PyPI · #4407 most downloaded on PyPI

onnx2tf

2.6.91.0M downloads/moPINTO0309/onnx2tf ↗New releases in your daily digest, changelog included.

A tool for converting ONNX files to LiteRT/TFLite/TensorFlow, PyTorch native code (nn.Module), TorchScript (.pt), state_dict (.pt), Exported Program (.pt2), and Dynamo ONNX. It also supports direct conversion from LiteRT to PyTorch.

Last release 10 days ago
14 Sep 2026
Release timing varies
gaps range from 8 days to 2 months
Nearly every release is documented
notes for 60 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
4 years old
629 releases · first in 2022

Release timeline

629 releases since 2022

202220232024

Releases

1.19.1023 Jan 2024
Modified to suppress broadcast when Gemm is immediately followed by a scalar Mul, Div, Sub, or Mod.
Mul, Div, Sub, or Mod

Modified to suppress broadcast when Gemm is immediately followed by a scalar Mul, Div, Sub, or Mod.

This is a very specific workaround for the special specifications of NNAPI.

Address the issue of NNAPI not allowing multi-dimensional bias.

The problem is not reproducible for variable batch sizes.

The problem reproduces only when changing to a fixed batch size.

pre_explicit_broadcast should not expend scalar tensor #573

What's Changed

Modified to suppress broadcast when Gemm is immediately followed by a scalar Mul, Div, Sub, or Mod. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/577

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.19.9...1.19.10
Open source →
1.19.922 Jan 2024
Improved GatherElements conversion stability.
Improved GatherElements conversion stability.

bodypix_mobilenet050_stride8_1x3x128x160.onnx.zip

What's Changed

Improved GatherElements conversion stability. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/576

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.19.8...1.19.9

One column per quarter.

1.19.819 Jan 2024

Fixed Abort bug when converting models with --optimization_for_gpu_delegate.

Gemm
- Fixed Abort bug when converting models with --optimization_for_gpu_delegate.
  
  onnx before tflite after tflite
pre_explicit_broadcast should not expend scalar tensor #573

What's Changed

Fixed Abort bug when converting models with --optimization_for_gpu_delegate. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/574

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.19.7...1.19.8

Open source →

1.19.711 Jan 2024

Improved the display of auxiliary error messages regarding flatc installation instructions.

Improved the display of auxiliary error messages regarding flatc installation instructions.
Full int8 quantized model doesn't work when "flatc" is installed on the system #567

What's Changed

improve flatc warning message by @spacycoder in https://github.com/PINTO0309/onnx2tf/pull/570

New Contributors

@spacycoder made their first contribution in https://github.com/PINTO0309/onnx2tf/pull/570

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.19.6...1.19.7

Open source →

1.19.69 Jan 2024

Support for STFT (Partially supported)

Support for STFT (Partially supported)
1. add onnx2tf/onnx2tf/ops/STFT.py. This is the main implementation of the STFT op.
2. provide example STFT onnx model

What's Changed

Add STFT op onnx2tf support by @guanerdan in https://github.com/PINTO0309/onnx2tf/pull/566

New Contributors

@guanerdan made their first contribution in https://github.com/PINTO0309/onnx2tf/pull/566

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.19.5...1.19.6

Open source →

1.19.56 Jan 2024

Implement workaround for cases where the input tensor contains multiple undefined dimensions.

Resize
- Implement workaround for cases where the input tensor contains multiple undefined dimensions.
- Addressed the issue of Resize aborting only when resizing by the scales parameter, not by the sizes parameter.
- fp32.onnx.zip
Failed to convert Real-ESRGAN x4-general (NCHW) #563

What's Changed

Implement workaround for cases where the input tensor contains multiple undefined dimensions Resize by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/564

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.19.4...1.19.5

Open source →

1.19.429 Dec 2023

Added ability to optimize patterns with consecutive ReLU and ReLU6

Added ability to optimize patterns with consecutive ReLU and ReLU6

ONNX TFLite (Unoptimized) TFLite (Optimized)

What's Changed

Added ability to optimize patterns with consecutive ReLU and ReLU6 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/562

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.19.3...1.19.4

Open source →

1.19.328 Dec 2023

Protocol Buffers serialization process was adapted to the new onnx==1.15.0 specification.

Support for onnx==1.15.0
Protocol Buffers serialization process was adapted to the new onnx==1.15.0 specification.

