MindSpore is a full-scenario deep learning framework designed to achieve easy development, efficient execution, and unified deployment across diverse environments.

Framework Architecture Overview

ModelZoo: A repository of pre-built deep learning models and algorithms. MindSpore Extend: Domain-specific extansions supporting new scenarios like Graph Neural Networks (GNN), deep probabilistic programming, and reinforcement learning. MindSpore Science: A scientific computing toolkit built on MindSpore, featuring datasets, base models, pre-trained high-precision models, and preprocessing tools to accelerate development in scientific applications. MindExpression: A unified Python-based API that integrates functional and object-oriented programming paradigms, combining AI with numerical computation, and unifying dynamic and static exprestions aswell as single-machine and distributed expressions. Third-Party Frontends: Support for multiple third-party language frontends, with plans to include C/C++ and other ecosystems in the future. MindSpore Data: A data processing layer providing efficient data handling, common dataset loading, and flexible user-defined processing with pipeline optimization. MindCompiler: The core graph-level compiler, based on the unified MindIR format, offering hardware-agnostic optimizations (type inference, automatic differentiation, expression simplification), hardware-specific optimizations (automatic parallelism, memory optimization, graph-fusion, pipeline execution), and deployment-related optimizations (quantization, pruning). MindRT: The runtime system for MindSpore, covering cloud-side host runtime, edge-side runtime, and lightweight runtime for IoT devices. MindSpore Insight: A visualization tool for debugging and optimization, enabling users to view training processes, model performance, precision issues, and interpret inference results. MindSpore Armour: Security and privacy enhancement features for enterprise-level deployments, including adversarial robustness, model security testing, differential privacy training, privacy leakage risk assessment, and data drift detection.

Execution Workflow

MindSpore supports diverse hardware across cloud, edge, and device scenarios, including Ascend series products, NVIDIA GPUs, Arm-based platforms like Qualcomm Snapdragon, and Huawei Kirin chips.

The main framework provides foundational APIs for neural network training and validation, with default features like automatic differentiation and parallelism.

Below the core framework, the MindSpore Data module handles data preprocessing, including sampling, iteration, and format conversion. During training, debugging and optimization issues are addressed by the MindSpore Insight module, which visualizes loss curves, operator execution, and parameter variations.

Above the core framework, components relevant to algorithm development include ModelZoo for AI models, MindSpore DevKit for domain-specific development tools, and the high-level extension library MindSpore Extend. Notably, MindSpore Extend includes MindScience, which integrates scientific computing with deep learning, combining numerical methods with AI to support applications like electromagnetic simulation and molecular drug simulation.

After training a neural network model, it can be exported or loaded from MindSpore Hub for pre-trained models. The unified IR format, MindIR, defines the logical structure and operator properties of networks, decoupling model files from hardware platforms to enable train-once, deploy-many capabilities. Through IR, models are exported to different modules for inference execution.

Quick Start Guide

Import MindSpore

import mindspore
from mindspore import nn
from mindspore.dataset import vision, transforms
from mindspore.dataset import MnistDataset

Process Dataset

MindSpore uses a pipeline-based data engine for efficient preprocessing via Dataset and Transforms.

from download import download
# Download dataset
url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/" \
      "notebook/datasets/MNIST_Data.zip"
path = download(url, "./", kind="zip", replace=True)

# Obtain dataset objects
train_data = MnistDataset('MNIST_Data/train')
test_data = MnistDataset('MNIST_Data/test')

# Transform image data and labels
def process_data(dataset, batch_num):
    img_transforms = [
        vision.Rescale(1.0 / 255.0, 0),
        vision.Normalize(mean=(0.1307,), std=(0.3081,)),
        vision.HWC2CHW()
    ]
    label_cast = transforms.TypeCast(mindspore.int32)

    dataset = dataset.map(img_transforms, 'image')
    dataset = dataset.map(label_cast, 'label')
    dataset = dataset.batch(batch_num)
    return dataset

# Process training and test sets
train_data = process_data(train_data, 64)
test_data = process_data(test_data, 64)

# Iterate through dataset
for img, lbl in test_data.create_tuple_iterator():
    print(f"Image shape [N, C, H, W]: {img.shape} {img.dtype}")
    print(f"Label shape: {lbl.shape} {lbl.dtype}")
    break