Converter for TFlite to Onnx

Description

Convert an TFlite model to Onnx. We parse TFlite model use:

1. flatc(https://google.github.io/flatbuffers/flatbuffers_guide_using_schema_compiler.html)
2. tensorflow lite schema file(https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/schema/schema.fbs)

Prerequisites

We test our script using:

• Python 3: Python==3.7.7
• ONNX: onnx==1.6
• Tensorflow: tensorflow==1.15.0
• FlatBuffer: flatbuffers==1.12
• Python-igraph: python-igraph==0.8.3

Basic Usage

1. run python path/to/tflite2onnx.py -h to check the parameter.

• 1.1. -tflite: the path to the tflite file
• 1.2. -save_path: final onnx file path to save

2. run python tflite2onnx.py -tflite YOUR_TFLITE_PATH -save_path ONNX_SAVE_PATH to convert the model,

Tested Model Table

tensorflow object detection model zoo:

https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf1_detection_zoo.md

( use export_tflite_ssd_graph.py with --add_postprocessing_op=false )

• Mobile_models/ssd_mobiledet_cpu_coco
• Mobile_models/ssd_mobilenet_v3_large_coco
• Mobile_models/ssd_mobilenet_v3_small_coco
• Mobile_models/ssd_mobilenet_v2_mnasfpn_coco

tflite model hub (Floating point models):

https://www.tensorflow.org/lite/guide/hosted_models

• Mobilenet_V1 series
• Mobilenet_V2 series
• Inception_V3
• Inception_V4
• Inception_ResNet_V2
• SqueezeNet
• DenseNet
• ResNet_V2_101
• EfficientNet-Lite

Now Supported Operators List

• ADD
• AVERAGE_POOL_2D
• CONCATENATION
• CONV_2D
• DEPTH_TO_SPACE
• DEPTHWISE_CONV_2D
• ELU
• FULLY_CONNECTED
• L2_NORMALIZATION
• LEAKY_RELU
• LOGISTIC
• MUL
• MEAN
• MAXIMUM
• MAX_POOL_2D
• PAD
• PRELU
• RELU
• RELU6
• RESHAPE
• RESIZE_BILINEAR
• RESIZE_NEAREST_NEIGHBOR
• SOFTMAX
• SPACE_TO_DEPTH
• SQUEEZE
• TRANSPOSE_CONV

Example 1: Convert model to the onnx optimized for Kneron Toolchain

Step0 (Optional). Download example tflite model

(Our provided example model is LFS-tracked file, sometimes 'git lfs pull' is needed)

git lfs pull

Step1. Convert tflite to onnx

(Originally, we add transpose nodes for the channel order difference between onnx and tflite. Use '-release_mode True' could ignore those transpose nodes)

python ./onnx_tflite/tflite2onnx.py -tflite ./example/example.tflite -save_path ./example/example.onnx -release_mode True

Step2. Convert onnx to the onnx which is optimized for Kneron Toolchain

python ../optimizer_scripts/onnx2onnx.py ./example/example.onnx

Example 2: Convert part of tflite model to onnx

python ./onnx_tflite/tflite2onnx.py -tflite ./example/example.tflite -save_path ./example/example.onnx -release_mode True -bottom_nodes sequential_1/model1/block_14_add/add

*Note 1:

• "./example/example.tflite" and "./onnx_tflite/flatc/flatc" are large file and not necessary. So we let them being tracked with GIT-LFS. You can use 'git lfs pull' to download them. Since the limit of lfs transfer size. They will not be available in the master branch. You can checkout the branch 'lfs' to get them. git checkout lfs

*Note 2 (convert tensorflow to tflite):

• Here is the official guide to convert tensorflow model and keras model to tflite: https://www.tensorflow.org/lite/guide/get_started#tensorflow_lite_converter

  Note that current tflite2onnx only support "float32" tflite,and current official tflite not support all tensorflow op
  ( ref: https://www.tensorflow.org/lite/guide/get_started#ops_compatibility )
  you might need to cut some unsupported nodes.