onnx2tf

• LSTM, GRU, RNN, MatMul
  • Fixed a malfunction in the processing of RNNs in general, affected by a significant improvement in the tool's overall axis correction functionality.
  • Bug fixes
    • crnn_lite_lstm.onnx.zip
    • LSTM.tanh.bidirectional_b10.onnx.zip
    • torch_lstm.onnx.zip
    • RNN.bidirectional.one_layer.relu.onnx.zip
    • RNN.bidirectional.one_layer.relu_b10.onnx.zip
    • GRU.onnx.zip
    • gru_14_b1.onnx.zip
    • gru_14_b2.onnx.zip
    • lstm_undefined_dim.onnx.zip
• TypeError: Failed to convert object of type <class 'list'> to Tensor. #688

What's Changed

• Bug fixes in LSTM, GRU, RNN, and MatMul by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/476

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.6...1.16.7

• Concat
  • Attempts to force axis correction when the number of axes in the combined tensor do not exactly match.
    • However, if more than 2 patterns of correct answers exist, give up the correction.
    • This workaround is useful when automatic axis correction is practically difficult, such as when all tensors to be combined originate from Transpose or Reshape.
    • Under very limited conditions, it is no longer necessary to correct the axis in JSON.
• [D3Net] Clarification regarding a specific input dimension type #473

What's Changed

• Attempts to force axis correction when the number of axes in the combined tensor do not exactly match by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/475

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.5...1.16.6

• Pad
  • Negative padding supported.
  • Force substitution to StridedSlice since the TensorFlow runtime does not support padding of negative numbers.
  • d3net_dnn_double_44.onnx.zip
    • onnx
    • tflite
    • results
• [D3Net] Clarification regarding a specific input dimension type #473

What's Changed

• Negative padding supported by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/474

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.4...1.16.5

• onnx2tf.py, common_functions.py
  • Significantly improved conversion stability by reducing the frequency of SWAPs due to lack of RAM for large models.
  • Change the maximum number of recursions in Python to the maximum number of recursions. For Python to abort when the model structure is too large.

    sys.setrecursionlimit(2147483647) # C int maximum
    

• InstanceNormalization
  • Workaround for inconsistent C position.
  • Improved processing stability of InstanceNormalization.
• Stable Diffusion v1.5
  • https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_430/stable_diffusion_v1.5.tar.gz
• [Stable Diffusion] Conversion error with stable diffusion model. #424
• [TODO] Switch to a mode that performs inference in slow delay shape estimation mode during very large size transformations such as diffusion models #430

What's Changed

• Significantly improved conversion stability by reducing the frequency of SWAPs due to lack of RAM for large models, and improved processing stability of InstanceNormalization. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/472

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.3...1.16.4

• Mul, Div
  • Prevents Mul -> Div patterns from creating redundant Mul -> Mul operation sets.
  • This optimization improves the overall computational efficiency of the model just a little bit.

    onnx	tflite

• [Demucs] Issue while converting demucs model from onnx to tflite in conv2d_transpose #470

What's Changed

• [experimental] Prevents Mul -> Div patterns from creating redundant Mul -> Mul operation sets. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/471

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.2...1.16.3

• PReLU
  • slope conversion bug fixed again
• onnx==1.13.1 -> onnx==1.14.0
• onnxruntime==1.15.0 -> onnxruntime==1.15.1

What's Changed

• PReLU slope conversion bug fixed again by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/468

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.1...1.16.2

• PReLU
  • Fixed slope conversion bug.
  • face_landmarks_detector_facemeshv2_1x3x256x256.onnx.zip

What's Changed

• Fixed PReLU slope conversion bug by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/467

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.16.0...1.16.1

• GeLU

  • Add logic to replace a set of primitive OPs with GeLUs to improve overall model processing efficiency. https://www.tensorflow.org/api_docs/python/tf/keras/activations/gelu https://www.tensorflow.org/api_docs/python/tf/nn/gelu
  • Automatically detect the range where the following OP combination is valid and replace it with the GeLU OP.
  • Add --rtpo, --replace_to_pseudo_operators options with GeLU.
  • If -rtpo, --replace_to_pseudo_operators option is set to GeLU, the internal processing of GeLU is replaced by an approximate calculation, speeding up the operation in exchange for a small loss of accuracy.

