nnoir-onnx 1.4.0

ONNX to NNOIR Converter

nnoir-onnx

nnoir-onnx is a converter from ONNX model to NNOIR model.

Install

From PyPI:

pip install nnoir-onnx

From Dockerhub:

docker pull idein/nnoir-tools:20240208

Example

wget https://www.cntk.ai/OnnxModels/mnist/opset_7/mnist.tar.gz
tar xvzf mnist.tar.gz
onnx2nnoir -o model.nnoir mnist/model.onnx

With docker:

docker run --rm -it -u $UID:$GID -v $(pwd):/work idein/nnoir-tools:20240208 onnx2nnoir --graph_name "mobilenet" -o mobilenetv2-1.0.nnoir mobilenetv2-1.0.onnx

Supported ONNX Operators

• Add
• AveragePool
• BatchNormalization
  • scale, B, mean, and var must be "constant"
• Clip
  • must be opset version 6 or 11
  • if opset version is 11
    • max must be "constant"
  • min must be 0
• Concat
• Conv
  • W must be Constant value or have initializer value
  • b must be Constant value or have initializer value
• Cos
• Div
  • 1st input must not be "constant"
• Dropout
  • equivalent identity function
• Elu
• Exp
• Flatten
• Gemm
  • B must be Constant value or have initializer value
  • C must be Constant value or have initializer value
• GlobalAveragePool
• HardSigmoid
• HardSwish
• LeakyRelu
• LRN
• LSTM
  • only seq_length == 1
  • direction must be forward
  • Supported activations are below
    • Sigmoid
    • Tanh
    • Relu
  • Not support clip and input_forget
• MatMul
• MaxPool
  • ceil_mode = 1 is not supported
  • dilations must be array of 1.
• Mul
• Pad
  • mode must be "constant"
• Pow
  • 2nd input must be 2.0
• PRelu
  • slope must be "constant" and a single value tensor
• ReduceMean
• ReduceSum
• Relu
• Reshape
• Resize
  • must be from opset version >= 11
  • mode must be "linear" or "nearest"
  • nearest_mode must be "floor"
  • coordinate_transformation_mode must be either "pytorch_half_pixel" or "align_corners" for "linear" mode
  • coordinate_transformation_mode must be either "asymmetric" for "nearest" mode
• Sigmoid
• Sin
• Slice
  • must be from opset version >= 10
  • starts must be Constant value or have initializer value
  • ends must be Constant value or have initializer value
  • axes must be Constant value or have initializer value
  • steps is not supported
• Softmax
• Split
• Squeeze
• Sub
• Sum
  • 2 inputs
• Tan
• Tanh
• Transpose
• Unsqueeze