MindSpore

什麼是 MindSpore？

MindSpore is an open-source deep learning framework for developers who write model code in Python. Its workflow connects datasets and tensor transformations with network definitions, loss functions, automatic differentiation and parameter updates.

You can evaluate a trained network, save its parameters as a checkpoint, and load them into a model for predictive inference. Installation depends on the selected version and hardware environment. The 2.10.0 Linux CPU and Ascend pip installations use Python 3.10–3.12.

MindSpore 能做什麼？

Prepare data in a pipeline

Load a dataset, apply image and label transformations with map operations, and batch the results for training and testing.

Define custom networks

Build network classes with mindspore.nn.Cell, define layers in __init__, and describe the forward calculation in construct.

Train with automatic differentiation

Compute a loss, obtain parameter gradients with value_and_grad, and pass the gradients to an optimizer. Run a separate evaluation loop to measure loss and prediction accuracy.

Save and reload model parameters

Write parameters to a .ckpt checkpoint, recreate the model structure, and load its saved weights before running predictions.

使用情境

“Learn an image-classification workflow”

Use the MNIST example to connect preprocessing, a fully connected network, training and test-set predictions in one Python project.

“Develop a custom neural network”

Customize the layers and forward computation in a Cell subclass while keeping explicit control over the loss function, optimizer and training loop.

“Reuse trained weights for inference”

Restore checkpoint parameters into the same model definition and run the network on input data without retraining it.

常見問題

What is required to install MindSpore 2.10.0?

Select an installation procedure for your hardware and operating system. The 2.10.0 Linux CPU and Ascend pip installations require Python 3.10–3.12 and GCC 9.5.0–11.3.0. Ascend additionally needs a compatible CANN package and runtime environment configuration.

How does automatic differentiation fit into training?

Define a forward function that returns the loss, use value_and_grad to compute the loss and parameter gradients, then apply an optimizer update. You write the model and training loop in Python.

Can I use a saved model for prediction?

Yes. Recreate the network, load the checkpoint and apply the parameters to the model. Switch it out of training mode before using it for predictive inference.

Is the core framework free to use?

The core code uses Apache License 2.0, whose copyright and patent grants are no-charge, subject to its terms. That software license covers the framework code. Hardware and cloud computing are separate resources.

MindSpore 流量分析

流量數據僅供參考。

月造訪量

10.6萬

全球排名

#359,035

使用者跳出率

39.1%

平均造訪時長

06:36

每次造訪頁數

3.46

網域評分

流量趨勢

月造訪量 - 3
5月	61602
6月	85061
7月	106439

流量來源

尚無流量分析數據。