-
Notifications
You must be signed in to change notification settings - Fork 9
/
utils.py
260 lines (217 loc) · 9.01 KB
/
utils.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
from theano import tensor as T
from collections import OrderedDict
from theano.ifelse import ifelse
import theano
from theano import config
from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams
import numpy as np
np.random.seed(23423)
rng = np.random.RandomState(897987)
srng = RandomStreams(rng.randint(2304234))
class ReverseGradient(theano.Op):
""" theano operation to reverse the gradients
Introduced in http://arxiv.org/pdf/1409.7495.pdf
"""
view_map = {0: [0]}
__props__ = ('hp_lambda', )
def __init__(self, hp_lambda):
super(ReverseGradient, self).__init__()
self.hp_lambda = hp_lambda
def make_node(self, x):
assert hasattr(self, '_props'), "Your version of theano is too old to support __props__."
x = theano.tensor.as_tensor_variable(x)
return theano.Apply(self, [x], [x.type()])
def perform(self, node, inputs, output_storage):
xin, = inputs
xout, = output_storage
xout[0] = xin
def grad(self, input, output_gradients):
return [-self.hp_lambda * output_gradients[0]]
def infer_shape(self, node, i0_shapes):
return i0_shapes
def hard_sigmoid(x):
return T.nnet.hard_sigmoid(x)
def log_softmax(x):
xdev = x - x.max(1, keepdims=True)
return xdev - T.log(T.sum(T.exp(xdev), axis=1, keepdims=True))
def categorical_crossentropy_logdomain(log_predictions, targets):
return -T.mean(targets * log_predictions, axis=1)
def normal(shape, scale=0.05):
return np.random.normal(0, scale, size=shape).astype('float32')
def get_fans(shape):
fan_in = shape[0] if len(shape) == 2 else np.prod(shape[1:])
fan_out = shape[1] if len(shape) == 2 else shape[0]
return fan_in, fan_out
def orthogonal(shape):
''' Reference: Glorot & Bengio, AISTATS 2010 glorot_normal
'''
fan_in, fan_out = get_fans(shape)
s = np.sqrt(2. / (fan_in * fan_out))
return normal(shape, s)
def he_normal(shape):
''' Reference: He et al., http://arxiv.org/abs/1502.01852
'''
fan_in, fan_out = get_fans(shape)
s = np.sqrt(2. / fan_in)
return normal(shape, s)
def glorot_uniform(shape):
fan_in, fan_out = get_fans(shape)
s = np.sqrt(6. / (fan_in + fan_out))
return uniform(shape, s)
def orthogonal_tmp2(shape):
fan_in, fan_out = get_fans(shape)
s = np.sqrt(6. / (fan_in + fan_out))
return uniform(shape, s)
def uniform(shape, scale=0.05):
return np.random.uniform(low=-scale, high=scale, size=shape).astype('float32')
def orthogonal_tmp(shape, scale=1.0):
''' From Lasagne. Reference: Saxe et al., http://arxiv.org/abs/1312.6120
'''
flat_shape = (shape[0], np.prod(shape[1:]))
a = np.random.normal(0.0, 1.0, flat_shape)
u, _, v = np.linalg.svd(a, full_matrices=False)
# pick the one with the correct shape
q = u if u.shape == flat_shape else v
q = q.reshape(shape)
return scale * q[:shape[0], :shape[1]]
def as_floatX(variable):
if isinstance(variable, float):
#return np.cast["float32"](variable)
return np.cast['float32'](variable)
elif isinstance(variable, np.ndarray):
#return np.cast["float32"](variable)
return np.cast['float32'](variable)
def rectify(X):
return T.maximum(X, 0.)
def cappedrectify(X):
return T.minimum(5., T.maximum(X, 0.))
def elu(X):
return T.switch(T.ge(X, 0), X, T.exp(X)-1.)
def snelu(X):
scale = 1.0507009873554804934193349852946
alpha = 1.6732632423543772848170429916717
return scale * T.switch(T.ge(X, 0), X, alpha*T.exp(X)-alpha)
def dropout(X, dropout_switch=1, p=0.):
retain_prob = 1 - p
mask = srng.binomial(X.shape, p=retain_prob, dtype='float32')
X = ifelse(T.lt(dropout_switch, 0.5), X*mask, (X*retain_prob).reshape(mask.shape))
return X
def dropout_scan(X, mask, dropout_switch=1, p=0.):
retain_prob = 1 - p
X = ifelse(T.lt(dropout_switch, 0.5), X*mask, (X*retain_prob).reshape(mask.shape))
return X
def clip_norm(g, c, n):
if c > 0:
g = T.switch(T.ge(n, c), g * c / n, g)
return g
def sgdm(cost, parameters, lr2=1., momentum=0.8):
lr = theano.shared(as_floatX(lr2).astype("float32"))
grads = T.grad(cost, parameters)
updates = OrderedDict()
for param,g2 in zip(parameters,grads):
grad = clip_norm(g2, 3, T.sum(g2 ** 2))
mparam = theano.shared(param.get_value()*0.)
