VGSLify: Variable-size Graph Specification for TensorFlow & PyTorch

VGSLify simplifies defining, training, and interpreting deep learning models using the Variable-size Graph Specification Language (VGSL). Inspired by Tesseract's VGSL specs, VGSLify enhances and streamlines the process for both TensorFlow and PyTorch.

Table of Contents

• Installation
• How VGSL Works
• Quick Start
  • Generating a Model with VGSLify
  • Creating Individual Layers with VGSLify
  • Converting Models to VGSL
• Additional Documentation
• Contributing
• License
• Acknowledgements

Installation

Basic Installation

To install VGSLify without any deep learning backend, run:

pip install vgslify

This installs only the core functionalities of VGSLify without torch or tensorflow.

Installing with a Specific Backend

VGSLify supports both TensorFlow and PyTorch. You can install it with the required backend:

# For TensorFlow (latest compatible version)
pip install vgslify[tensorflow]
# For PyTorch (latest compatible version)
pip install vgslify[torch]

By default, this will install:

• tensorflow (latest stable version)
• torch (latest stable version)

Controlling Backend Versions

If you need a specific version of torch or tensorflow, install VGSLify first and then manually install the backend:

pip install vgslify
pip install torch==2.1.0 # Example for PyTorch
pip install tensorflow==2.14 # Example for TensorFlow

Alternatively, you can specify the version during installation:

pip install vgslify[torch] torch==2.1.0
pip install vgslify[tensorflow] tensorflow==2.14

⚠ Note: If a different version of torch or tensorflow is already installed, pip may not downgrade it automatically. Use --force-reinstall or --upgrade if necessary:

pip install --upgrade --force-reinstall torch==2.1.0

Verify installation

To check that VGSLify is installed correctly, run:

python -c "import vgslify; print(vgslify.__version__)"

How VGSL Works

VGSL uses concise strings to define model architectures. For example:

None,None,64,1 Cr3,3,32 Mp2,2 Cr3,3,64 Mp2,2 Rc3 Fr64 D20 Lrs128 D20 Lrs64 D20 Fs92

Each part represents a layer: input, convolution, pooling, reshaping, fully connected, LSTM, and output. VGSL allows specifying activation functions for customization.

Quick Start

Generating a Model with VGSLify

fromvgslifyimportVGSLModelGenerator# Define the VGSL specificationvgsl_spec="None,None,64,1 Cr3,3,32 Mp2,2 Fs92"# Choose backend: "tensorflow", "torch", or "auto" (defaults to whichever is available)vgsl_gn=VGSLModelGenerator(backend="tensorflow") model=vgsl_gn.generate_model(vgsl_spec, model_name="MyModel")
model.summary()
vgsl_gn=VGSLModelGenerator(backend="torch") # Switch to PyTorchmodel=vgsl_gn.generate_model(vgsl_spec, model_name="MyTorchModel")
print(model)

Creating Individual Layers with VGSLify

fromvgslifyimportVGSLModelGeneratorvgsl_gn=VGSLModelGenerator(backend="tensorflow")
conv2d_layer=vgsl_gn.construct_layer("Cr3,3,64")
# Integrate into an existing model:# model = tf.keras.Sequential()# model.add(conv2d_layer) # ...# Example with generate_history:history=vgsl_gn.generate_history("None,None,64,1 Cr3,3,32 Mp2,2 Fs92")
forlayerinhistory:
print(layer)

Converting Models to VGSL

fromvgslifyimportmodel_to_specimporttensorflowastf# Or import torch.nn as nn# TensorFlow example:model=tf.keras.models.load_model("path_to_your_model.keras") # If loading from file# PyTorch example:# model = MyPyTorchModel() # Assuming MyPyTorchModel is defined elsewherevgsl_spec_string=model_to_spec(model)
print(vgsl_spec_string)

Note: Flatten/Reshape layers might require manual input shape adjustment in the generated VGSL.

Additional Documentation

See the VGSL Documentation for more details on supported layers and their specifications.

Contributing

Contributions are welcome! Fork the repository, set up your environment, make changes, and submit a pull request. Create issues for bugs or suggestions.

License

MIT License. See LICENSE file.

Acknowledgements

Thanks to the creators and contributors of the original VGSL specification.