ONNX Converter

ONNX_Convertor is an open-source project on Github. If there is any bugs in the ONNX_Convertor project inside the docker, don't hesitate to try git pull under the project folder to get the latest update. And if the problem persists, you can raise an issue there. We also welcome contributions to the project.

The general process for model conversion is as following:

  1. Convert the model from other platforms to onnx using the specific converters. See section 1 - 5.
  2. Optimize the onnx for Kneron toolchain using onnx2onnx.py. See section 6.
  3. Check the model and do further customization using editor.py (optional). See section 7


ONNX exported by Pytorch cannot skip step 1 and directly go into step 2. Please check section 2 for details.

If you're still confused reading the manual, please try our examples from https://github.com/kneron/ConvertorExamples

Table 1.2 shows supported operators conversion mapping table for every platform.

Table 1.2 Operators conversion table for every platform

type keras caffe tflite onnx(Opset 11)
add Add Eltwise ADD Add
average pooling AveragePooling2D Pooling AVERAGE_POOL_2D AveragePool
batchnormalization BatchNormalization BatchNorm BatchNormalization
concatenate Concatenate Concat CONCATENATION Concat
convolution Conv2D Convolution CONV_2D Conv
crop Cropping2D / Cropping1D Slice
deconvolution Conv2DTranspose Deconvolution TRANSPOSE_CONV ConvTranspose
dense Dense InnerProduct FULLY_CONNECTED Gemm
depthwise convolution DepthwiseConv2D DepthwiseConvolution DEPTHWISE_CONV_2D Conv (with group attribute)
flatten Flatten Flatten Flatten
global average pooling GlobalAveragePooling2D Pooling MEAN GlobalAveragePool
global max pooling GlobalMaxPooling2D Pooling GlobalMaxPool
leaky relu LeakyReLU LEAKY_RELU LeakyRelu
max pooling MaxPooling2D Pooling MAX_POOL_2D MaxPool
multiply Multiply Eltwise MUL Mul
padding ZeroPadding2D PAD Pad
prelu PReLU PReLU PRELU Prelu
relu ReLU ReLU RELU Relu
relu6 ReLU RELU6 Clip
separable conv2d SeparableConv2D Conv
sigmoid Sigmoid Sigmoid LOGISTIC Sigmoid
squeeze SQUEEZE Squeeze
tanh Tanh Tanh
roi pooling ROIPooling MaxRoiPool
  1. our tensorflow conversion tool is based on opensource "tf2onnx", please check "https://github.com/onnx/tensorflow-onnx/blob/r1.6/support_status.md" for the supported op information
  2. our pytorch conversion tool is based on onnx exporting api in torch.onnx, please check "https://pytorch.org/docs/stable/onnx.html#supported-operators" for the supported op information
  3. some operators could be replaced by onnx2onnx.py in order to fit the KL520/KL720 spec.

1 Keras to ONNX

For Keras, our converter support models from Keras 2.2.4. Note that tf.keras and Keras 2.3 is not supported. You may need to export the model as tflite model and see section 5 for TF Lite model conversion.

Suppose there is an hdf5 model exported By Keras, you need to convert it to onnx by the following command:

python /workspace/libs/ONNX_Convertor/keras-onnx/generate_onnx.py -o absolute_path_of_output_model_file -O --duplicate-shared-weights absolute_path_of_input_model_file 

For example, if the model is /docker_mount/onet.hdf5, and you want to convert it to /docker_mount/onet.onnx, the detailed command is:

python /workspace/libs/ONNX_Convertor/keras-onnx/generate_onnx.py -o /docker_mount/onet.onnx  -O --duplicate-shared-weights /docker_mount/onet.hdf5

There might be some warning log printed out by the TensorFlow backend, but we can ignore it since we do not actually run it. You can check whether the conversion succeed by checking whether the onnx file is generated.

You need to run the command in section 6 after this step.

Input shape change(optional)

If there’s customized input shape for the model file, you need to use -I flag in the command. Here is an example:

python /workspace/libs/ONNX_Convertor/keras-onnx/generate_onnx.py abs_path_of_input_model -o abs_path_of_output_model -I 1 model_input_width model_input_height num_of_channel -O --duplicate-shared-weights

Add RGBN to YYNN layer for Keras model(optional)

Some of the models might take gray scale images instead of RGB images as the input. Here is a small script to add an extra layer to let the model take RGB input.

cd /workspace/libs/ONNX_Convertor/keras-onnx && python rgba2yynn.py input_hdf5_file output_hdf5_file

2 Pytorch to ONNX

The pytorch2onnx.py script not only takes pth file as the input. It also takes Pytorch exported onnx as the input. In fact, we recommend using the Pytorch exported onnx file instead of the pth file, since the Pytorch do not has a very good model save and load API. You can check TIPS below on how to export models to onnx.

