def sg_summary_param(tensor, prefix='40. parameters'):
# defaults
prefix = '' if prefix is None else prefix + '/'
# summary name
name = prefix + _pretty_name(tensor)
# summary statistics
with tf.name_scope('summary'):
tf.scalar_summary(name + '/norm', tf.global_norm([tensor]))
tf.histogram_summary(name, tensor)
def sg_summary_metric(tensor, prefix='20. metric'):
# defaults
prefix = '' if prefix is None else prefix + '/'
# summary name
name = prefix + _pretty_name(tensor)
# summary statistics
with tf.name_scope('summary'):
tf.scalar_summary(name + '/avg', tf.reduce_mean(tensor))
tf.histogram_summary(name, tensor)
def sg_summary_gradient(tensor, gradient, prefix='50. gradient'):
# defaults
prefix = '' if prefix is None else prefix + '/'
# summary name
name = prefix + _pretty_name(tensor)
# summary statistics
with tf.name_scope('summary'):
tf.scalar_summary(name + '/norm', tf.global_norm([gradient]))
tf.histogram_summary(name, gradient)
def sg_summary_metric(tensor, prefix='20. metric'):
r"""Writes the average of `tensor` (=metric such as accuracy).
"""
# defaults
prefix = '' if prefix is None else prefix + '/'
# summary name
name = prefix + _pretty_name(tensor)
# summary statistics
with tf.name_scope('summary'):
tf.scalar_summary(name + '/avg', tf.reduce_mean(tensor))
tf.histogram_summary(name, tensor)
def sg_summary_activation(tensor, prefix='30. activation'):
# defaults
prefix = '' if prefix is None else prefix + '/'
# summary name
name = prefix + _pretty_name(tensor)
# summary statistics
with tf.name_scope('summary'):
tf.scalar_summary(name + '/norm', tf.global_norm([tensor]))
tf.scalar_summary(
name + '/ratio',
tf.reduce_mean(tf.cast(tf.greater(tensor, 0), tf.sg_floatx)))
tf.histogram_summary(name, tensor)
def sg_summary_gradient(tensor, gradient, prefix='50. gradient'):
r"""Writes the normalized gradient value
Args:
tensor: A `Tensor` variable.
gradient: A `Tensor`. Gradient of `tensor`.
"""
# defaults
prefix = '' if prefix is None else prefix + '/'
# summary name
name = prefix + _pretty_name(tensor)
# summary statistics
with tf.name_scope('summary'):
try:
tf.scalar_summary(name + '/norm', tf.global_norm([gradient]))
tf.histogram_summary(name, gradient)
except:
pass