Skip to content

Self-Created Tools to convert ONNX files (NCHW) to TensorFlow/TFLite/Keras format (NHWC). The purpose of this tool is to solve the massive Transpose extrapolation problem in onnx-tensorflow (onnx-tf). I don't need a Star, but give me a pull request.

License

Notifications You must be signed in to change notification settings

PINTO0309/onnx2tf

Repository files navigation

onnx2tf

Self-Created Tools to convert ONNX files (NCHW) to TensorFlow/TFLite/Keras format (NHWC). The purpose of this tool is to solve the massive Transpose extrapolation problem in onnx-tensorflow (onnx-tf). I don't need a Star, but give me a pull request. Since I am adding challenging model optimizations and fixing bugs almost daily, I frequently embed potential bugs that would otherwise break through CI's regression testing. Therefore, if you encounter new problems, I recommend that you try a package that is a few versions older, or try the latest package that will be released in a few days.

Incidentally, I have never used this tool in practice myself since I started working on it. It doesn't matter.

Downloads GitHub Python PyPI CodeQL Model Convert Test Status DOI

Note

  • The torch.script-based torch.onnx.export has already been moved to maintenance mode, and we recommend moving to the FX graph-based torch.onnx.dynamo_export starting with PyTorch v2.2.0.

  • The greatest advantage of ONNX generated by torch.onnx.dynamo_export would be that it directly references the PyTorch implementation, allowing for the conversion of any OP that was previously difficult to convert to ONNX.

  • The maintainers of ONNX and PyTorch have assured us that they will not add new OPs after opset=18 to the existing torch.onnx.export.

  • https://pytorch.org/docs/stable/onnx_dynamo.html#torch.onnx.dynamo_export

  • This can be converted directly into an ONNX graph using Pythonic code using onnxscript.

    image

  • For future model versatility, it would be a good idea to consider moving to torch.onnx.dynamo_export at an early stage.

  • Google AI Edge Torch AI Edge Torch is a python library that supports converting PyTorch models into a .tflite format, which can then be run with TensorFlow Lite and MediaPipe. This enables applications for Android, iOS and IOT that can run models completely on-device. AI Edge Torch offers broad CPU coverage, with initial GPU and NPU support. AI Edge Torch seeks to closely integrate with PyTorch, building on top of torch.export() and providing good coverage of Core ATen operators.

    https://github.com/google-ai-edge/ai-edge-torch?tab=readme-ov-file#pytorch-converter

    import torch
    import torchvision
    import ai_edge_torch
    
    # Use resnet18 with pre-trained weights.
    resnet18 = torchvision.models.resnet18(torchvision.models.ResNet18_Weights.IMAGENET1K_V1)
    sample_inputs = (torch.randn(1, 3, 224, 224),)
    
    # Convert and serialize PyTorch model to a tflite flatbuffer. Note that we
    # are setting the model to evaluation mode prior to conversion.
    edge_model = ai_edge_torch.convert(resnet18.eval(), sample_inputs)
    edge_model.export("resnet18.tflite")
  • Google for Developers Blog MAY 14, 2024 - AI Edge Torch: High Performance Inference of PyTorch Models on Mobile Devices

    https://developers.googleblog.com/en/ai-edge-torch-high-performance-inference-of-pytorch-models-on-mobile-devices/

  • Considering the compatibility of Pythonic code with TensorFlow/Keras/TFLite and the beauty of the conversion workflow, nobuco is the most optimal choice going forward.

  • The role of onnx2tf will end within the next one to two years. I don't intend to stop the maintenance of onnx2tf itself anytime soon, but I will continue to maintain it little by little as long as there is demand for it from everyone. The end of onnx2tf will be when TensorRT and other runtimes support porting from FX Graph based models.

