def higher_order_tensors(step):
# We're not limited to passing scalar tensors to the summary
# operation. If we pass a rank-1 or rank-2 tensor, it'll be visualized
# as a table in TensorBoard. (For higher-ranked tensors, you'll see
# just a 2D slice of the data.)
#
# To demonstrate this, let's create a multiplication table.
# First, we'll create the table body, a `step`-by-`step` array of
# strings.
numbers = tf.range(step)
numbers_row = tf.expand_dims(numbers, 0) # shape: [1, step]
numbers_column = tf.expand_dims(numbers, 1) # shape: [step, 1]
products = tf.matmul(numbers_column, numbers_row) # shape: [step, step]
table_body = tf.as_string(products)
# Next, we'll create a header row and column, and a little
# multiplication sign to put in the corner.
bold_numbers = tf.string_join(['**', tf.as_string(numbers), '**'])
bold_row = tf.expand_dims(bold_numbers, 0)
bold_column = tf.expand_dims(bold_numbers, 1)
corner_cell = tf.constant(u'\u00d7'.encode('utf-8')) # MULTIPLICATION SIGN
# Now, we have to put the pieces together. Using `axis=0` stacks
# vertically; using `axis=1` juxtaposes horizontally.
table_body_and_top_row = tf.concat([bold_row, table_body], axis=0)
table_left_column = tf.concat([[[corner_cell]], bold_column], axis=0)
table_full = tf.concat([table_left_column, table_body_and_top_row], axis=1)
# The result, `table_full`, is a rank-2 string tensor of shape
# `[step + 1, step + 1]`. We can pass it directly to the summary, and
# we'll get a nicely formatted table in TensorBoard.
tf.summary.text('multiplication_table', table_full)
def bisine_wahwah_wave(frequency):
"""Emit two sine waves with balance oscillating left and right."""
#
# This is clearly intended to build on the bisine wave defined above,
# so we can start by generating that.
waves_a = bisine_wave(frequency)
#
# Then, by reversing axis 2, we swap the stereo channels. By mixing
# this with `waves_a`, we'll be able to create the desired effect.
waves_b = tf.reverse(waves_a, axis=[2])
#
# Let's have the balance oscillate from left to right four times.
iterations = 4
#
# Now, we compute the balance for each sample: `ts` has values
# in [0, 1] that indicate how much we should use `waves_a`.
xs = tf.reshape(tf.range(_samples(), dtype=tf.float32), [1, _samples(), 1])
thetas = xs / _samples() * iterations
ts = (tf.sin(math.pi * 2 * thetas) + 1) / 2
#
# Finally, we can mix the two together, and we're done.
wave = ts * waves_a + (1.0 - ts) * waves_b
#
# Alternately, we can make the effect more pronounced by exaggerating
# the sample data. Let's emit both variations.
exaggerated_wave = wave**3.0
return tf.concat([wave, exaggerated_wave], axis=0)
def op(name,
images,
max_outputs=3,
display_name=None,
description=None,
collections=None):
"""Create an image summary op for use in a TensorFlow graph.
Arguments:
name: A unique name for the generated summary node.
images: A `Tensor` representing pixel data with shape `[k, h, w, c]`,
where `k` is the number of images, `h` and `w` are the height and
width of the images, and `c` is the number of channels, which
should be 1, 3, or 4. Any of the dimensions may be statically
unknown (i.e., `None`).
max_outputs: Optional `int` or rank-0 integer `Tensor`. At most this
many images will be emitted at each step. When more than
`max_outputs` many images are provided, the first `max_outputs` many
images will be used and the rest silently discarded.
display_name: Optional name for this summary in TensorBoard, as a
constant `str`. Defaults to `name`.
description: Optional long-form description for this summary, as a
constant `str`. Markdown is supported. Defaults to empty.
collections: Optional list of graph collections keys. The new
summary op is added to these collections. Defaults to
`[Graph Keys.SUMMARIES]`.
Returns:
A TensorFlow summary op.
"""
if display_name is None:
display_name = name
summary_metadata = metadata.create_summary_metadata(
display_name=display_name, description=description)
with tf.name_scope(name), \
tf.control_dependencies([tf.assert_rank(images, 4),
tf.assert_type(images, tf.uint8),
tf.assert_non_negative(max_outputs)]):
limited_images = images[:max_outputs]
encoded_images = tf.map_fn(tf.image.encode_png,
limited_images,
dtype=tf.string,
name='encode_each_image')
image_shape = tf.shape(images)
dimensions = tf.stack([
tf.as_string(image_shape[2], name='width'),
tf.as_string(image_shape[1], name='height')
],
name='dimensions')
tensor = tf.concat([dimensions, encoded_images], axis=0)
return tf.summary.tensor_summary(name='image_summary',
tensor=tensor,
collections=collections,
summary_metadata=summary_metadata)
def bisine_wave(frequency):
"""Emit two sine waves, in stereo at different octaves."""
