As already discussed, tensorflow doesn't provide its own way to cross-validate the model. The recommended way is to use KFold. It's a bit tedious, but doable. Here's a complete example of cross-validating MNIST model with tensorflow and KFold:

from sklearn.model_selection import KFold
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data

# Parameters
learning_rate = 0.01
batch_size = 500

# TF graph
x = tf.placeholder(tf.float32, [None, 784])
y = tf.placeholder(tf.float32, [None, 10])
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
pred = tf.nn.softmax(tf.matmul(x, W) + b)
cost = tf.reduce_mean(-tf.reduce_sum(y*tf.log(pred), reduction_indices=1))
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
init = tf.global_variables_initializer()

mnist = input_data.read_data_sets("data/mnist-tf", one_hot=True)
train_x_all = mnist.train.images
train_y_all = mnist.train.labels
test_x = mnist.test.images
test_y = mnist.test.labels

def run_train(session, train_x, train_y):
  print "\nStart training"
  session.run(init)
  for epoch in range(10):
    total_batch = int(train_x.shape[0] / batch_size)
    for i in range(total_batch):
      batch_x = train_x[i*batch_size:(i+1)*batch_size]
      batch_y = train_y[i*batch_size:(i+1)*batch_size]
      _, c = session.run([optimizer, cost], feed_dict={x: batch_x, y: batch_y})
      if i % 50 == 0:
        print "Epoch #%d step=%d cost=%f" % (epoch, i, c)

def cross_validate(session, split_size=5):
  results = []
  kf = KFold(n_splits=split_size)
  for train_idx, val_idx in kf.split(train_x_all, train_y_all):
    train_x = train_x_all[train_idx]
    train_y = train_y_all[train_idx]
    val_x = train_x_all[val_idx]
    val_y = train_y_all[val_idx]
    run_train(session, train_x, train_y)
    results.append(session.run(accuracy, feed_dict={x: val_x, y: val_y}))
  return results

with tf.Session() as session:
  result = cross_validate(session)
  print "Cross-validation result: %s" % result
  print "Test accuracy: %f" % session.run(accuracy, feed_dict={x: test_x, y: test_y})

Classes as objects

Before understanding metaclasses, you need to master classes in Python. And Python has a very peculiar idea of what classes are, borrowed from the Smalltalk language.

In most languages, classes are just pieces of code that describe how to produce an object. That's kinda true in Python too:

>>> class ObjectCreator(object):
...       pass
...

>>> my_object = ObjectCreator()
>>> print(my_object)
<__main__.ObjectCreator object at 0x8974f2c>

But classes are more than that in Python. Classes are objects too.

Yes, objects.

As soon as you use the keyword class, Python executes it and creates an object. The instruction

>>> class ObjectCreator(object):
...       pass
...

creates in memory an object with the name ObjectCreator.

This object (the class) is itself capable of creating objects (the instances), and this is why it's a class.

But still, it's an object, and therefore:

• you can assign it to a variable
• you can copy it
• you can add attributes to it
• you can pass it as a function parameter

e.g.:

>>> print(ObjectCreator) # you can print a class because it's an object
<class '__main__.ObjectCreator'>
>>> def echo(o):
...       print(o)
...
>>> echo(ObjectCreator) # you can pass a class as a parameter
<class '__main__.ObjectCreator'>
>>> print(hasattr(ObjectCreator, 'new_attribute'))
False
>>> ObjectCreator.new_attribute = 'foo' # you can add attributes to a class
>>> print(hasattr(ObjectCreator, 'new_attribute'))
True
>>> print(ObjectCreator.new_attribute)
foo
>>> ObjectCreatorMirror = ObjectCreator # you can assign a class to a variable
>>> print(ObjectCreatorMirror.new_attribute)
foo
>>> print(ObjectCreatorMirror())
<__main__.ObjectCreator object at 0x8997b4c>

Creating classes dynamically

Since classes are objects, you can create them on the fly, like any object.

First, you can create a class in a function using class:

>>> def choose_class(name):
...     if name == 'foo':
...         class Foo(object):
...             pass
...         return Foo # return the class, not an instance
...     else:
...         class Bar(object):
...             pass
...         return Bar
...
>>> MyClass = choose_class('foo')
>>> print(MyClass) # the function returns a class, not an instance
<class '__main__.Foo'>
>>> print(MyClass()) # you can create an object from this class
<__main__.Foo object at 0x89c6d4c>

But it's not so dynamic, since you still have to write the whole class yourself.

