- New API
getSupportedOperators()to know what operators have been supported. No longer need to maintain a list manually.
- FP16 quantization: fold FP16 tensors to unblock MediaPipe models. See this issue for details.
Thanks for the contribution of @briangrifiin!
- More operators, check the support list.
- Relax data type check, most for FP16 and INT8.
enableDebugLog()to dump log for debugging purpose.
Thanks for the contribution of @erizmr @briangrifiin and @IkbeomJeon!
- Now open source with annocement blog.
- Quantization support enabled, and tried quantized MobileNetV1 an MobileNetV2.
- Drop Transpose based layout handling to save effort of managing quantization.
- More operators added, and tested models:
- Operator support of MobileNetV2.
- Infrastructure improvements.
- Propagation based layout handling.
- Model converting Workflow.
- Basic operator support of MobileNetV1.
- Transpose based layout handling.