Conversion of models and restriction #

On ActDK, deep-learning models are in nnoir format. We provides the following tools to create nnoir:

• onnx2nnoir (onnx to nnoir)
• nnoir-chainer (chainer to nnoir)

You can install these tools like the following commands. onnx2nnoir is also included in the zip file that can be downloaded from Download ActDK .

$ pip3 install nnoir-onnx

$ pip3 install nnoir-chainer

See the following manuals for the details:

• nnoir-onnx: https://pypi.org/project/nnoir-onnx
• nnoir-chainer: https://pypi.org/project/nnoir-chainer

Restriction #

There are some restrictions as follows to convert deep-learning models.

• Restriction for arrays
  • The number of dimensions of an array must be lower than or equal to 16.
  • The length on each dimension must be lower than or equal to 2047.
• Conversion to nnoir
  • nnoir-onnx
    • Supported operators
      • See “Supported ONNX Operators” in https://pypi.org/project/nnoir-onnx/.
  • nnoir-chainer
    • Supported layers
      • See “These layers supported by the nnoir-chainer exporter.” in https://pypi.org/project/nnoir-chainer/.
  • If conversion is failed, use onnigiri and onnion included with ActDK.
• Conversion to C runtime libraries
  • All layers that can be converted by nnoir-onnx are supported.
    • Support to AveragePoll, MaxPool, and Convolution layers is limited to the 2D versions.
    • The use of LRN is not recommended.
  • Support of resize_image with nnoir-chainer.