def is_finished(self, step_idx, source_length, alive_log_probs, finished_scores, finished_in_finished):
"""
is_finished
"""
base_1 = layers.cast(source_length, 'float32') + 55.0
base_1 /= 6.0
max_length_penalty = layers.pow(base_1, self.alpha)
flat_alive_log_probs = layers.reshape(alive_log_probs, [-1])
lower_bound_alive_scores_1 = layers.gather(flat_alive_log_probs, [self.get_alive_index])
lower_bound_alive_scores = lower_bound_alive_scores_1 / max_length_penalty
lowest_score_of_finished_in_finish = layers.reduce_min(finished_scores * finished_in_finished, dim=1)
finished_in_finished = layers.cast(finished_in_finished, 'bool')
lowest_score_of_finished_in_finish += \
((1.0 - layers.cast(layers.reduce_any(finished_in_finished, 1), 'float32')) * -INF)
#print lowest_score_of_finished_in_finish
bound_is_met = layers.reduce_all(layers.greater_than(lowest_score_of_finished_in_finish,
lower_bound_alive_scores))
decode_length = source_length + 50
length_cond = layers.less_than(x=step_idx, y=decode_length)
return layers.logical_and(x=layers.logical_not(bound_is_met), y=length_cond)
def _push_to_stack(gmr_desc, gmr_pos, gmr_lens, gmr_stack_info):
"""push grammar id in gmr_desc from gmr_pos to gmr_lens to
gmr_stack. and update step_gmr_pos
Args:
gmr_desc (TYPE): NULL
gmr_pos (TYPE): NULL
gmr_lens (TYPE): NULL
gmr_stack_info (tuple): [in/out] (gmr_stack, gmr_stack_pos)
Returns: tuple (gmr_stack, gmr_stack_pos)
Raises: NULL
"""
gmr_stack, gmr_stack_pos = gmr_stack_info
mv_step = layers.cast(layers.greater_than(gmr_lens,
layers.zeros_like(gmr_lens)),
dtype=gmr_lens.dtype)
gmr_mv_pos = layers.elementwise_sub(gmr_lens, mv_step)
cond = layers.reduce_any(layers.greater_than(gmr_mv_pos, gmr_pos))
while_op = layers.While(cond)
with while_op.block():
gmr_ids = nn_utils.batch_gather(gmr_desc, gmr_mv_pos)
gmr_stack_tmp, gmr_stack_pos_tmp = data_structure.Stack.push(
gmr_stack_info, gmr_ids, in_place=False)
mv_cond = layers.greater_than(gmr_mv_pos, gmr_pos)
gmr_mv_pos_tmp = fluider.elementwise_sub(gmr_mv_pos,
mv_cond,
force=True)
new_gmr_stack, new_gmr_stack_pos = nn_utils.ifelse(
mv_cond, [gmr_stack_tmp, gmr_stack_pos_tmp],
[gmr_stack, gmr_stack_pos])
layers.utils.map_structure(layers.assign,
[new_gmr_stack, new_gmr_stack_pos],
[gmr_stack, gmr_stack_pos])
layers.assign(gmr_mv_pos_tmp, gmr_mv_pos)
layers.assign(
layers.reduce_any(layers.greater_than(gmr_mv_pos, gmr_pos)), cond)
return gmr_stack, gmr_stack_pos
def forward(self, pred, target):
target = 1 - target[:, 0]
batch_size, vector_size = pred.shape[0], pred.shape[1]
pred = L.l2_normalize(pred, axis=1, epsilon=1e-10)
square_norm = L.reduce_sum(L.square(pred), dim=1)
dist = L.elementwise_add(-2.0 * L.matmul(pred, pred, transpose_y=True),
square_norm,
axis=0)
dist = L.elementwise_add(dist, square_norm, axis=1)
dist = L.elementwise_max(dist, L.zeros_like(dist))
dist = L.sqrt(dist)
ap_dist = L.reshape(dist, (0, 0, 1))
an_dist = L.reshape(dist, (0, 1, -1))
loss = L.expand(ap_dist, (1, 1, batch_size)) - L.expand(
an_dist, (1, batch_size, 1)) + self.magin
indice_equal = L.diag(
L.fill_constant((batch_size, ), dtype='float32', value=1.0))
indice_not_equal = 1.0 - indice_equal
broad_matrix = L.expand(L.reshape(target, (-1, 1)),
(1, batch_size)) + L.expand(
L.reshape(target, (1, -1)),
(batch_size, 1))
pp = L.cast(L.equal(broad_matrix, L.zeros_like(broad_matrix)),
dtype='float32')
pp = L.reshape(indice_not_equal * pp, (0, 0, 1))
pn = L.cast(L.equal(broad_matrix,