We currently only support models exported by Pytorch version >=1.0.0, <=1.7.1, no matter it is a pth file or an onnx file exported by torch.onnx. The PyTorch version in the toolchain docker is 1.7.1.


You can use torch.onnx to export your model into onnx format. Here is the Pytorch 1.7.1 version document for onnx.torch. An example code for exporting the model is:

import torch.onnx
dummy_input = torch.randn(1, 3, 224, 224)
torch.onnx.export(model, dummy_input, 'output.onnx', opset_version=11)

In the example, (1, 3, 224, 224) are batch size, input channel, input height and input width. model is the model object you want to export. output.onnx is the output file.

Run pytorch2onnx with pth file

Suppose the input file is called /docker_mount/resnet34.pth and you want to save the onnx as /docker_mount/resnet34.onnx. The input channel, height, width for the model are (3, 224, 224). Here is the example command:

python /workspace/libs/ONNX_Convertor/optimizer_scripts/pytorch2onnx.py /docker_mount/resnet34.pth /docker_mount/resnet34.onnx --input-size 3 224 224 

You need to run the command in section 6 after this step.

Run pytorch_exported_onnx_preprocess with onnx file

Suppose the input file is called /docker_mount/resnet34.onnx and you want to save the optimized onnx as /docker_mount/resnet34.opt.onnx. Here is the example command:

python /workspace/libs/ONNX_Convertor/optimizer_scripts/pytorch_exported_onnx_preprocess.py /docker_mount/resnet34.onnx /docker_mount/resnet34.opt.onnx

You need to run the command in section 6 after this step.

Crash due to name conflict

If you meet the errors related to node not found or invalid input, this might be caused by a bug in the onnx library. Please try using --no-bn-fusion flag.

3 Caffe to ONNX

For caffe, we only support model which can be loaded by Intel Caffe 1.0.

Suppose you have model structure definition file /docker_mount/mobilenetv2.prototxt and model weight file /docker_mount/mobilenetv2.caffemodel and you want to output the result as /docker_mount/mobilenetv2.onnx, Here is the example command:

python /workspace/libs/ONNX_Convertor/caffe-onnx/generate_onnx.py -o /docker_mount/mobilenetv2.onnx -w /docker_mount/mobilenetv2.caffemodel -n /docker_mount/mobilenetv2.prototxt

You need to run the command in section 6 after this step.

4 Tensorflow to ONNX

Tensorflow to onnx script only support Tensorflow 1.x and the operator support is very limited. If it cannot work on your model, please try to export the model as tflite and convert it using section 5.

Suppose you want to convert the model /docker_mount/mnist.pb to /docker_mount/mnist.onnx, here is the example command:

python /workspace/libs/ONNX_Convertor/optimizer_scripts/tensorflow2onnx.py /docker_mount/mnist.pb /docker_mount/mnist.onnx

You need to run the command in section 6 after this step.

5 TF Lite to ONNX

python /workspace/libs/ONNX_Convertor/tflite-onnx/onnx_tflite/tflite2onnx.py -tflite path_of_input_tflite_model -save_path path_of_output_onnx_file -release_mode True

For the provided example model: model_unquant.tflite

python /workspace/libs/ONNX_Convertor/tflite-onnx/onnx_tflite/tflite2onnx.py -tflite /data1/tflite/model/model_unquant.tflite -save_path /data1/tflite/model/model_unquant.tflite.onnx -release_mode True

Then need to run the command in section 6.

6 ONNX to ONNX (ONNX optimization)

After converting models from other frameworks to onnx format, you need to run the following command to optimize the model for Kneron hardware, suppose your model is input.onnx and the output model is called output.onnx:

python /workspace/libs/ONNX_Convertor/optimizer_scripts/onnx2onnx.py input.onnx -o output.onnx --add-bn -t

Crash due to name conflict

If you meet the errors related to node not found or invalid input, this might be caused by a bug in the onnx library. Please try using --no-bn-fusion flag.