Since classes are objects, they must be generated by something.

When you use the class keyword, Python creates this object automatically. But as with most things in Python, it gives you a way to do it manually.

Remember the function type? The good old function that lets you know what type an object is:

>>> print(type(1))
<type 'int'>
>>> print(type("1"))
<type 'str'>
>>> print(type(ObjectCreator))
<type 'type'>
>>> print(type(ObjectCreator()))
<class '__main__.ObjectCreator'>

Well, type has also a completely different ability: it can create classes on the fly. type can take the description of a class as parameters, and return a class.

(I know, it's silly that the same function can have two completely different uses according to the parameters you pass to it. It's an issue due to backward compatibility in Python)

type works this way:

type(name, bases, attrs)

Where:

• name: name of the class
• bases: tuple of the parent class (for inheritance, can be empty)
• attrs: dictionary containing attributes names and values

e.g.:

>>> class MyShinyClass(object):
...       pass

can be created manually this way:

>>> MyShinyClass = type('MyShinyClass', (), {}) # returns a class object
>>> print(MyShinyClass)
<class '__main__.MyShinyClass'>
>>> print(MyShinyClass()) # create an instance with the class
<__main__.MyShinyClass object at 0x8997cec>

You'll notice that we use MyShinyClass as the name of the class and as the variable to hold the class reference. They can be different, but there is no reason to complicate things.

type accepts a dictionary to define the attributes of the class. So:

>>> class Foo(object):
...       bar = True

Can be translated to:

>>> Foo = type('Foo', (), {'bar':True})

And used as a normal class:

>>> print(Foo)
<class '__main__.Foo'>
>>> print(Foo.bar)
True
>>> f = Foo()
>>> print(f)
<__main__.Foo object at 0x8a9b84c>
>>> print(f.bar)
True

And of course, you can inherit from it, so:

>>>   class FooChild(Foo):
...         pass

would be:

>>> FooChild = type('FooChild', (Foo,), {})
>>> print(FooChild)
<class '__main__.FooChild'>
>>> print(FooChild.bar) # bar is inherited from Foo
True

Eventually, you'll want to add methods to your class. Just define a function with the proper signature and assign it as an attribute.

>>> def echo_bar(self):
...       print(self.bar)
...
>>> FooChild = type('FooChild', (Foo,), {'echo_bar': echo_bar})
>>> hasattr(Foo, 'echo_bar')
False
>>> hasattr(FooChild, 'echo_bar')
True
>>> my_foo = FooChild()
>>> my_foo.echo_bar()
True

And you can add even more methods after you dynamically create the class, just like adding methods to a normally created class object.

>>> def echo_bar_more(self):
...       print('yet another method')
...
>>> FooChild.echo_bar_more = echo_bar_more
>>> hasattr(FooChild, 'echo_bar_more')
True

You see where we are going: in Python, classes are objects, and you can create a class on the fly, dynamically.

This is what Python does when you use the keyword class, and it does so by using a metaclass.

What are metaclasses (finally)

Metaclasses are the 'stuff' that creates classes.

You define classes in order to create objects, right?

But we learned that Python classes are objects.

Well, metaclasses are what create these objects. They are the classes' classes, you can picture them this way:

MyClass = MetaClass()
my_object = MyClass()

You've seen that type lets you do something like this:

MyClass = type('MyClass', (), {})

It's because the function type is in fact a metaclass. type is the metaclass Python uses to create all classes behind the scenes.

Now you wonder "why the heck is it written in lowercase, and not Type?"

Well, I guess it's a matter of consistency with str, the class that creates strings objects, and int the class that creates integer objects. type is just the class that creates class objects.

You see that by checking the __class__ attribute.

Everything, and I mean everything, is an object in Python. That includes integers, strings, functions and classes. All of them are objects. And all of them have been created from a class:

>>> age = 35
>>> age.__class__
<type 'int'>
>>> name = 'bob'
>>> name.__class__
<type 'str'>
>>> def foo(): pass
>>> foo.__class__
<type 'function'>
>>> class Bar(object): pass
>>> b = Bar()
>>> b.__class__
<class '__main__.Bar'>

Now, what is the __class__ of any __class__ ?

>>> age.__class__.__class__
<type 'type'>
>>> name.__class__.__class__
<type 'type'>
>>> foo.__class__.__class__
<type 'type'>
>>> b.__class__.__class__
<type 'type'>

So, a metaclass is just the stuff that creates class objects.

