-
Notifications
You must be signed in to change notification settings - Fork 1
/
updateELMLiPoModel.p
317 lines (261 loc) · 11.2 KB
/
updateELMLiPoModel.p
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
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
import sys
import numpy as np
import pandas as pd
from joblib import dump, load # model persistance librny
# import mysql.connector
from sklearn.decomposition import PCA
import matplotlib.pyplot as plt
########ELM Dependencies########
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from elm import ELMRegressor
from random_hidden_layer import RBFRandomHiddenLayer, SimpleRandomHiddenLayer
########ELM Dependencies########
pd.set_option('display.max_rows', None)
# fallback csv training data if not using kdb data (for testing purposes)
csvTrainingData = 'trainingDataAbove100kph.csv'
# mysql update setup variables
# cannot use __file__ when running in KDB+
fileName = 'updateELMLiPoModel.p'
trainingSetName = 'trainingDataAbove100kph.csv'
comments = 'ELM Model'
def mse(pred, actual):
return ((pred-actual)**2).mean()
def rmse():
mse**0.5
def strFloat(floatVal):
return "{0:.2f}".format(round(floatVal, 2))
kdbSource = True
if 'trainingDataPDF' not in globals():
kdbSource = False
trainingDataPDF = pd.read_csv(csvTrainingData)
print("Training using csv input!")
# using else or try catch causes bugs with embedpy
if kdbSource:
trainingSetName = "KDB+ Input"
print("Training using KDB+ input!")
trainPercentage = 0.7
trainingDataTrain = trainingDataPDF[:int(trainPercentage*len(trainingDataPDF))]
trainingDataTest = trainingDataPDF[int(trainPercentage*len(trainingDataPDF)):]
# trainX <-- training observations [# points, # features]
# trainy <-- training labels [# points]
# testX <-- test observations [# points, # features]
# testy <-- test labels [# points]
trainX = trainingDataTrain.drop(['vbatLatestV'], axis=1, inplace=False)
#####APPLYING NORMALISATION TO DATASET AS REQUIRED BY ELM#####
trainXStandardScalar = StandardScaler()
yStandardScalar = StandardScaler()
trainX = trainXStandardScalar.fit_transform(trainX)
trainy = trainingDataTrain["vbatLatestV"].to_frame()
trainy = yStandardScalar.fit_transform(trainy)
# Index(['timeDeltaus', 'currentSampleHz', 'timeus', 'rcCommand0', 'rcCommand1',
# 'rcCommand2', 'rcCommand3', 'vbatLatestV', 'gyroADC0', 'gyroADC1',
# 'gyroADC2', 'accSmooth0', 'accSmooth1', 'accSmooth2', 'motor0',
# 'motor1', 'motor2', 'motor3'],
# dtype='object')
testX = trainingDataTest.drop(['vbatLatestV'], axis=1, inplace=False)
testXStandardScalar = StandardScaler()
testX = testXStandardScalar.fit_transform(testX)
testy = trainingDataTest["vbatLatestV"].to_frame()
#####APPLYING NORMALISATION TO DATASET AS REQUIRED BY ELM#####
######if using PCA, determine principal components######
usePCA = True
covarianceExplanation = 0.95
if 'usePCA' in globals():
if usePCA:
print("Using PCA!")
if 'pcaModel' not in globals():
if 'covarianceExplanation' not in globals():
covarianceExplanation = 1
pcaModel = PCA(n_components=covarianceExplanation)
pcaModel.fit(trainX)
principalComponents = pcaModel.components_
print("principalComponents:")
print(principalComponents)
trainX = pd.DataFrame(pcaModel.transform(trainX))
testX = pd.DataFrame(pcaModel.transform(testX))
if not 'usePCA' in globals():
print("Not using PCA!")
