Developer Documentation

Welcome to the Developer API page. Here you'll find documentation on functionality only available to developers via downloading the source code of nexport. Clone or fork this repository to get started!

Table of contents

Generic functions (generic.py)	Utility functions (utils.py)	Model classes (models.py)
append_extension()	detect_framework()	BFNetwork()
		FFNetwork()
		ICARNetwork()

Export functions (*/exporting.py)	Import functions (*/importing.py)
pytorch.exporting.create_layer_object()	pytorch.importing.import_from_file()
pytorch.exporting.create_model_metadata()
pytorch.exporting.create_model_object()
pytorch.exporting.create_parameter_arrays()
pytorch.exporting.export_to_file()
pytorch.exporting.export_to_json()

How nexport works

When you call a function - we'll call these "user functions" - in nexport, it'll call "hidden functions" that handle data input & output for you. nexport automatically detects which deep learning framework you are using from which modules you've imported in your current Python session. It uses this information to make decisions on which hidden functions to use, therefore dynamically selecting and executing code depending on environment variables. You can check which deep learning framework nexport thinks you're using by executing the following line after importing nexport:

print(nexport.__framework__)

When nexport calls hidden functions, it passes the arguments you provide when calling user functions across all functions called. This wiki page endeavors to explain where and how nexport calls hidden functions and which arguments are passed.

Utility functions

Utility functions are functions which can be called directly from nexport that don't involve exporting or importing deep neural networks. These are helper functions that can provide information about your neural networks and are useful when you need architecture information quickly.

Calculators

Generic functions

Generic functions are functions which are shared by the deep learning frameworks' nexport functions. These are not usually callable by the user but are separated from the framework-specific hidden functions to keep the module DRY.

append_extension

This function verifies that the provided file extension is supported by nexport and appends the extension to the provided file name. The function then returns the full filename. This function removes the reliance on the user to remember to append their own file extension and eliminates the error introduced with such a reliance.

Usage

filename = nexport.append_extension(filename="model",
                                    extension="txt"
)

Arguments

Argument	Data type	Default	Description
filename	string		Name to be used for exported file
extension	string		Extension of filetype to be used for exported file

Output

Output	Data type	Description
filename	string	File name including file extension

Export functions

export_to_file()

This function allows users to export an instantiated PyTorch model to .txt format. It is essentially a parameter dump of multi-dimensional arrays.

Usage

nexport.export_to_file(model=model,
                       filename="model file"
)

Arguments

Argument	Data type	Default	Description
model	object		Instantiated PyTorch model
filename	string	model	Name to be used for output file

Output

This function does not currently return anything, but a success boolean return is planned.

export_to_json()

This function allows users to export an instantiated PyTorch model to .json format. The file is structured to be as human-readable as possible and parameters are grouped into layers, then neurons, then weights or biases.

Usage

nexport.export_to_json(model=model,
                       filename="model file",
                       indent=2,
                       verbose=2,
                       include_metadata=True,
                       model_name="My Neural Network",
                       model_author="Andrew Ng"
)

Arguments

Argument	Data type	Default	Description
model	object		Instantiated PyTorch model
filename	string	model	Name to be used for output file
indent	integer	4	Number of spaces to be used to indent JSON object
verbose	integer	1	Degree of status output for model export
include_metadata	boolean	False	Whether to include arguments as metadata header in output file
model_name	string	My Model	Name of model to be used in metadata header
model_author	string	Name of logged-in user	Name of author of model to be used in metadata header

Accepted values for verbose: [0, 1, 2, 3]

Output

This function does not currently return anything, but a success boolean return is planned.

Import functions

import_from_file()

This function allows the user to import a PyTorch model by parsing the .txt file generated by export_to_file(). It outputs an instantiated PyTorch model with initialized parameters.

Usage

model = nexport.import_from_file(filepath="Desktop/model.txt",
                                 framework="PyTorch",
                                 architecture="linear"
)

Arguments

Argument	Data type	Default	Description
filepath	string		Path to file to be used for import
framework	string	PyTorch	Framework used to initially instantiate model
architecture	string	linear	Architecture used to connect model's hidden layers

Accepted values for framework: ['PyTorch']

Accepted values for architecture: ['linear']

Output

Output	Data type	Description
model	object	Instantiated model