• GeLU by itself can infer 28 times faster than before. gelu_11_float32_primitive.tflite.zip gelu_11_float32_approximate_false.tflite.zip gelu_11_float32_approximate_true.tflite.zip

  import tensorflow as tf
  import numpy as np
  np.random.seed(0)
  import time
  
  data = np.random.randn(1,512,512,3).astype(np.float32)
  loop = 10
  
  interpreter = tf.lite.Interpreter(
      model_path="gelu_11_float32_primitive.tflite",
      num_threads=20
  )
  tf_lite_model = interpreter.get_signature_runner()
  inputs = {
      'input': data,
  }
  # warmup
  _ = tf_lite_model(**inputs)
  # test
  total = 0.0
  for i in range(loop):
      start_time = time.perf_counter()
      _ = tf_lite_model(**inputs)
      elapsed_time = time.perf_counter() - start_time
      total += elapsed_time
  print(f"[TFLite] Primitive inf time: {total / loop}")
  
  interpreter = tf.lite.Interpreter(
      model_path="gelu_11_float32_approximate_false.tflite",
      num_threads=20
  )
  tf_lite_model = interpreter.get_signature_runner()
  inputs = {
      'input': data,
  }
  # warmup
  _ = tf_lite_model(**inputs)
  # test
  total = 0.0
  for i in range(loop):
      start_time = time.perf_counter()
      _ = tf_lite_model(**inputs)
      elapsed_time = time.perf_counter() - start_time
      total += elapsed_time
  print(f"[TFLite] Disable approximate inf time: {total / loop}")
  
  interpreter = tf.lite.Interpreter(
      model_path="gelu_11_float32_approximate_true.tflite",
      num_threads=20
  )
  tf_lite_model = interpreter.get_signature_runner()
  inputs = {
      'input': data,
  }
  # warmup
  _ = tf_lite_model(**inputs)
  # test
  total = 0.0
  for i in range(loop):
      start_time = time.perf_counter()
      _ = tf_lite_model(**inputs)
      elapsed_time = time.perf_counter() - start_time
      total += elapsed_time
  print(f"[TFLite] Enable approximate inf time: {total / loop}")
  

• Results gelu.onnx.zip

  onnx	tflite

• [TODO] Implemented forced replacement of GeLU processing with standard OP. #465

What's Changed

• Add logic to replace a set of primitive OPs with GeLUs to improve overall model processing efficiency by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/466

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.18...1.16.0

• common_functions.py
  • Bug fix for shape_unmatched_special_avoidance_workaround.
  • Fixed a problem in which expected transposition behavior was not performed under very limited conditions.
  • Fixed a problem in which the overall model structure was so broken due to transposition errors that when an accuracy check was performed, an inefficient and large np.ndarray was generated to the extent that RAM swapping occurred.
  • https://github.com/jinyeying/nighttime_dehaze

What's Changed

• Bug fix for shape_unmatched_special_avoidance_workaround by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/464

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.17...1.15.18

• common_functions.py
  • Experimental implementation.
  • To avoid generating unnecessary Reshape as much as possible, the process is terminated early when a simple operation of tensor A and tensor B is established.
  • mobilenetv3_large_pytorch.onnx.zip

    Before	After

• PRelu
  • Review processing of pre_explicit_broadcast and explicit_broadcast.

