Overview

nexport is a lightweight Python 3.10+ package which empowers deep learning developers in exporting the trainable parameters of deep neural networks to human-readable and transrerable file types.

Overview
Table of contents
Current support
Install & use
Objectives
History
Credits
License
Links

Current support

Filetype	PyTorch		Keras/TensorFlow
Filetype	Export	Import	Export	Import
Text (.txt)	✅	✅	🚧	❌
JSON (.json)	✅	🚧	❌	❌
CSV (.csv)	❌	❌	❌	❌
XML (.xml)	❌	❌	❌	❌

Install & use

From terminal: pip install nexport Or clone the repository to local and navigate to project folder pip install .
From python environment: import nexport
Call function via: nexport.export(...)

The example of using this function call is located in test/src/inference_example.py. To run it, you need to provide a network trained by PyTorch and place it under the test folder.

For current version, it is recommended to address all the following specified the parameters:

nexport.export(
              model = YOUR_PYTORCH_MODEL,
              filetype = "json_exp",
              input_size = 80, output_size = 31, intercept=0.0, slope=1.0,
              include_metadata=True, model_name="YOUR_PYTORCH_MODEL_NAME", model_author="YOUR_PYTORCH_MODEL_AUTHOR", activation_function="gelu",
              using_skip_connections=False)

model: Required. A PyTorch model instance.

acceptable_engine_tag: Required. A string version number indicates an acceptable inference engine release. Because nexport generates a JSON file that is intimately tied to Inference-Engine, the JSON keywords may change when Inference-Engine's file format changes. This parameter indicates an Inference-Engine git tag known to be able to read and write the nexport output. Other Inference-Engine revisions might also be compatible.

file_type: Required. Mandatory to put json_exp for now

input_size and output_size: Required. The input and output size of your model

intercept and slope: the normalization boundary for input and output data. If no normalization, you can leave it without specifying them. The formula is x = (slope - intercept) * y + intercept, where y is normalized value and x is original data.

include_metadata: Required. If True, the final json file will have the metadata. Mandatory to put True for now.

model_name, model_author, activation_function: not mandatory but recommend to put some text in there.

using_skip_connections: should be false in most cases.

Objectives

Export weights and biases to human-readable file
Ensure compatability with all popular neural network development software

History

This package is intended to be used in conjunction with inference-engine. As such, nexport was developed by the inference-engine developers to enable compatability between the two softwares. nexport does this by exporting the weights and biases from networks compiled in PyTorch, Keras, and TensorFlow into standardized human-readable files. These files can be read by inference-engine to instantiate the netwoks in Fortran 2018 for inference.