When NumPy is not fast enough - Step-by-Step Execution

Concept Flow - When NumPy is not fast enough

Start: NumPy operation

↓

Check: Is NumPy fast enough?

Yes→Use NumPy result

No↓

Try alternatives:

↓

Numba

↓

Faster computation

↓

Use result

Start with NumPy; if it's slow, try faster tools like Numba, Cython, or parallel processing to speed up.

Execution Sample

NumPy

import numpy as np
from numba import njit

@njit
def fast_sum(arr):
    total = 0
    for x in arr:
        total += x
    return total

arr = np.arange(1000000)
print(fast_sum(arr))

This code uses Numba to speed up summing a large NumPy array with a loop.

Execution Table

Step	Action	Evaluation	Result
1	Create NumPy array arr	np.arange(1000000)	Array of 1 million integers from 0 to 999999
2	Define fast_sum with @njit	Function compiled by Numba	fast_sum is ready for fast execution
3	Call fast_sum(arr)	Loop over arr adding elements	Sum calculated quickly using compiled code
4	Print result	Output sum of 0 to 999999	499999500000
5	End	All steps done	Program finishes

💡 Completed summing large array using Numba for speed

Variable Tracker

Variable	Start	After 1	After 2	After 3	Final
total	0	0 + 0 = 0	0 + 1 = 1	1 + 2 = 3	499999500000
x	N/A	0	1	2	999999

Key Moments - 3 Insights

Why does using a Python loop with NumPy arrays run slowly?

NumPy is fast with vectorized operations but slow with explicit Python loops, as shown in step 3 of the execution_table where looping in Python is replaced by Numba compiled code.

How does Numba speed up the loop?

Numba compiles the Python loop into fast machine code, avoiding Python overhead, as seen in step 2 and 3 where the function is compiled and runs faster.

Can we always replace NumPy operations with Numba?

No, Numba works best with simple loops and numerical code; complex NumPy functions or object arrays may not benefit, so choose based on your code.

Visual Quiz - 3 Questions

Look at the execution_table at step 3, what is happening inside fast_sum?

A A Python loop adding elements slowly

B A Numba compiled loop adding elements quickly

C A vectorized NumPy sum operation

D A recursive function call

💡 Hint

Check step 2 and 3 in execution_table where @njit compiles the function and step 3 shows the loop running fast.

According to variable_tracker, what is the value of 'total' after adding the first three elements?

A 6

B 0

C 3

D 1

💡 Hint

Look at 'total' values After 1, After 2, and After 3 in variable_tracker.

If we remove @njit decorator, how would the execution_table change at step 3?

A The loop would run slower in pure Python

B The loop would run faster

C The function would not run at all

D The sum would be incorrect

💡 Hint

Consider the role of Numba in speeding up loops as explained in key_moments.

Concept Snapshot

When NumPy loops are slow, use tools like Numba to compile loops into fast code.
Numba's @njit decorator speeds up numerical loops.
Not all NumPy operations need Numba; vectorized code is already fast.
Use alternatives like Cython or parallel processing for more speed.
Test performance to choose the best method.

Full Transcript

This visual execution shows how NumPy can be slow when using Python loops. We start by creating a large NumPy array. Then we define a function fast_sum that sums elements using a loop. Normally, Python loops are slow with NumPy arrays. But by adding the @njit decorator from Numba, the function is compiled into fast machine code. The execution table traces each step: creating the array, compiling the function, running the fast loop, and printing the result. The variable tracker shows how the total sum accumulates step by step. Key moments explain why loops are slow in Python and how Numba speeds them up. The quiz tests understanding of these steps and the effect of removing Numba. This helps beginners see when and how to speed up NumPy code beyond its built-in fast operations.

Practice

1. Why might NumPy sometimes be too slow for certain tasks?

easy

A. Because it does not support mathematical operations

B. Because it cannot handle arrays larger than 100 elements

C. Because it only works on single-core processors

D. Because it uses Python loops internally which are slow

Solution

1. Step 1: Understand NumPy's speed basis

   NumPy is fast because it uses optimized C code, but Python loops around NumPy operations can slow it down.

2. Step 2: Identify the cause of slowness

   When complex tasks require Python loops or non-vectorized code, speed drops because Python loops are slower than compiled code.