What's Changed

• [experimental] To avoid generating unnecessary Reshape as much as possible, the process is terminated early when a simple operation of tensor A and tensor B is established. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/463

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.16...1.15.17

• common_functions.py
  • Improved shape correction process.
  • Consider patterns where axis broadcasts occur.
  • Further improves the stability of the overall model transformation.
• [DNN-based_source_separation] Facing issue in Mul node during conversion #460

What's Changed

• Improved shape correction process by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/462

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.15...1.15.16

• Unsqueeze
  • Improved propagation of NHWC flags under limited conditions in Unsqueeze.
  • Reduces the frequency of shape unmatch errors in subsequent OPs.
• common_functions.py
  • Implemented workarounds to avoid shape mismatch errors in Mul, Sub, Div, and Mod to improve overall model transformation stability.

onnx2tf -i conv_tasnet_dnn.onnx -kat input -cotof

• [DNN-based_source_separation] Facing issue in Mul node during conversion #460

What's Changed

• Improved propagation of NHWC flags under limited conditions in Unsqueeze. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/461

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.14...1.15.15

• BatchNormalization
  • Workaround for inconsistent C position.
  • This is a special optimization measure especially for onnx generated using tf2onnx (tensorflow-onnx).
  • Reduction of conversion errors caused by missing the position of the C when there is a large number of unnecessary Transpose just before the BatchNormalization.
  • Improved stability of the overall model conversion.

What's Changed

• Added description of workaround to avoid accuracy errors due to TFLite runtime by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/457
• BatchNormalization Workaround for inconsistent "C" position by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/459

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.13...1.15.14

• onnx2tf.py, common_functions.py
  • Disable do_type_check when exporting ONNX graphs from onnx_graphsurgeon.
  • Ref: https://github.com/PINTO0309/soa4onnx/pull/1

What's Changed

• CI disk space further expanded from 53GB to 63GB by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/452
• Remove unused and useless lines by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/453
• Disable do_type_check when exporting ONNX graphs from onnx_graphsurgeon by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/456

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.12...1.15.13

• Conv
  • convert a Conv with pads [3,3,3,1] on onnx, than I receive an mismatch
  • when there's Asymmetrical pads, consider it as user define params and lead it to get_padding_as_op

What's Changed

• fix unusual pads of conv by @chanendy in https://github.com/PINTO0309/onnx2tf/pull/450
• v1.15.11 -> v1.15.12 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/451

New Contributors

• @chanendy made their first contribution in https://github.com/PINTO0309/onnx2tf/pull/450

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.11...1.15.12

• Gemm
  • Fixed to disable transposition when bias is one dimensional.
  • Fixed a problem in which tensors were erroneously transposed in the opposite direction when a one-dimensional bias was present.
  • https://github.com/woshidandan/TANet

What's Changed

• Fixed to disable transposition when bias is one dimensional. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/448

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.10...1.15.11

• MaxPool

  • Implemented a workaround to deal with the problem that padding with the minimum value causes the output error of MaxPool2D to be maximized only when quantizing with INT8 quantization. #444

    Float32 model outputs flattened shape: (512,)
    Int8 model outputs flattened shape: (512,)
    Euclidean Distance: 2.0942165851593018
    

• [OSNet] int8 tflite model - catastrophic accuracy degradation #444

What's Changed

• Implemented a workaround to deal with the problem that padding with the minimum value causes the output error of MaxPool2D to be maximized only when quantizing with INT8 quantization. #444 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/446

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.9...1.15.10

• onnx2tf.py
  • Fixed a lack of processing of OP name sanitizing when the -onimc and -osd or -oiqt options are specified in combination.
  • osnet_x1_0_fp_32_bs_1.onnx.zip

    onnx2tf \
    -i osnet_x1_0_fp_32_bs_1.onnx \
    -onimc /conv2/conv2.0/Relu_output_0 \
    -oiqt \
    -qt per-tensor
    

• [OSNet] int8 tflite model - catastrophic accuracy degradation #444

What's Changed

• Inference with dynamic tensors in TFLite by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/442
• Fixed a lack of processing of OP name sanitizing when the -onimc and -osd or -oiqt options are specified in combination. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/445

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.8...1.15.9

• Concat
  • Set NHWC flag to True if all input tensors are determined by NHWC
  • Significantly reduces the frequency of conversion errors in the overall model.
  • https://github.com/PINTO0309/PINTO_model_zoo/tree/main/303_FAN
  • Before
  • After

What's Changed

• Set NHWC flag to True if all input tensors are determined by NHWC by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/440

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.7...1.15.8

• Transpose
  • Improved conversion stability when the input tensor is np.ndarray.

