本文整理汇总了Python中util.projection方法的典型用法代码示例。如果您正苦于以下问题：Python util.projection方法的具体用法？Python util.projection怎么用？Python util.projection使用的例子？那么恭喜您, 这里精选的方法代码示例或许可以为您提供帮助。您也可以进一步了解该方法所在类util的用法示例。
在下文中一共展示了util.projection方法的13个代码示例，这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞，您的评价将有助于系统推荐出更棒的Python代码示例。
示例1: coarse_to_fine_pruning
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def coarse_to_fine_pruning(self, top_span_emb, top_span_mention_scores, c):
k = util.shape(top_span_emb, 0)
top_span_range = tf.range(k) # [k]
antecedent_offsets = tf.expand_dims(top_span_range, 1) - tf.expand_dims(top_span_range, 0) # [k, k]
antecedents_mask = antecedent_offsets >= 1 # [k, k]
fast_antecedent_scores = tf.expand_dims(top_span_mention_scores, 1) + tf.expand_dims(top_span_mention_scores, 0) # [k, k]
fast_antecedent_scores += tf.log(tf.to_float(antecedents_mask)) # [k, k]
fast_antecedent_scores += self.get_fast_antecedent_scores(top_span_emb) # [k, k]
if self.config['use_prior']:
antecedent_distance_buckets = self.bucket_distance(antecedent_offsets) # [k, c]
distance_scores = util.projection(tf.nn.dropout(tf.get_variable("antecedent_distance_emb", [10, self.config["feature_size"]]), self.dropout), 1, initializer=tf.truncated_normal_initializer(stddev=0.02)) #[10, 1]
antecedent_distance_scores = tf.gather(tf.squeeze(distance_scores, 1), antecedent_distance_buckets) # [k, c]
fast_antecedent_scores += antecedent_distance_scores
_, top_antecedents = tf.nn.top_k(fast_antecedent_scores, c, sorted=False) # [k, c]
top_antecedents_mask = util.batch_gather(antecedents_mask, top_antecedents) # [k, c]
top_fast_antecedent_scores = util.batch_gather(fast_antecedent_scores, top_antecedents) # [k, c]
top_antecedent_offsets = util.batch_gather(antecedent_offsets, top_antecedents) # [k, c]
return top_antecedents, top_antecedents_mask, top_fast_antecedent_scores, top_antecedent_offsets 示例2: coarse_to_fine_pruning
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def coarse_to_fine_pruning(self, top_span_emb, top_span_mention_scores, c):
k = util.shape(top_span_emb, 0)
top_span_range = tf.range(k) # [k]
antecedent_offsets = tf.expand_dims(top_span_range, 1) - tf.expand_dims(top_span_range, 0) # [k, k]
antecedents_mask = antecedent_offsets >= 1 # [k, k]
fast_antecedent_scores = tf.expand_dims(top_span_mention_scores, 1) + tf.expand_dims(top_span_mention_scores, 0) # [k, k]
fast_antecedent_scores += tf.log(tf.to_float(antecedents_mask)) # [k, k]
fast_antecedent_scores += self.get_fast_antecedent_scores(top_span_emb) # [k, k]
if self.config['use_prior']:
antecedent_distance_buckets = self.bucket_distance(antecedent_offsets) # [k, c]
distance_scores = util.projection(tf.nn.dropout(tf.get_variable("antecedent_distance_emb", [10, self.config["feature_size"]], initializer=tf.truncated_normal_initializer(stddev=0.02)), self.dropout), 1, initializer=tf.truncated_normal_initializer(stddev=0.02)) #[10, 1]
antecedent_distance_scores = tf.gather(tf.squeeze(distance_scores, 1), antecedent_distance_buckets) # [k, c]
fast_antecedent_scores += antecedent_distance_scores
_, top_antecedents = tf.nn.top_k(fast_antecedent_scores, c, sorted=False) # [k, c]
top_antecedents_mask = util.batch_gather(antecedents_mask, top_antecedents) # [k, c]
top_fast_antecedent_scores = util.batch_gather(fast_antecedent_scores, top_antecedents) # [k, c]
top_antecedent_offsets = util.batch_gather(antecedent_offsets, top_antecedents) # [k, c]
return top_antecedents, top_antecedents_mask, top_fast_antecedent_scores, top_antecedent_offsets 示例3: get_span_emb
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def get_span_emb(self, head_emb, context_outputs, span_starts, span_ends):