Model Conversion Status

https://github.com/PINTO0309/onnx2tf/wiki/model_status

Supported layers

  • https://github.com/onnx/onnx/blob/main/docs/Operators.md

  • ✔️: Supported ✅: Partial support Help wanted: Pull Request are welcome

    See the list of supported layers
    OP Status
    Abs ✔️
    Acosh ✔️
    Acos ✔️
    Add ✔️
    And ✔️
    ArgMax ✔️
    ArgMin ✔️
    Asinh ✔️
    Asin ✔️
    Atanh ✔️
    Atan ✔️
    AveragePool ✔️
    BatchNormalization ✔️
    Bernoulli ✔️
    BitShift ✔️
    BitwiseAnd Help wanted
    BitwiseNot Help wanted
    BitwiseOr Help wanted
    BitwiseXor Help wanted
    Cast ✔️
    Ceil ✔️
    Celu ✔️
    CenterCropPad Help wanted
    Clip ✔️
    Col2Im âś…
    Compress ✔️
    ConcatFromSequence ✔️
    Concat ✔️
    ConstantOfShape ✔️
    Constant ✔️
    Conv ✔️
    ConvInteger âś…
    ConvTranspose ✔️
    Cosh ✔️
    Cos ✔️
    CumSum ✔️
    DeformConv Help wanted
    DepthToSpace ✔️
    Det ✔️
    DequantizeLinear ✔️
    DFT Help wanted
    Div ✔️
    Dropout ✔️
    DynamicQuantizeLinear ✔️
    Einsum ✔️
    Elu ✔️
    Equal ✔️
    Erf ✔️
    Expand ✔️
    Exp ✔️
    EyeLike ✔️
    Flatten ✔️
    Floor ✔️
    FusedConv ✔️
    GatherElements ✔️
    GatherND ✔️
    Gather ✔️
    Gelu ✔️
    Gemm ✔️
    GlobalAveragePool ✔️
    GlobalLpPool ✔️
    GlobalMaxPool ✔️
    GreaterOrEqual ✔️
    Greater ✔️
    GridSample âś…
    GroupNormalization Help wanted
    GRU ✔️
    HammingWindow âś…
    HannWindow âś…
    Hardmax ✔️
    HardSigmoid ✔️
    HardSwish ✔️
    Identity ✔️
    If ✔️
    Input ✔️
    InstanceNormalization ✔️
    Inverse ✔️
    IsInf ✔️
    IsNaN ✔️
    LayerNormalization ✔️
    LeakyRelu ✔️
    LessOrEqual ✔️
    Less ✔️
    Log ✔️
    LogSoftmax ✔️
    Loop Help wanted
    LpNormalization ✔️
    LRN ✔️
    LSTM ✔️
    MatMul ✔️
    MatMulInteger ✔️
    MaxPool ✔️
    Max ✔️
    MaxRoiPool Help wanted
    MaxUnpool ✔️
    Mean ✔️
    MeanVarianceNormalization ✔️
    MelWeightMatrix ✔️
    Min ✔️
    Mish ✔️
    Mod ✔️
    Mul ✔️
    Multinomial ✔️
    Neg ✔️
    NonMaxSuppression ✔️
    NonZero ✔️
    Optional Help wanted
    OptionalGetElement ✔️
    OptionalHasElement ✔️
    Not ✔️
    OneHot ✔️
    Or ✔️
    Pad ✔️
    Pow ✔️
    PRelu ✔️
    QLinearAdd ✔️
    QLinearConcat ✔️
    QLinearConv ✔️
    QLinearLeakyRelu ✔️
    QLinearMatMul ✔️
    QLinearMul ✔️
    QLinearSigmoid ✔️
    QLinearSoftmax ✔️
    QuantizeLinear ✔️
    RandomNormalLike ✔️
    RandomNormal ✔️
    RandomUniformLike ✔️
    RandomUniform ✔️
    Range ✔️
    Reciprocal ✔️
    ReduceL1 ✔️
    ReduceL2 ✔️
    ReduceLogSum ✔️
    ReduceLogSumExp ✔️
    ReduceMax ✔️
    ReduceMean ✔️
    ReduceMin ✔️
    ReduceProd ✔️
    ReduceSum ✔️
    ReduceSumSquare ✔️
    Relu ✔️
    Reshape ✔️
    Resize ✔️
    ReverseSequence ✔️
    RNN ✔️
    RoiAlign ✔️
    Round ✔️
    ScaleAndTranslate ✔️
    Scatter ✔️
    ScatterElements ✔️
    ScatterND ✔️
    Scan Help wanted
    Selu ✔️
    SequenceAt ✔️
    SequenceConstruct ✔️
    SequenceEmpty ✔️
    SequenceErase ✔️
    SequenceInsert ✔️
    SequenceLength ✔️
    Shape ✔️
    Shrink ✔️
    Sigmoid ✔️
    Sign ✔️
    Sinh ✔️
    Sin ✔️
    Size ✔️
    Slice ✔️
    Softmax ✔️
    Softplus ✔️
    Softsign ✔️
    SpaceToDepth ✔️
    Split ✔️
    SplitToSequence ✔️
    Sqrt ✔️
    Squeeze ✔️
    STFT âś…
    StringNormalizer âś…
    Sub ✔️
    Sum ✔️
    Tanh ✔️
    Tan ✔️
    TfIdfVectorizer Help wanted
    ThresholdedRelu ✔️
    Tile ✔️
    TopK ✔️
    Transpose ✔️
    Trilu ✔️
    Unique ✔️
    Unsqueeze ✔️
    Upsample ✔️
    Where ✔️
    Xor ✔️