What's Changed

• Improved conversion stability when the input tensor is np.ndarray. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/439

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.6...1.15.7

• LpNormalization
  • LpNormalization conversion does not set parameters axis and p properly.

What's Changed

• LpNormalization.py bugfix - parameters by @radekzc in https://github.com/PINTO0309/onnx2tf/pull/438

New Contributors

• @radekzc made their first contribution in https://github.com/PINTO0309/onnx2tf/pull/438

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.5...1.15.6

• Expand
  • Improved the situation where FlexMul is generated when the input tensor of Expand is bool.
  • Stop using INT8 and use INT32 instead.
• Before
• After
• [ParSeq] [Error] convert error mismatch after saved_model output complete #436

What's Changed

• Improved the situation where FlexMul is generated when the input tensor of Expand is bool. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/437

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.4...1.15.5

• Eliminated the use_cuda option and modified it so that CUDA is automatically used when an OP that requires CUDA is present.
• [TODO] Add option to disable CUDAExecutionProvider during inference #412

What's Changed

• Eliminated the use_cuda option and modified it so that CUDA is automatically used when an OP that requires CUDA is present. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/435

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.3...1.15.4

• Added to read from local test image data file if available in current working dir instead of getting from remote

What's Changed

• added to read from local test image data file if available in current working dir instead of getting from remote by @dimalukas in https://github.com/PINTO0309/onnx2tf/pull/433
• v1.15.2 -> v1.15.3 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/434

New Contributors

• @dimalukas made their first contribution in https://github.com/PINTO0309/onnx2tf/pull/433

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.2...1.15.3

• Add verbosity flag to provide better configuration of prints

  • non_verbose just sets verbosity to "error" and is kept around for backwards compatibility.
  • Merge Color into logging file
    • Color objects are callable and wrap a string with the Color and Color.RESET
  • print statements use the proper logging method

• CI

  • Maximize free disk space

What's Changed

• Add verbosity flag and refactor logging by @democat3457 in https://github.com/PINTO0309/onnx2tf/pull/431
• Maximize free disk space by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/432

New Contributors

• @democat3457 made their first contribution in https://github.com/PINTO0309/onnx2tf/pull/431

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.1...1.15.2

• Improved stability when models exceeding the ProtocolBuffers size limit of 2 GB are targeted for conversion.
  • However, if the model size is about 4 GB, as in StableDiffusion, the conversion target is expected to require resources of about 128 GB RAM + 1 TB SWAP or 512 GB RAM or 1 TB RAM, etc...
  • Since it is not possible to prepare such a huge resource in the development environment, it is not possible to confirm that the conversion completes successfully.
  • TensorFlow's saved_model format uses the same file format as ONNX (Protocol Buffers), so it cannot generate models with file sizes larger than 2 GB.
• ConstantOfShape
  • Improved stability of ConstantOfShape conversions
• Conversion error with stable diffusion model. #424

What's Changed

• Improved stability when models exceeding the ProtocolBuffers size limit of 2 GB are targeted for conversion. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/428

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.15.0...1.15.1

• Support for Float32 inverse conversion of Float16 models.
• https://github.com/opencv/opencv_zoo
• text_recognition_CRNN_EN_2023feb_fp16.onnx.zip

What's Changed

• Support for Float32 inverse conversion of Float16 models by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/427

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.14.7...1.15.0

• acquisition_of_validation_data
  • Fix acquisition_of_validation_data to skip processing if the input tensor is a tf.Tensor and has the attribute numpy.
• Inverse
  • Fixed to convert to tf.Tensor when the input tensor is np.ndarray.
• Improved overall stability of model conversion behavior

What's Changed

• Fix acquisition_of_validation_data to skip processing if the input tensor is a tf.Tensor and has the attribute numpy. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/423

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.14.6...1.14.7

• Inverse
  • Fix Abort problem with Inverse transformations when the input tensor is np.ndarray.
• [TODO] Fix Abort problem with Inverse transformations when the input tensor is np.ndarray. #421

What's Changed

• Fix Abort problem with Inverse transformations when the input tensor is np.ndarray by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/422

