Pensieve
Pensieve is a Python library for organizing objects and dependencies in a graph structure.
"One simply siphons the excess thoughts from one's mind, pours them into the basin, and examines them at one's leisure. It becomes easier to spot patterns and links, you understand, when they are in this form."
—Albus Dumbledore (Harry Potter and the Goblet of Fire by J. K. Rowling)
Pensieve for Data
In J. K. Rowling's words:
"a witch or wizard can extract their own or another's memories, store them in the Pensieve,
and review them later. It also relieves the mind when it becomes cluttered with information.
Anyone can examine the memories in the Pensieve, which also allows viewers to fully immerse
themselves in the memories".
Dealing with data during data wrangling and model generation in data science is like dealing with memories
except that there is a lot more of back and forth and iteration when dealing with data.
You constantly update parameters of your models, improve your data wrangling,
and make changes to the ways you visualize or store data.
As with most processes in data science, each step along the way may take a long time to finish
which forces you to avoid rerunning everything from scratch; this approach is very error-prone as some
of the processes depend on others. To solve this problem I came up with the idea of a Computation Graph
where the nodes represent data objects and the direction of edges indicate the dependency between them.
After using Pensieve for some time myself, I have found it to be beneficial in several ways:
- error reduction, especially for data wrangling and model creation
- data object organization
- easy transfer of data
- coherent data processing and data pipelines
- data and model reproducibility
- parallel processing
- performance and cost analysis in terms of computation time and memory usage
- graphical visualization of data and processes
- most important of all: relieving the mind
Using pensieve is similar to using a dictionary:
from pensieve import Pensieve
from math import pi
pensieve = Pensieve()
pensieve['radius'] = 5
pensieve['circumference'] = lambda radius: 2 * pi * radius
print(pensieve['circumference'])
outputs:
31.41592653589793
Changing the radius, in this example, will affect the circumference
but it is only calculated when needed:
pensieve['radius'] = 6
print(pensieve['circumference'])
outputs
37.69911184307752
Installation
pip install pensieve
Usage
Pensieve stores memories and functions that define the relationship between memories.
Concepts
Memory
A Pensieve
is a computation graph where the nodes hold values and edges
show dependency between nodes. Each node is called a Memory
.
Every memory has two important attributes:
key
: the name of the memory which should be identicalcontent
: the object the memory holds
Some memories have two other attributes:
precursors
: other memories a memory depends onfunction
: a function that defines the relationship between a memory
and its precursors
There are two types of memories:
- independent memories (without precursors)
- dependent memories (with precursors)
Storing a Memory
As explained above, you can work with pensieve similar to how you use a
dictionary. Adding a new item, i.e., a memory and its content, to pensieve is
called storing. In fact the Pensieve
class has a store
method which
can be used for storing new memories. However, we only use it for advanced
functionality. We do not use it as frequently because a new simpler notation
introduced since version 2 makes working pensieve much more coherent.
We will explain the store
method and its notation in the Advanced Usage section.
Retrieving a Memory
Retrieving the content of a memory is like getting an item from a dictionary.
print(pensieve['circumference'])
Independent Memories
An independent memory is like a root node in pensieve. It holds an object and
it does not depend on any other memory.
from pensieve import Pensieve
pensieve = Pensieve()
pensieve['text'] = 'Hello World!'
pensieve['number'] = 1
pensieve['list_of_numbers'] = [1, 3, 2]
In the above example, text, number, and list are the names of three
independent memories and their contents are
the string 'Hello World'
,
the integer 1
,
and a list consisting of three integers.
Dependent Memories and Precursors
A dependent memory is created from running a function on other dependent or
independent memories as the function's arguments. We call those memories, precursors;
i.e., if a memory depends on another memory, the former is a dependent memory
and the latter is its precursor.
The easiest way to define a dependent memory is by passing a function to pensieve
whose arguments match the names of precursors.
def print_and_return_first_word(text):
words = text.split()
print(words[0])
return words[0]
pensieve['first_word'] = print_and_return_first_word
In the above example, the print_and_return_first_word
function accepts one argument:
text
which is the name of the precursor.
You can also use a lambda, when possible, to define a dependent memory.
pensieve['sorted_list'] = lambda list_of_numbers: sorted(list_of_numbers)
Successors
Memories that depend on a memory are its successors. If a precursor is like a
parent, a successor is like a child.