Demo

Video speed is adjusted approximately 50 times slower than actual speed. render1665941718294

Environment

  • Linux / Windows
  • onnx==1.16.1
  • onnxruntime==1.18.1
  • onnx-simplifier==0.4.33 or 0.4.30 (onnx.onnx_cpp2py_export.shape_inference.InferenceError: [ShapeInferenceError] (op_type:Slice, node name: /xxxx/Slice): [ShapeInferenceError] Inferred shape and existing shape differ in rank: (x) vs (y))
  • onnx_graphsurgeon
  • simple_onnx_processing_tools
  • tensorflow==2.17.0, Special bugs: #436
  • psutil==5.9.5
  • ml_dtypes==0.3.2
  • flatbuffers-compiler (Optional, Only when using the -coion option. Executable file named flatc.)
  • flatbuffers>=23.5.26
    # Custom flatc v23.5.26 binary for Ubuntu 20.04+
    # https://github.com/PINTO0309/onnx2tf/issues/196
    wget https://github.com/PINTO0309/onnx2tf/releases/download/1.16.31/flatc.tar.gz \
    && tar -zxvf flatc.tar.gz \
    && sudo chmod +x flatc \
    && sudo mv flatc /usr/bin/

Sample Usage

1. Install

Note:

1. If you are using TensorFlow v2.13.0 or earlier, use a version older than onnx2tf v1.17.5. onnx2tf v1.17.6 or later will not work properly due to changes in TensorFlow's API.

2. The latest onnx2tf implementation is based on Keras API 3 and will not work properly if you install TensorFlow v2.15.0 or earlier.

  • HostPC

    Click to expand
    # PAT authentication is required to pull from GHCR.
    docker login ghcr.io
    
    Username (xxxx): {Enter}
    Password: {Personal Access Token}
    Login Succeeded
    
    docker run --rm -it \
    -v `pwd`:/workdir \
    -w /workdir \
    ghcr.io/pinto0309/onnx2tf:1.26.2
    
    or
    
    # Authentication is not required for pulls from Docker Hub.
    docker run --rm -it \
    -v `pwd`:/workdir \
    -w /workdir \
    docker.io/pinto0309/onnx2tf:1.26.2
    
    or
    
    pip install -U onnx==1.16.1 \
    && pip install -U nvidia-pyindex \
    && pip install -U onnx-graphsurgeon \
    && pip install -U onnxruntime==1.18.1 \
    && pip install -U onnxsim==0.4.33 \
    && pip install -U simple_onnx_processing_tools \
    && pip install -U sne4onnx>=1.0.13 \
    && pip install -U sng4onnx>=1.0.4 \
    && pip install -U tensorflow==2.17.0 \
    && pip install -U protobuf==3.20.3 \
    && pip install -U onnx2tf \
    && pip install -U h5py==3.11.0 \
    && pip install -U psutil==5.9.5 \
    && pip install -U ml_dtypes==0.3.2 \
    && pip install -U tf-keras~=2.16 \
    && pip install flatbuffers>=23.5.26
    
    or
    
    pip install -e .