You can call it a 'class factory' if you wish.

type is the built-in metaclass Python uses, but of course, you can create your own metaclass.

The __metaclass__ attribute

In Python 2, you can add a __metaclass__ attribute when you write a class (see next section for the Python 3 syntax):

class Foo(object):
    __metaclass__ = something...
    [...]

If you do so, Python will use the metaclass to create the class Foo.

Careful, it's tricky.

You write class Foo(object) first, but the class object Foo is not created in memory yet.

Python will look for __metaclass__ in the class definition. If it finds it, it will use it to create the object class Foo. If it doesn't, it will use type to create the class.

Read that several times.

When you do:

class Foo(Bar):
    pass

Python does the following:

Is there a __metaclass__ attribute in Foo?

If yes, create in-memory a class object (I said a class object, stay with me here), with the name Foo by using what is in __metaclass__.

If Python can't find __metaclass__, it will look for a __metaclass__ at the MODULE level, and try to do the same (but only for classes that don't inherit anything, basically old-style classes).

Then if it can't find any __metaclass__ at all, it will use the Bar's (the first parent) own metaclass (which might be the default type) to create the class object.

Be careful here that the __metaclass__ attribute will not be inherited, the metaclass of the parent (Bar.__class__) will be. If Bar used a __metaclass__ attribute that created Bar with type() (and not type.__new__()), the subclasses will not inherit that behavior.

Now the big question is, what can you put in __metaclass__?

The answer is something that can create a class.

And what can create a class? type, or anything that subclasses or uses it.

Metaclasses in Python 3

The syntax to set the metaclass has been changed in Python 3:

class Foo(object, metaclass=something):
    ...

i.e. the __metaclass__ attribute is no longer used, in favor of a keyword argument in the list of base classes.

The behavior of metaclasses however stays largely the same.

One thing added to metaclasses in Python 3 is that you can also pass attributes as keyword-arguments into a metaclass, like so:

class Foo(object, metaclass=something, kwarg1=value1, kwarg2=value2):
    ...

Read the section below for how Python handles this.

Custom metaclasses

The main purpose of a metaclass is to change the class automatically, when it's created.

You usually do this for APIs, where you want to create classes matching the current context.

Imagine a stupid example, where you decide that all classes in your module should have their attributes written in uppercase. There are several ways to do this, but one way is to set __metaclass__ at the module level.

This way, all classes of this module will be created using this metaclass, and we just have to tell the metaclass to turn all attributes to uppercase.

Luckily, __metaclass__ can actually be any callable, it doesn't need to be a formal class (I know, something with 'class' in its name doesn't need to be a class, go figure... but it's helpful).

So we will start with a simple example, by using a function.

# the metaclass will automatically get passed the same argument
# that you usually pass to `type`
def upper_attr(future_class_name, future_class_parents, future_class_attrs):
    """
      Return a class object, with the list of its attribute turned
      into uppercase.
    """
    # pick up any attribute that doesn't start with '__' and uppercase it
    uppercase_attrs = {
        attr if attr.startswith("__") else attr.upper(): v
        for attr, v in future_class_attrs.items()
    }

    # let `type` do the class creation
    return type(future_class_name, future_class_parents, uppercase_attrs)

__metaclass__ = upper_attr # this will affect all classes in the module

class Foo(): # global __metaclass__ won't work with "object" though
    # but we can define __metaclass__ here instead to affect only this class
    # and this will work with "object" children
    bar = 'bip'

Let's check:

>>> hasattr(Foo, 'bar')
False
>>> hasattr(Foo, 'BAR')
True
>>> Foo.BAR
'bip'

Now, let's do exactly the same, but using a real class for a metaclass:

# remember that `type` is actually a class like `str` and `int`
# so you can inherit from it
class UpperAttrMetaclass(type):
    # __new__ is the method called before __init__
    # it's the method that creates the object and returns it
    # while __init__ just initializes the object passed as parameter
    # you rarely use __new__, except when you want to control how the object
    # is created.
    # here the created object is the class, and we want to customize it
    # so we override __new__
    # you can do some stuff in __init__ too if you wish
    # some advanced use involves overriding __call__ as well, but we won't
    # see this
    def __new__(upperattr_metaclass, future_class_name,
                future_class_parents, future_class_attrs):
        uppercase_attrs = {
            attr if attr.startswith("__") else attr.upper(): v
            for attr, v in future_class_attrs.items()
        }
        return type(future_class_name, future_class_parents, uppercase_attrs)

Let's rewrite the above, but with shorter and more realistic variable names now that we know what they mean:

class UpperAttrMetaclass(type):
    def __new__(cls, clsname, bases, attrs):
        uppercase_attrs = {
            attr if attr.startswith("__") else attr.upper(): v
            for attr, v in attrs.items()
        }
        return type(clsname, bases, uppercase_attrs)

You may have noticed the extra argument cls. There is nothing special about it: __new__ always receives the class it's defined in, as the first parameter. Just like you have self for ordinary methods which receive the instance as the first parameter, or the defining class for class methods.

