Performance Optimization with Arrays

Performance optimization with arrays involves improving the efficiency of operations on arrays, such as reducing computation time and memory usage.

We should optimize performance for the following reasons −

• Speed: Faster computations lead to quicker results and more responsive applications.
• Scalability: Optimized code can handle larger datasets and more complex operations efficiently.
• Resource Efficiency: Reduces memory usage and computational overhead.

Using Vectorized Operations

Vectorized operations refer to the ability to perform operations on entire arrays or matrices in a single step without using explicit loops.

This is achieved through broadcasting and internal optimization, making these operations faster and more efficient.

Example

In the following example, we are performing vectorized addition of two large arrays, “a” and “b”, using NumPy’s array operations. This operation calculates the element-wise sum of the arrays and stores the result in a new array “c” −

# Open Compiler
import numpy as np

# Create two large arrays
a = np.random.rand(1000000)
b = np.random.rand(1000000)

# Vectorized addition
c = a + b
print(c)

Utilizing Efficient Data Types

Choosing the appropriate data type for your arrays is important for optimizing performance and memory usage in NumPy.

For example, using np.float32 instead of np.float64 can significantly impact memory usage and performance, particularly when working with large datasets.

In NumPy, a data type (or dtype) defines the kind of elements that an array holds and how much space is required to store each element.

Example

In this example, we are demonstrating the usage of precision change by creating an array with double precision (64-bit) floating-point numbers and then converting it to single precision (32-bit) using the astype() method −

# Open Compiler
import numpy as np

# Create an array with double precision (64-bit)
arr_double = np.array([1.0, 2.0, 3.0], dtype=np.float64)

# Print the original double precision array
print("Original double precision array:")
print(arr_double)
print("Data type:", arr_double.dtype)

# Convert to single precision (32-bit)
arr_single = arr_double.astype(np.float32)

# Print the converted single precision array
print("\nConverted single precision array:")
print(arr_single)
print("Data type:", arr_single.dtype)

Avoiding Loops with NumPy Functions

In NumPy, one of the primary advantages is the ability to avoid explicit loops by using built-in functions and array operations. This approach is often referred to as vectorization.

By using NumPy functions, you can perform operations on entire arrays at once, which is more concise compared to using loops.

Example

In the example below, we calculate the mean of the array elements using the np.mean() function, without using any explicit loops −

# Open Compiler
import numpy as np

# Create an array
arr = np.array([1, 2, 3, 4, 5])

# Calculate the mean of array elements
mean = np.mean(arr)
print("mean:", mean)

Using Broadcasting for Vectorization

Broadcasting refers to the ability to perform element-wise operations on arrays with different shapes. It follows a set of rules to determine how arrays with different shapes can be aligned for operations −

• Same Dimensions: If the arrays have different dimensions, the smaller array’s shape is padded with ones on the left until both shapes have the same length.
• Dimension Compatibility: Two dimensions are compatible when they are equal or one of them is 1. For each dimension, if the sizes are different and, if neither of them is 1, then the broadcasting fails.
• Stretching: Arrays with a dimension of size 1 are stretched along that dimension to match the size of the other array’s dimension.

Example

In the following example, we are broadcasting “array_1d” to match the shape of “array_2d”, allowing element-wise addition −

# Open Compiler
import numpy as np

# Create a 2D array and a 1D array
array_2d = np.array([[1, 2, 3], [4, 5, 6]])
array_1d = np.array([10, 20, 30])

# Add the 1D array to each row of the 2D array
result = array_2d + array_1d
print(result)

In-place Operations for Vectorization

In-place operations in NumPy refer to modifying the data of an array directly, without creating a new array to store the result, saving memory and improving performance.

This is achieved by using operators and functions that alter the content of the original array. These operations generally use operators with an in-place suffix (e.g., +=, -=, *=, /=) or functions that support in-place modification.

Example: Using In-place Operators

In this example, we are applying arithmetic operation “+=” directly on an array without creating a new one −

# Open Compiler
import numpy as np

# Create an array
arr = np.array([1, 2, 3, 4, 5])

# Add 10 to each element in-place
arr += 10
print(arr)

Example: Using In-place Functions

Here, we are calculating the exponential value of each element in an array in-place using NumPy exp() function −

# Open Compiler
import numpy as np

# Create an array with a floating-point data type
arr = np.array([1, 2, 3, 4, 5], dtype=np.float64)

# Compute the exponential of each element in-place
np.exp(arr, out=arr)
print(arr)

Using Memory Views for Vectorization

Memory views refer to different ways of accessing or viewing the same underlying data in an array without duplicating it. This concept allows you to create different “views” or “slices” of the array that can operate on the same data in various ways −

• Slicing: When you slice an array, NumPy creates a view of the original array, not a copy. This view shares the same data buffer, so changes to the view affect the original array and vice versa.
• Reshaping: Reshaping an array creates a new view of the same data with a different shape. This does not alter the underlying data but changes how it is interpreted.

Example: Slicing

In the example below, we create a 2D NumPy array and a view (slice) of the original array. Modifying the view also affects the original array −

# Open Compiler
import numpy as np

# Create a 2D array
arr = np.array([[1, 2, 3], [4, 5, 6]])

# Create a view (slice) of the original array
view = arr[:, 1:]

# Modify the view
view[0, 0] = 99

print(arr)

Example: Reshaping

Here, we create a 1D NumPy array using the arange() function and then reshape it into a 2D array with 3 rows and 4 columns, changing its structure while preserving the original data −

# Open Compiler
import numpy as np

# Create a 1D array
arr = np.arange(12)

# Reshape to a 2D array
reshaped = arr.reshape((3, 4))

print(reshaped)

Using Strides for Vectorization

Strides are a tuple that indicates the number of bytes to step in each dimension when traversing an array. They determine how array elements are accessed in memory, providing insight into how data is laid out and accessed.

Strides give you the memory offset for each dimension. For instance, in a 2D array, the stride for the second dimension tells you how many bytes to move in memory to access the next element in that row.

Example

In the following example, we create a 2D NumPy array and use the strides attribute to retrieve the number of bytes to step in each dimension when traversing the array −

# Open Compiler
import numpy as np

# Create a 2D array
arr = np.array([[1, 2, 3], [4, 5, 6]])

# Print the strides of the array
print(arr.strides)

Key Takeaway: Master performance optimization with NumPy arrays at Vista Academy!