Model Building and Training

Programming Forms

• Programming Forms Overview
• Dynamic Graph
• Static Graph

Data Loading and Processing

• Data Processing Overview
• Data Sampling
• Converting Dataset to MindRecord
• Lightweight Data Processing
• Supporting Python Objects in Dataset Pipeline
• Auto Augmentation
• Single-Node Data Cache
• Optimizing the Data Processing

Model Building

• Overview
• Tensor and Parameter
• Functional and Cell

Training Process

• Training Process Overview
• Advanced Encapsulation: Model
• Optimizing Training Performance
• Advanced Automatic Differentiation
• Algorithm Optimization

Distributed Parallelism

• Distributed Parallelism Overview
• Distributed Parallel Startup Methods
• Data Parallel
• Semi-automatic Parallel
• Automatic Parallel
• Manually Parallelism
• Parameter Server
• Model Saving and Loading
• Fault Recovery
• Optimization Techniques
• Experimental Characteristics
• Distributed High-Level Configuration Case

Custom High-order Programming

• Overview of Custom Higher-Order Programming
• Custom Operators
• Custom Parameter Initialization
• Custom Loss Function
• Custom Optimizer
• Custom Fusion Pass
• Custom Neural Network Layers
• Hook Programming

Training High Availability

• Fault Recovery
• Training Process Exit Gracefully
• Interconnection with third-party storage systems

Model Debugging

• Overview of Model Debugging
• Using Dump in the Graph Mode
• Running Data Recorder
• Feature Value Detection
• Error Reporting Analysis
• ErrorMap↗

Model Tuning

• Overview of Model Tuning
• Graph Kernel Fusion
• Memory Reuse
• Ascend Optimization Engine (AOE)

Models Library

• Official Models