MXNet不同版本中nd与np数组的导入及差异疑问

Hey there! I totally get the confusion when switching between MXNet versions and navigating the two array APIs—let’s break this down clearly for you.

Why does MXNet 1.6.0b20190915 support from mxnet import np? #

MXNet introduced this NumPy-compatible API in its 1.6 preview versions to lower the learning curve for folks already familiar with NumPy. Before 1.6 (like your 1.5.1.post0 version), MXNet only had its native ndarray module (nd), which had its own set of APIs that weren’t aligned with NumPy. By adding the np namespace, MXNet let users write code that feels almost identical to NumPy while still getting all the MXNet benefits (like auto-grad, GPU acceleration, etc.).

What's the difference between arrays created with nd vs np? #

Even though they look different on the surface, they share a lot under the hood—here are the key distinctions:

• API Design:
  • nd is MXNet's original array module, with APIs tailored specifically for MXNet's workflow (e.g., some method names or parameter defaults might differ from NumPy).
  • np is a NumPy-compatible wrapper: most functions and methods match NumPy's syntax exactly. For example, np.reshape() works just like NumPy's, whereas nd.reshape() might have minor differences in how you pass arguments.
• Underlying Core:
  Both nd.array() and np.array() create MXNet tensors at their core. That’s why you can use MXNet-exclusive features like attach_grad() on both—they’re just different API layers on top of the same low-level tensor structure.
• Use Cases:
  • Use nd if you're working with legacy MXNet code, or need access to MXNet-specific ndarray features that aren’t replicated in the np module.
  • Use np if you're coming from a NumPy background, want to reuse existing NumPy-style code, or prefer the familiar NumPy API while leveraging MXNet's deep learning tools.

A quick note: The np module isn’t 100% identical to NumPy—some advanced NumPy features might not be implemented yet. If you hit a roadblock, check MXNet’s documentation for the np module to see what’s supported.

内容的提问来源于stack exchange，提问作者Carlo