CODEWITHSUNDEEP
pythontensorflowoptimization
Learning is a mess
Learning is a mess

Reputation: 8277

Tensorflow: 7x slower in eager mode

After doing a simple minimisation task (fitting optimal parameter for a hard sigmoid approximation) with tensorflow, I decided to translate it from graph mode to eager mode. My surprise was that it takes a lot longer to run in eager mode.

Here are the codes.

Graph mode code:

import tensorflow as tf
from time import time

beg = time()
a = tf.Variable(-10, name='a', dtype=tf.float32)
b = tf.Variable(10, name='b', dtype=tf.float32)

def g(x):
    return tf.clip_by_value( (x-a)/(b-a), 0, 1)

X = tf.lin_space(-20., 20., 2000)
loss = tf.reduce_sum( tf.square( tf.math.sigmoid(X) - g(X)))
opt = tf.train.AdamOptimizer(learning_rate=1e-3)
train_op = opt.minimize( loss)
init_op = tf.global_variables_initializer()

with tf.Session() as sess:
    sess.run(init_op)

    for _ in range( int(1e4)):
        sess.run( train_op)

print( 'Non-eager run in %.1f seconds' %(time()-beg))

printing Non-eager run in 3.5 seconds

Eager mode code:

import tensorflow as tf
from time import time

tf.enable_eager_execution()

beg = time()

a = tf.Variable(-10, name='a', dtype=tf.float32)
b = tf.Variable(10, name='b', dtype=tf.float32)

def g(x):
    return tf.clip_by_value( (x-a)/(b-a), 0, 1)

X = tf.lin_space(-20., 20., 2000)

opt = tf.train.AdamOptimizer(learning_rate=1e-3)

for _ in range( int(1e4)):
    with tf.GradientTape() as tape:
        loss = tf.reduce_sum( tf.square( tf.math.sigmoid(X) - g(X)))
        grads = tape.gradient(loss, [a,b])
    opt.apply_gradients(zip(grads, [a,b]), global_step=tf.train.get_or_create_global_step())
print( 'Eager run in %.1f seconds' %(time()-beg))

printing Eager run in 20.9 seconds

My bet is that my eager code is sub-optimal, and as tensorflow seems to be shifting to eager execution in its next big release I would like to know how to optimise this code to performances at least matching the first version.

Upvotes: 1

Views: 677

Answers (1)

nessuno
nessuno

Reputation: 27050

Your code, in tensorflow 2.0 will look something like (please note you can already try the nightly build of tensorflow 2.0 (https://pypi.org/project/tf-nightly-2.0-preview/))

import tensorflow as tf
from time import time

tf.enable_eager_execution()

beg = time()


@tf.function
def train():
    a = tf.Variable(-10, name='a', dtype=tf.float32)
    b = tf.Variable(10, name='b', dtype=tf.float32)

    def g(x):
        return tf.clip_by_value((x - a) / (b - a), 0, 1)

    X = tf.lin_space(-20., 20., 2000)
    opt = tf.train.AdamOptimizer(learning_rate=1e-3)

    for _ in range(int(1e4)):
        with tf.GradientTape() as tape:
            loss = tf.reduce_sum(tf.square(tf.math.sigmoid(X) - g(X)))
            grads = tape.gradient(loss, [a, b])
        opt.apply_gradients(
            zip(grads, [a, b]),
            global_step=tf.train.get_or_create_global_step())


train()
print('Eager run in %.1f seconds' % (time() - beg))

Please note that tf.contrib.eager.defun and Autograph (available in 1.12 and above), that are the base of @tf.session are still under active development and are experimental, hence implementation is a little bit buggy right now; so if it fails to run or it is slower probably is worth opening an issue on Github.

In 2.0 @tf.session will merge the pros of both defun and autograd

Upvotes: 2

Related Questions