NetsPresso Trainer

NetsPresso Trainer is an open-source project that provides various model architectures, training, evaluation, and inference for AI models compatible with the NetsPresso Pipeline. Using NetsPresso Trainer, you can create models that support all of NetsPresso’s capabilities, including initial training, compression, conversion to other intermediate representations, and benchmarking on actual devices in the cloud.

Key Features

NetsPresso Trainer focuses on model compression and device deployment, ensuring that the models adhere to the following criteria:

• Compatible with torch.fx conversion.
• Compressible using the pruning methods provided by NetsPresso.
• Easily deployable across a variety of edge devices.
• NetsPresso Trainer can be used through the PyNetsPresso API. https://github.com/Nota-NetsPresso/PyNetsPresso/tree/v1.10.0.b9?tab=readme-ov-file#trainer

Model Structure

To offer a wide range of models that meet these requirements in various forms, NetsPresso Trainer defines models using four key components: full, backbone, neck, and head. This modular approach allows users to configure backbones, necks, and heads as needed. For models that do not easily fit into these three segments, we provide them as full models.

You can find more information on the official NetsPresso Trainer page below.

https://nota-netspresso.github.io/netspresso-trainer/