numpy-indexing

---

name: numpy-indexing description: Advanced indexing techniques including slicing, fancy indexing, and boolean masks, along with memory implications of views vs. copies. Triggers: indexing, slicing, fancy indexing, boolean mask, np.where, np.ix_.

Overview

Indexing in NumPy ranges from basic slicing (zero-copy) to advanced "fancy" indexing (always creates a copy). Understanding the distinction is vital for memory management and avoiding unintended side effects in data analysis.

When to Use

• Extracting sub-regions of arrays for processing.
• Filtering data based on complex conditional logic (boolean masking).
• Selecting arbitrary elements using coordinate lists.
• Managing memory when dealing with large datasets that have small regions of interest.

Decision Tree

1. Do you need a view or a copy?
   • View: Use basic slicing (arr[0:5]).
   • Copy: Use advanced indexing (arr[[0, 1, 2]]) or .copy().
2. Are you filtering by value?
   • Use a boolean mask: arr[arr > threshold].
3. Selecting a grid of values across axes?
   • Use np.ix_ to construct the selection mesh.

Workflows

1. Filtering Data with Boolean Masks

   • Apply a comparison operator (e.g., x > 0) to an array to create a boolean mask.
   • Pass the mask into the array's indexing brackets: x[mask].
   • Operate on the resulting array (note that this is a copy, not a view).

2. Memory-Efficient Sub-array Extraction

   • Slice a small portion from a large ndarray.
   • Call .copy() on the slice to create a new independent array.
   • Delete the original large array to free system memory.

3. Cross-Axis Selection with np.ix_

   • Define row indices and column indices as separate lists.
   • Pass them into np.ix_ to construct the appropriate broadcasting meshes.
   • Apply the resulting objects to the array to select a sub-grid of values.

Non-Obvious Insights

• Memory Leak Risks: Small views of large arrays prevent garbage collection of the entire base array; always copy small slices of massive data.
• Copy vs. View Rule: Basic slicing always returns a view; advanced indexing (using non-tuple sequences or arrays) always returns a copy.
• Adjacent Indexing: Mixing basic and advanced indexing behavior changes significantly based on whether the advanced indices are adjacent in the index tuple.

Evidence

• "All arrays generated by basic slicing are always views of the original array." Source
• "Advanced indexing always returns a copy of the data (contrast with basic slicing that returns a view)." Source

Scripts

• scripts/numpy-indexing_tool.py: Demonstrates boolean masking and sub-array extraction.
• scripts/numpy-indexing_tool.js: Simulated coordinate selection logic.

Dependencies

• numpy (Python)

References

• references/README.md