def __init__(self, rng, input, input_shape, activation=tanh, n_kernels=4, n_out=300, pool_out=False,
mean_pool=True, sum_out=False, concat_out=False, skip_gram=False, mask=None, use_bias=False):
"""
Allocate a BigramLayer with shared variable internal parameters.
"""
self.input = input
self.activation = activation
# initialize weights with random weights
n_in, n_out, fan_in, fan_out = get_input_info(input_shape, sum_out=sum_out, ngram=3, n_out=n_out)
W_values = get_W_values(rng=rng, activation=activation, fan_in=fan_in, fan_out=fan_out, n_in=n_in, n_out=n_out,
n_kernels=n_kernels)
self.T1 = theano.shared(W_values, borrow=True, name="T1")
self.T2 = theano.shared(W_values, borrow=True, name="T2")
self.T3 = theano.shared(W_values, borrow=True, name="T3")
self.b = theano.shared(np.zeros(shape=(n_out, ), dtype=theano.config.floatX), borrow=True, name="b_cov")
offset = 4 if skip_gram else 2
self.mask = mask[:, : -offset] if mask is not None else None
left = T.dot(input, self.T1)[:, : -offset]
center = T.dot(input, self.T2)[:, offset / 2: -offset / 2]
right = T.dot(input, self.T3)[:, offset:]
cov_out = left + center + right + self.b if use_bias else left + center + right
activation_out = activation(cov_out)
# concatenate the output of each kernel
if concat_out:
trigram_sum = T.sum(activation_out, axis=1)
self.output = trigram_sum.flatten(2)
else:
pooled_out = T.mean(activation_out, axis=2) if mean_pool else T.max(activation_out, axis=2)
if self.mask is not None:
trigram_sum = (pooled_out * self.mask[:, :, None]).sum(axis=1)
trigram_avg = trigram_sum / self.mask.sum(axis=1)[:, None]
else:
trigram_sum = T.sum(pooled_out, axis=1)
trigram_avg = trigram_sum / input_shape[0]
if pool_out:
self.output = pooled_out
elif sum_out:
self.output = trigram_sum
else:
self.output = trigram_avg
self.params = [self.T1, self.T2, self.T3, self.b] if use_bias else [self.T1, self.T2, self.T3]
def __init__(self, rng, input, input_shape, activation=relu, n_kernels=4, n_out=300, pool_out=False,
sum_out=True, mean_pool=True, concat_out=False, mask=None, skip_gram=False, use_bias=False):
"""
Allocate a UnigramLayer with shared variable internal parameters.
"""
self.input = input
self.activation = activation
self.mask = mask
# initialize weights with random weights
n_in, n_out, fan_in, fan_out = get_input_info(input_shape, sum_out=sum_out, ngram=1, n_out=n_out)
W_values = get_W_values(rng=rng, activation=activation, fan_in=fan_in, fan_out=fan_out, n_in=n_in, n_out=n_out,
n_kernels=n_kernels)
self.W = theano.shared(W_values, borrow=True, name="W_cov")
self.b = theano.shared(np.zeros(shape=(n_out, ), dtype=theano.config.floatX), borrow=True, name="b_cov")
cov_out = T.dot(input, self.W) + self.b if use_bias else T.dot(input, self.W)
activation_out = activation(cov_out)
if concat_out:
unigram_sum = T.sum(activation_out, axis=1)
self.output = unigram_sum.flatten(2)
else:
pooled_out = T.mean(activation_out, axis=2) if mean_pool else T.max(activation_out, axis=2)
if self.mask is not None:
unigram_sum = (pooled_out * self.mask[:, :, None]).sum(axis=1)
unigram_avg = unigram_sum / self.mask.sum(axis=1)[:, None]
else:
unigram_sum = T.sum(pooled_out, axis=1)
unigram_avg = unigram_sum / input_shape[0]
if pool_out:
self.output = pooled_out
elif sum_out:
self.output = unigram_sum
else:
self.output = unigram_avg
self.params = [self.W, self.b] if use_bias else [self.W]
