-
Notifications
You must be signed in to change notification settings - Fork 0
/
vw-hyperopt.py
393 lines (318 loc) · 16 KB
/
vw-hyperopt.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
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
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
#!/usr/bin/env python
# coding: utf-8
"""
Github version of hyperparameter optimization for Vowpal Wabbit via hyperopt
"""
__author__ = 'kurtosis'
import os
from hyperopt import hp, fmin, tpe, rand, Trials, STATUS_OK
from sklearn.metrics import roc_curve, auc, log_loss, average_precision_score, hinge_loss, \
mean_squared_error
import numpy as np
from datetime import datetime as dt
import subprocess, shlex
from math import exp, log
import argparse
import re
import logging
import json
from matplotlib import pyplot as plt
try:
import seaborn as sns
except ImportError:
print ("Warning: seaborn is not installed. "
"Without seaborn, standard matplotlib plots will not look very charming. "
"It's recommended to install it via pip install seaborn")
def read_arguments():
parser = argparse.ArgumentParser()
parser.add_argument('--searcher', type=str, default='tpe', choices=['tpe', 'rand'])
parser.add_argument('--max_evals', type=int, default=100)
parser.add_argument('--train', type=str, required=True, help="training set")
parser.add_argument('--holdout', type=str, required=True, help="holdout set")
parser.add_argument('--vw_space', type=str, required=True, help="hyperparameter search space (must be 'quoted')")
parser.add_argument('--outer_loss_function', default='logistic',
choices=['logistic', 'roc-auc', 'pr-auc', 'hinge', 'squared']) # TODO: implement quantile
parser.add_argument('--regression', action='store_true', default=False, help="""regression (continuous class labels)
or classification (-1 or 1, default value).""")
parser.add_argument('--plot', action='store_true', default=False, help=("Plot the results in the end. "
"Requires matplotlib and "
"(optionally) seaborn to be installed."))
args = parser.parse_args()
return args
class HyperoptSpaceConstructor(object):
"""
Takes command-line input and transforms it into hyperopt search space
An example of command-line input:
--algorithms=ftrl,sgd --l2=1e-8..1e-4~LO -l=0.01..10~L --ftrl_beta=0.01..1 --passes=1..10~I -q=SE+SZ+DR,SE~O
"""
def __init__(self, command):
self.command = command
self.space = None
self.algorithm_metadata = {
'ftrl': {'arg': '--ftrl', 'prohibited_flags': set()},
'sgd': {'arg': '', 'prohibited_flags': {'--ftrl_alpha', '--ftrl_beta'}}
}
self.range_pattern = re.compile("[^~]+") # re.compile("(?<=\[).+(?=\])")
self.distr_pattern = re.compile("(?<=~)[IOL]*") # re.compile("(?<=\])[IOL]*")
self.only_continuous = re.compile("(?<=~)[IL]*") # re.compile("(?<=\])[IL]*")
def _process_vw_argument(self, arg, value, algorithm):
try:
distr_part = self.distr_pattern.findall(value)[0]
except IndexError:
distr_part = ''
range_part = self.range_pattern.findall(value)[0]
is_continuous = '..' in range_part
ocd = self.only_continuous.findall(value)
if not is_continuous and len(ocd)> 0 and ocd[0] != '':
raise ValueError(("Need a range instead of a list of discrete values to define "
"uniform or log-uniform distribution. "
"Please, use [min..max]%s form") % (distr_part))
if is_continuous and arg == '-q':
raise ValueError(("You must directly specify namespaces for quadratic features "
"as a list of values, not as a parametric distribution"))
hp_choice_name = "_".join([algorithm, arg.replace('-', '')])
try_omit_zero = 'O' in distr_part
distr_part = distr_part.replace('O', '')
if is_continuous:
vmin, vmax = [float(i) for i in range_part.split('..')]
