nncase is a cross-platform neural network optimization toolkit for fast inference.

Download prebuilt binaries from Release.

ncc -i <input format> -o <output format> [--dataset <dataset path>] [--postprocess <dataset postprocess>] [--weights-bits <weights quantization bits>] <input path> <output path>

• -i Input format

• -o Output format

• --inference-type Inference type

• --dataset Dataset path, required when the output format is inference or k210model with inference type equals uint8.

• --postprocess Dataset postprocess method

• --weights-bits Weights quantization bits

• --float-fc Use float fullyconnected kernels.

• --channelwise-output Use channelwise quantization for output layers.

• Convert TFLite model to K210 model.

  ncc -i tflite -o k210model --dataset ./images ./mbnetv1.tflite ./mbnetv1.kmodel

• Convert PaddlePaddle model to TensorFlow model.

  ncc -i paddle -o tf ./MobileNetV1_pretrained ./mbnetv1.pb

• Inference K210 model and get output binaries.

  ncc -i k210model -o inference --dataset ./images ./mbnetv1.kmodel ./output

• Tutorials

  • 20 classes object detection
  • Iris flowers classification

  See https://github.com/kendryte/nncase/tree/master/examples

* When using TensorFlow Conv2d/DepthwiseConv2d kernel=3x3 stride=2 padding=same, you must first use tf.pad([[0,0],[1,1],[1,1],[0,0]]) to pad the input and then use Conv2d/DepthwiseConv2d with valid padding.