or

  • Google Colaboratory Python3.10

    Click to expand
    !sudo apt-get -y update
    !sudo apt-get -y install python3-pip
    !sudo apt-get -y install python-is-python3
    !wget https://github.com/PINTO0309/onnx2tf/releases/download/1.16.31/flatc.tar.gz \
      && tar -zxvf flatc.tar.gz \
      && sudo chmod +x flatc \
      && sudo mv flatc /usr/bin/
    !pip install -U pip \
      && pip install tensorflow==2.17.0 \
      && pip install -U onnx==1.16.1 \
      && python -m pip install onnx_graphsurgeon \
            --index-url https://pypi.ngc.nvidia.com \
      && pip install -U onnxruntime==1.18.1 \
      && pip install -U onnxsim==0.4.33 \
      && pip install -U simple_onnx_processing_tools \
      && pip install -U onnx2tf \
      && pip install -U protobuf==3.20.3 \
      && pip install -U h5py==3.11.0 \
      && pip install -U psutil==5.9.5 \
      && pip install -U ml_dtypes==0.3.2 \
      && pip install -U tf-keras~=2.16 \
      && pip install flatbuffers>=23.5.26
    

2. Run test

Only patterns that are considered to be used particularly frequently are described. In addition, there are several other options, such as disabling Flex OP and additional options to improve inference performance. See: CLI Parameter

# Float32, Float16
# This is the fastest way to generate tflite.
# Improved to automatically generate `signature` without `-osd` starting from v1.25.3.
# Also, starting from v1.24.0, efficient TFLite can be generated
# without unrolling `GroupConvolution`. e.g. YOLOv9, YOLOvN
# Conversion to other frameworks. e.g. TensorFlow.js, CoreML, etc
# https://github.com/PINTO0309/onnx2tf#19-conversion-to-tensorflowjs
# https://github.com/PINTO0309/onnx2tf#20-conversion-to-coreml
wget https://github.com/PINTO0309/onnx2tf/releases/download/0.0.2/resnet18-v1-7.onnx
onnx2tf -i resnet18-v1-7.onnx

ls -lh saved_model/

assets
fingerprint.pb
resnet18-v1-7_float16.tflite
resnet18-v1-7_float32.tflite
saved_model.pb
variables

TF_CPP_MIN_LOG_LEVEL=3 \
saved_model_cli show \
--dir saved_model \
--signature_def serving_default \
--tag_set serve

The given SavedModel SignatureDef contains the following input(s):
  inputs['data'] tensor_info:
      dtype: DT_FLOAT
      shape: (-1, 224, 224, 3)
      name: serving_default_data:0
The given SavedModel SignatureDef contains the following output(s):
  outputs['output_0'] tensor_info:
      dtype: DT_FLOAT
      shape: (-1, 1000)
      name: PartitionedCall:0
Method name is: tensorflow/serving/predict

# In the interest of efficiency for my development and debugging of onnx2tf,
# the default configuration shows a large amount of debug level logs.
# However, for most users, a large number of debug logs are unnecessary.
# If you want to reduce the amount of information displayed in the conversion log,
# you can change the amount of information in the log by specifying the
# `--verbosity` or `-v` option as follows.
# Possible values are "debug", "info", "warn", and "error".
wget https://github.com/PINTO0309/onnx2tf/releases/download/0.0.2/resnet18-v1-7.onnx
onnx2tf -i resnet18-v1-7.onnx -v info

