THIS LIBRARY HAS BEEN DEPRECATED

With the release of TF2.0, this library is largely broken and we have no plans on fixing it. Most of the tests (especially the "make sure all variables are modified during training") are already included in Keras.

mltest

Machine learning testing framework for tensorflow. If you are looking for a pytorch version, go check out the port suriyadeepan made!

How to install

You can either use it directly from pip

sudo pip install mltest

Or you can copy and paste the mltest.py file. Everthing is self contained and opensourced, so feel free to do what you want!

How to use

See the medium post for an overview of features.

The most basic test is to use all of the default settings of test_suite(). This will run all of the tests described in the medium post.

import mltest
import your_model_file
import tensorflow as tf
import numpy as np

def setup():
  mltest.setup()

# Build your test function.
def test_mltest_suite():
  # Make your model input a placeholder.
  input_tensor = tf.placeholder(tf.float32, (None, 100))
  label_tensor = tf.placeholder(tf.int32, (None))
  # Build your model.
  model = your_model_file.build_model(input_tensor, label_tensor)
  # Give it some random input (Be sure to seed it!!).
  feed_dict = {
      input_tensor: np.random.normal(size=(10, 100)),
      label_tensor: np.random.randint((100))
  }
  # Run the test suite!
  mltest.test_suite(
      model.prediction,
      model.train_op,
      feed_dict=feed_dict)

Setup

The function mltest.setup() will automatically reset the default graph and seed all of the random values for tensorflow, numpy, and python's random.

Variables change/don't change.

You can also specify which variables you expect to train with each training op. A major use case is for GAN training.

def test_descriminator():
  model = Model()
  mltest.test_suite(
    model.descriminator_output,
    model.train_descriminator,
    scope="descriminator")

def test_generator():
  model = Model()
  mltest.test_suite(
    model.generator_output,
    model.train_generator,
    scope="generator")

You can also define the variables to test against directly with a list.

def test_list_of_variables():
  model = Model()
  vars_that_should_train = [var1, var2, ...]
  mltest.test_suite(
    model.generator_output,
    model.train_generator,
    var_list=vars_that_should_train)

Logits Range

If your model has a specific output range rather than linear, you can test to make sure that range stays consistent. In this example, we assume our logits has a tanh output, so all of our values should fall between 0 and 1.

def test_range():
  model = Model()
  mltest.test_suite(
    model.logits,
    model.train_op,
    output_range=(0,1))

Input Dependencies

The last test makes sure all of the variables in feed_dict affect the train_op

def test_range():
  model = Model()
  mltest.test_suite(
    model.logits,
    model.train_op,
    feed_dict={
      input_1: [1.0],
      input_2: [2.0]
    })

NaN and Inf Tensors

Badly initalized or normalized layers can cause NaN or Inf ouputs that a developer may not have been prepared to handle. This isn't caught with < or > comparisons, so we must test for each directly. These have been added to the test suite aswell.

def test_nan():
  x = tf.Variable(-1.0)
  y = tf.log(x) # This will output a NaN value.
  mltest.assert_never_nan(y)

def test_inf():
  a = tf.Variable(1.0)
  b = tf.Variable(0.0)
  c = a / b # This will output inf.
  mltest.assert_never_inf(c)

Turn off tests

Finally, you can turn off any of these tests if they do not suit your needs.

mltest.test_suite(
    test_all_inputs_dependent=True, # Test that all vars in feed_dict 
                                    # affect train_op.
    test_other_vars_dont_change=True, # Test that all vars outside 
                                      # of scope or var_list do not change.
    test_output_range=True, # Test output range of the first argument.
    test_nan_vals=True, # Check for NaN values in all tensors.
    test_inf_vals=True # Check for Inf values in all tensors.
    )

Contributors

Thanks to everyone who has helped!

dekromp