What does the # symbol do in Python?
An # symbol at the beginning of a line is used for class and function decorators:
PEP 318: Decorators
Python Decorators
The most common Python decorators are:
#property
#classmethod
#staticmethod
An # in the middle of a line is probably matrix multiplication:
# as a binary operator.
Example
class Pizza(object):
def __init__(self):
self.toppings = []
def __call__(self, topping):
# When using '#instance_of_pizza' before a function definition
# the function gets passed onto 'topping'.
self.toppings.append(topping())
def __repr__(self):
return str(self.toppings)
pizza = Pizza()
#pizza
def cheese():
return 'cheese'
#pizza
def sauce():
return 'sauce'
print pizza
# ['cheese', 'sauce']
This shows that the function/method/class you're defining after a decorator is just basically passed on as an argument to the function/method immediately after the # sign.
First sighting
The microframework Flask introduces decorators from the very beginning in the following format:
from flask import Flask
app = Flask(__name__)
#app.route("/")
def hello():
return "Hello World!"
This in turn translates to:
rule = "/"
view_func = hello
# They go as arguments here in 'flask/app.py'
def add_url_rule(self, rule, endpoint=None, view_func=None, **options):
pass
Realizing this finally allowed me to feel at peace with Flask.
In Python 3.5 you can overload # as an operator. It is named as __matmul__, because it is designed to do matrix multiplication, but it can be anything you want. See PEP465 for details.
This is a simple implementation of matrix multiplication.
class Mat(list):
def __matmul__(self, B):
A = self
return Mat([[sum(A[i][k]*B[k][j] for k in range(len(B)))
for j in range(len(B[0])) ] for i in range(len(A))])
A = Mat([[1,3],[7,5]])
B = Mat([[6,8],[4,2]])
print(A # B)
This code yields:
[[18, 14], [62, 66]]
This code snippet:
def decorator(func):
return func
#decorator
def some_func():
pass
Is equivalent to this code:
def decorator(func):
return func
def some_func():
pass
some_func = decorator(some_func)
In the definition of a decorator you can add some modified things that wouldn't be returned by a function normally.
What does the “at” (#) symbol do in Python?
In short, it is used in decorator syntax and for matrix multiplication.
In the context of decorators, this syntax:
#decorator
def decorated_function():
"""this function is decorated"""
is equivalent to this:
def decorated_function():
"""this function is decorated"""
decorated_function = decorator(decorated_function)
In the context of matrix multiplication, a # b invokes a.__matmul__(b) - making this syntax:
a # b
equivalent to
dot(a, b)
and
a #= b
equivalent to
a = dot(a, b)
where dot is, for example, the numpy matrix multiplication function and a and b are matrices.
How could you discover this on your own?
I also do not know what to search for as searching Python docs or Google does not return relevant results when the # symbol is included.
If you want to have a rather complete view of what a particular piece of python syntax does, look directly at the grammar file. For the Python 3 branch:
~$ grep -C 1 "#" cpython/Grammar/Grammar
decorator: '#' dotted_name [ '(' [arglist] ')' ] NEWLINE
decorators: decorator+
--
testlist_star_expr: (test|star_expr) (',' (test|star_expr))* [',']
augassign: ('+=' | '-=' | '*=' | '#=' | '/=' | '%=' | '&=' | '|=' | '^=' |
'<<=' | '>>=' | '**=' | '//=')
--
arith_expr: term (('+'|'-') term)*
term: factor (('*'|'#'|'/'|'%'|'//') factor)*
factor: ('+'|'-'|'~') factor | power
We can see here that # is used in three contexts:
decorators
an operator between factors
an augmented assignment operator
Decorator Syntax:
A google search for "decorator python docs" gives as one of the top results, the "Compound Statements" section of the "Python Language Reference." Scrolling down to the section on function definitions, which we can find by searching for the word, "decorator", we see that... there's a lot to read. But the word, "decorator" is a link to the glossary, which tells us:
decorator
A function returning another function, usually applied as a function transformation using the #wrapper syntax. Common
examples for decorators are classmethod() and staticmethod().