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.14.5...1.14.6

• Improved stability when slope is less than the number of axes in the input tensor and slope is np.ndarray and the input tensor is fixed in NHWC.
• Reduces the problem of transposition errors when all dimensions except batch size have the same value.
• example.onnx.zip
• PRelu "not compatible with provided weight shape" conversion problem #418

What's Changed

• Revision of environment construction procedures due to Goole Colab backend OS version upgrade to Ub Ubuntu22.04. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/417
• Improved stability when slope is less than the number of axes in the input tensor and slope is np.ndarray and the input tensor is fixed in NHWC by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/419

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.14.4...1.14.5

• Pad
  • Improved stability of constant transformations when paddings is Numpy.ndarray
• Dummy inference with CUDAExcecutionProvider causes large errors in inference results, so fixed to infer using CUDA only when --use_cuda is explicitly specified.
• Upgrade onnxsim to 0.4.33
• [TODO] Add option to disable CUDAExecutionProvider during inference #412
• Error in tf.pad operation #413

What's Changed

• Improved stability of constant transformations when paddings is Numpy.ndarray by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/414

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.14.3...1.14.4

• DequantizeLinear
  • Improved DequantizeLinear conversion stability

What's Changed

• tf v2.13.0rc0 -> v2.13.0 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/408
• Improved DequantizeLinear conversion stability by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/409

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.14.2...1.14.3

• GridSample
  • Fixed a transposition error when the input tensor of grid in GridSample is a constant.
• Tile
  • Added NHWC information propagation process to Tile.
  • Fail to understand the GridSample error message #402

What's Changed

• Fixed a transposition error when the input tensor of grid in GridSample is a constant. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/407

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.14.1...1.14.2

• Support for docker.io
  • docker.io/pinto0309/onnx2tf:latest

What's Changed

• Support for docker.io by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/405
• Fix Meta by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/406

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.14.0...1.14.1

• TensorFlow v2.13.0
• Improved detection of CUDAExecutionProvider.

What's Changed

• TensorFlow v2.13.0 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/404

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.13...1.14.0

• onnxruntime-gpu support
• Support for com.microsoft.GroupNorm
• com.microsoft.GroupNorm.onnx.zip
• com.microsoft.GroupNorm support for Stable Diffusion model support #401

What's Changed

• Add OS information by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/400
• Support for com.microsoft.GroupNorm by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/403

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.12...1.13.13

• MaxPool
  • MaxPoolWithArgmax support.
  • MaxPoolWithArgmax with dilations is not yet implemented.
  • Since TFLite runtime does not support MaxPoolWithArgmax, the OP in the .tflite file is FlexMaxPoolWithArgmax.
  • Also, the replacement of MaxPoolWithArgmax with the standard OP was not performed because the computational complexity of ArgMax would be unrealistically large.
  • Partial support.
  • https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_397/hair_segmenter.onnx
• onnx2tf.py
  • Fixed a problem where models were not generated correctly if the model output name contained :.
• [TODO] [hair_segmenter] Implementation of MaxPoolWithArgmax #397

What's Changed

• Support for MaxPoolWithArgmax by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/398

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.11...1.13.12

• BatchNormalization
  • Improved BatchNormalization processing stability
  • CNN_AUTOENCODER.onnx.zip
  • [CNN_AUTOENCODER] Error in cnn autoencoder output value validation
• Gather
  • Fixed a transposition problem when indices is less than one dimension.
  • [PaDiM] The order of the Gather indices list may change. #391

What's Changed

• Improved BatchNormalization processing stability by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/395
• Fixed a transposition problem when indices is less than one dimension by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/393

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.10...1.13.11

• Split
  • Modified Split to take over NHWC information to improve stability of the conversion process.
  • crestereo_combined_iter2_240x320.onnx.zip
• CREStereo conversion from onnx to tensorflow failed #380

What's Changed

• Modified Split to take over NHWC information to improve stability of the conversion process. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/386