if distr_part == 'L':
distrib = hp.loguniform(hp_choice_name, log(vmin), log(vmax))
elif distr_part == '':
distrib = hp.uniform(hp_choice_name, vmin, vmax)
elif distr_part == 'I':
distrib = hp.quniform(hp_choice_name, vmin, vmax, 1)
elif distr_part in {'LI', 'IL'}:
distrib = hp.qloguniform(hp_choice_name, log(vmin), log(vmax), 1)
else:
raise ValueError("Cannot recognize distribution: %s" % (distr_part))
else:
possible_values = range_part.split(',')
if arg == '-q':
possible_values = [v.replace('+', ' -q ') for v in possible_values]
distrib = hp.choice(hp_choice_name, possible_values)
if try_omit_zero:
hp_choice_name_outer = hp_choice_name + '_outer'
distrib = hp.choice(hp_choice_name_outer, ['omit', distrib])
return distrib
def string_to_pyll(self):
line = shlex.split(self.command)
algorithms = ['sgd']
for arg in line:
arg, value = arg.split('=')
if arg == '--algorithms':
algorithms = set(self.range_pattern.findall(value)[0].split(','))
if tuple(self.distr_pattern.findall(value)) not in {(), ('O',)}:
raise ValueError(("Distribution options are prohibited for --algorithms flag. "
"Simply list the algorithms instead (like --algorithms=ftrl,sgd)"))
elif self.distr_pattern.findall(value) == ['O']:
algorithms.add('sgd')
for algo in algorithms:
if algo not in self.algorithm_metadata:
raise NotImplementedError(("%s algorithm is not found. "
"Supported algorithms by now are %s")
% (algo, str(self.algorithm_metadata.keys())))
break
self.space = {algo: {'type': algo, 'argument': self.algorithm_metadata[algo]['arg']} for algo in algorithms}
for algo in algorithms:
for arg in line:
arg, value = arg.split('=')
if arg == '--algorithms':
continue
if arg not in self.algorithm_metadata[algo]['prohibited_flags']:
distrib = self._process_vw_argument(arg, value, algo)
self.space[algo][arg] = distrib
else:
pass
self.space = hp.choice('algorithm', self.space.values())
class HyperOptimizer(object):
def __init__(self, train_set, holdout_set, command, max_evals=100,
outer_loss_function='logistic',
searcher='tpe', is_regression=False):
self.train_set = train_set
self.holdout_set = holdout_set
# self.train_model = './current.model'
# self.holdout_pred = './holdout.pred'
# self.trials_output = './trials.json'
# self.hyperopt_progress_plot = './hyperopt_progress.png'
# self.log = './log.log'
prefix = './' + os.path.split(train_set)[-1].split('.')[0]
self.train_model = prefix + '_current.model'
self.holdout_pred = prefix + '_holdout.pred'
self.trials_output = prefix + '_trials.json'
self.hyperopt_progress_plot = prefix + '_hyperopt_progress.png'
self.log = prefix + '_log.log'
self.logger = self._configure_logger()
# hyperopt parameter sample, converted into a string with flags
self.param_suffix = None
self.train_command = None
self.validate_command = None
self.y_true_train = []
self.y_true_holdout = []
self.outer_loss_function = outer_loss_function
self.space = self._get_space(command)
self.max_evals = max_evals
self.searcher = searcher
self.is_regression = is_regression
self.trials = Trials()
self.current_trial = 0
def _get_space(self, command):
hs = HyperoptSpaceConstructor(command)
hs.string_to_pyll()
return hs.space
def _configure_logger(self):
LOGGER_FORMAT = "%(asctime)s,%(msecs)03d %(levelname)-8s [%(name)s/%(module)s:%(lineno)d]: %(message)s"
LOGGER_DATEFMT = "%Y-%m-%d %H:%M:%S"
LOGFILE = self.log
logging.basicConfig(format=LOGGER_FORMAT,
datefmt=LOGGER_DATEFMT,
level=logging.DEBUG)
formatter = logging.Formatter(LOGGER_FORMAT, datefmt=LOGGER_DATEFMT)
file_handler = logging.FileHandler(LOGFILE)
file_handler.setFormatter(formatter)
logger = logging.getLogger()
logger.addHandler(file_handler)
return logger
def get_hyperparam_string(self, **kwargs):
for arg in ['--passes']: #, '--rank', '--lrq']:
if arg in kwargs:
kwargs[arg] = int(kwargs[arg])
#print 'KWARGS: ', kwargs
flags = [key for key in kwargs if key.startswith('-')]
for flag in flags:
if kwargs[flag] == 'omit':
del kwargs[flag]
self.param_suffix = ' '.join(['%s %s' % (key, kwargs[key]) for key in kwargs if key.startswith('-')])
self.param_suffix += ' %s' % (kwargs['argument'])
def compose_vw_train_command(self):
data_part = ('vw -d %s -f %s --holdout_off -c '
% (self.train_set, self.train_model))
self.train_command = ' '.join([data_part, self.param_suffix])
def compose_vw_validate_command(self):
data_part = 'vw -t -d %s -i %s -p %s --holdout_off -c' \
% (self.holdout_set, self.train_model, self.holdout_pred)
self.validate_command = data_part
def fit_vw(self):
self.compose_vw_train_command()
self.logger.info("executing the following command (training): %s" % self.train_command)
subprocess.call(shlex.split(self.train_command))
def validate_vw(self):
self.compose_vw_validate_command()
self.logger.info("executing the following command (validation): %s" % self.validate_command)
subprocess.call(shlex.split(self.validate_command))
def get_y_true_train(self):
self.logger.info("loading true train class labels...")
yh = open(self.train_set, 'r')
self.y_true_train = []
for line in yh:
self.y_true_train.append(int(line.strip()[0:2]))
if not self.is_regression:
self.y_true_train = [(i + 1.) / 2 for i in self.y_true_train]
self.logger.info("train length: %d" % len(self.y_true_train))
def get_y_true_holdout(self):
self.logger.info("loading true holdout class labels...")