span_emb_list = []
span_start_emb = tf.gather(context_outputs, span_starts) # [k, emb]
span_emb_list.append(span_start_emb)
span_end_emb = tf.gather(context_outputs, span_ends) # [k, emb]
span_emb_list.append(span_end_emb)
span_width = 1 + span_ends - span_starts # [k]
if self.config["use_features"]:
span_width_index = span_width - 1 # [k]
span_width_emb = tf.gather(tf.get_variable("span_width_embeddings", [self.config["max_span_width"], self.config["feature_size"]]), span_width_index) # [k, emb]
span_width_emb = tf.nn.dropout(span_width_emb, self.dropout)
span_emb_list.append(span_width_emb)
if self.config["model_heads"]:
span_indices = tf.expand_dims(tf.range(self.config["max_span_width"]), 0) + tf.expand_dims(span_starts, 1) # [k, max_span_width]
span_indices = tf.minimum(util.shape(context_outputs, 0) - 1, span_indices) # [k, max_span_width]
span_text_emb = tf.gather(head_emb, span_indices) # [k, max_span_width, emb]
with tf.variable_scope("head_scores"):
self.head_scores = util.projection(context_outputs, 1) # [num_words, 1]
span_head_scores = tf.gather(self.head_scores, span_indices) # [k, max_span_width, 1]
span_mask = tf.expand_dims(tf.sequence_mask(span_width, self.config["max_span_width"], dtype=tf.float32), 2) # [k, max_span_width, 1]
span_head_scores += tf.log(span_mask) # [k, max_span_width, 1]
span_attention = tf.nn.softmax(span_head_scores, 1) # [k, max_span_width, 1]
span_head_emb = tf.reduce_sum(span_attention * span_text_emb, 1) # [k, emb]
span_emb_list.append(span_head_emb)
span_emb = tf.concat(span_emb_list, 1) # [k, emb]
return span_emb # [k, emb] 示例4: get_fast_antecedent_scores
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def get_fast_antecedent_scores(self, top_span_emb):
with tf.variable_scope("src_projection"):
source_top_span_emb = tf.nn.dropout(util.projection(top_span_emb, util.shape(top_span_emb, -1)), self.dropout) # [k, emb]
target_top_span_emb = tf.nn.dropout(top_span_emb, self.dropout) # [k, emb]
return tf.matmul(source_top_span_emb, target_top_span_emb, transpose_b=True) # [k, k] 示例5: lstm_contextualize
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def lstm_contextualize(self, text_emb, text_len, text_len_mask):
num_sentences = tf.shape(text_emb)[0]
current_inputs = text_emb # [num_sentences, max_sentence_length, emb]
for layer in range(self.config["contextualization_layers"]):
with tf.variable_scope("layer_{}".format(layer)):
with tf.variable_scope("fw_cell"):
cell_fw = util.CustomLSTMCell(self.config["contextualization_size"], num_sentences, self.lstm_dropout)
with tf.variable_scope("bw_cell"):
cell_bw = util.CustomLSTMCell(self.config["contextualization_size"], num_sentences, self.lstm_dropout)
state_fw = tf.contrib.rnn.LSTMStateTuple(tf.tile(cell_fw.initial_state.c, [num_sentences, 1]), tf.tile(cell_fw.initial_state.h, [num_sentences, 1]))
state_bw = tf.contrib.rnn.LSTMStateTuple(tf.tile(cell_bw.initial_state.c, [num_sentences, 1]), tf.tile(cell_bw.initial_state.h, [num_sentences, 1]))
(fw_outputs, bw_outputs), _ = tf.nn.bidirectional_dynamic_rnn(
cell_fw=cell_fw,
cell_bw=cell_bw,
inputs=current_inputs,
sequence_length=text_len,
initial_state_fw=state_fw,
initial_state_bw=state_bw)
text_outputs = tf.concat([fw_outputs, bw_outputs], 2) # [num_sentences, max_sentence_length, emb]
text_outputs = tf.nn.dropout(text_outputs, self.lstm_dropout)
if layer > 0:
highway_gates = tf.sigmoid(util.projection(text_outputs, util.shape(text_outputs, 2))) # [num_sentences, max_sentence_length, emb]
text_outputs = highway_gates * text_outputs + (1 - highway_gates) * current_inputs
current_inputs = text_outputs