# Override undefined batch size or other dimensions with static values.
# If the model has undefined dimensions, rewriting them to a static size will significantly
# improve the success rate of the conversion.
# The `-b` option overwrites the zero-dimensional batch size with the number specified
# without input OP name.
# Note that if there are multiple input OPs, the zero dimension of all input OPs is
# forced to be rewritten.
# The `-ois` option allows undefined dimensions in all dimensions, including
# the zero dimensionality, to be overwritten to a static shape, but requires
# the input OP name to be specified.
# e.g. -ois data1:1,3,224,224 data2:1,255 data3:1,224,6
wget https://github.com/PINTO0309/onnx2tf/releases/download/0.0.2/resnet18-v1-7.onnx
onnx2tf -i resnet18-v1-7.onnx -b 1
or
onnx2tf -i resnet18-v1-7.onnx -ois data:1,3,224,224

# Suppress automatic transposition of input OPs from NCW, NCHW, NCDHW to NWC, NHWC, NDHWC.
# onnx2tf is a specification that automatically transposes the input OP to [N,H,W,C] format
# before converting the model. However, since onnx2tf cannot determine from the structure of
# the model whether the input data is image, audio data, or something else, it unconditionally
# transposes the channels. Therefore, it is the models of STT/TTS models where the input is
# not NHWC that tend to have particular problems with the automatic transposition of the
# input OP.
# If you do not want input OPs to be automatically transposed, you can disable automatic
# transposition of input OPs by specifying the `-kat` option.
wget https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/double_gru.onnx
# INPUT OPs: "spec": float32[1,3,257,1], "states_in": float32[2,1,32]
# The following command suppresses the automatic transposition of "states_in" and converts it.
onnx2tf -i double_gru.onnx -kat states_in

# Keras h5 format
# .h5, .json, .keras, .weights.h5, .weights.keras, .data-00000-of-00001, .index
wget https://github.com/PINTO0309/onnx2tf/releases/download/0.0.2/resnet18-v1-7.onnx
onnx2tf -i resnet18-v1-7.onnx -oh5

# Keras keras_v3 format (TensorFlow v2.12.0 or later only)
wget https://github.com/PINTO0309/onnx2tf/releases/download/0.0.2/resnet18-v1-7.onnx
onnx2tf -i resnet18-v1-7.onnx -okv3

# TensorFlow v1 (.pb) format
wget https://github.com/PINTO0309/onnx2tf/releases/download/0.0.2/resnet18-v1-7.onnx
onnx2tf -i resnet18-v1-7.onnx -otfv1pb

# INT8 Quantization, Full INT8 Quantization
# INT8 Quantization with INT16 activation, Full INT8 Quantization with INT16 activation
# Dynamic Range Quantization
wget https://github.com/PINTO0309/onnx2tf/releases/download/1.1.1/emotion-ferplus-8.onnx
# INT8 Quantization (per-channel)
onnx2tf -i emotion-ferplus-8.onnx -oiqt
# INT8 Quantization (per-tensor)
onnx2tf -i emotion-ferplus-8.onnx -oiqt -qt per-tensor

# Split the model at the middle position for debugging
# Specify the input name of the OP
wget https://github.com/PINTO0309/onnx2tf/releases/download/1.25.0/cf_fus.onnx
onnx2tf -i cf_fus.onnx -inimc 448

# Split the model at the middle position for debugging
# Specify the output name of the OP
wget https://github.com/PINTO0309/onnx2tf/releases/download/1.25.0/cf_fus.onnx
onnx2tf -i cf_fus.onnx -onimc dep_sec

# Split the model at the middle position for debugging
# Specify the input/output name of the OP
wget https://github.com/PINTO0309/onnx2tf/releases/download/1.25.0/cf_fus.onnx
onnx2tf -i cf_fus.onnx -inimc 448 -onimc velocity

