def summarize_weights_biases():
with tf.variable_scope('weights'):
for l in tf.get_collection('weights'):
tf.summary.histogram(hem.tensor_name(l), l)
tf.summary.scalar(hem.tensor_name(l) + '/sparsity', tf.nn.zero_fraction(l))
# montage_summary(l, name=tensor_name(l) + '/montage')
with tf.variable_scope('biases'):
for l in tf.get_collection('biases'):
tf.summary.histogram(hem.tensor_name(l), l)
tf.summary.scalar(hem.tensor_name(l) + '/sparsity', tf.nn.zero_fraction(l))
def summarize_activations(scope=None, summarize_montages=True):
with tf.variable_scope('activations'):
for l in tf.get_collection('conv_layers', scope=scope):
layer_name = hem.tensor_name(l)
l = tf.transpose(l, [0, 2, 3, 1])
tf.summary.histogram(layer_name, l)
tf.summary.scalar(layer_name + '/sparsity', tf.nn.zero_fraction(l))
tf.summary.scalar(layer_name + '/mean', tf.reduce_mean(l))
if summarize_montages:
montage(tf.transpose(l[0], [2, 0, 1]), name=layer_name + '/montage')
for l in tf.get_collection('dense_layers', scope=scope):
# l = tf.transpose(l, [0, 2, 3, 1])
layer_name = hem.tensor_name(l)
tf.summary.histogram(layer_name, l)
tf.summary.scalar(layer_name + '/sparsity', tf.nn.zero_fraction(l))
tf.summary.scalar(layer_name + '/mean', tf.reduce_mean(l))
def activation_summaries():
layers = tf.get_collection('conv_layers')
d_layers = [l for l in layers if 'discriminator' in l.name]
g_layers = [l for l in layers if 'generator' in l.name]
with tf.variable_scope('discriminator_activations'):
for l in d_layers:
layer_name = hem.tensor_name(l)
layer_name = '/'.join(layer_name.split('/')[0:2])
l = tf.transpose(l, [0, 2, 3, 1])
tf.summary.histogram(layer_name, l)
tf.summary.scalar(layer_name + '/sparsity', tf.nn.zero_fraction(l))
tf.summary.scalar(layer_name + '/mean', tf.reduce_mean(l))
with tf.variable_scope('generator_activations'):
for l in g_layers:
layer_name = hem.tensor_name(l)
layer_name = '/'.join(layer_name.split('/')[0:2])
l = tf.transpose(l, [0, 2, 3, 1])
tf.summary.histogram(layer_name, l)
tf.summary.scalar(layer_name + '/sparsity', tf.nn.zero_fraction(l))
tf.summary.scalar(layer_name + '/mean', tf.reduce_mean(l))
def summarize_gradients(grads_and_vars, name=None):
"""Adds histogram summaries for input list.
Args:
grads_and_vars: List, a list of tuples in form (grad, var).e
name: String, name to use for var scope.
"""
with tf.name_scope(name, 'gradients', grads_and_vars):
for g, v in grads_and_vars:
n = hem.tensor_name(v) + '/gradient'
tf.summary.histogram(n, g)
tf.summary.scalar(n, tf.reduce_mean(g))
def summarize_layers(scope, layers, montage=False):
with tf.variable_scope(scope):
# activations
for l in layers:
layer_name = hem.tensor_name(l)
layer_name = '/'.join(layer_name.split('/')[0:2])
l = tf.transpose(l, [0, 2, 3, 1])
tf.summary.histogram(layer_name, l)
tf.summary.scalar(layer_name + '/sparsity', tf.nn.zero_fraction(l))
tf.summary.scalar(layer_name + '/mean', tf.reduce_mean(l))
if montage:
hem.montage(tf.transpose(l[0], [2, 0, 1]), name=layer_name + '/montage')
def _summaries(g, x, args):
"""Add specialized summaries to the graph.
This adds summaries that:
- Track variability in generator samples given random noise vectors.
- Track how much of the noise vector is used.
- Generate examples from the real and learned distributions.
Args:
g: Tensor, the generator's output. i.e., the fake images.
x: Tensor, the real (input) images.
args: Argparse structure.
