Skip to content

Latest commit

 

History

History
30 lines (19 loc) · 1.25 KB

README.md

File metadata and controls

30 lines (19 loc) · 1.25 KB

Model Compiler

  1. Support optimization for models from different kinds of deep learning architecture, e.g. TensorFlow/Caffe/PyTorch.

  2. Support compiling models for different runtime, e.g. OpenVINO IR/TensorFlow Serving/TensorFlow Lite/TensorRT.

  3. Simplified interfaces for the workflow.

Usage

  1. Prepare a model file in the specified format, such as H5, Checkpoint, Frozen Graph, ONNX.

  2. Create a json file which must match config_schema.json.

  3. Install model_compiler and compile the model, you can refer to benchmark link or examples link:

cd {Adlik_root_dir}/model_compiler
python3 -m pip install .

Quantify model

  1. When you compile the model to tensorflow lite runtime, you can quantify the model by set the parameter optimization， supported_types， supported_ops，inference_input_type，inference_output_type in json file or environment variable, you can refer to quantify tensorflow lite model.

  2. When you compiler the model to tensorrt runtime, you can quantify the model to FP16 by set the parameter enable_fp16 and enable_strict_types in json file or environment variable.