def dropout(X, p=0.):
"""
dropout using activation scaling to avoid test time weight rescaling
"""
if p > 0:
retain_prob = 1 - p
X *= t_rng.binomial(X.shape, p=retain_prob, dtype=theano.config.floatX)
X /= retain_prob
return X
def dropout(X, p=0.):
"""
dropout using activation scaling to avoid test time weight rescaling
"""
if p > 0:
retain_prob = 1 - p
X *= t_rng.binomial(X.shape, p=retain_prob, dtype=theano.config.floatX)
X /= retain_prob
return X
def op(self, state):
X = self.l_in.op(state=state)
retain_prob = 1 - self.p_drop
if state['dropout']:
X = X / retain_prob * t_rng.binomial(X.shape, p=retain_prob, dtype=theano.config.floatX)
return X
def op(self, state):
X = self.l_in.op(state=state)
retain_prob = 1 - self.p_drop
if state['dropout']:
X = X / retain_prob * t_rng.binomial(X.shape, p=retain_prob, dtype=theano.config.floatX)
return X