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.9...1.13.10

• MatMul
  • Improved stability of transformations when MatMul input tensor contains undefined dimensions in cases where onnxsim (shape_inference) fails.
• InstanceNormalization
  • Since the error between ONNX and TF in InstanceNormalization was originally large, the criteria for accuracy check was relaxed.
• common_functions (explicit_broadcast)
  • Fixed explicit_broadcast to skip subsequent processing if input_1 and input_2 have exactly the same shape from the beginning.
• CREStereo conversion from onnx to tensorflow failed #380

What's Changed

• Improved stability of MatMul, explicit_broadcast, and InstanceNormalization transformations by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/385

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.8...1.13.9

• common_functions
  • Improve conversion stability of explicit_broadcast
  • gaze_estimation_adas_0002_zero_remove.onnx.zip

What's Changed

• Improve conversion stability of explicit_broadcast by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/384

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.7...1.13.8

• Rewriting of tflite input/output OP names and signature_defs If you do not like tflite input/output names such as serving_default_*:0 or StatefulPartitionedCall:0, you can rewrite them using the following tools and procedures. It can be rewritten from any name to any name, so it does not have to be serving_default_*:0 or StatefulPartitionedCall:0.

  https://github.com/PINTO0309/tflite-input-output-rewriter

  # Install custom flatc
  $ wget https://github.com/PINTO0309/onnx2tf/releases/download/1.7.3/flatc.tar.gz \
      && tar -zxvf flatc.tar.gz \
      && sudo chmod +x flatc \
      && sudo mv flatc /usr/bin/ \
      && rm flatc.tar.gz
  
  # Path check
  $ which flatc
  /usr/bin/flatc
  
  # Install tfliteiorewriter
  $ pip install -U tfliteiorewriter
  

  • Before

    $ tfliteiorewriter \
    -i xxxx.tflite \
    -r serving_default_input_1:0 aaa \
    -r StatefulPartitionedCall:0 bbb
    

  • After

What's Changed

• Rewriting of tflite input/output OP names and signature_defs by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/383

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.6...1.13.7

• Conv1D, Conv2D, Conv3D, DepthwiseConv2D, GroupConv1D, GroupConv2D, GroupConv3D, SeparableConv
  • Improved transposition error when receiving tensors with all axes having the same numerical value as input, thus reducing accuracy degradation.
  • https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_381/monodepth_op15.onnx

    onnx2tf -i monodepth_op15.onnx -cotof
    

• Before
• After
• Conversion issue when trying to run model with Vx delegate #381

What's Changed

• Improved transposition error when receiving tensors with all axes having the same numerical value as input, thus reducing accuracy degradation by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/382

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.5...1.13.6

• Slice

  • Added optimization of wasted Slice where the shape of the input tensor is exactly the same as the shape of the output tensor.

  • lite-model_rosetta_float16_1_part.onnx.zip

    Before (ONNX)	After (TFLite)

What's Changed

• Added optimization of wasted Slice where the shape of the input tensor is exactly the same as the shape of the output tensor by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/378

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.4...1.13.5

• Slice
  • Improved stability of transformations when axes is None and slices are for multiple axes
  • https://discuss.tensorflow.org/t/how-to-convert-tflite-to-model-json-weights/17174
  • lite-model_rosetta_float16_1_part.onnx.zip

What's Changed

• Improved stability of transformations when axes is None and slices are for multiple axes by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/377

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.3...1.13.4

• InstanceNormalization
  • Added automatic correction function for accuracy degradation.
  • For models with strict_mode enabled (Default is enabled) and including InstanceNormalization, the conversion is slightly slower, but at the same time minimizes the post-conversion inference error as much as possible.

What's Changed

• Added automatic correction function for accuracy degradation by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/376

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.2...1.13.3

• onnx2tf
  • Improvements to reduce processing time for CI
  • Added option to disable model saving process after conversion.

    -dms, --disable_model_save
      Does not save the converted model. For CIs RAM savings.
    

What's Changed

• Improvements to reduce processing time for CI by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/375

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.13.1...1.13.2

Re-release

• onnxruntime
  • Upgrade to onnxruntime 1.15.0
  • Increase inference speed with onnxruntime
    • (upper limit of the number of logical cores in the CPU - 1) threads are launched for parallel inference, improving the speed of dummy inference.