The decorator syntax is merely syntactic sugar, the following two
function definitions are semantically equivalent:
def f(...):
...
f = staticmethod(f)
#staticmethod
def f(...):
...
The same concept exists for classes, but is less commonly used there.
See the documentation for function definitions and class definitions
for more about decorators.
So, we see that
#foo
def bar():
pass
is semantically the same as:
def bar():
pass
bar = foo(bar)
They are not exactly the same because Python evaluates the foo expression (which could be a dotted lookup and a function call) before bar with the decorator (#) syntax, but evaluates the foo expression after bar in the other case.
(If this difference makes a difference in the meaning of your code, you should reconsider what you're doing with your life, because that would be pathological.)
Stacked Decorators
If we go back to the function definition syntax documentation, we see:
#f1(arg)
#f2
def func(): pass
is roughly equivalent to
def func(): pass
func = f1(arg)(f2(func))
This is a demonstration that we can call a function that's a decorator first, as well as stack decorators. Functions, in Python, are first class objects - which means you can pass a function as an argument to another function, and return functions. Decorators do both of these things.
If we stack decorators, the function, as defined, gets passed first to the decorator immediately above it, then the next, and so on.
That about sums up the usage for # in the context of decorators.
The Operator, #
In the lexical analysis section of the language reference, we have a section on operators, which includes #, which makes it also an operator:
The following tokens are operators:
+ - * ** / // % #
<< >> & | ^ ~
< > <= >= == !=
and in the next page, the Data Model, we have the section Emulating Numeric Types,
object.__add__(self, other)
object.__sub__(self, other)
object.__mul__(self, other)
object.__matmul__(self, other)
object.__truediv__(self, other)
object.__floordiv__(self, other)
[...]
These methods are called to implement the binary arithmetic operations (+, -, *, #, /, //, [...]
And we see that __matmul__ corresponds to #. If we search the documentation for "matmul" we get a link to What's new in Python 3.5 with "matmul" under a heading "PEP 465 - A dedicated infix operator for matrix multiplication".
it can be implemented by defining __matmul__(), __rmatmul__(), and
__imatmul__() for regular, reflected, and in-place matrix multiplication.
(So now we learn that #= is the in-place version). It further explains:
Matrix multiplication is a notably common operation in many fields of
mathematics, science, engineering, and the addition of # allows
writing cleaner code:
S = (H # beta - r).T # inv(H # V # H.T) # (H # beta - r)
instead of:
S = dot((dot(H, beta) - r).T,
dot(inv(dot(dot(H, V), H.T)), dot(H, beta) - r))
While this operator can be overloaded to do almost anything, in numpy, for example, we would use this syntax to calculate the inner and outer product of arrays and matrices:
>>> from numpy import array, matrix
>>> array([[1,2,3]]).T # array([[1,2,3]])
array([[1, 2, 3],
[2, 4, 6],
[3, 6, 9]])
>>> array([[1,2,3]]) # array([[1,2,3]]).T
array([[14]])
>>> matrix([1,2,3]).T # matrix([1,2,3])
matrix([[1, 2, 3],
[2, 4, 6],
[3, 6, 9]])
>>> matrix([1,2,3]) # matrix([1,2,3]).T
matrix([[14]])
Inplace matrix multiplication: #=
While researching the prior usage, we learn that there is also the inplace matrix multiplication. If we attempt to use it, we may find it is not yet implemented for numpy:
>>> m = matrix([1,2,3])
>>> m #= m.T
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: In-place matrix multiplication is not (yet) supported. Use 'a = a # b' instead of 'a #= b'.
When it is implemented, I would expect the result to look like this:
>>> m = matrix([1,2,3])
>>> m #= m.T
>>> m
matrix([[14]])
What does the “at” (#) symbol do in Python?
# symbol is a syntactic sugar python provides to utilize decorator,
to paraphrase the question, It's exactly about what does decorator do in Python?
Put it simple decorator allow you to modify a given function's definition without touch its innermost (it's closure).
It's the most case when you import wonderful package from third party. You can visualize it, you can use it, but you cannot touch its innermost and its heart.