But this is not proper OOP. We are calling type directly and we aren't overriding or calling the parent's __new__. Let's do that instead:

class UpperAttrMetaclass(type):
    def __new__(cls, clsname, bases, attrs):
        uppercase_attrs = {
            attr if attr.startswith("__") else attr.upper(): v
            for attr, v in attrs.items()
        }
        return type.__new__(cls, clsname, bases, uppercase_attrs)

We can make it even cleaner by using super, which will ease inheritance (because yes, you can have metaclasses, inheriting from metaclasses, inheriting from type):

class UpperAttrMetaclass(type):
    def __new__(cls, clsname, bases, attrs):
        uppercase_attrs = {
            attr if attr.startswith("__") else attr.upper(): v
            for attr, v in attrs.items()
        }

        # Python 2 requires passing arguments to super:
        return super(UpperAttrMetaclass, cls).__new__(
            cls, clsname, bases, uppercase_attrs)

        # Python 3 can use no-arg super() which infers them:
        return super().__new__(cls, clsname, bases, uppercase_attrs)

Oh, and in Python 3 if you do this call with keyword arguments, like this:

class Foo(object, metaclass=MyMetaclass, kwarg1=value1):
    ...

It translates to this in the metaclass to use it:

class MyMetaclass(type):
    def __new__(cls, clsname, bases, dct, kwargs1=default):
        ...

That's it. There is really nothing more about metaclasses.

The reason behind the complexity of the code using metaclasses is not because of metaclasses, it's because you usually use metaclasses to do twisted stuff relying on introspection, manipulating inheritance, vars such as __dict__, etc.

Indeed, metaclasses are especially useful to do black magic, and therefore complicated stuff. But by themselves, they are simple:

• intercept a class creation
• modify the class
• return the modified class

Why would you use metaclasses classes instead of functions?

Since __metaclass__ can accept any callable, why would you use a class since it's obviously more complicated?

There are several reasons to do so:

• The intention is clear. When you read UpperAttrMetaclass(type), you know what's going to follow
• You can use OOP. Metaclass can inherit from metaclass, override parent methods. Metaclasses can even use metaclasses.
• Subclasses of a class will be instances of its metaclass if you specified a metaclass-class, but not with a metaclass-function.
• You can structure your code better. You never use metaclasses for something as trivial as the above example. It's usually for something complicated. Having the ability to make several methods and group them in one class is very useful to make the code easier to read.
• You can hook on __new__, __init__ and __call__. Which will allow you to do different stuff, Even if usually you can do it all in __new__, some people are just more comfortable using __init__.
• These are called metaclasses, damn it! It must mean something!

Why would you use metaclasses?

Now the big question. Why would you use some obscure error-prone feature?

Well, usually you don't:

Metaclasses are deeper magic that 99% of users should never worry about it. If you wonder whether you need them, you don't (the people who actually need them know with certainty that they need them, and don't need an explanation about why).

Python Guru Tim Peters

The main use case for a metaclass is creating an API. A typical example of this is the Django ORM. It allows you to define something like this:

class Person(models.Model):
    name = models.CharField(max_length=30)
    age = models.IntegerField()

But if you do this:

person = Person(name='bob', age='35')
print(person.age)

It won't return an IntegerField object. It will return an int, and can even take it directly from the database.

This is possible because models.Model defines __metaclass__ and it uses some magic that will turn the Person you just defined with simple statements into a complex hook to a database field.

Django makes something complex look simple by exposing a simple API and using metaclasses, recreating code from this API to do the real job behind the scenes.

The last word

First, you know that classes are objects that can create instances.

Well, in fact, classes are themselves instances. Of metaclasses.

>>> class Foo(object): pass
>>> id(Foo)
142630324

Everything is an object in Python, and they are all either instance of classes or instances of metaclasses.