What's Changed

• Upgrade to onnxruntime 1.15.0 and increase inference speed with onnxruntime by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/374

• onnxruntime
  • Upgrade to onnxruntime 1.15.0
  • Increase inference speed with onnxruntime
    • (upper limit of the number of logical cores in the CPU - 1) threads are launched for parallel inference, improving the speed of dummy inference.

What's Changed

• Upgrade to onnxruntime 1.15.0 and increase inference speed with onnxruntime by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/374

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.12.7...1.13.0

• common_functions
  • Improved conversion stability for models where ONNX shape estimation has not yet been performed

What's Changed

• Improved conversion stability for models where ONNX shape estimation has not yet been performed by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/373

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.12.6...1.12.7

• OptionalGetElement, OptionalHasElement
  • Support for OptionalGetElement, OptionalHasElement
  • It is an error if the input value does not have an element and the behavior is undefined in this case.
  • https://github.com/onnx/onnx/blob/main/docs/Changelog.md#optionalgetelement-15
  • https://www.tensorflow.org/api_docs/python/tf/experimental/Optional

What's Changed

• OptionalGetElement, OptionalHasElement by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/372

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.12.5...1.12.6

• HannWindow, HammingWindow
  • Support for HannWindow, HammingWindow

What's Changed

• HannWindow, HammingWindow by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/371

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.12.4...1.12.5

• StringNormalizer
  • Experimental implementation.
  • Test patterns are severely lacking, waiting for sample ONNX files from the community to cover test patterns.
  • Since TensorFlow does not have an equivalent for locale, I have implemented it ignoring the locale process for now.
  • I don't want to install an additional add-on package tensorflow_text, so for now, I've decided to keep the implementation without tensorflow_text. However, it is my understanding that tensorflow_text is inherently considerably more flexible and capable of various text processing.

What's Changed

• [Experimental / Partial] Support for StringNormalizer by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/370

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.12.3...1.12.4

• Support for MelWeightMatrix

What's Changed

• Support for MelWeightMatrix by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/369

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.12.2...1.12.3

• Fix issue related to shape of Resize input 5D
  • Support 5D input for operator Resize #362

What's Changed

• fix resize operator for 5D input by @nguyencse in https://github.com/PINTO0309/onnx2tf/pull/366
• Only adjust variable names and remove comments by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/367

New Contributors

• @nguyencse made their first contribution in https://github.com/PINTO0309/onnx2tf/pull/366

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.12.1...1.12.2

• common_functions
  • Correction of 3D Resize shape acquisition process
  • upsampling3d_bicubic, upsampling3d_nearest
• Support 5D input for operator Resize #362

What's Changed

• Correction of 3D Resize shape acquisition process by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/364

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.12.0...1.12.1

• TF v2.12.0 -> TF v2.13.0-rc0
• https://github.com/tensorflow/tensorflow/releases/tag/v2.13.0-rc0
• [TODO] Verification of operation when upgrading to TensorFlow v1.13.0 #348

What's Changed

• TF v2.12.0 -> TF v2.13.0-rc0 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/363

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.11.11...1.12.0

• onnx2tf
  • Fixed to sanitize all error messages to be more understandable to avoid confusion caused by error messages related to TensorFlow bug that does not allow GroupConvolution to be output to saved_model.
• [PaddleOCR] Unable to convert paddleOCR model to tensorflow/keras only tflite #360

What's Changed

• Fixed to sanitize all error messages to be more understandable to avoid confusion caused by error messages related to TensorFlow bug that does not allow GroupConvolution to be output to saved_model. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/361

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.11.10...1.11.11

• InstanceNormalization
  • Improved processing efficiency of InstanceNormalization
• Added JSON for AnimeGANv2
• https://github.com/Kazuhito00/AnimeGANv2-ONNX-Sample/raw/main/model/face_paint_512_v2_0.onnx

What's Changed

• Improved processing efficiency of InstanceNormalization and added JSON for AnimeGANv2 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/359

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.11.9...1.11.10