Thank you for sending your enquiry! One of our team members will contact you shortly.
Thank you for sending your booking! One of our team members will contact you shortly.
Course Outline
Introduction to CANN and Ascend AI Processors
- Definition of CANN and its function within Huawei’s AI computing ecosystem
- Summary of Ascend processor architectures (e.g., 310, 910)
- Overview of supported AI frameworks and the associated toolchain
Model Conversion and Compilation
- Application of the ATC tool for converting models from TensorFlow, PyTorch, and ONNX
- Generation and verification of OM model files
- Addressing unsupported operators and typical conversion challenges
Deployment via MindSpore and Alternative Frameworks
- Implementing model deployment with MindSpore Lite
- Integrating OM models with Python APIs or C++ SDKs
- Utilization of the Ascend Model Manager
Performance Optimization and Profiling
- Grasping optimizations related to AI Cores, memory management, and tiling
- Analyzing model execution using CANN profiling tools
- Best practices for boosting inference speed and optimizing resource consumption
Error Management and Debugging
- Identification and resolution of frequent deployment errors
- Interpretation of logs and application of error diagnostic utilities
- Conducting unit tests and functional validation for deployed models
Edge and Cloud Deployment Contexts
- Deployment to Ascend 310 for edge-based applications
- Integration with cloud-based APIs and microservice architectures
- Real-world case studies in computer vision and Natural Language Processing (NLP)
Recap and Future Steps
Requirements
- Proficiency in Python-based deep learning frameworks, including TensorFlow or PyTorch
- Solid grasp of neural network architectures and the model training process
- Fundamental knowledge of Linux command-line interface (CLI) and scripting
Intended Audience
- AI engineers focused on model deployment strategies
- Machine learning specialists aiming for hardware acceleration
- Deep learning developers constructing inference-based solutions
14 Hours
