Chapter 12 – Distributed TensorFlow
This notebook contains all the sample code and solutions to the exercises in chapter 12.
First, let's make sure this notebook works well in both python 2 and 3, import a few common modules, ensure MatplotLib plots figures inline and prepare a function to save the figures:
In [1]:
# To support both python 2 and python 3
from __future__ import division, print_function, unicode_literals
# Common imports
import numpy as np
import os
# to make this notebook's output stable across runs
def reset_graph(seed=42):
tf.reset_default_graph()
tf.set_random_seed(seed)
np.random.seed(seed)
# To plot pretty figures
%matplotlib inline
import matplotlib
import matplotlib.pyplot as plt
plt.rcParams['axes.labelsize'] = 14
plt.rcParams['xtick.labelsize'] = 12
plt.rcParams['ytick.labelsize'] = 12
# Where to save the figures
PROJECT_ROOT_DIR = "."
CHAPTER_ID = "distributed"
def save_fig(fig_id, tight_layout=True):
path = os.path.join(PROJECT_ROOT_DIR, "images", CHAPTER_ID, fig_id + ".png")
print("Saving figure", fig_id)
if tight_layout:
plt.tight_layout()
plt.savefig(path, format='png', dpi=300)
In [2]:
import tensorflow as tf
In [3]:
c = tf.constant("Hello distributed TensorFlow!")
server = tf.train.Server.create_local_server()
In [4]:
with tf.Session(server.target) as sess:
print(sess.run(c))
In [5]:
cluster_spec = tf.train.ClusterSpec({
"ps": [
"127.0.0.1:2221", # /job:ps/task:0
"127.0.0.1:2222", # /job:ps/task:1
],
"worker": [
"127.0.0.1:2223", # /job:worker/task:0
"127.0.0.1:2224", # /job:worker/task:1
"127.0.0.1:2225", # /job:worker/task:2
]})
In [6]:
task_ps0 = tf.train.Server(cluster_spec, job_name="ps", task_index=0)
task_ps1 = tf.train.Server(cluster_spec, job_name="ps", task_index=1)
task_worker0 = tf.train.Server(cluster_spec, job_name="worker", task_index=0)
task_worker1 = tf.train.Server(cluster_spec, job_name="worker", task_index=1)
task_worker2 = tf.train.Server(cluster_spec, job_name="worker", task_index=2)
In [7]:
reset_graph()
with tf.device("/job:ps"):
a = tf.Variable(1.0, name="a")
with tf.device("/job:worker"):
b = a + 2
with tf.device("/job:worker/task:1"):
c = a + b
In [8]:
with tf.Session("grpc://127.0.0.1:2221") as sess:
sess.run(a.initializer)
print(c.eval())
In [9]:
reset_graph()
with tf.device(tf.train.replica_device_setter(
ps_tasks=2,
ps_device="/job:ps",
worker_device="/job:worker")):
v1 = tf.Variable(1.0, name="v1") # pinned to /job:ps/task:0 (defaults to /cpu:0)
v2 = tf.Variable(2.0, name="v2") # pinned to /job:ps/task:1 (defaults to /cpu:0)
v3 = tf.Variable(3.0, name="v3") # pinned to /job:ps/task:0 (defaults to /cpu:0)
s = v1 + v2 # pinned to /job:worker (defaults to task:0/cpu:0)
with tf.device("/task:1"):
p1 = 2 * s # pinned to /job:worker/task:1 (defaults to /cpu:0)
with tf.device("/cpu:0"):
p2 = 3 * s # pinned to /job:worker/task:1/cpu:0
config = tf.ConfigProto()
config.log_device_placement = True
with tf.Session("grpc://127.0.0.1:2221", config=config) as sess:
v1.initializer.run()
In [10]:
reset_graph()
In [11]:
default1 = tf.constant([5.])
