tomasatdatabricks commented 6 years ago

Re-generated models for DeepImageFeaturizer with new version of Keras and tensorflow.
Tweaked some tests due to the updated keras api.
"Dockerized" tests - tf 1.6 requires Ubuntu version >=16.04 and currently the only way to get 16.04 into Travis is via Docker.
Had to decrease numeric precision in several tests. Traced this issue down to changes in tf. Following simple python example demonstrates tf getting different result than numy on a simple expression (tf and numpy matched exactly in tf 1.4):

import tensorflow as tf tf.version '1.6.0' import numpy as np def simple_graph(): ... g = tf.Graph() ... with g.as_default(): ... x = tf.placeholder(tf.float32, [1], name = 'x') ... x /= 127.5 ... x -= 1 ... return (g, x) ...

def make_and_run_simple_graph(): ... g, x = simple_graph() ... return tf.Session(graph = g).run(x, feed_dict = {'x:0':np.array([55], dtype = np.float32)}) ...

tst = make_and_run_simple_graph() 2018-03-26 14:29:40.269905: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA a = np.array([55], dtype= np.float32) b = np.array([127.5], dtype= np.float32) c = np.array([1], dtype= np.float32) tst = make_and_run_simple_graph() tst - (a/b - c) array([ 5.96046448e-08], dtype=float32)

codecov-io commented 6 years ago

Codecov Report

Merging #110 into master will not change coverage. The diff coverage is n/a.

@@           Coverage Diff           @@
##           master     #110   +/-   ##
=======================================
  Coverage   82.84%   82.84%           
=======================================
  Files          34       34           
  Lines        1999     1999           
  Branches       44       44           
=======================================
  Hits         1656     1656           
  Misses        343      343

Impacted Files	Coverage Δ
src/main/scala/com/databricks/sparkdl/Models.scala	`85% <ø> (ø)`	:arrow_up:

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update e3c5876...db484d3. Read the comment docs.

tomasatdatabricks commented 6 years ago

Thanks for the review Sid!

Yes, we need to do that and yes we should get all of the #106, #108 and #109 in before releasing 1.0. All of this makes the name of this PR confusing => I'll change that :)

tomasatdatabricks commented 6 years ago

The model files are different but they do produce identical results. At least on our test images.

databricks / spark-deep-learning

Update to tensorflow 1.6 and Keras 2.1.5 #110

Codecov Report