In the above example, sorted_list
is a successor of list_of_numbers
.
Staleness
If one or more precursors of a memory change, the memory and all its successors becomes stale.
A stale memory is only refreshed when needed and if after calculation, it is found out
that the content has not changed, the successors go back to being up-to-date, but if
the content has in fact changed, the stay stale and will be updated when needed.
Note: if a memory is stale, retrieving its content will update it.
Visualization
from pensieve import Pensieve
from pandas import DataFrame, concat
from numpy.random import randint, seed
seed(17)
pensieve = Pensieve(graph_direction='tb')
pensieve['number_of_columns'] = 9
pensieve['column_names'] = lambda number_of_columns: [
f'x_{i + 1}' for i in range(number_of_columns)
]
pensieve['value_range'] = {'low': 1, 'high': 5}
def create_dataframe(column_names, value_range, number_of_rows):
return DataFrame({
column: randint(
low=value_range['low'],
high=value_range['high'],
size=number_of_rows
)
for column in column_names
})
pensieve['data_1'] = lambda column_names, value_range: create_dataframe(
column_names=column_names, value_range=value_range, number_of_rows=5
)
pensieve['data_2'] = lambda column_names, value_range: create_dataframe(
column_names=column_names, value_range=value_range, number_of_rows=3
)
pensieve['data_1_and_2'] = lambda data_1, data_2: concat(
objs=[data_1, data_2],
sort=False
)
pensieve['coefficient'] = 5
def sum_and_multiply(data_1_and_2, coefficient):
data = data_1_and_2.copy()
data['summation'] = data.apply(sum, axis=1)
data['coefficient'] = coefficient
data['y'] = data['summation'] * data['coefficient']
return data
pensieve['result'] = sum_and_multiply
display(pensieve)
Advanced Usage
Parallel Processing
from pensieve import Pensieve
from time import sleep
from datetime import datetime
start_time = datetime.now()
pensieve = Pensieve(num_threads=-1, evaluate=False)
pensieve['x'] = 1
pensieve['y'] = 10
pensieve['z'] = 2
pensieve['w'] = 20
def add_with_delay(x, y):
print(f'adding {x} and {y}, slowly, at {datetime.now()}')
sleep(1)
return x + y
pensieve['x_plus_y'] = add_with_delay
pensieve['z_plus_w'] = lambda z, w: add_with_delay(x=z, y=w)
pensieve['all_the_four'] = lambda x_plus_y, z_plus_w: add_with_delay(x=x_plus_y, y=z_plus_w)
elapsed = datetime.now() - start_time
print('Nothing has been calculated yet. Elapsed time:', elapsed)
print('Getting all_the_four forces the calculation of everything')
start_time = datetime.now()
print('Result of adding the four numbers:', pensieve['all_the_four'])
elapsed = datetime.now() - start_time
print('Elapsed time:', elapsed)
The above code produces the following output:
Nothing has been calculated yet. Elapsed time: 0:00:00.000716
Getting all_the_four forces the calculation of everything
adding 2 and 20, slowly, at 2019-12-15 21:33:55.063888
adding 1 and 10, slowly, at 2019-12-15 21:33:55.064526
adding 11 and 22, slowly, at 2019-12-15 21:33:56.188258
Result of adding the four numbers: 33
Elapsed time: 0:00:02.341677
Two of the calculations were executed in parallel: x + y
and z + w
.
With an overhead of 0.34
seconds, the three calculations took 2.34
seconds.
Let's see what happens if we do it the ordinary way:
start_time = datetime.now()
x = 1
y = 10
z = 2
w = 20
x_plus_y = add_with_delay(x, y)
z_plus_w = add_with_delay(z, w)
all_the_four = add_with_delay(x_plus_y, z_plus_w)
print('Result of adding the four numbers:', all_the_four)
elapsed = datetime.now() - start_time
print('Elapsed time:', elapsed)
This time the following output is produced:
adding 1 and 10, slowly, at 2019-12-15 21:38:11.618910
adding 2 and 20, slowly, at 2019-12-15 21:38:12.620105
adding 11 and 22, slowly, at 2019-12-15 21:38:13.625195
Result of adding the four numbers: 33
Elapsed time: 0:00:03.011291
With an overhead of 0.01
seconds, the three calculations
ran one after the other and took 3.01
seconds.
The store
Method
TBD