TensorFlow 1.4.0 released – Google’s Second-Generation Machine Learning System

TensorFlow is Google’s second-generation machine learning system. According to Google, in some test, TensorFlow performance than the first generation of DistBelief almost double.

TensorFlow has extended support for deep learning built in, and any calculation that can be expressed in a computational flow graph can use TensorFlow. Any gradient-based machine learning algorithm can benefit from the auto-differentiation of TensorFlow. With the flexible Python interface, it’s easy to express ideas in TensorFlow.

TensorFlow also makes sense for the actual product. Move ideas seamlessly from desktop GPU training to mobile phones.

Sample code:

import tensorflow as tf
import numpy as np

# Create 100 phony x, y data points in NumPy, y = x * 0.1 + 0.3
x_data = np.random.rand(100).astype(np.float32)
y_data = x_data * 0.1 + 0.3

# Try to find values for W and b that compute y_data = W * x_data + b
# (We know that W should be 0.1 and b 0.3, but TensorFlow will
# figure that out for us.)
W = tf.Variable(tf.random_uniform([1], -1.01.0))
b = tf.Variable(tf.zeros([1]))
y = W * x_data + b

# Minimize the mean squared errors.
loss = tf.reduce_mean(tf.square(y - y_data))
optimizer = tf.train.GradientDescentOptimizer(0.5)
train = optimizer.minimize(loss)

# Before starting, initialize the variables.  We will 'run' this first.
init = tf.global_variables_initializer()

# Launch the graph.
sess = tf.Session()

# Fit the line.
for step in range(201):
    if step % 20 == 0:
        print(step, sess.run(W), sess.run(b))

# Learns best fit is W: [0.1], b: [0.3]

TensorFlow 1.4.0 released now.

Main features and improvements:

  • tf.keras is now part of the core TensorFlow API
  • tf.data is now part of the core TensorFlow API
  • Add train_and_evaluate for simple, distributed Estimator processing
  • Add tf.spectral.dct to calculate DCT-II

Bug fixes and other changes:

  • Fix tf.contrib.distributions.Affine incorrectly calculates log-det-jacobian.
  • Fix tf.random_gamm a to handle non-batch, scalar rendering incorrectly.
  • Resolved the race condition issue in TensorForest TreePredictionsV4Op.
  • Google Cloud Storage file system, Amazon S3 file system, and Hadoop files, system support is now the default build option.

API major changes:

  • The signature of the tf.contrib.data.rejection_resample () function has changed, and now it returns a function that can be used as a parameter, Dataset.apply ().
  • Use the Dataset.make_initializable_iterator () method instead of the tf.contrib.data.Iterator.from_dataset () method.
  • Remove rarely used and unnecessary tf.contrib.data.Iterator.dispose_op () method.
  • Rearrange some TFGAN missing functions in a backward compatible way.

Download TensorFlow :