updates[param] = param - lr * mparam
updates[mparam] = mparam*momentum + (1.-momentum)*grad
return updates, lr
def sgd(cost, parameters, lr, updates=None):
grads = T.grad(cost,parameters)
updates = OrderedDict({})
for param,grad in zip(parameters,grads):
updates[param] = param - lr*grad
return updates
#def Adam(cost, params, lr=0.0002, b1=0.1, b2=0.001, e=1e-8):
#def Adam(cost, params, lr2=0.001, b1=0.1, b2=0.001, e=1e-8):
#def Adam(cost, params, lr2=0.001, b1=0.5, b2=0.001, e=1e-8):
def Adam(cost, params, lr2=0.001, b1=0.1, b2=0.001, e=1e-8):
updates = []
lr = theano.shared(as_floatX(lr2).astype("float32"))
grads = T.grad(cost, params)
i = theano.shared(as_floatX(0.))
i_t = i + as_floatX(1.)
fix1 = as_floatX(1.) - (as_floatX(1.) - as_floatX(b1))**i_t
fix2 = as_floatX(1.) - (as_floatX(1.) - as_floatX(b2))**i_t
#lr_t = as_floatX(lr) * (T.sqrt(fix2) / fix1)
lr_t = lr * (T.sqrt(fix2) / fix1)
for p, g2 in zip(params, grads):
'''
if p.name != 'Words' and p.name != 'Pos' and p.name != 'lang':
g = clip_norm(g2, 1, T.sum(g2 ** 2))
else:
g = g2
'''
#g = clip_norm(g2, 3, T.sum(g2 ** 2))
g = g2
#g = g2.clip(-.5, .5)
#g = clip_norm(g, 3, T.sqrt(T.sum(g**2)))
m = theano.shared(p.get_value() * as_floatX(0.))
v = theano.shared(p.get_value() * as_floatX(0.))
m_t = (as_floatX(b1) * g) + ((as_floatX(1.) - as_floatX(b1)) * m)
v_t = (as_floatX(b2) * T.sqr(g)) + ((as_floatX(1.) - as_floatX(b2)) * v)
g_t = m_t / (T.sqrt(v_t) + as_floatX(e))
p_t = p - (lr_t * g_t)
updates.append((m, m_t))
updates.append((v, v_t))
updates.append((p, p_t))
updates.append((i, i_t))
return updates, lr
def RMSprop(cost, params, lr=0.001, rho=0.9, epsilon=1e-6):
grads = T.grad(cost=cost, wrt=params)
'''
norm = T.sqrt(sum([T.sum(g ** 2) for g,p in zip(grads, params) if p.name != 'Words' and p.get_value(borrow=True).ndim == 2 and p.name != 'label_embeddings']))
tmp_grads = []
for g,p in zip(grads, params):
if p.name != 'Words' and (p.get_value(borrow=True).ndim == 2) and p.name != 'label_embeddings':
tmp_grads.append(clip_norm(g, 5, norm))
else:
tmp_grads.append(g)
grads = tmp_grads
'''
norm = T.sqrt(sum([T.sum(g ** as_floatX(2.)) for g in grads]))
grads = [clip_norm(g, as_floatX(5.), norm) for g in grads]
updates = []
for p, g in zip(params, grads):
acc = theano.shared(p.get_value() * as_floatX(0.))
acc_new = as_floatX(rho) * acc + (as_floatX(1.) - as_floatX(rho)) * g ** as_floatX(2.)
gradient_scaling = T.sqrt(acc_new + as_floatX(epsilon))
g = g / gradient_scaling
updates.append((acc, acc_new))
updates.append((p, p - as_floatX(lr) * g))
return updates
def adagrad(cost, params, lr=0.001, eps=1e-8, sparse=False):
lr = theano.shared(as_floatX(lr).astype("float32"))
eps = as_floatX(eps).astype("float32")
gsums = [theano.shared(np.zeros_like(param.get_value(borrow=True))+0.1) for param in params]
#gsums = [theano.shared(np.zeros_like(param.get_value(borrow=True))) for param in params]
xsums = [None for param in params]
gparams = T.grad(cost, params)
updates = OrderedDict()
for gparam, param, gsum in zip(gparams, params, gsums):
updates[gsum] = T.cast(gsum + (gparam ** as_floatX(2.)), "float32")
updates[param] = T.cast(param - lr * (gparam / (T.sqrt(updates[gsum] + eps))), "float32")
return updates, lr
def sgd_updates_adadelta(params,cost,rho=0.95,epsilon=1e-6,norm_lim=9,word_vec_name='Words'):
"""
adadelta update rule, mostly from
https://groups.google.com/forum/#!topic/pylearn-dev/3QbKtCumAW4 (for Adadelta)
"""
updates = OrderedDict({})
exp_sqr_grads = OrderedDict({})
exp_sqr_ups = OrderedDict({})
gparams = []
for param in params:
empty = np.zeros_like(param.get_value())
exp_sqr_grads[param] = theano.shared(value=as_floatX(empty),name="exp_grad_%s" % param.name)
gp = T.grad(cost, param)
exp_sqr_ups[param] = theano.shared(value=as_floatX(empty), name="exp_grad_%s" % param.name)
gparams.append(gp)
for param, gp in zip(params, gparams):
exp_sg = exp_sqr_grads[param]
exp_su = exp_sqr_ups[param]
up_exp_sg = rho * exp_sg + (1 - rho) * T.sqr(gp)
updates[exp_sg] = up_exp_sg
step = -(T.sqrt(exp_su + epsilon) / T.sqrt(up_exp_sg + epsilon)) * gp
updates[exp_su] = rho * exp_su + (1 - rho) * T.sqr(step)
stepped_param = param + step
updates[param] = stepped_param
return updates