return self.flatten_emb_by_sentence(text_outputs, text_len_mask) 示例6: get_span_emb
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def get_span_emb(self, head_emb, context_outputs, span_starts, span_ends):
span_emb_list = []
span_start_emb = tf.gather(context_outputs, span_starts) # [k, emb]
span_emb_list.append(span_start_emb)
span_end_emb = tf.gather(context_outputs, span_ends) # [k, emb]
span_emb_list.append(span_end_emb)
span_width = 1 + span_ends - span_starts # [k]
if self.config["use_features"]:
span_width_index = span_width - 1 # [k]
span_width_emb = tf.gather(
tf.get_variable("span_width_embeddings", [self.config["max_span_width"], self.config["feature_size"]]),
span_width_index) # [k, emb]
span_width_emb = tf.nn.dropout(span_width_emb, self.dropout)
span_emb_list.append(span_width_emb)
if self.config["model_heads"]:
span_indices = tf.expand_dims(tf.range(self.config["max_span_width"]), 0) + tf.expand_dims(span_starts,
1) # [k, max_span_width]
span_indices = tf.minimum(util.shape(context_outputs, 0) - 1, span_indices) # [k, max_span_width]
span_text_emb = tf.gather(head_emb, span_indices) # [k, max_span_width, emb]
with tf.variable_scope("head_scores"):
self.head_scores = util.projection(context_outputs, 1) # [num_words, 1]
span_head_scores = tf.gather(self.head_scores, span_indices) # [k, max_span_width, 1]
span_mask = tf.expand_dims(tf.sequence_mask(span_width, self.config["max_span_width"], dtype=tf.float32),
2) # [k, max_span_width, 1]
span_head_scores += tf.log(span_mask) # [k, max_span_width, 1]
span_attention = tf.nn.softmax(span_head_scores, 1) # [k, max_span_width, 1]
span_head_emb = tf.reduce_sum(span_attention * span_text_emb, 1) # [k, emb]
span_emb_list.append(span_head_emb)
span_emb = tf.concat(span_emb_list, 1) # [k, emb]
return span_emb # [k, emb] 示例7: get_fast_antecedent_scores
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def get_fast_antecedent_scores(self, top_span_emb):
with tf.variable_scope("src_projection"):
source_top_span_emb = tf.nn.dropout(util.projection(top_span_emb, util.shape(top_span_emb, -1)),
self.dropout) # [k, emb]
target_top_span_emb = tf.nn.dropout(top_span_emb, self.dropout) # [k, emb]
return tf.matmul(source_top_span_emb, target_top_span_emb, transpose_b=True) # [k, k] 示例8: lstm_contextualize
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def lstm_contextualize(self, text_emb, text_len, text_len_mask):
num_sentences = tf.shape(text_emb)[0]
current_inputs = text_emb # [num_sentences, max_sentence_length, emb]
for layer in range(self.config["contextualization_layers"]):
with tf.variable_scope("layer_{}".format(layer)):
with tf.variable_scope("fw_cell"):
cell_fw = util.CustomLSTMCell(self.config["contextualization_size"], num_sentences,
self.lstm_dropout)
with tf.variable_scope("bw_cell"):
cell_bw = util.CustomLSTMCell(self.config["contextualization_size"], num_sentences,
self.lstm_dropout)
state_fw = tf.contrib.rnn.LSTMStateTuple(tf.tile(cell_fw.initial_state.c, [num_sentences, 1]),
tf.tile(cell_fw.initial_state.h, [num_sentences, 1]))
state_bw = tf.contrib.rnn.LSTMStateTuple(tf.tile(cell_bw.initial_state.c, [num_sentences, 1]),
tf.tile(cell_bw.initial_state.h, [num_sentences, 1]))
(fw_outputs, bw_outputs), _ = tf.nn.bidirectional_dynamic_rnn(
cell_fw=cell_fw,
cell_bw=cell_bw,
inputs=current_inputs,
sequence_length=text_len,
initial_state_fw=state_fw,
initial_state_bw=state_bw)
text_outputs = tf.concat([fw_outputs, bw_outputs], 2) # [num_sentences, max_sentence_length, emb]
text_outputs = tf.nn.dropout(text_outputs, self.lstm_dropout)
if layer > 0:
highway_gates = tf.sigmoid(util.projection(text_outputs, util.shape(text_outputs,
2))) # [num_sentences, max_sentence_length, emb]
text_outputs = highway_gates * text_outputs + (1 - highway_gates) * current_inputs
current_inputs = text_outputs
return self.flatten_emb_by_sentence(text_outputs, text_len_mask) 示例9: get_span_emb
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def get_span_emb(self, head_emb, context_outputs, span_starts, span_ends):
span_emb_list = []
span_start_emb = tf.gather(context_outputs, span_starts) # [k, emb]
span_emb_list.append(span_start_emb)
span_end_emb = tf.gather(context_outputs, span_ends) # [k, emb]
span_emb_list.append(span_end_emb)
span_width = 1 + span_ends - span_starts # [k]
if self.config["use_features"]:
span_width_index = span_width - 1 # [k]
span_width_emb = tf.gather(
tf.get_variable("span_width_embeddings", [self.config["max_span_width"], self.config["feature_size"]],
initializer=create_initializer()),
span_width_index) # [k, emb]
span_width_emb = tf.nn.dropout(span_width_emb, self.dropout)
span_emb_list.append(span_width_emb)
if self.config["model_heads"]:
span_indices = tf.expand_dims(tf.range(self.config["max_span_width"]), 0) + tf.expand_dims(span_starts,
1) # [k, max_span_width]
span_indices = tf.minimum(util.shape(context_outputs, 0) - 1, span_indices) # [k, max_span_width]
span_text_emb = tf.gather(head_emb, span_indices) # [k, max_span_width, emb]
with tf.variable_scope("head_scores"):
self.head_scores = util.projection(context_outputs, 1) # [num_words, 1]
span_head_scores = tf.gather(self.head_scores, span_indices) # [k, max_span_width, 1]
span_mask = tf.expand_dims(tf.sequence_mask(span_width, self.config["max_span_width"], dtype=tf.float32),
2) # [k, max_span_width, 1]
span_head_scores += tf.log(span_mask) # [k, max_span_width, 1]
span_attention = tf.nn.softmax(span_head_scores, 1) # [k, max_span_width, 1]
span_head_emb = tf.reduce_sum(span_attention * span_text_emb, 1) # [k, emb]
span_emb_list.append(span_head_emb)
span_emb = tf.concat(span_emb_list, 1) # [k, emb]
return span_emb # [k, emb] 示例10: lstm_contextualize
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def lstm_contextualize(text_emb, text_len, config, lstm_dropout):
num_sentences = tf.shape(text_emb)[0]
current_inputs = text_emb # [num_sentences, max_sentence_length, emb]
for layer in xrange(config["contextualization_layers"]):
with tf.variable_scope("layer_{}".format(layer)):
with tf.variable_scope("fw_cell"):
cell_fw = util.CustomLSTMCell(config["contextualization_size"], num_sentences, lstm_dropout)
with tf.variable_scope("bw_cell"):
cell_bw = util.CustomLSTMCell(config["contextualization_size"], num_sentences, lstm_dropout)
state_fw = tf.contrib.rnn.LSTMStateTuple(tf.tile(cell_fw.initial_state.c, [num_sentences, 1]),
tf.tile(cell_fw.initial_state.h, [num_sentences, 1]))
state_bw = tf.contrib.rnn.LSTMStateTuple(tf.tile(cell_bw.initial_state.c, [num_sentences, 1]),
tf.tile(cell_bw.initial_state.h, [num_sentences, 1]))
(fw_outputs, bw_outputs), _ = tf.nn.bidirectional_dynamic_rnn(
cell_fw=cell_fw,
cell_bw=cell_bw,
inputs=current_inputs,
sequence_length=text_len,
initial_state_fw=state_fw,
initial_state_bw=state_bw)
text_outputs = tf.concat([fw_outputs, bw_outputs], 2) # [num_sentences, max_sentence_length, emb]
text_outputs = tf.nn.dropout(text_outputs, lstm_dropout)
if layer > 0:
highway_gates = tf.sigmoid(util.projection(
text_outputs, util.shape(text_outputs, 2))) # [num_sentences, max_sentence_length, emb]
text_outputs = highway_gates * text_outputs + (1 - highway_gates) * current_inputs
current_inputs = text_outputs
return text_outputs # [num_sentences, max_sentence_length, emb] 示例11: get_masked_mention_word_scores
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def get_masked_mention_word_scores(self, encoded_doc, span_starts, span_ends):
num_words = util.shape(encoded_doc, 0) # T
num_c = util.shape(span_starts, 0) # NC
doc_range = tf.tile(tf.expand_dims(tf.range(0, num_words), 0), [num_c, 1]) # [K, T]
mention_mask = tf.logical_and(doc_range >= tf.expand_dims(span_starts, 1), doc_range <= tf.expand_dims(span_ends, 1)) #[K, T]
with tf.variable_scope("mention_word_attn"):
word_attn = tf.squeeze(util.projection(encoded_doc, 1, initializer=tf.truncated_normal_initializer(stddev=0.02)), 1)
mention_word_attn = tf.nn.softmax(tf.log(tf.to_float(mention_mask)) + tf.expand_dims(word_attn, 0))
return mention_word_attn 示例12: combine_passes
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def combine_passes(self, original_doc, input_ids, input_mask, overlap_doc, overlap_ids, overlap_mask):
overlap_mask, input_mask = tf.equal(overlap_mask, 1), tf.equal(input_mask, 1)
org_content_mask = tf.logical_and(input_mask, tf.logical_and(tf.not_equal(input_ids, self.cls), tf.not_equal(input_ids, self.sep)))
overlap_content_mask = tf.logical_and(overlap_mask, tf.logical_and(tf.not_equal(overlap_ids, self.cls), tf.not_equal(overlap_ids, self.sep)))
flat_org_doc = self.flatten_emb_by_sentence(original_doc, org_content_mask)
flat_overlap_doc = self.flatten_emb_by_sentence(overlap_doc, overlap_content_mask)
with tf.variable_scope("combo"):
f = tf.sigmoid(util.projection(tf.concat([flat_org_doc, flat_overlap_doc], -1), util.shape(flat_org_doc, -1))) # [n, emb]
combo = f * flat_org_doc + (1 - f) * flat_overlap_doc
return combo, org_content_mask 示例13: get_span_emb
# 需要导入模块: import util [as 别名]
# 或者: from util import projection [as 别名]
def get_span_emb(head_emb, context_outputs, span_starts, span_ends, config, dropout):
"""Compute span representation shared across tasks.
Args:
head_emb: Tensor of [num_words, emb]
context_outputs: Tensor of [num_words, emb]
span_starts: [num_spans]
span_ends: [num_spans]
"""
text_length = util.shape(context_outputs, 0)
num_spans = util.shape(span_starts, 0)
span_start_emb = tf.gather(context_outputs, span_starts) # [num_words, emb]
span_end_emb = tf.gather(context_outputs, span_ends) # [num_words, emb]
span_emb_list = [span_start_emb, span_end_emb]
span_width = 1 + span_ends - span_starts # [num_spans]
max_arg_width = config["max_arg_width"]
num_heads = config["num_attention_heads"]
if config["use_features"]:
span_width_index = span_width - 1 # [num_spans]
span_width_emb = tf.gather(
tf.get_variable("span_width_embeddings", [max_arg_width, config["feature_size"]]),
span_width_index) # [num_spans, emb]
span_width_emb = tf.nn.dropout(span_width_emb, dropout)
span_emb_list.append(span_width_emb)
head_scores = None
span_text_emb = None
span_indices = None
span_indices_log_mask = None
if config["model_heads"]:
span_indices = tf.minimum(
tf.expand_dims(tf.range(max_arg_width), 0) + tf.expand_dims(span_starts, 1),
text_length - 1) # [num_spans, max_span_width]
span_text_emb = tf.gather(head_emb, span_indices) # [num_spans, max_arg_width, emb]
span_indices_log_mask = tf.log(
tf.sequence_mask(span_width, max_arg_width, dtype=tf.float32)) # [num_spans, max_arg_width]
with tf.variable_scope("head_scores"):
head_scores = util.projection(context_outputs, num_heads) # [num_words, num_heads]
span_attention = tf.nn.softmax(
tf.gather(head_scores, span_indices) + tf.expand_dims(span_indices_log_mask, 2),
dim=1) # [num_spans, max_arg_width, num_heads]
span_head_emb = tf.reduce_sum(span_attention * span_text_emb, 1) # [num_spans, emb]
span_emb_list.append(span_head_emb)
span_emb = tf.concat(span_emb_list, 1) # [num_spans, emb]
return span_emb, head_scores, span_text_emb, span_indices, span_indices_log_mask