Here is a quick example,
suppose I define a read_a_book function on Ipython
In [9]: def read_a_book():
...: return "I am reading the book: "
...:
In [10]: read_a_book()
Out[10]: 'I am reading the book: '
You see, I forgot to add a name to it.
How to solve such a problem? Of course, I could re-define the function as:
def read_a_book():
return "I am reading the book: 'Python Cookbook'"
Nevertheless, what if I'm not allowed to manipulate the original function, or if there are thousands of such function to be handled.
Solve the problem by thinking different and define a new_function
def add_a_book(func):
def wrapper():
return func() + "Python Cookbook"
return wrapper
Then employ it.
In [14]: read_a_book = add_a_book(read_a_book)
In [15]: read_a_book()
Out[15]: 'I am reading the book: Python Cookbook'
Tada, you see, I amended read_a_book without touching it inner closure. Nothing stops me equipped with decorator.
What's about #
#add_a_book
def read_a_book():
return "I am reading the book: "
In [17]: read_a_book()
Out[17]: 'I am reading the book: Python Cookbook'
#add_a_book is a fancy and handy way to say read_a_book = add_a_book(read_a_book), it's a syntactic sugar, there's nothing more fancier about it.
If you are referring to some code in a python notebook which is using Numpy library, then # operator means Matrix Multiplication. For example:
import numpy as np
def forward(xi, W1, b1, W2, b2):
z1 = W1 # xi + b1
a1 = sigma(z1)
z2 = W2 # a1 + b2
return z2, a1
Decorators were added in Python to make function and method wrapping (a function that receives a function and returns an enhanced one) easier to read and understand. The original use case was to be able to define the methods as class methods or static methods on the head of their definition. Without the decorator syntax, it would require a rather sparse and repetitive definition:
class WithoutDecorators:
def some_static_method():
print("this is static method")
some_static_method = staticmethod(some_static_method)
def some_class_method(cls):
print("this is class method")
some_class_method = classmethod(some_class_method)
If the decorator syntax is used for the same purpose, the code is shorter and easier to understand:
class WithDecorators:
#staticmethod
def some_static_method():
print("this is static method")
#classmethod
def some_class_method(cls):
print("this is class method")
General syntax and possible implementations
The decorator is generally a named object ( lambda expressions are not allowed) that accepts a single argument when called (it will be the decorated function) and returns another callable object. "Callable" is used here instead of "function" with premeditation. While decorators are often discussed in the scope of methods and functions, they are not limited to them. In fact, anything that is callable (any object that implements the _call__ method is considered callable), can be used as a decorator and often objects returned by them are not simple functions but more instances of more complex classes implementing their own __call_ method.
The decorator syntax is simply only a syntactic sugar. Consider the following decorator usage:
#some_decorator
def decorated_function():
pass
This can always be replaced by an explicit decorator call and function reassignment:
def decorated_function():
pass
decorated_function = some_decorator(decorated_function)
However, the latter is less readable and also very hard to understand if multiple decorators are used on a single function.
Decorators can be used in multiple different ways as shown below:
As a function
There are many ways to write custom decorators, but the simplest way is to write a function that returns a subfunction that wraps the original function call.
The generic patterns is as follows:
def mydecorator(function):
def wrapped(*args, **kwargs):
# do some stuff before the original
# function gets called
result = function(*args, **kwargs)
# do some stuff after function call and
# return the result
return result
# return wrapper as a decorated function
return wrapped
As a class
While decorators almost always can be implemented using functions, there are some situations when using user-defined classes is a better option. This is often true when the decorator needs complex parametrization or it depends on a specific state.
The generic pattern for a nonparametrized decorator as a class is as follows:
class DecoratorAsClass:
def __init__(self, function):
self.function = function
def __call__(self, *args, **kwargs):
# do some stuff before the original
# function gets called
result = self.function(*args, **kwargs)
# do some stuff after function call and
# return the result
return result
Parametrizing decorators
In real code, there is often a need to use decorators that can be parametrized. When the function is used as a decorator, then the solution is simple—a second level of wrapping has to be used. Here is a simple example of the decorator that repeats the execution of a decorated function the specified number of times every time it is called:
def repeat(number=3):
"""Cause decorated function to be repeated a number of times.
