1.22.0

• Docker Image (arm64, Apple Silicon Mac)
  Distributed docker image is not compatible with ARM environments such as Apple Silicon Mac.
  Although Docker offers emulation for x86/amd64 environments, onnx2tf within this emulation mode results in an error, as shown below:

  $ root@39d07181ce27:/# onnx2tf -h
  > Illegal instruction

  This issue is suspected to be caused by the dependency of PyPI TensorFlow on x86-specific instruction sets.

  To address this,
  I've augmented the release process with GitHub Actions to include building and pushing Docker images for Arm64 architecture. This make it possible to execute onnx2tf with docker on Arm64 hosts.

  • Update Dockerfile
    • Extra dependencies (pkg-config and libhdf5-dev) are required to install h5py for arm64 docker build.
  • Add Arm64 build and push in GitHub Actions
    • Referenced the following URL for guidance: https://docs.docker.com/build/ci/github-actions/multi-platform/

What's Changed

• Add arm64 docker image by @ysohma in #636

Full Changelog: 1.21.6...1.22.0

1.21.6

• MatMulInteger
  Currently, MatMulInteger is implemented as tf matmul with int32 inputs/outputs, which leads to generation of Flex(Batch)MatMul ops.

  When -rtpo MatMulInteger is specified, inputs of MatMulInteger are casted to float32 instead, allowing the node to be converted to the builtin FullyConnected or BatchMatMul ops.

  ONNX input:
  

  Before:
  

  After:
  

What's Changed

• support suppressing flex ops for MatMulInteger by @DDoSolitary in #635

Full Changelog: 1.21.5...1.21.6

1.21.4

• Add -odrqt, --output-dynamic-range-quantized-tflite option.
  While output_integer_quantized_tflite already enables dynamic range quantization output, the option also triggers checks for calibration data, which is only required for full integer quantization, and causes errors when no calibration data is provided.

  This is undesirable if only dynamic quantization is wanted.

  A new option (-odrqt, --output-dynamic-range-quantized-tflite) is added to only enable dynamic range quant output, which doesn't need calibration data.

  Before:

  $ onnx2tf -i some_model_with_non_regular_input_shape.onnx -oiqt
  (other output omitted)
  Model conversion started ============================================================
  INFO: input_op_name: input shape: [1] dtype: float32
  ERROR: For INT8 quantization, the input data type must be Float32. Also, if --custom_input_op_name_np_data_path is not specified, all input OPs must assume 4D tensor image data. INPUT Name: input INPUT Shape: [1] INPUT dtype: float32
  

  After:

  $ onnx2tf -i some_model_with_non_regular_input_shape.onnx -odrqt
  (other output omitted)
  saved_model output started ==========================================================
  saved_model output complete!
  WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
  W0000 00:00:1715853625.734342    7691 tf_tfl_flatbuffer_helpers.cc:390] Ignored output_format.
  W0000 00:00:1715853625.734397    7691 tf_tfl_flatbuffer_helpers.cc:393] Ignored drop_control_dependency.
  Float32 tflite output complete!
  W0000 00:00:1715853629.274694    7691 tf_tfl_flatbuffer_helpers.cc:390] Ignored output_format.
  W0000 00:00:1715853629.274724    7691 tf_tfl_flatbuffer_helpers.cc:393] Ignored drop_control_dependency.
  Float16 tflite output complete!
  W0000 00:00:1715853631.535535    7691 tf_tfl_flatbuffer_helpers.cc:390] Ignored output_format.
  W0000 00:00:1715853631.535568    7691 tf_tfl_flatbuffer_helpers.cc:393] Ignored drop_control_dependency.
  Dynamic Range Quantization tflite output complete!
  

What's Changed

• add option to enable only dynamic range quant by @DDoSolitary in #633

Full Changelog: 1.21.3...1.21.4

1.21.3

• Significantly faster flatbuffer update speed with -coion option
  Currently, the copy_onnx_input_output_names_to_tflite flag converts the tflite model to json for modification, and then convert it back. For large models, the conversions take a long time and consume a large amount of disk space.