L.zeros_like(broad_matrix) + 1),
dtype='float32')
pn = L.reshape(indice_not_equal * pn, (1, 0, -1))
apn = L.expand(pp,
(1, 1, batch_size)) * L.expand(pn, (batch_size, 1, 1))
loss = loss * L.cast(apn, dtype='float32')
loss = L.elementwise_max(loss, L.zeros_like(loss))
num_tri = L.reduce_sum(
L.cast(L.greater_than(loss, L.zeros_like(loss)), dtype='float32'))
loss = L.reduce_sum(loss) * self.loss_weight / (num_tri + 1e-16)
return loss
def pairwise_hinge(self):
"""pairwise model"""
poi_repr = L.split(self.poi_repr, 2, dim=0)
pos_repr, neg_repr = poi_repr
pos_pred = L.cos_sim(self.query_repr, pos_repr)
neg_pred = L.cos_sim(self.query_repr, neg_repr)
mode = 'hinge_loss'
# log(1 + e-z), max(0, 1 - z)
if 'hinge_loss' == mode:
theta_z = L.relu(1 + neg_pred - pos_pred)
elif 'logistic_loss' == mode:
theta_z = L.log(1 + L.exp(neg_pred - pos_pred))
self.loss = L.reduce_mean(theta_z)
pos_cnt = L.reduce_sum(L.cast(L.greater_than(pos_pred, neg_pred), dtype="float32"))
neg_cnt = L.reduce_sum(L.cast(L.less_than(pos_pred, neg_pred), dtype="float32"))
self.order = pos_cnt / (1e-5 + neg_cnt)
self.metrics = [self.loss, self.order]
def pairwise_loss(self):
"""pairwise model"""
# TODO: for neg_num neg poi, split num should be (neg_num + 1) on dim 0
poi_repr = L.split(self.poi_repr, [1 * self.batch_size, self.neg_num * self.batch_size], dim=0)
pos_repr, neg_repr = poi_repr
# size [-1 x emb_size]
# size [-1*n x emb_size]
prefix_expand = L.reshape(L.expand(self.query_repr, [1, self.neg_num]), [-1, self.hidden_size])
# size [-1*n x 1]
neg_pred_n = self.safe_cosine_sim(neg_repr, prefix_expand)
# size [-1 x 1]
pos_pred = self.safe_cosine_sim(pos_repr, self.query_repr)
cost = self.loss_neg_log_of_pos(pos_pred, L.reshape(neg_pred_n, [-1, self.neg_num]), 15)
self.loss = L.mean(x=cost)
# size [-1 x 1]
neg_avg = L.reduce_mean(L.reshape(neg_pred_n, [-1, self.neg_num]), dim=1, keep_dim=True)
pos_cnt = L.reduce_sum(L.cast(L.greater_than(pos_pred, neg_avg), dtype="float32"))
neg_cnt = L.reduce_sum(L.cast(L.less_than(pos_pred, neg_avg), dtype="float32"))
# equal to positive and negative order
self.order = pos_cnt / (1e-5 + neg_cnt)
self.metrics = [self.loss, self.order]
def early_finish(alive_log_probs, finished_scores,
finished_in_finished):
max_length_penalty = np.power(((5. + max_len) / 6.), alpha)
# The best possible score of the most likely alive sequence
lower_bound_alive_scores = alive_log_probs[:,
0] / max_length_penalty
# Now to compute the lowest score of a finished sequence in finished
# If the sequence isn't finished, we multiply it's score by 0. since
# scores are all -ve, taking the min will give us the score of the lowest
# finished item.
lowest_score_of_fininshed_in_finished = layers.reduce_min(
finished_scores * finished_in_finished, 1)
# If none of the sequences have finished, then the min will be 0 and
# we have to replace it by -ve INF if it is. The score of any seq in alive
# will be much higher than -ve INF and the termination condition will not
# be met.
lowest_score_of_fininshed_in_finished += (
1. - layers.reduce_max(finished_in_finished, 1)) * -inf
bound_is_met = layers.reduce_all(
layers.greater_than(lowest_score_of_fininshed_in_finished,
lower_bound_alive_scores))
return bound_is_met
def get_mask(seq, padding_idx=0):
pix = layers.unsqueeze(layers.ones_like(seq) * padding_idx, axes=2)
mask = layers.cast(layers.greater_than(layers.unsqueeze(seq, axes=2), pix), 'float32')
return mask