Last value of original function call is returned as a result
:param number: number of repetitions, 3 if not specified
"""
def actual_decorator(function):
def wrapper(*args, **kwargs):
result = None
for _ in range(number):
result = function(*args, **kwargs)
return result
return wrapper
return actual_decorator
The decorator defined this way can accept parameters:
>>> #repeat(2)
... def foo():
... print("foo")
...
>>> foo()
foo
foo
Note that even if the parametrized decorator has default values for its arguments, the parentheses after its name is required. The correct way to use the preceding decorator with default arguments is as follows:
>>> #repeat()
... def bar():
... print("bar")
...
>>> bar()
bar
bar
bar
Finally lets see decorators with Properties.
Properties
The properties provide a built-in descriptor type that knows how to link an attribute to a set of methods. A property takes four optional arguments: fget , fset , fdel , and doc . The last one can be provided to define a docstring that is linked to the attribute as if it were a method. Here is an example of a Rectangle class that can be controlled either by direct access to attributes that store two corner points or by using the width , and height properties:
class Rectangle:
def __init__(self, x1, y1, x2, y2):
self.x1, self.y1 = x1, y1
self.x2, self.y2 = x2, y2
def _width_get(self):
return self.x2 - self.x1
def _width_set(self, value):
self.x2 = self.x1 + value
def _height_get(self):
return self.y2 - self.y1
def _height_set(self, value):
self.y2 = self.y1 + value
width = property(
_width_get, _width_set,
doc="rectangle width measured from left"
)
height = property(
_height_get, _height_set,
doc="rectangle height measured from top"
)
def __repr__(self):
return "{}({}, {}, {}, {})".format(
self.__class__.__name__,
self.x1, self.y1, self.x2, self.y2
)
The best syntax for creating properties is using property as a decorator. This will reduce the number of method signatures inside of the class
and make code more readable and maintainable. With decorators the above class becomes:
class Rectangle:
def __init__(self, x1, y1, x2, y2):
self.x1, self.y1 = x1, y1
self.x2, self.y2 = x2, y2
#property
def width(self):
"""rectangle height measured from top"""
return self.x2 - self.x1
#width.setter
def width(self, value):
self.x2 = self.x1 + value
#property
def height(self):
"""rectangle height measured from top"""
return self.y2 - self.y1
#height.setter
def height(self, value):
self.y2 = self.y1 + value
Starting with Python 3.5, the '#' is used as a dedicated infix symbol for MATRIX MULTIPLICATION (PEP 0465 -- see https://www.python.org/dev/peps/pep-0465/)
# can be a math operator or a DECORATOR but what you mean is a decorator.
This code:
def func(f):
return f
func(lambda :"HelloWorld")()
using decorators can be written like:
def func(f):
return f
#func
def name():
return "Hello World"
name()
Decorators can have arguments.
You can see this GeeksforGeeks post: https://www.geeksforgeeks.org/decorators-in-python/
It indicates that you are using a decorator. Here is Bruce Eckel's example from 2008.
Python decorator is like a wrapper of a function or a class. It’s still too conceptual.
def function_decorator(func):
def wrapped_func():
# Do something before the function is executed
func()
# Do something after the function has been executed
return wrapped_func
The above code is a definition of a decorator that decorates a function.
function_decorator is the name of the decorator.
wrapped_func is the name of the inner function, which is actually only used in this decorator definition. func is the function that is being decorated.
In the inner function wrapped_func, we can do whatever before and after the func is called. After the decorator is defined, we simply use it as follows.
#function_decorator
def func():
pass
Then, whenever we call the function func, the behaviours we’ve defined in the decorator will also be executed.
EXAMPLE :
from functools import wraps
def mydecorator(f):
#wraps(f)
def wrapped(*args, **kwargs):
print "Before decorated function"
r = f(*args, **kwargs)
print "After decorated function"
return r
return wrapped
#mydecorator
def myfunc(myarg):
print "my function", myarg
return "return value"
r = myfunc('asdf')
print r
Output :
Before decorated function
my function asdf
After decorated function
return value
To say what others have in a different way: yes, it is a decorator.
