def innerapply(self, x, _trainmode=False): # x is probabilities??
if _trainmode or self._debug:
rng = RVal(self.seed)
shap = self._shape if self._shape is not None else x.shape
g = rng.gumbel(shap)
y = (T.log(x) + g) / self.temp
ret = T.softmax(y, x.mask)
ret.mask = x.mask
return ret
else:
return T.softmax(x, x.mask, temperature=self._det_sm_temp)
def __init__(self,
p=0.3,
seed=None,
rescale=True,
_alwaysrandom=False,
**kw):
super(Dropout, self).__init__(**kw)
if seed is None:
seed = np.random.randint(0, 1e6)
self.p = 0.0 if (p is False or p is None) else 0.3 if p is True else p
self.rescale = rescale
self.seed = seed
self._debug = _alwaysrandom
self.rval = RVal(self.seed)
def apply(self, x, _trainmode=False):
if (_trainmode or self._debug) and self.p > 0:
xmask = x.mask
if self.rescale:
one = T.constant(1)
x /= one - self.p
rng = RVal(self.seed)
rv = rng.binomial(x.shape, p=1 - self.p, dtype=x.dtype)
x = x * rv
#print "done dropout"
x.mask = xmask
# x.push_extra_outs({"dropout{}".format(np.random.randint(100, 199)): rv})
return x
else:
return x
class Dropout(Block):
def __init__(self,
p=0.3,
seed=None,
rescale=True,
_alwaysrandom=False,
**kw):
super(Dropout, self).__init__(**kw)
if seed is None:
seed = np.random.randint(0, 1e6)
self.p = 0.0 if (p is False or p is None) else 0.3 if p is True else p
self.rescale = rescale
self.seed = seed
self._debug = _alwaysrandom
self.rval = RVal(self.seed)
def apply(self, x, _trainmode=False):
if (_trainmode or self._debug) and self.p > 0:
#print "Dropout: YES"
xmask = x.mask
if self.rescale:
one = T.constant(1)
x /= one - self.p
#rng = RVal(self.seed)
rv = self.rval.binomial(x.shape, p=1 - self.p, dtype=x.dtype)
x = x * rv
#print "done dropout"
x.mask = xmask
# x.push_extra_outs({"dropout{}".format(np.random.randint(100, 199)): rv})
return x
else:
#print "Dropout: NO"
return x
def __init__(self, **kw):
super(RandomSequence, self).__init__(**kw)
self.randval = RVal().normal((5, ))
def rec(self, x_t):
return RVal().normal(x_t.shape)
def apply(self, x): # (batsize,)
m = self.W_m[x]
s = self.W_s[x]
z = RVal(seed=self.seed).normal(m.shape) * s + m
o = T.dot(z, self.O)
return Softmax()(o)
def test_value_shape(self):
rv = RVal().binomial((25, ), p=0.5)
self.assertEqual(rv.d.shape.eval(), (25, ))