7 Model Editor

KL720 NPU supports most of the compute extensive OPs, such as Conv, BatchNormalization, Fully Connect/GEMM, in order to speed up the model inference run time. On the other hand, there are some OPs that KL720 NPU cannot support well, such as Softmax or Sigmod. However, these OPs usually are not compute extensive and they are better to execute in CPU. Therefore, Kneron provides a model editor which is editor.py to help user modify the model so that KL720 NPU can run the model more efficiently.

General Editor Guideline

Here are some general guideline to edit a model to take full advantage of KL720 NPU MAC efficiency:

Step 1: Start from each output node of the model, we should trace back to an operator that has significant compute workload, such as GlobalAveragePool, Gemm(fully connect), or Conv. Then, we could cut to the output of that OP. For example, there is a model looks Figure 4, and it has two output nodes: 350 and 349. From output node 349, we can trace back to the Gemm above the red line because the Div, Clip, Add and Mul only have 1x5 dimension, and these OPs are not very heavy computation. Since Mul and Div are not support in NPU, so it is recommend to cut the rest of the OPs and let the model finish at the output of the Gemm (red line). For the other output node 350, since it is the output of a Gemm, there is no need to do any more edition.

Step 2: If both input nodes and output nodes are channel last, and there is a Transpose after the input and before the output, then the model is transposed into channel first. We can use the model editor to safely remove the Transpose.

Step 3: If the input shape is not availble or invalid, we can usethe editor to give it a valid shape.

Step 4: The model need to pass onnx2onnx.py again after running the editor. See section 6.

Figure 4. Pre-edited model

Feature The script called editor.py is under the folder /workspace/libs/ONNX_Convertor/optimizer_scripts. It is a simple ONNX editor which achieves the following functions:

  1. Add nop BN or Conv nodes.
  2. Delete specific nodes or inputs.
  3. Cut the graph from certain node (Delete all the nodes following the node).
  4. Reshape inputs and outputs
  5. Rename the output.


editor.py [-h] [-c CUT_NODE [CUT_NODE ...]]
             [--cut-type CUT_TYPE [CUT_TYPE ...]]
             [-d DELETE_NODE [DELETE_NODE ...]]
             [--delete-input DELETE_INPUT [DELETE_INPUT ...]]
             [-i INPUT_CHANGE [INPUT_CHANGE ...]]
             [-o OUTPUT_CHANGE [OUTPUT_CHANGE ...]]
             [--add-conv ADD_CONV [ADD_CONV ...]]
             [--add-bn ADD_BN [ADD_BN ...]]
             in_file out_file

Edit an ONNX model. The processing sequense is 'delete nodes/values' -> 'add nodes' -> 'change shapes' -> 'cut node'.
Cutting is not recommended to be done with other operations together.

positional arguments:

in_file   input ONNX FILE
out_file  ouput ONNX FILE

optional arguments:

        -h, --help            show this help message and exit
        -c CUT_NODE [CUT_NODE ...], --cut CUT_NODE [CUT_NODE ...]
        remove nodes from the given nodes(inclusive)

        --cut-type CUT_TYPE [CUT_TYPE ...]
        remove nodes by type from the given nodes(inclusive)

        -d DELETE_NODE [DELETE_NODE ...], --delete DELETE_NODE [DELETE_NODE ...]
        delete nodes by names and only those nodes

        --delete-input DELETE_INPUT [DELETE_INPUT ...]
        delete inputs by names

        change input shape (e.g. -i 'input_0 1 3 224 224')

        change output shape (e.g. -o 'input_0 1 3 224 224')

        --add-conv ADD_CONV [ADD_CONV ...]
        add nop conv using specific input

        --add-bn ADD_BN [ADD_BN ...]
        add nop bn using specific input

8 ONNX Updater

For existed onnx models with the previous version of the opset, we provide the following scripts to update your model. Note that after running any of the following scripts, you may still need to run onnx2onnx.py.

# For ONNX 1.3 to 1.4 (opset 7/8 to opset 9):
python /workspace/libs/ONNX_Convertor/optimizer_scripts/onnx1_3to1_4.py input.onnx output.onnx

# For ONNX 1.4 to 1.6 (opset 9 to opset 11):
python /workspace/libs/ONNX_Convertor/optimizer_scripts/onnx1_4to1_6.py input.onnx output.onnx