forked from Zvezdin/blockchain-predictor
-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset_generator.py
283 lines (209 loc) · 8.85 KB
/
dataset_generator.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
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
import sys
import os
from datetime import timezone, timedelta, datetime as dt
import time
import dateutil.parser
import argparse
import pickle
import io
import codecs
import math
import pandas as pd
from arctic.date import CLOSED_OPEN
import numpy as np
sys.path.insert(0, os.path.realpath('dataset_models'))
from dataset_model import DatasetModel
from matrix_model import MatrixModel
from stacked_model import StackedModel
import database_tools as db
chunkStore = db.getChunkstore()
models = [MatrixModel(), StackedModel()]
save = True
debug = False
labelKey = 'closePrice'
def generateDataset(modelName, propertyNames, targetNames, labelsType='full', start=None, end=None, args = {}, preprocess = {}):
print("Generating dataset for properties %s, targets %s, model %s and range from %s to %s." % (str(propertyNames), str(targetNames), modelName, str(start), str(end)))
for arr in [propertyNames, targetNames]:
while '' in arr:
arr.remove('')
model = None
#get the model instance
for mod in models:
if mod.name == modelName:
model = mod
if model is None:
print("Error: Couldn't find model ", modelName)
return
properties = []
targets = []
#make sure we don't go off bounds for any property
start, end = db.getMasterInterval(chunkStore, propertyNames+targetNames, start, end)
#load the needed properties
for dataType, inputData in [('property', propertyNames), ('target', targetNames)]:
for prop in inputData:
data = db.loadData(chunkStore, prop, start, end, True, CLOSED_OPEN)
if type(data.iloc[0][prop]) == str: #if the property values have been encoded, decode them
print("Running numpy array Arctic workaround for prop %s..." % prop)
data[prop] = data[prop].apply(lambda x: db.decodeObject(x))
if prop in preprocess:
settings = preprocess[prop]
if 'scale' in settings:
if settings['scale'] == 'log2':
scaleF = np.log2
elif settings['scale'] == 'log10':
scaleF = np.log10
else:
raise ValueError("Unsupported scale type %s for preprocessing of property %s!" % (settings['scale'], prop))
def scale(val):
global globalMin
if globalMin < 0: #if we have relative values
val -= globalMin #turn all negatives to positives
val = scaleF(val)
val[val<0] = 0 #log if 0 is -inf
return val
else:
scale = lambda x: x #no scaling
xAxis = ':'
yAxis = ':'
if 'slices' in settings:
xAxis, yAxis = settings['slices']
strToSlice = lambda string: slice(*map(lambda x: int(x.strip()) if x.strip() else None, string.split(':')))
xAxis = strToSlice(xAxis)
yAxis = strToSlice(yAxis)
print("Slicing data by %s and %s." % (str(xAxis), str(yAxis)))
data[prop] = data[prop].apply(lambda x: x[yAxis, xAxis]) # trim
global globalMin #we need the minimum single value, to see if the property is realtive or not
globalMin = 0
def findMin(x):
global globalMin
globalMin = min(globalMin, np.min(x))
return x
data[prop].apply(findMin)
data[prop] = data[prop].apply(lambda x: scale(x)) # scale
if dataType == 'property':
properties.append(data)
if dataType == 'target':
targets.append(data)
for prop in properties:
if len(properties[0]) != len(prop):
raise ValueError("Error: Length mismatch in the data properties.")
#feed the model the properties and let it generate
dataset, dates, nextPrices, targetNorms = model.generate(properties, targets, args)
labels, dates = generateLabels(dates, nextPrices, db.loadData(chunkStore, labelKey, start, None, True), labelsType)
if len(dataset) != len(labels): #if we have a length mismatch, probably due to insufficient data for the last label
print("Mismatch in lengths of dataset and labels, removing excessive entries")
dataset = dataset[:len(labels)] #remove dataframes for which we have no labels
package = {
'dataset': dataset,
'dates': dates,
'labels': nextPrices,
'normalization': targetNorms
}
return package
def generateLabels(dates, nextPrices, ticks, labelsType):
"""Generates dataset labels for each passed date, getting data from ticks. dates MUST BE CHRONOLOGICALLY ORDERED. """
if labelsType == "boolean":
labels = []
i=0
indices = ticks.index.values
for date in dates:
while ticks.get_value(indices[i], 'date') != date:
i+=1
try:
currPrice = ticks.get_value(indices[i], 'closePrice')
nextPrice = ticks.get_value(indices[i+1], 'closePrice')
except (ValueError, IndexError, KeyError):
print("Failed to load the date after", date, ". Probably end of data. Will remove one dataset entry.")
