Tensor Slicing and Indexing

1. Background Knowledge

PyTorch tensors are multi-dimensional arrays similar to NumPy arrays, supporting advanced indexing and slicing operations that enable efficient data extraction without copying data (views are returned when possible). Key concepts include:

• Basic Indexing: tensor[i, j] accesses element at row i, column j.
• Slicing: Uses Python slice notation start:stop:step.
• Rows: tensor (first row), tensor[-1] (last row).
• Columns: tensor[:, -1] (last column, : selects all rows).
• Sub-matrices: tensor[row_slice, col_slice].
• Tensor Shape: For a 4×4 tensor, shape = (4, 4). Center 2×2 is rows/cols [1:3, 1:3] (indices 1,2).
• Views vs. Copies: Slicing creates views (zero-copy), preserving gradients for autograd.

These operations are $O(1)$ for view creation, essential for memory-efficient ML pipelines.

2. Algorithm Approach

This problem uses direct indexing/slicing, a fundamental tensor operation in PyTorch (NumPy-compatible). No complex algorithms needed:

• First row: Single-index slice on dimension 0.
• Last column: Full slice on dim 0, single index on dim 1.
• Center sub-matrix: 2D slice on both dimensions.

For fixed-size (4×4) tensors, compute indices statically:

• Last column index: tensor.size(1) - 1.
• Center: (n//2-1): (n//2+1) where $n=4$.

3. Step-by-Step Strategy

1. Inspect input: Confirm tensor.shape == (4, 4) (per constraints).
2. Extract first row: first_row = tensor.
3. Extract last column: last_col = tensor[:, -1] (or tensor[:, 3]).
4. Extract center 2×2: center = tensor[1:3, 1:3].
5. Package results: Return {"first_row": first_row, "last_column": last_col, "center": center}.

Complete Solution:

import torch

def extract_parts(tensor):
 """
 Extract specific parts from a 4x4 tensor.

 Args:
 tensor: 4x4 PyTorch tensor