeTokenizedTexts(authorA_Tokens)
authorB_Corpora <- MergeTokenizedTexts(authorB_Tokens)
bowAuthorAVsB <- MakeBOWModel( list(AuthorA = authorA_Corpora, AuthorB = authorB_Corpora) )
# - calculate per author TF-IDF to identify author specific words:
weights <- CalculateTFIDFOnBOW(bowAuthorAVsB, omitZeroWeightTerms=TRUE)
newAllBOW <- KeepOnlyGivenColumns( allBOW, names(weights) )
# 5. Prepare datasets & Train
newAllBOW.Target <- FirstColNameWordsToColumn(newAllBOW, "AuthorTarget")
## [2] Prepare datasets
datasets <- PrepareTrainAndTest(newAllBOW.Target, "AuthorTarget", 2/3, scaleBy="binarize", convertToFactors=FALSE)
datasets <- PrepareDatasetsForNeuralNetwork(datasets)
## [3] Prepare & Train Neural Network:
myNN <-(
mx.symbol.Variable("data") %>%
mx.symbol.FullyConnected(num_hidden=100) %>%
mx.symbol.Activation(act_type="relu") %>%
mx.symbol.FullyConnected(num_hidden=10) %>%
mx.symbol.Activation(act_type="relu") %>%
mx.symbol.FullyConnected(num_hidden=2) %>%
mx.symbol.SoftmaxOutput()
)
mx.set.seed(0)
model <- mx.model.FeedForward.create(myNN,
X = datasets$Train$X,
y = datasets$Train$Y,
eval.data=list(data=datasets$Test$X, label=datasets$Test$Y),
optimizer="adam",
ctx=mx.cpu(),
num.round=50,
array.batch.size=2,
learning.rate=0.00001,
wd=0.0001,
eval.metric=mx.metric.accuracy,
epoch.end.callback=plotMetric)
EvaluateModelAndPlot(model, datasets$Train, datasets$Test, GetSoftmaxResult)
####################################################################################################
## EXAMPLE 3: Sentiment BOW
####################################################################################################
goodTokens <- TokenizeTexts( GetFilesContentsFromFolder("C:/DATA/NLP/Sentiment/GOOD", "GOOD") ) # 1000 files with positive reviews
badTokens <- TokenizeTexts( GetFilesContentsFromFolder("C:/DATA/NLP/Sentiment/BAD", "BAD") ) # 1000 files with negative reviews
goodCorpora <- MergeTokenizedTexts(goodTokens)
badCorpora <- MergeTokenizedTexts(badTokens)
# BOW per Texts & per sentiment, apply TFIDF on sentiment
allBOW <- MakeBOWModel( append( goodTokens, badTokens) )
bowGoodVsBad <- MakeBOWModel( list(Goods = goodCorpora, Bads = badCorpora) )
weights <- CalculateTFIDFOnBOW(bowGoodVsBad, omitZeroWeightTerms=TRUE)
newAllBOW <- KeepOnlyGivenColumns(allBOW, names(weights))
# Prepare dataset & Train & Eval
newAllBOW.Target <- FirstColNameWordsToColumn(newAllBOW, "Sentiment")
datasets <- PrepareTrainAndTest(newAllBOW.Target, "Sentiment", 2/3, scaleBy="z-score")
datasets <- PrepareDatasetsForNeuralNetwork(datasets)
net <-( mx.symbol.Variable( "data" ) %>%
mx.symbol.Dropout( p=0.3 ) %>%
mx.symbol.FullyConnected( num_hidden=30 ) %>%
mx.symbol.Dropout( p=0.1 ) %>%
mx.symbol.Activation( act_type="relu" ) %>%
mx.symbol.FullyConnected( num_hidden=5 ) %>%
mx.symbol.Activation( act_type="relu" ) %>%
mx.symbol.FullyConnected( num_hidden=2 ) %>%
mx.symbol.SoftmaxOutput()
)
model <- mx.model.FeedForward.create(net,
X = datasets$Train$X,
y = datasets$Train$Y,
eval.data = list(data=datasets$Test$X, label=datasets$Test$Y),
optimizer = "adam",
ctx = mx.cpu(),
num.round = 100,
array.batch.size= 250,
learning.rate = 0.001,
wd = 0.001,
eval.metric = mx.metric.accuracy,
epoch.end.callback = plotMetric)
EvaluateModelAndPlot(model, datasets$Train, datasets$Test, GetSoftmaxResult)