Except for type.

type is actually its own metaclass. This is not something you could reproduce in pure Python, and is done by cheating a little bit at the implementation level.

Secondly, metaclasses are complicated. You may not want to use them for very simple class alterations. You can change classes by using two different techniques:

• monkey patching
• class decorators

99% of the time you need class alteration, you are better off using these.

But 98% of the time, you don't need class alteration at all.

I solved this by:

1. implementing a custom py.test collector that grafts the necessary suite of tests from my package,
2. requiring the consumer to instantiate a custom class of mine, my_pkg.MyDriverSuite, in one of their test files, then,
3. writing a custom py.test plugin that uses the pytest_pycollect_makeitem hook to intercept the custom instance and inject my collector.

(Note: This approach obviated the use of parametrized fixtures.)

---

The custom collector was by far the trickiest part to figure out, as it required lots of digging through the py.test internals and, ultimately, re-implementing a stripped-down version of a core py.test collector, Session, with a custom session Config whose rootdir was set to that of my package.

As of 2.7 (per the documentation) you get setUpClass and tearDownClass which execute before and after the tests in a given class are run, respectively. Alternatively, if you have a group of them in one file, you can use setUpModule and tearDownModule (documentation).

Otherwise your best bet is probably going to be to create your own derived TestSuite and override run(). All other calls would be handled by the parent, and run would call your setup and teardown code around a call up to the parent's run method.

if you don't mind using NumPy (which comes with your Python(x,y)), you may want to look at the np.testing module which defines, among others, a assert_almost_equal function.

The signature is np.testing.assert_almost_equal(actual, desired, decimal=7, err_msg='', verbose=True)

>>> x = 1.000001
>>> y = 1.000002
>>> np.testing.assert_almost_equal(x, y)
AssertionError: 
Arrays are not almost equal to 7 decimals
ACTUAL: 1.000001
DESIRED: 1.000002
>>> np.testing.assert_almost_equal(x, y, 5)
>>> np.testing.assert_almost_equal([x, x, x], [y, y, y], 5)
>>> np.testing.assert_almost_equal((x, x, x), (y, y, y), 5)

I figured it out finally. I had this environment variable set:

PYTEST_ADDOPTS=-m networks and level1

I was experimenting with env variables a while back but never had an issue until I did a reboot. Hopefully this helps someone in the future not be an idiot like myself lol.

Putting hpaulj's answer into actual code, something like this works:

class CustomHelpFormatter(argparse.HelpFormatter):
    def _format_action_invocation(self, action):
        if not action.option_strings or action.nargs == 0:
            return super()._format_action_invocation(action)
        default = self._get_default_metavar_for_optional(action)
        args_string = self._format_args(action, default)
        return ', '.join(action.option_strings) + ' ' + args_string

fmt = lambda prog: CustomHelpFormatter(prog)
parser = argparse.ArgumentParser(formatter_class=fmt)

To additionally extend the default column size for help variables, add constructor to CustomHelpFormatter:

def __init__(self, prog):
    super().__init__(prog, max_help_position=40, width=80)

Seeing it in action:

usage: bk set [-h] [-p] [-s r] [-f] [-c] [-b c] [-t x y] [-bs s] [-bc c]
              [--crop x1 y1 x2 y2] [-g u r d l]
              monitor [path]

positional arguments:
  monitor                    monitor number
  path                       input image path

optional arguments:
  -h, --help                 show this help message and exit
  -p, --preview              previews the changes without applying them
  -s, --scale r              scales image by given factor
  -f, --fit                  fits the image within monitor work area
  -c, --cover                makes the image cover whole monitor work area
  -b, --background c         selects background color
  -t, --translate x y        places the image at given position
  -bs, --border-size s       selects border width
  -bc, --border-color c      selects border size
  --crop x1 y1 x2 y2         selects crop area
  -g, --gap, --gaps u r d l  keeps "border" around work area

Open up Ganache and check which port you have at the start. Use the same port for the mainnet-fork. I used port: 7545

Actually, this is not a design flaw, and it is not because of internals or performance. It comes simply from the fact that functions in Python are first-class objects, and not only a piece of code.

As soon as you think of it this way, then it completely makes sense: a function is an object being evaluated on its definition; default parameters are kind of "member data" and therefore their state may change from one call to the other - exactly as in any other object.

In any case, the effbot (Fredrik Lundh) has a very nice explanation of the reasons for this behavior in Default Parameter Values in Python. I found it very clear, and I really suggest reading it for a better knowledge of how function objects work.