default2 = tf.constant([6])
default3 = tf.constant([7])
dec = tf.decode_csv(tf.constant("1.,,44"),
record_defaults=[default1, default2, default3])
with tf.Session() as sess:
print(sess.run(dec))
In [12]:
reset_graph()
test_csv = open("my_test.csv", "w")
test_csv.write("x1, x2 , target\n")
test_csv.write("1.,, 0\n")
test_csv.write("4., 5. , 1\n")
test_csv.write("7., 8. , 0\n")
test_csv.close()
filename_queue = tf.FIFOQueue(capacity=10, dtypes=[tf.string], shapes=[()])
filename = tf.placeholder(tf.string)
enqueue_filename = filename_queue.enqueue([filename])
close_filename_queue = filename_queue.close()
reader = tf.TextLineReader(skip_header_lines=1)
key, value = reader.read(filename_queue)
x1, x2, target = tf.decode_csv(value, record_defaults=[[-1.], [-1.], [-1]])
features = tf.stack([x1, x2])
instance_queue = tf.RandomShuffleQueue(
capacity=10, min_after_dequeue=2,
dtypes=[tf.float32, tf.int32], shapes=[[2],[]],
name="instance_q", shared_name="shared_instance_q")
enqueue_instance = instance_queue.enqueue([features, target])
close_instance_queue = instance_queue.close()
minibatch_instances, minibatch_targets = instance_queue.dequeue_up_to(2)
with tf.Session() as sess:
sess.run(enqueue_filename, feed_dict={filename: "my_test.csv"})
sess.run(close_filename_queue)
try:
while True:
sess.run(enqueue_instance)
except tf.errors.OutOfRangeError as ex:
print("No more files to read")
sess.run(close_instance_queue)
try:
while True:
print(sess.run([minibatch_instances, minibatch_targets]))
except tf.errors.OutOfRangeError as ex:
print("No more training instances")
In [13]:
#coord = tf.train.Coordinator()
#threads = tf.train.start_queue_runners(coord=coord)
#filename_queue = tf.train.string_input_producer(["test.csv"])
#coord.request_stop()
#coord.join(threads)
In [14]:
reset_graph()
filename_queue = tf.FIFOQueue(capacity=10, dtypes=[tf.string], shapes=[()])
filename = tf.placeholder(tf.string)
enqueue_filename = filename_queue.enqueue([filename])
close_filename_queue = filename_queue.close()
reader = tf.TextLineReader(skip_header_lines=1)
key, value = reader.read(filename_queue)
x1, x2, target = tf.decode_csv(value, record_defaults=[[-1.], [-1.], [-1]])
features = tf.stack([x1, x2])
instance_queue = tf.RandomShuffleQueue(
capacity=10, min_after_dequeue=2,
dtypes=[tf.float32, tf.int32], shapes=[[2],[]],
name="instance_q", shared_name="shared_instance_q")
enqueue_instance = instance_queue.enqueue([features, target])
close_instance_queue = instance_queue.close()
minibatch_instances, minibatch_targets = instance_queue.dequeue_up_to(2)
n_threads = 5
queue_runner = tf.train.QueueRunner(instance_queue, [enqueue_instance] * n_threads)
coord = tf.train.Coordinator()
with tf.Session() as sess:
sess.run(enqueue_filename, feed_dict={filename: "my_test.csv"})
sess.run(close_filename_queue)
enqueue_threads = queue_runner.create_threads(sess, coord=coord, start=True)
try:
while True:
print(sess.run([minibatch_instances, minibatch_targets]))
except tf.errors.OutOfRangeError as ex:
print("No more training instances")
In [15]:
reset_graph()
def read_and_push_instance(filename_queue, instance_queue):
reader = tf.TextLineReader(skip_header_lines=1)
key, value = reader.read(filename_queue)
x1, x2, target = tf.decode_csv(value, record_defaults=[[-1.], [-1.], [-1]])
features = tf.stack([x1, x2])
enqueue_instance = instance_queue.enqueue([features, target])
return enqueue_instance
filename_queue = tf.FIFOQueue(capacity=10, dtypes=[tf.string], shapes=[()])
filename = tf.placeholder(tf.string)
enqueue_filename = filename_queue.enqueue([filename])
close_filename_queue = filename_queue.close()
instance_queue = tf.RandomShuffleQueue(
capacity=10, min_after_dequeue=2,
dtypes=[tf.float32, tf.int32], shapes=[[2],[]],
name="instance_q", shared_name="shared_instance_q")
minibatch_instances, minibatch_targets = instance_queue.dequeue_up_to(2)
read_and_enqueue_ops = [read_and_push_instance(filename_queue, instance_queue) for i in range(5)]
queue_runner = tf.train.QueueRunner(instance_queue, read_and_enqueue_ops)
with tf.Session() as sess:
sess.run(enqueue_filename, feed_dict={filename: "my_test.csv"})
sess.run(close_filename_queue)
coord = tf.train.Coordinator()
enqueue_threads = queue_runner.create_threads(sess, coord=coord, start=True)
try:
while True:
print(sess.run([minibatch_instances, minibatch_targets]))
except tf.errors.OutOfRangeError as ex:
print("No more training instances")
In [16]:
reset_graph()
q = tf.FIFOQueue(capacity=10, dtypes=[tf.float32], shapes=[()])
v = tf.placeholder(tf.float32)
enqueue = q.enqueue([v])
dequeue = q.dequeue()
output = dequeue + 1
config = tf.ConfigProto()
config.operation_timeout_in_ms = 1000
with tf.Session(config=config) as sess:
sess.run(enqueue, feed_dict={v: 1.0})
sess.run(enqueue, feed_dict={v: 2.0})
sess.run(enqueue, feed_dict={v: 3.0})
print(sess.run(output))
print(sess.run(output, feed_dict={dequeue: 5}))
print(sess.run(output))
print(sess.run(output))
try:
print(sess.run(output))
except tf.errors.DeadlineExceededError as ex:
print("Timed out while dequeuing")
The Data API, introduced in TensorFlow 1.4, makes reading data efficiently much easier.