What's Changed

-nodafsc, -eatfp16 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/560
Support for onnx==1.15.0 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/561

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.19.2...1.19.3

Open source →

1.19.218 Dec 2023

Support for Gelu https://github.com/onnx/onnx/blob/main/docs/Changelog.md#gelu-20

Support for Gelu https://github.com/onnx/onnx/blob/main/docs/Changelog.md#gelu-20

What's Changed

Support for Gelu by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/559

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.19.1...1.19.2

Open source →

1.19.115 Dec 2023

Fixed a bug that caused -nodafsc processing to be unconditionally skipped.

Fixed a bug that caused -nodafsc processing to be unconditionally skipped.
Conversion of YoloV5, conversion of StridedSlice from 5d to 4d #557

What's Changed

Fixed a bug that caused -nodafsc processing to be unconditionally skipped. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/558

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.19.0...1.19.1

Open source →

1.19.010 Dec 2023

tf v2.15.0 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/554

TensorFlow v2.15.0
onnxruntime v1.16.3

What's Changed

tf v2.15.0 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/554

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.15...1.19.0

Open source →

1.18.158 Dec 2023

Half-precision Inference Doubles On-Device Inference Performance.

Half-precision Inference Doubles On-Device Inference Performance.

On certain ARM cores, specifying a hint that the model is a Float16 precision model doubles the conventional inference speed.

-eatfp16, --enable_accumulation_type_float16 ENABLE_ACCUMULATION_TYPE_FLOAT16
  Hint for XNNPack fp16 inference on float16 tflite model.

https://blog.tensorflow.org/2023/11/half-precision-inference-doubles-on-device-inference-performance.html

This is result in macbook.
- Chip: Apple M1 Pro (ArmV8 processor)
- Python 3.9
- Tensorflow 2.15.0 (from pip install tensorflow==2.15.0)
  
  <img width="811" alt="image" src="https://github.com/PINTO0309/onnx2tf/assets/74748700/51799aff-b006-46e1-a372-bd8b2195b854">
Regarding on x86, AVX2 is necessary and rebuild python package in PyPI seems be NOT enabled AVX2. According to the blog, AVX2 emulation in x86 is for precision check and its is slow.

What's Changed

Improve macOS compatibility via docker by @ysohma in https://github.com/PINTO0309/onnx2tf/pull/548
Add option for tflite float16 model by @ysohma in https://github.com/PINTO0309/onnx2tf/pull/553

New Contributors

@ysohma made their first contribution in https://github.com/PINTO0309/onnx2tf/pull/548

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.14...1.18.15

Open source →

1.18.1417 Oct 2023

Fixed a problem with conversion abort when K is set to a primitive integer.

Trilu
- Fixed a problem with conversion abort when K is set to a primitive integer.
- trilu_14.onnx.zip
- Trilu dtype error when using k greater than tensor dimension. #538

What's Changed

Bug fixes in Trilu by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/539

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.13...1.18.14

Open source →

1.18.139 Oct 2023

Bug fix for --number_of_dimensions_after_flextranspose_compression option.

Reshape, common_functions.py
- Bug fix for --number_of_dimensions_after_flextranspose_compression option.

What's Changed

Bug fix for --number_of_dimensions_after_flextranspose_compression option. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/537

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.12...1.18.13

Open source →

1.18.123 Oct 2023

Improved stability of accuracy check process for unsearchable OPs in subgraphs.

onnx2tf.py
- Improved stability of accuracy check process for unsearchable OPs in subgraphs.
- Workaround for processing main graphs extracted from subgraphs in the process of model structure optimization.
- ssdlite320_mobilenet_v3_large.onnx.zip

What's Changed

Improved stability of accuracy check process for unsearchable OPs in subgraphs by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/535

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.11...1.18.12

Open source →

1.18.113 Oct 2023

If all axes are of different sizes and the axis sizes specified in axis are the same in onnx and TensorFlow, skip the accuracy check.

Softmax, LayerNormalization
- If all axes are of different sizes and the axis sizes specified in axis are the same in onnx and TensorFlow, skip the accuracy check.
- Faster model transformations in some architectures, such as Transformer.
- OPs with multiple axes of the same size are corrected for accuracy as usual, so the conversion speed is not accelerated.