# Suppress generation of Flex OP and replace with Pseudo-Function
# [
#     Asin, Acos, Atan, Abs, PReLU,
#     LeakyReLU, Power, GatherND,
#     Neg, HardSwish, Erf, GeLU, MatMulInteger,
# ]
# Below is a sample of replacing Erf with another set of operations.
wget https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_readme/Erf_11.onnx
onnx2tf -i Erf_11.onnx -rtpo Erf

# High-dimensional Transpose decomposition
# If you do not like FlexTranspose being generated, try `-nodaftc`.
# Suppresses the generation of FlexTranspose by decomposing Transpose
# to the specified number of dimensions.
# In TensorFlow v2.12.0 and later, up to 6 dimensions are converted to normal Transpose;
# in v2.11.0 and earlier, up to 5 dimensions are converted to normal Transpose.
# Note that specifying `2` for the `-nodaftc` option causes all Transpose OPs to disappear
# from the model structure.
# Below is an example of decomposing a Transpose of 5 or more dimensions into a Transpose
# of 4 dimensions.
onnx2tf -i xxxx.onnx -nodaftc 4

# High-dimensional Slice(StridedSlice) decomposition
# If your special circumstances do not allow you to deploy a `StridedSlice` with more than
# 5 dimensions to a device, you can use the `-nodafsc` option to decompose the `StridedSlice`
# into a process with 4 or fewer dimensions.
# Below is an example of decomposing a `StridedSlice` of 5 or more dimensions into a
# `StridedSlice` of 4 dimensions.
onnx2tf -i xxxx.onnx -nodafsc 4

# Float16 inference doubling on devices with ARM64 ARMv8.2 or higher instruction set
# Double the inference speed with Float16 precision tflite models on devices with
# high-performance CPUs such as Snapdragon.
# (Pixel 3a, Pixel 5a, Pixel 7, Galaxy M12 and Galaxy S22, ...)
# XNNPACK float16 inference on certain ARM64 cores is 2x faster.
# Unfortunately, Float16 inference cannot be accelerated when using the RaspberryPi4's
# ARM64 CPU.
onnx2tf -i xxxx.onnx -eatfp16

# Parameter replacement (Resize,Transpose,Softmax)
rm replace.json
wget https://github.com/PINTO0309/onnx2tf/releases/download/1.1.27/human_segmentation_pphumanseg_2021oct.onnx
wget https://github.com/PINTO0309/onnx2tf/releases/download/1.1.27/replace.json
onnx2tf -i human_segmentation_pphumanseg_2021oct.onnx -prf replace.json

3. Accuracy check

Click to expand

Perform error checking of ONNX output and TensorFlow output. Verify that the error of all outputs, one operation at a time, is below a certain threshold. Automatically determines before and after which OPs the tool's automatic conversion of the model failed. Know where dimensional compression, dimensional expansion, and dimensional transposition by Reshape and Traspose are failing. Once you have identified the problem area, you can refer to the tutorial on Parameter replacement to modify the tool's behavior.

After many upgrades, the need for JSON parameter correction has become much less common, but there are still some edge cases where JSON correction is required. If the PC has sufficient free space in its RAM, onnx2tf will convert the model while carefully performing accuracy checks on all OPs. Thus, at the cost of successful model conversion, the conversion speed is a little slower. If the amount of RAM required for the accuracy check is expected to exceed 80% of the total available RAM capacity of the entire PC, the conversion operation will be performed without an accuracy check. Therefore, if the accuracy of the converted model is found to be significantly degraded, the accuracy may be automatically corrected by re-conversion on a PC with a large amount of RAM. For example, my PC has 128GB of RAM, but the StableDiffusion v1.5 model is too complex in its structure and consumed about 180GB of RAM in total with 50GB of SWAP space.

-ois an option to overwrite the input OP to a static size if it has undefined dimensions. -cotof option checks the accuracy of all OPs one by one. -cotoa is the error value of the threshold for determining an accuracy error. If there are undefined dimensions in the input OP, it is better to fix them to the static geometry to improve the accuracy of the accuracy measurement.

Also, you can use the -cind option to specify custom input for -cotof, instead of using the default dummy input. Otherwise, all input values will be set to 1. For more information about the -cind option, please refer to here.

