Supported Models and Quantization
This section outlines the supported model types for code generation and the quantization configurations available in the DEEPCRAFT™ Model Converter.
Model Converter does not support code generation for PyTorch models with multiple inputs and/or outputs.
Supported Models
You can generate code for TensorFlow, Keras or PyTorch models with the following specifications:
TensorFlow Lite or TensorFlow Keras
You can use TensorFlow models built with Keras saved in (.h5) or (.keras) format. If you use TensorFlow 2.16+ and Keras 3 to build a model, save the model in the new (.keras) file format.
PyTorch
You must use torch==2.6.0+, when exporting models to the .pt2 format. Models exported with other PyTorch versions are not supported. For information on exporting pytorch models to .pt2, refer to torch.export and serialization section in the PyTorch documentation.
Supported Quantization
The table below lists the supported quantization configurations in Model Converter:
| Supported Quantization | TensorFlow Keras | TensorFlow Lite | PyTorch |
|---|---|---|---|
| Float32 activations + Float32 weights | ✓ | ✓ | ✓ |
| Int8 activations + Int8 weights | ✓ | ✓ | ✓ |
| Int16 activations + Int8 weights | ✓ | ✓ | x |
You can quantize Keras (.h5) or (.keras) models to Int8×8 or Int16×8 by providing representative input samples for calibration.
Models must have a single input and output. The input and output must use the same data type:
- Float32 in → Float32 out
- Int8 in→ Int8 out
- Int16 in → Int16 out