  Flatbuffers provides Python API allowing reading and writing model files as Python objects directly. We can run flatc --python --gen-object-api to generate Python object API from the downloaded schema file and use it to read the model, add signature defs, and write it back.

What's Changed

• use flatbuffers python api for coion by @DDoSolitary in #631
• flatbuffers #631 by @PINTO0309 in #632

New Contributors

• @DDoSolitary made their first contribution in #631

Full Changelog: 1.21.2...1.21.3

1.21.2

• GatherElements
  • Added automatic error correction.
  • convnext-det.onnx.zip
    
• Quantized ConvNext #629

What's Changed

• Added automatic error correction for GatherElements by @PINTO0309 in #630

Full Changelog: 1.21.1...1.21.2

1.21.1

• Constant

  • Bring Constant layers unconnected to the model into the model.

  • It is assumed that the -nuo option is specified because running onnxsim will remove constants from the ONNX file.

  • Wrap constants in a Lambda layer and force them into the model.

  • toy_with_constant.onnx.zip

  • Convert test

    onnx2tf -i toy_with_constant.onnx -nuo -cotof

    ONNX	TFLite

    

  • Inference test

    import tensorflow as tf
    import numpy as np
    from pprint import pprint
    
    interpreter = tf.lite.Interpreter(model_path="saved_model/toy_with_constant_float32.tflite")
    interpreter.allocate_tensors()
    
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()
    
    interpreter.set_tensor(
        tensor_index=input_details[0]['index'],
        value=np.ones(tuple(input_details[0]['shape']), dtype=np.float32)
    )
    interpreter.invoke()
    
    variable_output = interpreter.get_tensor(output_details[0]['index'])
    constant_output = interpreter.get_tensor(output_details[1]['index'])
    
    print("=================")
    print("Variable Output:")
    pprint(variable_output)
    print("=================")
    print("Constant Output:")
    pprint(constant_output)

    =================
    Variable Output:
    array([[-0.02787317, -0.05505124,  0.05421712,  0.03526559, -0.14131774,
             0.0019211 ,  0.08399964,  0.00433664, -0.00984338, -0.03370604]],
          dtype=float32)
    =================
    Constant Output:
    array([1., 2., 3., 4., 5.], dtype=float32)
    

• Constant outputs removed from ONNX during conversion #627

1.21.0

• Fix Typo
• Change API parameter names

What's Changed

• Fix Typo by @PINTO0309 in #625

Full Changelog: 1.20.10...1.21.0

1.20.10

• YOLOvN ONNX
• Fixed to restore metadata
• pip install -U sng4onnx>=1.0.4 sne4onnx>=1.0.13
  

What's Changed

• Fixed to restore metadata by @PINTO0309 in #624

Full Changelog: 1.20.9...1.20.10

1.20.9

• Tile, Reshape
  • Tensor processing of dynamic shapes is supported.
  • Due to the experimental implementation, regression testing is still insufficient. However, this ONNX is ready for conversion.
  • daquexian/onnx-simplifier#323
  • src_ocr.onnx.zip
  • https://github.com/PaddlePaddle/PaddleOCR
  • Input: [N, 3, 32, W]]
    
    
    

What's Changed

• [experimental] Support for dynamic Tile, dynamic Reshape by @PINTO0309 in #623

Full Changelog: 1.20.8...1.20.9

1.20.8

• If
  • Improved conversion stability of subgraphs of If operations.
  • onnx/onnx-tensorflow#1056
  • https://github.com/PINTO0309/onnx2tf/releases/download/1.20.7/maskrcnn_resnet50_fpn.onnx

onnx2tf \
-i maskrcnn_resnet50_fpn.onnx \
-onimc boxes.55 onnx::Shape_3316 3315 onnx::Loop_3751

What's Changed

• Improved conversion stability of subgraphs of If operations. by @PINTO0309 in #622

Full Changelog: 1.20.7...1.20.8