What's Changed

If all axes are of different sizes and the axis sizes specified in axis are the same in onnx and TensorFlow, skip the accuracy check. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/534

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.10...1.18.11

Open source →

1.18.102 Oct 2023

Fixed an error when the input tensor was 1-dimensional during Concat accuracy correction.

Concat
- Fixed an error when the input tensor was 1-dimensional during Concat accuracy correction.
Clip
- Support for special max and min values for Clip.

What's Changed

Fixed an error when the input tensor was 1-dimensional during Concat accuracy correction. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/532
Support for special max and min values for Clip by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/533

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.9...1.18.10

Open source →

1.18.92 Oct 2023

Much stricter accuracy correction conditions for Reshape.

Reshape
- Much stricter accuracy correction conditions for Reshape.
- The conversion speed is greatly reduced at the cost of not causing accuracy degradation.
- mobileformer.onnx.zip
Concat
- Fixed to skip accuracy check if one of the multiple input tensors contains np.ndarray (constant).

What's Changed

Fixed to skip accuracy check if one of the multiple input tensors contains np.ndarray (constant) by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/531
Much stricter accuracy correction conditions for Reshape by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/530

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.8...1.18.9

Open source →

1.18.82 Oct 2023

https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_529/latest.opset17.onnx

https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_529/latest.opset17.onnx
Reshape
- Fixed a critical bug in Reshape accuracy correction process.
- Removed special workaround processing to accommodate ShuffleNet, which is no longer needed.
LayerNormalization
- Added accuracy correction process.

What's Changed

Fixed a critical bug in Reshape, LayerNormalization accuracy correction process by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/529

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.7...1.18.8

Open source →

1.18.71 Oct 2023

Improved processing speed of MatMul accuracy correction process.

MatMul
- Improved processing speed of MatMul accuracy correction process.
Concat
- Improved the high likelihood of accuracy degradation in Concat input tensors where all axes except the batch size are the same size.

What's Changed

Improved processing speed of MatMul accuracy correction process by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/527
Improved the high likelihood of accuracy degradation in Concat input tensors where all axes except the batch size are the same size. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/528

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.6...1.18.7

Open source →

1.18.61 Oct 2023

Improved BatchMatMul transformation stability in the presence of undefined dimensions in the input tensor.

MatMul
- Improved BatchMatMul transformation stability in the presence of undefined dimensions in the input tensor.

What's Changed

Improved BatchMatMul transformation stability in the presence of undefined dimensions in the input tensor. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/526

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.5...1.18.6

Open source →

1.18.51 Oct 2023

Reduces the frequency of output shape mismatch.

common_functions.py
- Reduces the frequency of output shape mismatch.
- Improved conversion stability when the structure of the model is so complex that dummy inference is not possible.

What's Changed

Reduces the frequency of output shape mismatch by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/525

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.4...1.18.5

Open source →

1.18.430 Sept 2023

Pilot implementation of Col2Im.

Col2Im
- Pilot implementation of Col2Im.
- Large numbers of constants are generated, resulting in a bloated model size.
- Not all patterns are supported, and I have only verified the sample models I have on hand.
- mobilenetmss.onnx.zip
[Music Source Separation] Is it possible to use onnx2tf lite to convert audio based onnx models to tflite models #443

What's Changed

Pilot implementation of Col2Im by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/524

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.3...1.18.4

Open source →

1.18.330 Sept 2023

Improved stability of the overall model transformation when the input tensor contains two or more undefined dimensions.

Improved stability of the overall model transformation when the input tensor contains two or more undefined dimensions.
Resize
- Added optimization process for Resize with undefined dimensions.
- Valid only if the input tensor is fixed in NHWC.
- Axis transposition correction by JSON in this pattern is no longer necessary.
- https://github.com/PINTO0309/onnx2tf/files/12758261/retinaface_onnx_dynamic.onnx.zip
- convert
```
onnx2tf \
-i retinaface_onnx_dynamic.onnx \
-osd \
-cotof
```
  onnx Before tflite After tflite
Cannot use converted model with dynamic input shape #521

What's Changed

Added optimization process for Resize with undefined dimensions by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/523

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.2...1.18.3

Open source →

1.18.229 Sept 2023

Improved conversion stability when multiple undefined dimensions exist in the input tensor.