dates = dates[:len(labels)] #keep only the labeled dates
break
if debug:
print(ticks.loc[indices[i] : indices[i+1]])
label = nextPrice > currPrice
if debug:
print("Label for dataframe at %s is %s for prices curr/next : %s and %s" % (date, label, currPrice, nextPrice) )
labels.append([label])
#make numpy array
labels = np.array(labels)
return (labels, dates)
elif labelsType == 'full': #nothing to do, the prices are already given and are normalized
return (nextPrices, dates)
def randomizeDataset(dataset):
main = dataset['dataset']
permutation = np.random.permutation(main)
for key in dataset:
if type(dataset[key]) != np.ndarray or len(dataset[key]) != len(main):
print("Unable to shuffle key %s. Leaving it as is." % key)
continue
dataset[key] = dataset[key][permutation]
return dataset
def saveDataset(filename, data):
if save:
#save the dataset to a file
try:
with open(filename, 'wb') as f:
pickle.dump(data, f, pickle.HIGHEST_PROTOCOL)
except Exception as e:
print('Unable to save data to', filename, ':', e)
def run(model, properties, targets, filename, start=None, end=None, ratio=[1], shuffle=False, args={}, preprocess={}):
if type(properties) != list:
properties = [properties]
if type(targets) != list:
targets = [targets]
#generate the dataset
dataset = generateDataset(model, properties, targets, start=start, end=end, args=args, preprocess=preprocess)
if shuffle:
#randomize it
dataset = randomizeDataset(dataset)
print("Randomized dataset and labels.")
if len(ratio) == 1:
data = dataset
else:
data = []
split = [] #the lenghts of the dataset pieces
mainLen = len(dataset['dataset'])
for rat in ratio:
split.append( int((rat * mainLen) / np.sum(ratio)) ) #calculate the length by keeping the given ratio
print(split, ratio)
index = 0
for i, spl in enumerate(split):
end = (spl + index) if i != len(split) -1 else None #because of integer division, add anything left on the last iteration
newDataset = {}
for key in dataset:
if type(dataset[key]) != np.ndarray or len(dataset[key]) != mainLen:
newDataset[key] = dataset[key]
print("Unable to split key %s. Leaving it as is." % key)
else:
newDataset[key] = dataset[key][index:end]
data.append(newDataset)
index += spl
#save it
if save:
saveDataset(filename, data)
print("saved dataset and labels as %a." % filename)
def init():
parser = argparse.ArgumentParser(description="Generates a dataset by compiling generated data properties using a certain dataset model")
parser.add_argument('--model', type=str, default='matrix', help='The name of the dataset model to use. Defaults to matrix.')
parser.add_argument('properties', type=str, default='openPrice,closePrice,gasPrice', help='A list of the names of the properties to use, separated by a comma.')
parser.add_argument('targets', type=str, default='highPrice', help='A list of target property names, separated by a comma.')
parser.add_argument('--start', type=str, default=None, help='The start date. YYYY-MM-DD-HH')
parser.add_argument('--end', type=str, default=None, help='The end date. YYYY-MM-DD-HH')
parser.add_argument('--filename', type=str, default=None, help='The target filename / dir to save the pickled dataset to. Defaults to "data/dataset.pickle"')
parser.add_argument('--ratio', type=str, default='1', help='On how many fragments to split the main dataset. For example, "1:2:3" will create three datasets with sizes proportional to what given.')
parser.add_argument('--shuffle', dest='shuffle', action="store_true", help="Shuffle the generated dataset and labels.")
parser.set_defaults(shuffle=False)
args, _ = parser.parse_known_args()
if args.filename == None:
filename = "data/dataset_" + str(args.start) + "-" + str(args.end) + ".pickle"
else: filename = args.filename
start = args.start
end = args.end
start = dateutil.parser.parse(start) if start is not None else None
end = dateutil.parser.parse(end) if end is not None else None
try:
ratio = [int(x) for x in args.ratio.split(':')]
except ValueError:
print("Error while reading the given ratio. Did you format it in the correct way?")
return
run(args.model, args.properties.split(','), args.targets.split(','), filename, start=start, end=end, ratio=ratio, shuffle=args.shuffle)
if __name__ == "__main__":
init()