######if using PCA, determine principal components######
#########test rbf kernel#########
# #########from plot_elm_comparison.p#########
# kernel_names = ["default", "tanh", "tribas", "hardlim", "rbf(0.1)"]
# model_names = list(map(lambda name: name+"LiPoVoltageModel", kernel_names))
# nh = 10
# # pass user defined transfer func
# sinsq = (lambda x: np.power(np.sin(x), 2.0))
# srhl_sinsq = SimpleRandomHiddenLayer(n_hidden=nh,
# activation_func=sinsq,
# random_state=0)
# # use internal transfer funcs
# srhl_tanh = SimpleRandomHiddenLayer(n_hidden=nh,
# activation_func='tanh',
# random_state=0)
# srhl_tribas = SimpleRandomHiddenLayer(n_hidden=nh,
# activation_func='tribas',
# random_state=0)
# srhl_hardlim = SimpleRandomHiddenLayer(n_hidden=nh,
# activation_func='hardlim',
# random_state=0)
# # use gaussian RBF
# srhl_rbf = RBFRandomHiddenLayer(n_hidden=nh*2, gamma=0.1, random_state=0)
# log_reg = LogisticRegression(solver='liblinear')
# #ELMRegressor(tanh, regressor = log_reg) and ELMRegressor(srhl_sinsq) transfer functions are not compatible
# #########from plot_elm_comparison.p#########
# print("Testing ELM RBF Kernel")
# #test model
# model = load('elmLiPoVoltageModel.joblib')
# # print(testX.columns)
# y_pred= model.predict(testX)
# #calculate mean square error
# MSE = mse(y_pred,testy)
# RMSE = MSE**0.5
# #display mean square error
# # print("Actual vs Predictions:")
# # testy = testy.to_numpy()
# # for i in range(len(y_pred)):
# # print(strFloat(testy[i]) + " || " + strFloat(y_pred[i]))
# print("MSE:")
# print(strFloat(MSE))
# print("RMSE:")
# print(strFloat(RMSE))
#########test rbf kernel#########
#########train all elm models#########
# iterate through different numbers of hidden layers
allTrainedELMModels = dict()
maxHiddenLayers = 20
kernel_names = ["default", "tanh", "tribas", "hardlim", "rbf(0.1)"]
model_names = list(map(lambda name: name+"LiPoVoltageModel", kernel_names))
for numHiddenLayers in range(1, 1 + maxHiddenLayers):
print("Number of hidden layers: " + str(numHiddenLayers))
#########from plot_elm_comparison.p#########
# pass user defined transfer func
sinsq = (lambda x: np.power(np.sin(x), 2.0))
srhl_sinsq = SimpleRandomHiddenLayer(n_hidden=numHiddenLayers,
activation_func=sinsq,
random_state=0)
# use internal transfer funcs
srhl_tanh = SimpleRandomHiddenLayer(n_hidden=numHiddenLayers,
activation_func='tanh',
random_state=0)
srhl_tribas = SimpleRandomHiddenLayer(n_hidden=numHiddenLayers,
activation_func='tribas',
random_state=0)
srhl_hardlim = SimpleRandomHiddenLayer(n_hidden=numHiddenLayers,
activation_func='hardlim',
random_state=0)
# use gaussian RBF
srhl_rbf = RBFRandomHiddenLayer(
n_hidden=numHiddenLayers*2, gamma=0.1, random_state=0)
log_reg = LogisticRegression(solver='liblinear')
# ELMRegressor(tanh, regressor = log_reg) and ELMRegressor(srhl_sinsq) transfer functions are not compatible
#########from plot_elm_comparison.p#########
regressors = [ELMRegressor(), ELMRegressor(srhl_tanh), ELMRegressor(
srhl_tribas), ELMRegressor(srhl_hardlim), ELMRegressor(srhl_rbf)]
TrainedELMModels = dict()
# iterate through each kernel
# for modeltype, model_name, kernel_name in zip(classifiers,model_names,kernel_names):
for modeltype, model_name, kernel_name in zip(regressors, model_names, kernel_names):
print("Training kernel: " + kernel_name)
model = modeltype
model.fit(trainX, trainy)
# add trained model to allTrainedELMModels dictionary
TrainedELMModels[kernel_name] = model
print()
# save models from current iteration of hidden layers
allTrainedELMModels[numHiddenLayers] = TrainedELMModels
#########train all elm models#########
model = None # erase model from memory
#########test all ELM kernels#########
print("\nTesting all kernels...")
