MRT Documentation ¶

Contents

MRT Documentation

MRT, short of Model Representation Tool, aims at transforming floating model into a deterministic and non-data-overflow(containging upflow and downflow) format that CVM defines. Currently the tool only supported AI framework - MxNet, and we have developed the transformation program from TensorFlow into MxNet for google developer. PyTorch is also in our consideration, just wait for some time.

Pre-train Model ¶

There have been many pre-trained models in Gluon Model Zoo, these models are wide-spread used and regarded as an reference to various situations. Thus we do the quantization for those common floating models and expose an whole sequence of MRT execution as the mnist tutorial.

Pre-quantized Model List ¶

All the available models, which have benn quantized and tested accuracy in MxNet Gluon Zoo, are located in the python/mrt/model_zoos directory for reference.

Note

The pre-quantized models are executed via MRT’s configuration file settings. Refer the Configuration File for more details please.

Pre-quantized Model Introduction

Mnist Tutorial ¶

Mnist Training & Quantization

Quantization Documentation ¶

Quantization Documentation

V3 Documentation ¶

V3 Documentation
- V3 Architecture
- Benchmark Quantization Results

API Documentation ¶

Core python API briefly involves model quantization, compilation and evaluation. For standardized implementation of MRT stages, please refer to MRT Main2 API.

For other detailed classes and methods, please refer to the API links.