In [17]:
tf.reset_default_graph()
Let's start with a simple dataset composed of three times the integers 0 to 9, in batches of 7:
In [18]:
dataset = tf.data.Dataset.from_tensor_slices(np.arange(10))
dataset = dataset.repeat(3).batch(7)
The first line creates a dataset containing the integers 0 through 9. The second line creates a new dataset based on the first one, repeating its elements three times and creating batches of 7 elements. As you can see, we start with a source dataset, then we chain calls to various methods to apply transformations to the data.
Next, we create a one-shot-iterator to go through this dataset just once, and we call its get_next() method to get a tensor that represents the next element.
In [19]:
iterator = dataset.make_one_shot_iterator()
next_element = iterator.get_next()
Let's repeatedly evaluate next_element to go through the dataset. When there are not more elements, we get an OutOfRangeError:
In [20]:
with tf.Session() as sess:
try:
while True:
print(next_element.eval())
except tf.errors.OutOfRangeError:
print("Done")
Great! It worked fine.
Note that, as always, a tensor is only evaluated once each time we run the graph (sess.run()): so even if we evaluate multiple tensors that all depend on next_element, it is only evaluated once. This is true as well if we ask for next_element to be evaluated twice in just one run:
In [21]:
with tf.Session() as sess:
try:
while True:
print(sess.run([next_element, next_element]))
except tf.errors.OutOfRangeError:
print("Done")
The interleave() method is powerful but a bit tricky to grasp at first. The easiest way to understand it is to look at an example:
In [22]:
tf.reset_default_graph()
In [23]:
dataset = tf.data.Dataset.from_tensor_slices(np.arange(10))
dataset = dataset.repeat(3).batch(7)
dataset = dataset.interleave(
lambda v: tf.data.Dataset.from_tensor_slices(v),
cycle_length=3,
block_length=2)
iterator = dataset.make_one_shot_iterator()
next_element = iterator.get_next()
In [24]:
with tf.Session() as sess:
try:
while True:
print(next_element.eval(), end=",")
except tf.errors.OutOfRangeError:
print("Done")
Because cycle_length=3, the new dataset starts by pulling 3 elements from the previous dataset: that's [0,1,2,3,4,5,6], [7,8,9,0,1,2,3] and [4,5,6,7,8,9,0]. Then it calls the lambda function we gave it to create one dataset for each of the elements. Since we use Dataset.from_tensor_slices(), each dataset is going to return its elements one by one. Next, it pulls two items (since block_length=2) from each of these three datasets, and it iterates until all three datasets are out of items: 0,1 (from 1st), 7,8 (from 2nd), 4,5 (from 3rd), 2,3 (from 1st), 9,0 (from 2nd), and so on until 8,9 (from 3rd), 6 (from 1st), 3 (from 2nd), 0 (from 3rd). Next it tries to pull the next 3 elements from the original dataset, but there are just two left: [1,2,3,4,5,6,7] and [8,9]. Again, it creates datasets from these elements, and it pulls two items from each until both datasets are out of items: 1,2 (from 1st), 8,9 (from 2nd), 3,4 (from 1st), 5,6 (from 1st), 7 (from 1st). Notice that there's no interleaving at the end since the arrays do not have the same length.
Instead of using a source dataset based on from_tensor_slices() or from_tensor(), we can use a reader dataset. It handles most of the complexity for us (e.g., threads):
In [25]:
tf.reset_default_graph()
In [26]:
filenames = ["my_test.csv"]
In [27]:
dataset = tf.data.TextLineDataset(filenames)
We still need to tell it how to decode each line:
In [28]:
def decode_csv_line(line):
x1, x2, y = tf.decode_csv(
line, record_defaults=[[-1.], [-1.], [-1.]])
X = tf.stack([x1, x2])
return X, y
Next, we can apply this decoding function to each element in the dataset using map():
In [29]:
dataset = dataset.skip(1).map(decode_csv_line)
Finally, let's create a one-shot iterator:
In [30]:
it = dataset.make_one_shot_iterator()
X, y = it.get_next()
In [31]:
with tf.Session() as sess:
try:
while True:
X_val, y_val = sess.run([X, y])
print(X_val, y_val)
except tf.errors.OutOfRangeError as ex:
print("Done")
Coming soon
In [ ]: