本文整理汇总了Python中utils.util.KNNIdsWithDistances方法的典型用法代码示例。如果您正苦于以下问题：Python util.KNNIdsWithDistances方法的具体用法？Python util.KNNIdsWithDistances怎么用？Python util.KNNIdsWithDistances使用的例子？那么, 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utils.util的用法示例。
在下文中一共展示了util.KNNIdsWithDistances方法的2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: compute_average_alignment
# 需要导入模块: from utils import util [as 别名]
# 或者: from utils.util import KNNIdsWithDistances [as 别名]
def compute_average_alignment(
seqname_to_embeddings, num_views, summary_writer, training_step):
"""Computes the average cross-view alignment for all sequence view pairs.
Args:
seqname_to_embeddings: Dict, mapping sequence name to a
[num_views, embedding size] numpy matrix holding all embedded views.
num_views: Int, number of simultaneous views in the dataset.
summary_writer: A `SummaryWriter` object.
training_step: Int, the training step of the model used to embed images.
Alignment is the scaled absolute difference between the ground truth time
and the knn aligned time.
abs(|time_i - knn_time|) / sequence_length
"""
all_alignments = []
for _, view_embeddings in seqname_to_embeddings.iteritems():
for idx_i in range(num_views):
for idx_j in range(idx_i+1, num_views):
embeddings_view_i = view_embeddings[idx_i]
embeddings_view_j = view_embeddings[idx_j]
seq_len = len(embeddings_view_i)
times_i = np.array(range(seq_len))
# Get the nearest time_index for each embedding in view_i.
times_j = np.array([util.KNNIdsWithDistances(
q, embeddings_view_j, k=1)[0][0] for q in embeddings_view_i])
# Compute sequence view pair alignment.
alignment = np.mean(
np.abs(np.array(times_i)-np.array(times_j))/float(seq_len))
all_alignments.append(alignment)
print 'alignment so far %f' % alignment
average_alignment = np.mean(all_alignments)
print 'Average alignment %f' % average_alignment
summ = tf.Summary(value=[tf.Summary.Value(
tag='validation/alignment', simple_value=average_alignment)])
summary_writer.add_summary(summ, int(training_step)) 示例2: compute_average_alignment
# 需要导入模块: from utils import util [as 别名]
# 或者: from utils.util import KNNIdsWithDistances [as 别名]
def compute_average_alignment(
seqname_to_embeddings, num_views, summary_writer, training_step):
"""Computes the average cross-view alignment for all sequence view pairs.
Args:
seqname_to_embeddings: Dict, mapping sequence name to a
[num_views, embedding size] numpy matrix holding all embedded views.
num_views: Int, number of simultaneous views in the dataset.
summary_writer: A `SummaryWriter` object.
training_step: Int, the training step of the model used to embed images.
Alignment is the scaled absolute difference between the ground truth time
and the knn aligned time.
abs(|time_i - knn_time|) / sequence_length
"""
all_alignments = []
for _, view_embeddings in seqname_to_embeddings.iteritems():
for idx_i in range(num_views):
for idx_j in range(idx_i+1, num_views):
embeddings_view_i = view_embeddings[idx_i]
embeddings_view_j = view_embeddings[idx_j]
seq_len = len(embeddings_view_i)
times_i = np.array(range(seq_len))
# Get the nearest time_index for each embedding in view_i.
times_j = np.array([util.KNNIdsWithDistances(
q, embeddings_view_j, k=1)[0][0] for q in embeddings_view_i])
# Compute sequence view pair alignment.
alignment = np.mean(
np.abs(np.array(times_i)-np.array(times_j))/float(seq_len))
all_alignments.append(alignment)
print('alignment so far %f' % alignment)
average_alignment = np.mean(all_alignments)
print('Average alignment %f' % average_alignment)
summ = tf.Summary(value=[tf.Summary.Value(
tag='validation/alignment', simple_value=average_alignment)])
summary_writer.add_summary(summ, int(training_step))