yh = open(self.holdout_set, 'r')
self.y_true_holdout = []
for line in yh:
self.y_true_holdout.append(float(line.split()[0]))
if not self.is_regression:
self.y_true_holdout = [int((i + 1.) / 2) for i in self.y_true_holdout]
self.logger.info("holdout length: %d" % len(self.y_true_holdout))
def validation_metric_vw(self):
v = open('%s' % self.holdout_pred, 'r')
y_pred_holdout = []
for line in v:
y_pred_holdout.append(float(line.split()[0].strip()))
if self.outer_loss_function == 'logistic':
y_pred_holdout_proba = [1. / (1 + exp(-i)) for i in y_pred_holdout]
loss = log_loss(self.y_true_holdout, y_pred_holdout_proba)
elif self.outer_loss_function == 'squared':
loss = mean_squared_error(self.y_true_holdout, y_pred_holdout)
elif self.outer_loss_function == 'hinge':
loss = hinge_loss(self.y_true_holdout, y_pred_holdout)
elif self.outer_loss_function == 'pr-auc':
loss = -average_precision_score(self.y_true_holdout, y_pred_holdout)
elif self.outer_loss_function == 'roc-auc':
y_pred_holdout_proba = [1. / (1 + exp(-i)) for i in y_pred_holdout]
fpr, tpr, _ = roc_curve(self.y_true_holdout, y_pred_holdout_proba)
loss = -auc(fpr, tpr)
else:
raise KeyError('Invalide outer loss function')
self.logger.info('parameter suffix: %s' % self.param_suffix)
self.logger.info('loss value: %.6f' % loss)
return loss
def hyperopt_search(self, parallel=False): # TODO: implement parallel search with MongoTrials
def objective(kwargs):
start = dt.now()
self.current_trial += 1
self.logger.info('\n\nStarting trial no.%d' % self.current_trial)
self.get_hyperparam_string(**kwargs)
self.fit_vw()
self.validate_vw()
loss = self.validation_metric_vw()
finish = dt.now()
elapsed = finish - start
self.logger.info("evaluation time for this step: %s" % str(elapsed))
# clean up
subprocess.call(shlex.split('rm %s %s' % (self.train_model, self.holdout_pred)))
to_return = {'status': STATUS_OK,
'loss': loss, # TODO: include also train loss tracking in order to prevent overfitting
'eval_time': elapsed.seconds,
'train_command': self.train_command,
'current_trial': self.current_trial
}
return to_return
self.trials = Trials()
if self.searcher == 'tpe':
algo = tpe.suggest
elif self.searcher == 'rand':
algo = rand.suggest
else:
raise KeyError('Invalid searcher')
logging.debug("starting hypersearch...")
best_params = fmin(objective, space=self.space, trials=self.trials, algo=algo, max_evals=self.max_evals)
self.logger.debug("the best hyperopt parameters: %s" % str(best_params))
json.dump(self.trials.results, open(self.trials_output, 'w'))
self.logger.info('All the trials results are saved at %s' % self.trials_output)
best_configuration = self.trials.results[np.argmin(self.trials.losses())]['train_command']
best_loss = self.trials.results[np.argmin(self.trials.losses())]['loss']
self.logger.info("\n\nA full training command with the best hyperparameters: \n%s\n\n" % best_configuration)
self.logger.info("\n\nThe best holdout loss value: \n%s\n\n" % best_loss)
return best_configuration, best_loss
def plot_progress(self):
try:
sns.set_palette('Set2')
sns.set_style("darkgrid", {"axes.facecolor": ".95"})
except:
pass
self.logger.debug('plotting...')
plt.figure(figsize=(15,10))
plt.subplot(211)
plt.plot(self.trials.losses(), '.', markersize=12)
plt.title('Per-Iteration Outer Loss', fontsize=16)
plt.ylabel('Outer loss function value')
if self.outer_loss_function in ['logloss']:
plt.yscale('log')
xticks = [int(i) for i in np.linspace(plt.xlim()[0], plt.xlim()[1], min(len(self.trials.losses()), 11))]
plt.xticks(xticks, xticks)
plt.subplot(212)
plt.plot(np.minimum.accumulate(self.trials.losses()), '.', markersize=12)
plt.title('Cumulative Minimum Outer Loss', fontsize=16)
plt.xlabel('Iteration number')
plt.ylabel('Outer loss function value')
xticks = [int(i) for i in np.linspace(plt.xlim()[0], plt.xlim()[1], min(len(self.trials.losses()), 11))]
plt.xticks(xticks, xticks)
plt.tight_layout()
plt.savefig(self.hyperopt_progress_plot)
self.logger.info('The diagnostic hyperopt progress plot is saved: %s' % self.hyperopt_progress_plot)
def main():
args = read_arguments()
h = HyperOptimizer(train_set=args.train, holdout_set=args.holdout, command=args.vw_space,
max_evals=args.max_evals,
outer_loss_function=args.outer_loss_function,
searcher=args.searcher, is_regression=args.regression)
h.get_y_true_holdout()
h.hyperopt_search()
if args.plot:
h.plot_progress()
if __name__ == '__main__':
main()