3. Final Answer:

   Because it uses Python loops internally which are slow -> Option D

4. Quick Check:

   Python loops slow NumPy = Because it uses Python loops internally which are slow [OK]

Hint: NumPy is fast with vectorized code, slow with Python loops [OK]

Common Mistakes:

• Thinking NumPy can't handle large arrays
• Believing NumPy only works on single-core CPUs
• Assuming NumPy lacks math operation support

2. Which of the following is the correct way to import Numba to speed up NumPy code?

easy

A. import numba as nb

B. import numba.speedup

C. from numba import jit

D. import numba.jit

Solution

1. Step 1: Recall Numba import syntax

   Numba's just-in-time compiler is accessed by importing jit from numba: from numba import jit.

2. Step 2: Check other options

   Options A, C, and D are invalid imports; 'import numba as nb' imports the whole module but requires nb.jit to use the decorator, while C and D do not exist.

3. Final Answer:

   from numba import jit -> Option C

4. Quick Check:

   Correct Numba import = from numba import jit [OK]

Hint: Use 'from numba import jit' to access speedup decorator [OK]

Common Mistakes:

• Trying to import submodules that don't exist
• Importing whole numba without jit
• Using incorrect import syntax

3. What will be the output of this code snippet?

import numpy as np
from numba import jit

@jit(nopython=True)
def sum_squares(n):
    total = 0
    for i in range(n):
        total += i * i
    return total

print(sum_squares(5))

medium

A. 30

B. 55

C. 10

D. Error due to jit decorator

Solution

1. Step 1: Understand the function logic

   The function sums squares of numbers from 0 to n-1. For n=5, it sums 0²+1²+2²+3²+4² = 0+1+4+9+16 = 30.

2. Step 2: Check jit decorator effect

   The jit decorator speeds up the function but does not change output. No error occurs.

3. Final Answer:

   30 -> Option A

4. Quick Check:

   Sum of squares 0 to 4 = 30 [OK]

Hint: Sum squares 0 to 4 = 30; jit doesn't change output [OK]

Common Mistakes:

• Confusing sum of squares with sum of numbers
• Expecting jit to cause errors
• Miscounting range end

4. Identify the error in this code that tries to speed up a NumPy operation using Numba:

import numpy as np
from numba import jit

@jit
def multiply_array(arr):
    return arr * 2

print(multiply_array(np.array([1, 2, 3])))

medium

A. Numba jit cannot speed up NumPy array operations directly

B. The jit decorator is missing parentheses

C. The function multiply_array must use a loop inside

D. NumPy arrays cannot be passed to jit functions

Solution

1. Step 1: Understand Numba's limitations

   Numba jit speeds up Python loops and scalar operations but does not speed up vectorized NumPy operations directly.

2. Step 2: Analyze the code behavior

   The function returns arr * 2, which is a NumPy vectorized operation. Numba jit won't speed this up; it just calls NumPy internally.

3. Final Answer:

   Numba jit cannot speed up NumPy array operations directly -> Option A

4. Quick Check:

   Numba jit doesn't speed vectorized NumPy ops = Numba jit cannot speed up NumPy array operations directly [OK]

Hint: Numba jit speeds loops, not vectorized NumPy ops [OK]

Common Mistakes:

• Thinking jit decorator syntax is wrong
• Believing jit requires loops inside
• Assuming NumPy arrays can't be used with jit

5. You have a large NumPy array and want to speed up a custom function that applies a complex calculation element-wise. Which approach is best?

hard

A. Use pure Python loops without any acceleration

B. Rewrite the function using Numba jit with a loop over array elements

C. Convert the array to a list and use list comprehensions

D. Use NumPy vectorized operations only, no jit

Solution

1. Step 1: Understand the problem context

   Complex element-wise calculations may not be vectorizable easily with NumPy, so pure vectorized operations might not work.

2. Step 2: Choose the best speedup method

   Using Numba jit with a loop over array elements compiles the loop to fast machine code, speeding up the custom calculation.

3. Step 3: Evaluate other options

   Pure Python loops and list comprehensions are slow; vectorized NumPy ops may not support complex logic.

4. Final Answer:

   Rewrite the function using Numba jit with a loop over array elements -> Option B

5. Quick Check:

   Numba jit + loops = best speedup for complex element-wise ops [OK]

Hint: Use Numba jit with loops for complex element-wise speedup [OK]

Common Mistakes:

• Relying on pure Python loops for speed
• Converting arrays to lists unnecessarily
• Assuming all complex ops can be vectorized