Improved conversion stability when multiple undefined dimensions exist in the input tensor.
Gather
- Fix parameter substitution logic.
- Cannot use converted model with dynamic input shape #521
- retinaface_onnx_dynamic.onnx.zip
- https://github.com/PINTO0309/onnx2tf/blob/main/json_samples/replace_retinaface_dynamic_shape.json
  
  onnx tflite
onnx2tf.py
- Fix parameter replacement logic.
Concat, common_functions.py
- Added the ability to swap axes using Gather instead of Transpose when the tensor to be concatenated is a 1D tensor.

What's Changed

Fix Gather parameter substitution logic by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/522

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.1...1.18.2

Open source →

1.18.127 Sept 2023

Improved stability of Float16 Pow conversion.

Pow
- Improved stability of Float16 Pow conversion.

What's Changed

Improved stability of Float16 Pow conversion by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/519

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.18.0...1.18.1

Open source →

1.18.027 Sept 2023

CLIP ViT Large https://github.com/onnx/models/blob/new-models/models/mlagility/openai_clip-vit-large-patch14/openai_clip-vit-large-patch14-16.onnx !im

tensorflow==2.14.0
onnx==1.14.1
onnxruntime==1.16.0
CLIP ViT Large https://github.com/onnx/models/blob/new-models/models/mlagility/openai_clip-vit-large-patch14/openai_clip-vit-large-patch14-16.onnx

What's Changed

tensorflow==2.14.0, onnx==1.14.1, onnxruntime==1.16.0 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/518

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.17.9...1.18.0

Open source →

1.17.926 Sept 2023

Improved conversion stability for tensors with None geometry.

Slice, Gather, Concat
- Improved conversion stability for tensors with None geometry.
- Improved transformation stability for models with undefined dimensions.
- fasterrcnn_resnet50_fpn.onnx
- https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_517/fasterrcnn_resnet50_fpn.onnx.tar.gz
ValueError: as_list() is not defined on an unknown TensorShape #1067

What's Changed

Improved conversion stability for tensors with None geometry by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/517

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.17.8...1.17.9

Open source →

1.17.824 Sept 2023

Fixed a conversion bug in the Dropout, GRU, LSTM, RNN OP for a very old opset.

Dropout, GRU, LSTM, RNN
- Fixed a conversion bug in the Dropout, GRU, LSTM, RNN OP for a very old opset.

What's Changed

Fixed a conversion bug in the Dropout, GRU, LSTM, RNN OP for a very old opset by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/516

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.17.7...1.17.8

Open source →

1.17.723 Sept 2023

converter._experimental_disable_batchmatmul_unfold -> converter.unfold_batchmatmul

onnx2tf.py
- converter._experimental_disable_batchmatmul_unfold -> converter.unfold_batchmatmul
- From TensorFlow v2.14.0rc1

What's Changed

converter._experimental_disable_batchmatmul_unfold -> converter.unfold_batchmatmul by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/514

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.17.6...1.17.7

Open source →

1.17.623 Sept 2023

Bernoulli, Resize, common_functions.py

Bernoulli, Resize, common_functions.py
- Fixed reference paths for Lambda functions.
- This is in preparation for using proto_splitter, which will be implemented in TensorFlow v2.15.0 or later.
[Stable Diffusion] Conversion error with stable diffusion model. #424
[TODO] Implementation of Proto Splitter (breakthrough of 2GB contract for Protocol Buffers) #455

What's Changed

Fixed reference paths for Lambda functions. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/513

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.17.5...1.17.6

Open source →

1.17.523 Sept 2023

Softmax, InstanceNormalization, Add, Sub, Div, Mul, Mod

Softmax, InstanceNormalization, Add, Sub, Div, Mul, Mod
- Significantly faster Softmax, InstanceNormalization, Add, Sub, Div, Mul and Mod conversions when converting models too large to perform dummy inference.
- However, because the size of the model is so large that dummy inference cannot be performed, no accuracy correction is performed and accuracy is not guaranteed.
- Stable Diffusion - stable-diffusion-v1-5
  
  https://github.com/PINTO0309/onnx2tf/assets/33194443/98c86e36-9ae1-4b59-b53c-2827502e0532
[Stable Diffusion] Conversion error with stable diffusion model. #424

What's Changed

Significantly faster Softmax, InstanceNormalization, Add, Sub, Div, Mul and Mod conversions when converting models too large to perform dummy inference by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/512

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.17.4...1.17.5

Open source →

1.17.422 Sept 2023

Significantly improved the efficiency of MatMul error correction process when dummy inference is not possible.