Returns:
None
"""
# 1. generate multiple samples using a single image
with tf.variable_scope(tf.get_variable_scope()):
gpu_id = 0
with tf.device(hem.variables_on_cpu(gpu_id)):
with tf.name_scope('tower_{}'.format(gpu_id)) as scope:
# print('input shape', x.shape)
with tf.variable_scope('generator'):
# using first image in batch, form new one with just this image
x_repeated = tf.stack([x[0]] * args.batch_size)
x_rgb, x_depth = _split(x_repeated)
# then create a new path for the generator using just this dataset
import copy
args_copy = copy.copy(args)
args_copy.batch_norm = False
d = generator(x_rgb, args_copy, reuse=True)
# scale to [0, 1]
sampler_rgb, sampler_depth = _split(d)
sampler_rgb = hem.rescale(sampler_rgb, (-1, 1), (0, 1))
sampler_depth = hem.rescale(sampler_depth, (-1, 1), (0, 1))
with tf.variable_scope('sampler'):
hem.montage(sampler_rgb[0:args.examples], 8, 8, name='images')
hem.montage(sampler_depth[0:args.examples], 8, 8, name='depths')
# and track the variance of the depth predictions
mean, var = tf.nn.moments(sampler_depth, axes=[0])
hem.scalars(('predicted_depth_mean', tf.reduce_mean(mean)),
('predicted_depth_var', tf.reduce_mean(var)))
# and the images (temporary, for calibration/sanity check)
x_rgb, x_depth = _split(x)
mean, var = tf.nn.moments(x_depth, axes=[0])
hem.scalars(('real_depth_mean', tf.reduce_mean(mean)),
('real_depth_var', tf.reduce_mean(var)))
sampler_rgb = tf.transpose(sampler_rgb, [0, 2, 3, 1])
sampler_depth = tf.transpose(sampler_depth, [0, 2, 3, 1])
# mean, var = tf.nn.moments(tf.image.rgb_to_grayscale(sampler_rgb), axes=[0])
axis = [0, -1]
mean, var = tf.nn.moments(tf.image.rgb_to_grayscale(sampler_rgb),
axes=[0])
hem.scalars(('image_mean', tf.reduce_mean(mean)),
('image_var', tf.reduce_mean(var)))
mean, var = tf.nn.moments(sampler_depth, axes=[0])
hem.scalars(('depth_mean', tf.reduce_mean(mean)),
('depth_var', tf.reduce_mean(var)))
# 2. generate summaries for real and fake images
with tf.variable_scope('examples'):
hem.histograms(('fake', g), ('real', x))
# rescale, split, and colorize
x = hem.rescale(x[0:args.examples], (-1, 1), (0, 1))
g = hem.rescale(g[0:args.examples], (-1, 1), (0, 1))
hem.histograms(('fake_rescaled', g), ('real_rescaled', x))
x_rgb, x_depth = _split(x)
g_rgb, g_depth = _split(g)
# note: these are rescaled to (0, 1)
hem.histograms(('real_depth', x_depth), ('fake_depth', g_depth),
('real_rgb', x_rgb), ('fake_rgb', g_rgb))
# add montages
# TODO shouldn't be fixed to 8, but to ceil(sqrt(args.examples))
hem.montage(x_rgb, 8, 8, name='real/images')
hem.montage(x_depth, 8, 8, name='real/depths')
hem.montage(g_rgb, 8, 8, name='fake/images')
hem.montage(g_depth, 8, 8, name='fake/depths')
# 3. add additional summaries for weights and biases in e_c1 (the initial noise layer)
# TODO don't iterate through list, but grab directly by full name
with tf.variable_scope('noise'):
for l in tf.get_collection('weights'):
if 'e_c1' in hem.tensor_name(l):
print(l, hem.tensor_name(l))
x_rgb, x_noise = _split(l)
hem.histograms(('rgb/weights', x_rgb),
('noise/weights', x_noise))
hem.scalars(
('rgb/weights/sparsity', tf.nn.zero_fraction(x_rgb)),
('noise/weights/sparsity', tf.nn.zero_fraction(x_noise)),
('noise/weights/mean', tf.reduce_mean(x_noise)),
('rgb/weights/mean', tf.reduce_mean(x_rgb)))
break
def gradient_summaries(grads, name='gradients'):
with tf.variable_scope(name):
for g, v in grads:
n = hem.tensor_name(v)
tf.summary.histogram(n, g)
tf.summary.scalar(n, tf.reduce_mean(g))
def loss_summaries():
for l in tf.get_collection('losses'):
tf.summary.scalar(hem.tensor_name(l), l)
tf.summary.histogram(hem.tensor_name(l), l)
def summarize_collection(name, scope):
"""Add a scalar summary for every tensor in a collection."""
collection = tf.get_collection(name, scope)
for x in collection:
tf.summary.scalar(hem.tensor_name(x), x)
return collection
def summarize_losses():
with tf.variable_scope('loss'):
for l in tf.get_collection('losses'):
tf.summary.scalar(hem.tensor_name(l), l)
tf.summary.histogram(hem.tensor_name(l), l)