bestHiddenLayerCount, bestModel, bestModelName, bestMSE, bestPrediction = None, None, None, None, None
savedLiPoVoltageModel = None
RMSElist = []
hiddenLayerAndKernelRMSEs = dict()
print(kernel_names)
for kernel_name in kernel_names:
hiddenLayerAndKernelRMSEs[kernel_name] = []
for numHiddenLayers in range(1, 1 + maxHiddenLayers):
print("\nLayer: " + str(numHiddenLayers))
for kernel_name in kernel_names:
print("Testing kernel: " + kernel_name)
# test model
model = allTrainedELMModels[numHiddenLayers][kernel_name]
y_pred = model.predict(testX) # shape: (215,)
y_pred = yStandardScalar.inverse_transform(y_pred)
# calculate mean square error
MSE = mse(y_pred, np.asarray(testy))
RMSE = MSE**0.5
hiddenLayerAndKernelRMSEs[kernel_name].append(RMSE)
print("MSE:")
print(MSE)
print("RMSE:")
print(strFloat(RMSE))
improvement = 5 # set improvement percentage here
improvement /= 100
improvement = 1 - improvement
if bestMSE != None:
if MSE < (bestMSE*improvement):
bestMSE = MSE
bestHiddenLayerCount = numHiddenLayers
bestModel = model
bestPrediction = y_pred
else:
bestMSE = MSE
bestHiddenLayerCount = numHiddenLayers
bestModel = model
bestPrediction = y_pred
modelSave = {"model": bestModel, "usePCA": usePCA,
"yStandardScalar": yStandardScalar}
if usePCA:
modelSave["pcaModel"] = pcaModel
savedLiPoVoltageModel = dump(modelSave, './models/elmLiPoVoltageModel.model')
#####Plot RMSE performance for kernels and hidden layers#####
showPlots = True
if not kdbSource:
colours = ['b', 'r', 'c', 'm', 'y']
for kernelName, colour in zip(hiddenLayerAndKernelRMSEs.keys(), colours):
fig = plt.figure()
ax = fig.add_subplot(111)
ax.set(ylim=(0, 2))
plotTitle = "ELM Regressor " + kernelName
if usePCA:
plotTitle += " PCA"
else:
plotTitle += " no PCA"
plotTitle += " Lowest RMSE: " + \
strFloat(min(hiddenLayerAndKernelRMSEs[kernelName]))
plt.title(plotTitle)
plt.xlabel('Hidden layers')
plt.ylabel('RMSE')
ax.scatter(list(range(1, maxHiddenLayers+1)),
hiddenLayerAndKernelRMSEs[kernelName], s=10, c=colour, marker="s", label=kernelName)
# plt.legend(loc='upper left');
figureName = kernelName + " " + str(maxHiddenLayers) + " layers"
if usePCA:
figureName += " PCA"
else:
figureName += " no PCA"
fileExt = ".png"
plt.savefig("./results/" + figureName + fileExt)
if showPlots:
plt.show()
#####Plot RMSE performance for kernels and hidden layers#####
print("Optimal model:")
print(bestModel)
print("bestHiddenLayerCount:")
print(bestHiddenLayerCount)
print("MSE: " + str("{:.2f}".format(bestMSE)))
print("RMSE: " + str("{:.2f}".format(bestMSE**0.5)))
if not kdbSource:
print("Predictions || Actual Voltage || Error:")
Errors = np.asarray(testy) - bestPrediction
comparisons = []
for prediction, actual, Error in zip(bestPrediction, np.asarray(testy), Errors):
comparisons.append((prediction[0], actual[0], Error[0]))
print(comparisons)
comparisons.sort(key=lambda x: abs(x[2]))
for sample in comparisons:
print(strFloat(sample[0]) + "V || " + strFloat(sample[1]
) + "V || " + strFloat(sample[2]) + "V")