MatMul
- Significantly improved the efficiency of MatMul error correction process when dummy inference is not possible.
- Fixed so that the axis correction process is performed only when an axis malformed error occurs during the MatMul calculation.

What's Changed

Significantly improved the efficiency of MatMul error correction process when dummy inference is not possible. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/511

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.17.3...1.17.4

Open source →

1.17.322 Sept 2023

Added display of log messages by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/510

onnx2tf.py
- Added display of log messages.

What's Changed

Added display of log messages by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/510

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.17.2...1.17.3

Open source →

1.17.222 Sept 2023

Improved efficiency of dummy shape estimation process.

MatMul
- Improved efficiency of dummy shape estimation process.
  - np.dot -> np.matmul
  - np.float64 -> np.float32

What's Changed

np.dot -> np.matmul, np.float64 -> np.float32 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/509

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.17.1...1.17.2

Open source →

1.17.122 Sept 2023

Fixed to output all TF, h5, and keras formats when outputting .h5. !image

onnx2tf.py
- Fixed to output all TF, h5, and keras formats when outputting .h5.

What's Changed

Fixed to output all TF, h5, and keras formats when outputting .h5 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/508

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.17.0...1.17.1

Open source →

1.17.021 Sept 2023

Ubuntu 20.04 is no longer supported.

Ubuntu22.04
tensorflow v2.14.0rc1
- Ubuntu 20.04 is no longer supported.
ml_dtypes 0.2.0
flatc v23.5.26
- This adds support for larger buffers > 2 GiB by adding a 64-bit offsets for vectors and strings. This is only for C++ at the moment.
- https://github.com/google/flatbuffers/pull/7935

What's Changed

Ubuntu22.04, tensorflow v2.14.0rc1, ml_dtypes, flatc v23.5.26 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/506

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.31...1.17.0

Open source →

1.16.3121 Sept 2023

Added conversion progress indicator. !image

common_functions.py
- Added conversion progress indicator.

What's Changed

Added conversion progress indicator by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/505

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.30...1.16.31

Open source →

1.16.3021 Sept 2023

Improved stability of MatMul transformations when the output tensor is too large to allow dummy inference

MatMul
- Improved stability of MatMul transformations when the output tensor is too large to allow dummy inference
- For models that consume so much RAM that dummy inference cannot be performed to guarantee accuracy after the transformation, the model is changed to only check that the shape after the inner product of MatMul's trailing two axes matches ONNX's output shape.
- https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_248/bertsquad-12.onnx
- The figure below shows that the total size of the Numpy.ndarray after dummy inference is expected to exceed 93 GB.
- The specification is to skip dummy inference if it is expected to consume more than 80% of the PC's RAM space.
Even if the conversion is successful, the accuracy of the converted model is not guaranteed.

What's Changed

Improved stability of MatMul transformations when the output tensor is too large to allow dummy inference by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/504

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.29...1.16.30

Open source →

1.16.2920 Sept 2023

Improved transposition determination algorithm.

ScatterND
- Improved transposition determination algorithm.
- Improved overall model conversion stability.
Concat
- Added a remedy for cases where it is difficult to determine the correct axis.
https://github.com/PINTO0309/onnx2tf/releases/download/1.16.29/diffusion_low_light_1x3x192x320.onnx

What's Changed

Improved ScatterND transposition determination algorithm by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/503

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.28...1.16.29

Open source →

1.16.2817 Sept 2023

Improved optimization of non-optimizable OneHot with onnx-optimizer and onnxsim.

OneHot
- Improved optimization of non-optimizable OneHot with onnx-optimizer and onnxsim.
- Further improves the stability of the overall model conversion.

Before	After

What's Changed

Improved optimization of non-optimizable OneHot with onnx-optimizer and onnxsim by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/502

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.27...1.16.28

Open source →

1.16.2716 Sept 2023

Added NHWC information propagation processing to DepthToSparse and improved conversion stability.

DepthToSparse
- Added NHWC information propagation processing to DepthToSparse and improved conversion stability.
- Improved overall model conversion stability.
- https://github.com/apple/ml-cvnets
- mobilevitv2-0.5.onnx.zip

What's Changed

Added NHWC information propagation processing to DepthToSparse and improved conversion stability by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/501

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.26...1.16.27

Open source →

1.16.2616 Sept 2023

The logic for early termination of the accuracy verification step was dropped.

