Open project to directly convert models from the .onnx to .tflite format. Currently a reduced subset of ONNX operators is supported. Verified models:
- Alexnet (https://github.com/onnx/models/blob/main/vision/classification/alexnet/model/bvlcalexnet-12.onnx)
- TinyYOLO (https://github.com/onnx/models/blob/main/vision/object_detection_segmentation/tiny-yolov2/model/tinyyolov2-8.onnx)
- DUC (https://github.com/onnx/models/blob/main/vision/object_detection_segmentation/duc/model/ResNet101-DUC-12.onnx)
- Speech Classifier (https://github.com/Pop-korn/ONNX2TFLite/blob/main/data/onnx/speech_command_classifier_trained.onnx)
The models can be downloaded using make get-default-models and then converted to TFLite using make convert-default-models from the root directory.
Project is available on GitHub (https://github.com/Pop-korn/ONNX2TFLite).
The program was developped using python 3.10.8. Some earlier versions will not work because of the use of type hints.
The program was developped and verified using Ubuntu 22.04. It should work on most Linux distributions.
The pip tool is required for instalation of Python packages.
To install just the Python packages required for model conversion, use make install-essential in the root directory. To install all used python packages (including the ones needed for testing), run make install. There is a requirement confilct between tensorflow and onnx over protobuf version. This program doesn't require that part of tensorflow functionality, so if there is a problem during instalation, run pip install onnx==1.13.0. Conversion testing should now work.
Project comes with ONNX and TFLite schemas pre-installed and libraries pre-compiled. If you want, you can optionally delete the data/schemas and lib directories and run make regenerate-lib in the root directory to get the schemas and complie them yourself. THIS MAT CAUSE make test-model-conversion TO NOT WORK, because of limitations of the onnx library.
python onnx2tflite.py <input_model.onnx> [ -o/--output <output_model.tflite> --verbose]
The program is used via a command line interface of the onnx2tflite.py module. It takes 1 required argument, which is the input .onnx model to be converted. The output .tflite file can be optionally specified using the -o/--output option. By default, it will be generated in the current working directory and have the same name as the input file.
The program will only print error and warning messages, if the input model cannot be converted exactly. If you wish to see more detailed logging of the internal conversion process, use the --verbose option.
Example use:
- python onnx2tflite.py data/onnx/tinyyolov2-8.onnx --output test/tinyyolov2-8.tflite
- python onnx2tflite.py data/onnx/bvlcalexnet-12.onnx -o test/bvlcalexnet-12.tflite --verbose
The module conversion_test.py implements functions for testing the accuracy of converted models. Tests for the default models are at the end of the file. Uncomment the lines which test the model you are interested in and run make test-model-conversion from the root directory.
Two testing functions are implemented. runAndTestOperators() allows testing of conversion of a selected portion of an ONNX model. It prints a short summary of the output of the original model, the output of the converted model and their difference. A more thorough testing function is testConversion(). It converts an entire ONNX model to TFLite and runs them on equal random data for multiple iterations. It prints a set of statistics describing the output errors.
The module generator_test.py implements a simple .tflite model in code. It generates and runs the model and prints its output. Also prints the output of the original model. Use make test-tflite-file-generation to run this test.
The tensorflow library sometimes prints error messages when loaded. This happens when GPU acceleration is not supported. These errors can be ignored.
The entry point of the program is the onnx2tflite.py file in the root directory. The conversion_test.py modulde is used for testing of the accuracy of converted models. The module generator_test.py implements a simple .tflite model in code and is used to test the /src/generator functionality.
The /data directory contains pre-trained models, tensor values, input images for inference testing and schemas. Some models are not included on GitHub because of size restrictions.
The code in the /src diroctory is split into multiple subdirectories.
-
/src/generator contains classes used for internal representation of a TFLite model and for subsequent .tflie file generation.
-
/src/parser conatins classes used for internal representation of an ONNX model and for loading it's data from a .onnx file.
-
/src/convertor contains files for converting from the parser object model to the generator model.
The /lib directory contains files generated by compiling TFLite an ONNX schemas. They provide a low level interface for parsing a .onnx file and for generating a .tflite file.
-
/lib/tflite is a library generated from the /data/schemas/tflite/schema.bfs file, using the flatc -p ../data/schema.fbs command.
-
/lib/onnx is a library generated from the files in the /data/schemas/onnx/ directory.
The /test directory is used for analyzing the generated .tflite models.
This software is covered by the MIT license.