Stage 1: Model Development

Model Zoo

• A curated collection of pre-trained models across various tasks such as classification, detection, and segmentation.
• It enables users to start optimization without training from scratch.

📘

NetsPresso advantages:

• Offers a wide range of lightweight, mobile-friendly backbones ready for immediate training
• Seamless integration with the full optimization and deployment workflow

Trainer

• Used to train new models or fine-tune existing ones.
• It supports transfer learning and hyperparameter control.
  • Trainer: Available Models
  • NetsPresso Trainer Doc

📘

NetsPresso advantage:

• Hardware-aware training workflow; ensures all models are compatible with downstream compression and export
• Training graph conversion (via torch.fx) maximizes flexibility and future optimization
• Supports local datasets and Hugging Face integration

Compressor

• Reduces the size and complexity of a trained model through methods like Structured Pruning and Filter Decomposition.
• These techniques help maintain accuracy while significantly reducing model size and inference time.
  • How to use compressor
  • Compressor: Supported models
• Method details
  • Method: Filter Decomposition
  • Method: Structured Pruning

📘

NetsPresso advantages:

• Hardware-aware compression: Structured pruning and filter decomposition methods are tailored for target hardware, maximizing real-world efficiency on edge devices.
• Visual compression profiling: Instantly see which layers can be compressed and preview the impact, using Studio’s visual interface.
• Automatic fine-tuning support: After compression, models can be fine-tuned to recover accuracy with just a few clicks or a single command.
• Seamless workflow: Compressed models remain fully compatible with the next steps (quantization, conversion, benchmarking) with no manual rework needed.
• Extensive model support: Wide range of architectures supported, including MobileNet, ResNet, YOLO, EfficientFormer, and more.