Softmax
- The logic for early termination of the accuracy verification step was dropped.
- Support for edge cases where Softmax is calculated on axes other than channels.
GridSample
- Workaround define_accurate_gridsample to a problem where the accuracy of define_fast_gridsample degrades significantly when more than 3 channels of tensors are input.
- Instead of maintaining accuracy, inference is sacrificed.
https://github.com/lyuwenyu/RT-DETR
https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_500/rtdetr_r18vd_6x_coco_192x320.onnx

What's Changed

Added correction process for Softmax and GridSample accuracy degradation by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/500

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.25...1.16.26

Open source →

1.16.2516 Sept 2023

Improved model conversion stability by adding an initialization process for useless output shapes.

Improved model conversion stability by adding an initialization process for useless output shapes.
This is a special workaround for ONNX generated from PaddlePaddle.
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.27/human_segmentation_pphumanseg_2021oct.onnx

What's Changed

Improved model conversion stability by adding an initialization process for useless output shapes. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/499

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.24...1.16.25

Open source →

1.16.2415 Sept 2023

Fails to export BatchNorm1d layers from Pytorch #5590

Selu
- Optimization of Selu.
- selu.onnx.zip
  
  Before After
Fails to export BatchNorm1d layers from Pytorch #5590

What's Changed

Optimization of Selu by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/498

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.23...1.16.24

Open source →

1.16.2314 Sept 2023

Significantly improved conversion success rate for models containing Einsum.

Einsum
- Significantly improved conversion success rate for models containing Einsum.
- Special techniques significantly optimize model structures for which model optimization with onnx.shape_inference and onnxsim fails.
- However, relief is provided only when there are no undefined dimensions in the input tensor.
- https://huggingface.co/SmilingWolf/wd-v1-4-moat-tagger-v2
- https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_497/wd-v1-4-moat-tagger-v2.onnx

What's Changed

Significantly improved conversion success rate for models containing Einsum by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/497

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.22...1.16.23

Open source →

1.16.2214 Sept 2023

Implemented workaround for shape correction when final output shapes do not match.

BatchNormalization
- Implemented workaround for shape correction when final output shapes do not match.
- This is a special treatment for unstable onnx files generated using tensorflow-onnx.

What's Changed

Implemented workaround for shape correction when final output shapes do not match by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/496

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.21...1.16.22

Open source →

1.16.2112 Sept 2023

Add handling of cases where the first or second tensor is 1-dimensional.

MatMul
- Add handling of cases where the first or second tensor is 1-dimensional.
- https://numpy.org/doc/stable/reference/generated/numpy.matmul.html

What's Changed

Add handling of cases where the first or second tensor is 1-dimensional by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/495

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.20...1.16.21

Open source →

1.16.2010 Sept 2023

Fix Slice degrades by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/494

Slice
- Fix Slice degrades

What's Changed

Fix Slice degrades by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/494

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.19...1.16.20

Open source →

1.16.1910 Sept 2023

Improved processing stability for Where and GatherND by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/493

Where, GatherND
- Improved processing stability.
superpoint.onnx.zip

What's Changed

Improved processing stability for Where and GatherND by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/493

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.18...1.16.19

Open source →

1.16.1810 Sept 2023

Mul, Div, Add, Sub, Mod, Less, LessOrEqual, Greater, GreaterOrEqual, Equal ,And, Or, Xor

Mul, Div, Add, Sub, Mod, Less, LessOrEqual, Greater, GreaterOrEqual, Equal ,And, Or, Xor
- Quite a fatal bug.
- Fixed a problem in which tensors were automatically transposed when the input tensor consisted of multiple values in one dimension.
- This is the cause of the precision degradation that has been occurring in the four arithmetic operations up to now.
- Improvement in overall model conversion speed due to elimination of accuracy degradation.
Softmax, Conv, GroupNorm, Reshape, ConvInteger, InstanceNormalization
- Bug fixes for R-CNN only.

What's Changed

Fixed a problem in which tensors were automatically transposed when the input tensor consisted of multiple values in one dimension by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/492

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.17...1.16.18

Open source →

1.16.179 Sept 2023

Improve accuracy of NHWC decision in Concat.

