MindSpore Programming Guide

Overview

• Overall Architecture
• MindSpore API Overview

Quickstart

• Implementing Simple Linear Function Fitting
• Implementing an Image Classification Application
• Hands-on Installation and Experience

Basic Concepts

• dtype
• Tensor
• Operators
• Cell

Numpy

• NumPy Interfaces in MindSpore

Execution Management

• Configuring Running Information

Data Pipeline

• Quick Start of Dataset
• Loading Dataset
• Processing Data
• Advanced Usage of Pipeline

Build the Network

• Initialization of Network Parameters
• Updating Network Parameters
• Model Layers
• Loss Function
• Optimization Algorithms
• Building a Customized Network
• Common Network Components

Train Models

• Running Mode
• Callback Mechanism
• Saving Models
• Loading a Model for Inference and Transfer Learning
• Advanced Usage of Training

Inference

• Inference Model Overview
• Inference on the Ascend 910 AI processor
• Inference on Ascend 310
• Inference on a GPU
• Inference on a CPU

Distributed Training

• Distributed Training Overview
• Parallel Distributed Training (Ascend)
• Distributed Parallel Training (GPU)
• Pipeline Parallelism Application
• Applying Host&Device Hybrid Training
• Training with Parameter Server
• Saving and Loading Models in Hybrid Parallel Mode
• Distributed Inference With Multi Devices
• Parallel Distributed Training Interfaces

Function Debugging

• Debugging in PyNative Mode
• Using Dump in the Graph Mode
• Custom Debugging Information
• Evaluating the Model during Training
• Incremental Operator Build

Performance Optimization

• Optimizing the Data Processing
• Enabling Mixed Precision
• Enabling Graph Kernel Fusion
• Applying a Gradient Accumulation Algorithm
• Applying Quantization Aware Training
• Applying Post Training Quantization