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
|
import numpy
import numpy.random
import copy
import itertools
import dill
import tqdm
def minibatch(train_X, train_y, size=16, nr_update=1000):
with tqdm.tqdm(total=nr_update * size, leave=False) as pbar:
while nr_update >= 0:
indices = numpy.arange(len(train_X))
numpy.random.shuffle(indices)
j = 0
while j < indices.shape[0]:
slice_ = indices[j : j + size]
X = _take_slice(train_X, slice_)
y = _take_slice(train_y, slice_)
yield X, y
j += size
nr_update -= 1
if nr_update <= 0:
break
pbar.update(size)
def _take_slice(data, slice_):
if isinstance(data, list) or isinstance(data, tuple):
return [data[int(i)] for i in slice_]
else:
return data[slice_]
class BestFirstFinder(object):
def __init__(self, **param_values):
self.queue = []
self.limit = 16
self.params = param_values
self.best_acc = 0.0
self.best_i = 0
self.i = 0
self.j = 0
self.best_model = None
self.temperature = 0.0
@property
def configs(self):
keys, value_groups = zip(*self.params.items())
for values in itertools.product(*value_groups):
config = dict(zip(keys, values))
yield config
def enqueue(self, model, train_acc, check_acc):
fom = check_acc * min(check_acc / train_acc, 1.0)
self.queue.append([fom, self.i, 0, model])
if check_acc >= self.best_acc:
self.best_acc = check_acc
self.best_i = self.i
self.best_model = model
self.temperature = 0.0
else:
self.temperature += 0.01
self.j = 0
self.queue.sort(reverse=True)
self.queue = self.queue[:self.limit]
def __iter__(self):
self.queue.sort(reverse=True)
self.queue = self.queue[:self.limit]
for i in range(len(self.queue)):
self.queue[i][0] = self.queue[i][0] - 0.01
self.queue[i][-1][2]['parent'] = self.queue[i][2]
self.queue[i][2] += 1
yield self.queue[i][-1]
@property
def best(self):
return self.best_model
def resample_hyper_params(hparams, temperature):
hparams = dict(hparams)
hparams['epochs'] = hparams.get('epochs', 0) + 1
hparams['learn_rate'] = resample(hparams['learn_rate'], 1e-6, 0.1, temperature)
#hparams['beta1'] = resample(hparams.get('beta1', 0.9), 0.8, 1.0, temperature)
#hparams['beta2'] = resample(hparams.get('beta2', 0.9), 0.8, 1.0, temperature)
#hparams['L2'] = resample(hparams['L2'], 0.0, 1e-3, temperature)
#hparams['batch_size'] = int(resample(hparams['batch_size'], 10, 256, temperature))
#hparams['dropout'] = resample(hparams['dropout'], 0.05, 0.7, temperature)
return hparams
def resample(curr, min_, max_, temperature):
if temperature == 0.0:
return curr
scale = (max_ - min_) * temperature
next_ = numpy.random.normal(loc=curr, scale=scale)
return min(max_, max(min_, next_))
def train_epoch(model, sgd, hparams, train_X, train_y, dev_X, dev_y, device_id=-1,
temperature=0.0):
model, sgd, hparams = dill.loads(dill.dumps((model, sgd, hparams)))
if device_id >= 0:
device = model.to_gpu(device_id)
sgd.ops = model.ops
sgd.to_gpu()
if isinstance(train_y, numpy.ndarray):
train_y = model.ops.asarray(train_y)
dev_y = model.ops.asarray(dev_y)
hparams = resample_hyper_params(hparams, temperature)
sgd.learn_rate = hparams['learn_rate']
sgd.beta1 = hparams['beta1']
sgd.beta2 = hparams['beta2']
sgd.L2 = hparams['L2']
train_acc = 0.
train_n = 0
for X, y in minibatch(train_X, train_y, size=hparams['batch_size'], nr_update=hparams['nr_update']):
yh, finish_update = model.begin_update(X, drop=hparams['dropout'])
if hasattr(y, 'shape'):
dy = (yh-y) / y.shape[0]
train_acc += (y.argmax(axis=1) == yh.argmax(axis=1)).sum()
train_n += y.shape[0]
else:
n_y = sum(len(y_i) for y_i in y)
dy = [(yh[i]-y[i])/n_y for i in range(len(yh))]
for i in range(len(y)):
train_acc += (y[i].argmax(axis=1) == yh[i].argmax(axis=1)).sum()
train_n += n_y
finish_update(dy, sgd=sgd)
train_acc /= train_n
with model.use_params(sgd.averages):
dev_acc = model.evaluate(dev_X, dev_y)
model.to_cpu()
sgd.to_cpu()
return device_id, ((model, sgd, hparams), float(train_acc), float(dev_acc))
class DevicePool(object):
"""Synchronize GPU usage"""
def __init__(self, n):
self.devices = {i: None for i in range(n)}
def acquire(self):
for i, device in self.devices.items():
if device is None:
self.devices[i] = True
return i
else:
return None
def release(self, i):
if i in self.devices:
self.devices[i] = None
#
#def best_first_sgd(initials, train_X, train_y, dev_X, dev_y,
# get_new_model=None, get_score=None):
# if get_new_model is None:
# get_new_model = _get_new_model
# if get_score is None:
# get_score = _get_score
#
# queue = []
# for i, model in enumerate(initials):
# train_acc, model = get_new_model(model, train_X, train_y)
# check_acc = get_score(model, dev_X, dev_y)
# ratio = min(check_acc / train_acc, 1.0)
# print((model[-1], train_acc, check_acc))
# queue.append([check_acc * ratio, i, model])
#
# train_acc = 0
# limit = 8
# i = 0
# best_model = None
# best_acc = 0.0
# best_i = 0
# while best_i > (i - 100) and train_acc < 0.999:
# queue.sort(reverse=True)
# queue = queue[:limit]
# prev_score, parent, model = queue[0]
# queue[0][0] -= 0.001
# yield prev_score, parent, model
# train_acc, new_model = get_new_model(model, train_X, train_y)
# check_acc = get_score(new_model, dev_X, dev_y)
# ratio = min(check_acc / train_acc, 1.0)
#
# i += 1
# queue.append([check_acc * ratio, i, new_model])
#
# if check_acc >= best_acc:
# best_acc = check_acc
# best_i = i
# best_model = new_model
# progress = {
# 'i': i,
# 'parent': parent,
# 'prev_score': prev_score,
# 'this_score': queue[-1][0],
# 'train_acc': train_acc,
# 'check_acc': check_acc,
# 'best_acc': best_acc,
# 'hparams': new_model[-1]
# }
# yield best_model, progress
#
#
#
|