Concat
- Improve accuracy of NHWC decision in Concat.
Slice
- Support edge cases in Slice.
- Address a situation where the axes of a Slice transformed using tensorflow-onnx are in a disjointed state.
face_blendshapes.onnx.zip

What's Changed

Improve accuracy of NHWC decision in Concat and support edge cases in Slice by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/491

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.16...1.16.17

Open source →

1.16.168 Sept 2023

Major overhaul of ScatterND process.

ScatterND
- Major overhaul of ScatterND process.
- Disruptive changes in algorithms.
- alike_t_opset11_192x320.onnx.zip
Einsum
- Improved Einsum conversion stability.

What's Changed

Major overhaul of ScatterND process by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/489
Improved Einsum conversion stability by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/490

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.15...1.16.16

Open source →

1.16.158 Sept 2023

Improved stability of the process of calculating expected RAM consumption.

common_functions.py
- Improved stability of the process of calculating expected RAM consumption.

What's Changed

Improved stability of the process of calculating expected RAM consumption by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/488

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.14...1.16.15

Open source →

1.16.147 Sept 2023

Revised the method of determining the position of the axis of the Unsqueeze target.

Unsqueeze
- Revised the method of determining the position of the axis of the Unsqueeze target.
- alike_t_opset11_192x320.onnx.zip

What's Changed

[experimental] Revised the method of determining the position of the axis of the Unsqueeze target. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/487

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.13...1.16.14

Open source →

1.16.137 Sept 2023

Official tf.nn.softmax doesn't support input logits with dimension > 5. This limit would fail onnx2tf correction_process_for_accuracy_errors() check,

Softmax
- Official tf.nn.softmax doesn't support input logits with dimension > 5. This limit would fail onnx2tf correction_process_for_accuracy_errors() check, when tf subgraph contains softmax op with input dimension > 5.
- Before tf.nn.softmax, reshape the input logits to [reduce_prod(shape[:axis]), shape[axis], reduce_prod(shape[axis + 1:])]. after tf.nn.softmax, reshape back to original logits shape.
- Test code: https://github.com/PINTO0309/onnx2tf/blob/main/make_test_op/test_softmax_convert.py
- https://github.com/princeton-vl/RAFT-Stereo
- parameter replacement: raftstereo-middlebury.zip

What's Changed

tf.nn.softmax doesn't support Dim(logits) > 5 by @babiking in https://github.com/PINTO0309/onnx2tf/pull/486

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.12...1.16.13

Open source →

1.16.127 Sept 2023

Improved overall model transformation processing stability.

Improved overall model transformation processing stability.
ScatterND
- Fixed to determine if the trailing index needs to be transposed and force transposition and retry if necessary.
Where
- Improved NHWC information to be passed on to subsequent tensors.
yolox_nano_ti_lite_26p1_41p8.onnx.zip

What's Changed

Improved stability of ScatterND and Where conversions by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/485

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.11...1.16.12

Open source →

1.16.116 Sept 2023

Reduces post-concat accuracy degradation when all axes except batch size have the same value.

Concat
- Reduces post-concat accuracy degradation when all axes except batch size have the same value.
- Only supported when there are two input tensors, as the process is redundant.
- vttrack.onnx.zip

What's Changed

Correction of missing judgment conditions for parameter replacement by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/484
Reduces post-concat accuracy degradation when all axes except batch size have the same value by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/483

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.10...1.16.11

Open source →

1.16.106 Sept 2023

Reduce error frequency of Concat of two tensors containing constants.

Concat
- Reduce error frequency of Concat of two tensors containing constants.
- It is a simple correction and does not cover all patterns.
common_functions.py
- Disable tensor swapping during Sub, Div and Mod.

What's Changed

Reduce error frequency of Concat of two tensors containing constants and disable tensor swapping during Sub, Div and Mod. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/482

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.9...1.16.10

Open source →

1.16.95 Sept 2023

Eliminates unnecessary narrowing of the search for the correct answer axis.

MatMul
- Eliminates unnecessary narrowing of the search for the correct answer axis.
- Reduction of accuracy degradation due to transposition errors and overall time required for model conversion.
- parseq_tiny.onnx.zip

What's Changed

Eliminates unnecessary narrowing of the search for the correct answer axis by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/481

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.8...1.16.9

Open source →

Your coding agent can read these notes before it upgrades. Set up the MCP server →