In Python, it's like:
Creating a function (follows under the # call)
Calling another function to operate on your created function. This returns a new function. The function that you call is the argument of the #.
Replacing the function defined with the new function returned.
This can be used for all kinds of useful things, made possible because functions are objects and just necessary just instructions.
# symbol is also used to access variables inside a plydata / pandas dataframe query, pandas.DataFrame.query.
Example:
df = pandas.DataFrame({'foo': [1,2,15,17]})
y = 10
df >> query('foo > #y') # plydata
df.query('foo > #y') # pandas
I have written a scikit-learn estimator. It has a parameter and a model_ attribute that is set by fit.
class MyEstimator(BaseEstimator, TransformerMixin):
def __init__(self, param="default"):
self.param = param
self.model_ = None
def fit(self, x, y):
# Sets the value of self.model_
I want to be able to pickle MyEstimator, but the model_ object I create cannot be serialized with pickle because it is a keras model. Following the example of the blog post "Pickling Keras Models" I added the following pickling handler methods to my class.
class MyEstimator(BaseEstimator, TransformerMixin):
def __getstate__(self):
state = super().__getstate__().copy()
with tempfile.NamedTemporaryFile(suffix=".hdf5", delete=True) as fd:
keras.models.save_model(self.model_, fd.name, overwrite=True)
state["model_"] = fd.read()
return state
def __setstate__(self, state):
super().__setstate__(state)
with tempfile.NamedTemporaryFile(suffix=".hdf5", delete=True) as fd:
fd.write(state["model_"])
fd.flush()
self.__dict__["model_"] = keras.models.load_model(fd.name)
This replaces the unpickleable model_ member with a representation generated by keras' serializer that can be pickled. Using this customization I can call fit, serialize and deserialize, and get back my original model. Everything works.
e = MyEstimator()
e.fit(x, y)
with open("myfile.pk", mode="wb") as f:
pickle.dump(e, f)
with open("myfile.pk", mode="rb") as f:
pickle.load(f) # Returns a copy of e
However, serialization does not work when I try to put MyEstimator in a pipeline and pickle the result of a GridSearchCV.
s = GridSearchCV(Pipeline([
# ...
("estimator", MyEstimator())
# ...
]))
s.fit(x, y)
with open("myfile.pk", mode="wb") as f:
pickle.dump(s, f)
During the pickle.dump call I expect to see MyEstimator.__getstate__ get called with a fitted self.model_ object. (This is what happens when I serialize the model by itself, outside the grid search.) Instead self.model_ is None, so I am unable to serialize the best_estimator_ generated by my grid search.
It looks like grid search serialization is instantiating a new MyEstimator object instead of using the one that was in the pipeline. This seems wrong to me. I've looked through the scikit-learn code, but can't see where this is happening.
Is this a bug in scikit-learn, or am I doing something wrong?
(Note: keras does have a wrapper layer that can convert some keras models into scikit-learn estimators, but I can't use that here for other reasons and I'm not sure it wouldn't just have the same problem.)
The search object contains a mixed of MyEstimator objects, some of which have not had fit called on them. The fix is to check if model_ is None before trying to serialize it with the keras tools.
class MyEstimator(BaseEstimator, TransformerMixin):
def __getstate__(self):
state = super().__getstate__().copy()
if self.model_ is not None:
with tempfile.NamedTemporaryFile(suffix=".hdf5", delete=True) as fd:
keras.models.save_model(self.model_, fd.name, overwrite=True)
state["model_"] = fd.read()
return state
def __setstate__(self, state):
super().__setstate__(state)
if self.model_ is not None:
with tempfile.NamedTemporaryFile(suffix=".hdf5", delete=True) as fd:
fd.write(state["model_"])
fd.flush()
self.__dict__["model_"] = keras.models.load_model(fd.name)
I don't know why there would be any unfitted models in the search object after the grid search had completed, but there are.
Relu function as defined in keras/activation.py is:
def relu(x, alpha=0., max_value=None):
return K.relu(x, alpha=alpha, max_value=max_value)
It has a max_value which can be used to clip the value. Now how can this be used/called in the code?
