PyTorch张量语境下‘view’的正式定义是什么？

First, let's recap the context from Deep Learning with PyTorch:

Python数字列表或元组是由单独分配内存的Python对象组成的集合，如图3.3左侧所示。而PyTorch张量或NumPy数组则是（通常）对包含未装箱C语言数值类型而非Python对象的连续内存块的视图（view）。在本例中，每个元素都是32位（4字节）浮点数，如图3.3右侧所示。这意味着存储一个包含100万个浮点数的1D张量恰好需要400万个连续字节，再加上少量元数据（如维度和数值类型）的开销。

In this context, view is best thought of as an "interpretation layer" or a "window" into a single contiguous block of raw memory. Here's a straightforward breakdown:

• The actual numerical data lives in one unbroken chunk of memory (no extra Python object overhead—just raw binary values like 32-bit floats). For that 1M-element tensor example, that's exactly 4MB of contiguous bytes.
• The tensor itself doesn't store this raw data—it holds metadata that tells PyTorch how to read that memory block. This metadata includes key details like:
  • The tensor's shape (e.g., a 1D (1000000,) tensor or a 2D (1000, 1000) tensor)
  • Data type (float32, int64, etc.)
  • Strides (how many bytes to jump to get the next element in each dimension)
  • The starting offset in the memory block

To make this concrete: you can create two totally different tensors from the same underlying memory block. One could be a 1D tensor that reads elements in sequence, and another could be a 2D tensor that interprets the same bytes as rows and columns. Both tensors reference the exact same memory—modify an element in one, and the other will show the change immediately—because they're just two different "views" of the same data.

This approach is incredibly efficient because it avoids copying large chunks of memory when reshaping or slicing tensors. Instead, you just tweak the metadata to create a new view of the existing data.

内容的提问来源于stack exchange，提问作者hanugm