Releases: kendryte/nncase
Releases · kendryte/nncase
Release list
nncase v1.0.0 beta1
Major changes
- Upgrade to kmodel v5
- Python compiler interface
- Support preprocess methods:
normalizeletter box - More PTQ methods:
l2kld_m0kld_m1kld_m2cdf - More TFLite ONNX & Caffe Ops
- Plugable compiler backends
- Ready for K510
New examples
NNCase v0.2.0 Beta4
Major changes
- Add
--weights-quantize-threshold,--output-quantize-thresholdand--no-quantized-binary(See usage)
Bugs fixes
- Fix a bug in binary ops optimization
- Fix a bug in k210 pooling optimization
- Some bug fixes in tflite and onnx importer
NNCase v0.2.0 Beta3
Major changes
- Add ONNX importer (thanks to @aclex 's contribution)
- Add fused unary op for uint8 inference
- Add any unary activation fusion for KPU ops
- Add KPU matmul
- Support accelerating dilated conv2d by KPU
Bugs fixes
- Fix a bug in binary ops optimization
NNCase v0.2.0 Beta2
Major changes
- Add L2 based calibration methods
- Support Slice
- Support Mac OS
Bugs fixes
- Fix a bug when compiling multiple outputs model on Linux
Tips
- Beta2 version support tflite and caffe importer
NNCase v0.2.0 Beta1
Major changes
- Add quantized version of Conv2D, MatMul, Binary, ResizeImage ops
- Add Conv2DTranspose op
- Add best fit memory allocation solver (enabled by default, see usage)
- Update documentation
Bugs fixes
- Fix a bug in TFLite Softmax lowering
Tips
- Beta1 version support tflite and caffe importer
NNCase v0.2.0 Alpha2
Major changes
- Add tflite ops lowering: Relu LeakyRelu Logistic
Bugs fixes
- Fix a bug in strided slice codegen
- Fix reduce axis calculation
- Fix shape getter in tflite importer
Tips
- Alpha2 version only support tflite importer
NNCase v0.2.0 Alpha1
Major changes
- Rewrite in C++
- Generate kmodel v4
- Support more ops
Tips
- Alpha1 version only support tflite importer
NNCase Converter v0.1.0 RC5
Major changes
- Add K210 emulator
- Support float inference type
NNCase Converter v0.1.0 RC4
Major changes
- Support Sigmoid
- Add quantized version of ResizeNearestNeighbor
- Support Mean (axis=H,W)
Other changes
- Optimize error message
NNCase Converter v0.1.0 RC3
Major changes
- Support ResizeNearestNeighbor (yolov3)
- Support Flatten
- Support Channelwise quantization for output layers.