Lattice sensAI Studio is a collection of tools that assist developers to experiment with different DNN models, provides training scripts, simulation to validate functions when deployed on an FPGA without the need for HW platforms, and a flexible post training quantization package to enable converting floating point models to FPGA compatible models.
Lattice sensAI™ Neural Networks Training Environment has been designed to standardize DNN training code to eliminate the repetitive task of writing boilerplate code whenever a new DNN model is developed. By providing a unified framework, the training environment tool streamlines the development process, ensuring consistency and efficiency across various projects.
This tool is first and foremost a Python package providing a set of Python classes and utility functions that simplifies the development of the training process of DNN models. It is also a command line program that will helps perform the three most common tasks that need to be done when it comes to DNN models: training, testing, and converting models.
Features:
- Separation of hyper-parameters (YAML / JSON) and Python code
- Automatic checkpoint management (save after each epoch, restore when restarting training)
- Automatic logging of losses and metrics in Tensorboard
- Support for Weights and Biases (automatic logging of losses, metrics, artefacts, configurations, checkpoints, converted models, etc.)
- Integration with Lattice sensAI™ Neural Networks Quantizer (LSCQuant) to perform quantization-aware training
- Integration with Lattice sensAI™ ML Engine Simulator to test quantized models as if they were run on FPGA
- Support model conversion to HDF5 (Keras .h5 files), frozen PB, ONNX, or TFLite
Lattice sensAI™ ML Engine Simulator is a tool that allows running networks in software and getting the same results as if they were run on a Lattice Semiconductor FPGA. With the Simulator, we can analyze not only the final output of the network, but also the results of each layer individually.
The ML Engine Simulator is implemented in C++, but with Python bindings that allow it to be imported as a Python package for use in Python scripts.
- Can be run in both FPGA fixed-point and floating-point number modes.
- Supports more than 15 different layer types.
- Supports multiple input layers (experimental feature) and multiple output layers.
- Comes in an easy-to-use Python package that can be imported and used in any script.
- Can be used both on Windows and Linux.
Lattice sensAI™ Neural Networks Quantizer package provides a set of Keras layers that can be used to train quantized DNN suitable for Lattice Semiconductor FPGAs.
Lattice Semiconductor FPGAs support a subset of Keras layers, which means they do not require a quantized version of the layer. Some other layers require a quantized version, which must be used instead of the original Keras layer. These layers support two different quantization methods:
- Fixed Step-size Quantization (FPQ)
- Learned Step-size Quantization (LSQ)
This tool also provides a way to easily transform a non-quantized Keras models to a quantized model by substituting layers for their corresponding quantizable counterparts while applying a specified quantization scheme.