Stage 2: Model Optimization

Graph Optimizer

• Simplifies and restructures computational graphs for faster and more efficient inference.
• This includes layer fusion and operator replacement.

📘

NetsPresso advantages:

• Automatic graph optimization after compression and quantization
• Improved inference latency and deployment stability

Quantizer

• Converts model weights and activations into lower-precision representations such as INT8.
• It supports calibration with sample datasets to maintain accuracy after quantization.
  • INT8 quantization with Optimization Studio

📘

NetsPresso advantages:

• Cutting-edge post-training quantization with minimal accuracy loss
• Built-in calibration workflow for robust performance on various datasets
• One-click INT8 export for TensorRT, TFLite, and more

IR Converter

• Converts models into intermediate representations suitable for cross-framework transformation.
• Features & Scope of support

📘

NetsPresso advantage:

• Multi-platform support (NVIDIA, Qualcomm, Renesas, Intel, and more)
• Automated model structure adjustment for maximum device compatibility
• No need for manual conversion or re-writing