#
# We can first our existing sine generator to generate two different
# waves.
f_hi = frequency
f_lo = frequency / 2.0
with tf.name_scope('hi'):
sine_hi = sine_wave(f_hi)
with tf.name_scope('lo'):
sine_lo = sine_wave(f_lo)
#
# Now, we have two tensors of shape [1, _samples(), 1]. By concatenating
# them along axis 2, we get a tensor of shape [1, _samples(), 2]---a
# stereo waveform.
return tf.concat([sine_lo, sine_hi], axis=2)
def run_all(logdir, verbose=False):
"""Generate a bunch of histogram data, and write it to logdir."""
del verbose
tf.set_random_seed(0)
k = tf.placeholder(tf.float32)
# Make a normal distribution, with a shifting mean
mean_moving_normal = tf.random_normal(shape=[1000], mean=(5*k), stddev=1)
# Record that distribution into a histogram summary
histogram_summary.op("normal/moving_mean",
mean_moving_normal,
description="A normal distribution whose mean changes "
"over time.")
# Make a normal distribution with shrinking variance
shrinking_normal = tf.random_normal(shape=[1000], mean=0, stddev=1-(k))
# Record that distribution too
histogram_summary.op("normal/shrinking_variance", shrinking_normal,
description="A normal distribution whose variance "
"shrinks over time.")
# Let's combine both of those distributions into one dataset
normal_combined = tf.concat([mean_moving_normal, shrinking_normal], 0)
# We add another histogram summary to record the combined distribution
histogram_summary.op("normal/bimodal", normal_combined,
description="A combination of two normal distributions, "
"one with a moving mean and one with "
"shrinking variance. The result is a "
"distribution that starts as unimodal and "
"becomes more and more bimodal over time.")
# Add a gamma distribution
gamma = tf.random_gamma(shape=[1000], alpha=k)
histogram_summary.op("gamma", gamma,
description="A gamma distribution whose shape "
"parameter, α, changes over time.")
# And a poisson distribution
poisson = tf.random_poisson(shape=[1000], lam=k)
histogram_summary.op("poisson", poisson,
description="A Poisson distribution, which only "
"takes on integer values.")
# And a uniform distribution
uniform = tf.random_uniform(shape=[1000], maxval=k*10)
histogram_summary.op("uniform", uniform,
description="A simple uniform distribution.")
# Finally, combine everything together!
all_distributions = [mean_moving_normal, shrinking_normal,
gamma, poisson, uniform]
all_combined = tf.concat(all_distributions, 0)
histogram_summary.op("all_combined", all_combined,
description="An amalgamation of five distributions: a "
"uniform distribution, a gamma "
"distribution, a Poisson distribution, and "
"two normal distributions.")
summaries = tf.summary.merge_all()
# Setup a session and summary writer
sess = tf.Session()
writer = tf.summary.FileWriter(logdir)
# Setup a loop and write the summaries to disk
N = 400
for step in xrange(N):
k_val = step/float(N)
summ = sess.run(summaries, feed_dict={k: k_val})
writer.add_summary(summ, global_step=step)
def start_runs(logdir,
steps,
run_name,
thresholds,
mask_every_other_prediction=False):
"""Generate a PR curve with precision and recall evenly weighted.
Arguments:
logdir: The directory into which to store all the runs' data.
steps: The number of steps to run for.
run_name: The name of the run.
thresholds: The number of thresholds to use for PR curves.
mask_every_other_prediction: Whether to mask every other prediction by
alternating weights between 0 and 1.
"""
tf.reset_default_graph()
tf.set_random_seed(42)
# Create a normal distribution layer used to generate true color labels.
distribution = tf.distributions.Normal(loc=0., scale=142.)