I have tried the following:
(a)
model.add(Dense(512,input_dim=1))
model.add(Activation('relu',max_value=250))
assert kwarg in allowed_kwargs, 'Keyword argument not understood:
' + kwarg
AssertionError: Keyword argument not understood: max_value
(b)
Rel = Activation('relu',max_value=250)
same error
(c)
from keras.layers import activations
uu = activations.relu(??,max_value=250)
The problem with this is that it expects the input to be present in the first value. The error is 'relu() takes at least 1 argument (1 given)'
So how do I make this a layer?
model.add(activations.relu(max_value=250))
has the same issue 'relu() takes at least 1 argument (1 given)'
If this file cannot be used as layer, then there seems to be no way of specifying a clip value to Relu. This implies that the comment here https://github.com/fchollet/keras/issues/2119 closing a proposed change is wrong...
Any thoughts? Thanks!
You can use the ReLU function of the Keras backend. Therefore, first import the backend:
from keras import backend as K
Then, you can pass your own function as activation using backend functionality.
This would look like
def relu_advanced(x):
return K.relu(x, max_value=250)
Then you can use it like
model.add(Dense(512, input_dim=1, activation=relu_advanced))
or
model.add(Activation(relu_advanced))
Unfortunately, you must hard code additional arguments.
Therefore, it is better to use a function, that returns your function and passes your custom values:
def create_relu_advanced(max_value=1.):
def relu_advanced(x):
return K.relu(x, max_value=K.cast_to_floatx(max_value))
return relu_advanced
Then you can pass your arguments by either
model.add(Dense(512, input_dim=1, activation=create_relu_advanced(max_value=250)))
or
model.add(Activation(create_relu_advanced(max_value=250)))
That is as easy as one lambda :
from keras.activations import relu
clipped_relu = lambda x: relu(x, max_value=3.14)
Then use it like this:
model.add(Conv2D(64, (3, 3)))
model.add(Activation(clipped_relu))
When reading a model saved in hdf5 use custom_objects dictionary:
model = load_model(model_file, custom_objects={'<lambda>': clipped_relu})
Tested below, it'd work:
import keras
def clip_relu (x):
return keras.activations.relu(x, max_value=1.)
predictions=Dense(num_classes,activation=clip_relu,name='output')
This is what I did using Lambda layer to implement clip relu:
Step 1: define a function to do reluclip:
def reluclip(x, max_value = 20):
return K.relu(x, max_value = max_value)
Step 2: add Lambda layer into model:
y = Lambda(function = reluclip)(y)
def rnn_seq2seq(encoder_inputs, decoder_inputs, cell, output_projection=None,feed_previous=False, dtype=tf.float32, scope=None):
with tf.variable_scope(scope or "rnn_seq2seq"):
_, enc_states = rnn.rnn(cell, encoder_inputs,dtype=dtype)
def extract_argmax(prev, i):
if output_projection is not None:
prev = tf.nn.xw_plus_b(prev, output_projection[0], output_projection[1])
return tf.to_float(tf.equal(prev,tf.reduce_max(prev,reduction_indices=[1],keep_dims=True)))
loop_function = None
if feed_previous:
loop_function = extract_argmax
#seq2seq.rnn_decoder is provided in tensorflow/models/rnn/seq2seq.py
return seq2seq.rnn_decoder(decoder_inputs, enc_states[-1], cell, loop_function=loop_function)
I want to create two RNN models, one for training and another for testing. For that, I can call the function twice passing the feed_previous to True or False.
train_op,train_states = rnn_seq2seq(enc_inp,dec_inp,cell,output_projection=op,feed_previous=False)
test_op,_ = rnn_seq2seq(enc_inp,dec_inp,cell,output_projection=op,feed_previous=True)
But if I call the above function twice, wouldn't it create two different RNNs ? I am wondering if they would be able to share the weights.
Both functions operate on the same default graph and so can reuse the variables, check out variable scopes tutorial and see if your variables are created with reuse=True parameter
As a sanity check, try following snippet to list all variables in the default graph:
[v.name for v in tf.get_default_graph().as_graph_def().node if v.op=='Variable']