PAI-EasyVision provides training and prediction capabilities for computer vision (CV) models, helping you build and deploy visual AI applications.
Building CV models with deep learning presents several challenges:
-
Deep learning code is costly to develop and debug.
-
Models evolve rapidly, making it time-consuming to understand their principles.
-
Optimizing training and inference performance requires specialized system knowledge.
-
Data annotation costs are high.
-
Adapting open-source algorithms to PAI has a learning curve.
PAI-EasyVision is a user-friendly CV framework that helps developers quickly build and deploy visual models. Core advantages:
-
Ease of use
Supports multiple visual tasks through modular, pluggable interfaces. Covers the full pipeline from data I/O and preprocessing to training and offline prediction. Works in Designer and Data Science Workshop (DSW).
-
Performance
Leverages PAI-TensorFlow optimization engines for distributed training, compilation optimization, and mixed precision. Achieve high performance with simple configuration files. Compatible with open-source TensorFlow.
-
Model variety
Offers pre-trained models on open-source datasets and integrates PAI models such as OCR, reducing development and training costs.
Architecture
PAI-EasyVision provides an extensive Model Zoo with flexible invocation methods including PAI-VIP, PAI commands, and DSW. It uses a high-availability distributed pipeline for offline prediction that processes hundreds of millions of records. PAI system and model optimization produce smaller, faster models for deployment on Elastic Algorithm Service (EAS). Custom training and prediction interfaces let you reuse existing optimizations.
Features
-
Ease of use
Supports interactive training, scheduled tasks, and module-level customization. Invoke PAI-EasyVision through PAI-VIP, PAI commands, or DSW.
-
Performance optimization
Built on PAI-TensorFlow for high-performance distributed training across single and multi-GPU servers. Supports inference-stage optimization including graph optimization and model compression.
-
Connection to PAI Smart Labeling
Integrates with PAI Smart Labeling. Convert labeled data to TFRecord format for training with the built-in conversion tool. Includes data augmentation modules to dynamically expand training datasets.
-
Efficient offline prediction
Multi-server pipeline system for offline prediction with multi-threaded acceleration at each step. Asynchronous pipeline processing maximizes throughput. Each step is customizable.
-
Connection to EAS
Training produces a SavedModel file for integration with your online prediction system. EAS provides online prediction services through the PAI-EasyVision EAS Python Processor. Configure the model address and type in the configuration file for real-time processing.