# Sample the distribution to generate colors. Lets generate different numbers
# of each color. The first dimension is the count of examples.
# The calls to sample() are given fixed random seed values that are "magic"
# in that they correspond to the default seeds for those ops when the PR
# curve test (which depends on this code) was written. We've pinned these
# instead of continuing to use the defaults since the defaults are based on
# node IDs from the sequence of nodes added to the graph, which can silently
# change when this code or any TF op implementations it uses are modified.
# TODO(nickfelt): redo the PR curve test to avoid reliance on random seeds.
# Generate reds.
number_of_reds = 100
true_reds = tf.clip_by_value(
tf.concat([
255 - tf.abs(distribution.sample([number_of_reds, 1], seed=11)),
tf.abs(distribution.sample([number_of_reds, 2], seed=34))
],
axis=1), 0, 255)
# Generate greens.
number_of_greens = 200
true_greens = tf.clip_by_value(
tf.concat([
tf.abs(distribution.sample([number_of_greens, 1], seed=61)),
255 - tf.abs(distribution.sample([number_of_greens, 1], seed=82)),
tf.abs(distribution.sample([number_of_greens, 1], seed=105))
],
axis=1), 0, 255)
# Generate blues.
number_of_blues = 150
true_blues = tf.clip_by_value(
tf.concat([
tf.abs(distribution.sample([number_of_blues, 2], seed=132)),
255 - tf.abs(distribution.sample([number_of_blues, 1], seed=153))
],
axis=1), 0, 255)
# Assign each color a vector of 3 booleans based on its true label.
labels = tf.concat([
tf.tile(tf.constant([[True, False, False]]), (number_of_reds, 1)),
tf.tile(tf.constant([[False, True, False]]), (number_of_greens, 1)),
tf.tile(tf.constant([[False, False, True]]), (number_of_blues, 1)),
],
axis=0)
# We introduce 3 normal distributions. They are used to predict whether a
# color falls under a certain class (based on distances from corners of the
# color triangle). The distributions vary per color. We have the distributions
# narrow over time.
initial_standard_deviations = [v + FLAGS.steps for v in (158, 200, 242)]
iteration = tf.placeholder(tf.int32, shape=[])
red_predictor = tf.distributions.Normal(
loc=0.,
scale=tf.cast(initial_standard_deviations[0] - iteration,
dtype=tf.float32))
green_predictor = tf.distributions.Normal(
loc=0.,
scale=tf.cast(initial_standard_deviations[1] - iteration,
dtype=tf.float32))
blue_predictor = tf.distributions.Normal(
loc=0.,
scale=tf.cast(initial_standard_deviations[2] - iteration,
dtype=tf.float32))
# Make predictions (assign 3 probabilities to each color based on each color's
# distance to each of the 3 corners). We seek double the area in the right
# tail of the normal distribution.
examples = tf.concat([true_reds, true_greens, true_blues], axis=0)
probabilities_colors_are_red = (1 - red_predictor.cdf(
tf.norm(examples - tf.constant([255., 0, 0]), axis=1))) * 2
probabilities_colors_are_green = (1 - green_predictor.cdf(
tf.norm(examples - tf.constant([0, 255., 0]), axis=1))) * 2
probabilities_colors_are_blue = (1 - blue_predictor.cdf(
tf.norm(examples - tf.constant([0, 0, 255.]), axis=1))) * 2
predictions = (probabilities_colors_are_red,
probabilities_colors_are_green,
probabilities_colors_are_blue)
# This is the crucial piece. We write data required for generating PR curves.
# We create 1 summary per class because we create 1 PR curve per class.
for i, color in enumerate(('red', 'green', 'blue')):
description = (
'The probabilities used to create this PR curve are '
'generated from a normal distribution. Its standard '
'deviation is initially %0.0f and decreases over time.' %
initial_standard_deviations[i])
weights = None
if mask_every_other_prediction:
# Assign a weight of 0 to every even-indexed prediction. Odd-indexed
# predictions are assigned a default weight of 1.
consecutive_indices = tf.reshape(tf.range(tf.size(predictions[i])),
tf.shape(predictions[i]))
weights = tf.cast(consecutive_indices % 2, dtype=tf.float32)
summary.op(name=color,
labels=labels[:, i],
predictions=predictions[i],
num_thresholds=thresholds,
weights=weights,
display_name='classifying %s' % color,
description=description)
merged_summary_op = tf.summary.merge_all()
events_directory = os.path.join(logdir, run_name)
sess = tf.Session()
writer = tf.summary.FileWriter(events_directory, sess.graph)
for step in xrange(steps):
feed_dict = {
iteration: step,
}
merged_summary = sess.run(merged_summary_op, feed_dict=feed_dict)
writer.add_summary(merged_summary, step)
writer.close()