The -cotof option only compares the original ONNX and converted TensorFlow (Keras) models at Float32 precision, not at Float16 or INT8 precision.

onnx2tf -i mobilenetv2-12.onnx -ois input:1,3,224,224 -cotof -cotoa 1e-1

or

onnx2tf -i mobilenetv2-12.onnx -b 1 -cotof -cotoa 1e-1

or

onnx2tf -i mobilenetv2-12.onnx -cotof -cotoa 1e-1 -cind "input" "/your/path/x.npy"

image

Kazam_screencast_00108_

4. Match tflite input/output names and input/output order to ONNX

Click to expand

If you want to match tflite's input/output OP names and the order of input/output OPs with ONNX, you can use the interpreter.get_signature_runner() to infer this after using the -coion / --copy_onnx_input_output_names_to_tflite option to output tflite file. See: #228

onnx2tf automatically compares the final input/output shapes of ONNX and the generated TFLite and tries to automatically correct the input/output order as much as possible if there is a difference. However, if INT8 quantization is used and there are multiple inputs and outputs with the same shape, automatic correction may fail. This is because TFLiteConverter shuffles the input-output order by itself only when INT8 quantization is performed.

import torch
import onnxruntime
import numpy as np
import onnx2tf
import tensorflow as tf
from tensorflow.lite.python import interpreter as tflite_interpreter

class Model(torch.nn.Module):
    def forward(self, x, y):
        return {
            "add": x + y,
            "sub": x - y,
        }

# Let's double check what PyTorch gives us
model = Model()
pytorch_output = model.forward(10, 2)
print("[PyTorch] Model Predictions:", pytorch_output)

# First, export the above model to ONNX
torch.onnx.export(
    Model(),
    {"x": 10, "y": 2},
    "model.onnx",
    opset_version=16,
    input_names=["x", "y"],
    output_names=["add", "sub"],
)

# And check its output
session = onnxruntime.InferenceSession("model.onnx")
onnx_output = session.run(["add", "sub"], {"x": np.array(10), "y": np.array(2)})
print("[ONNX] Model Outputs:", [o.name for o in session.get_outputs()])
print("[ONNX] Model Predictions:", onnx_output)

# Now, let's convert the ONNX model to TF
onnx2tf.convert(
    input_onnx_file_path="model.onnx",
    output_folder_path="model.tf",
    copy_onnx_input_output_names_to_tflite=True,
    non_verbose=True,
)

# Now, test the newer TFLite model
interpreter = tf.lite.Interpreter(model_path="model.tf/model_float32.tflite")
tf_lite_model = interpreter.get_signature_runner()
inputs = {
  'x': np.asarray([10], dtype=np.int64),
  'y': np.asarray([2], dtype=np.int64),
}
tf_lite_output = tf_lite_model(**inputs)
print("[TFLite] Model Predictions:", tf_lite_output)
[PyTorch] Model Predictions:
  {
    'add': 12,
    'sub': 8
  }
[ONNX] Model Outputs:
  [
    'add',
    'sub'
  ]
[ONNX] Model Predictions:
  [
    array(12, dtype=int64),
    array(8, dtype=int64)
  ]
[TFLite] Model Predictions:
  {
    'add': array([12]),
    'sub': array([8])
  }

image

5. Rewriting of tflite input/output OP names and signature_defs

Click to expand

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

    01

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

    02

  • After

    03

6. Embed metadata in tflite

Click to expand

If you want to embed label maps, quantization parameters, descriptions, etc. into your tflite file, you can refer to the official tutorial and try it yourself. For now, this tool does not plan to implement the ability to append metadata, as I do not want to write byte arrays to the tflite file that are not essential to its operation.

7. If the accuracy of the INT8 quantized model degrades significantly

Click to expand

It is a matter of model structure. The activation function (SiLU/Swish), kernel size and stride for Pooling, and kernel size and stride for Conv should be completely revised. See: #269

If you want to see the difference in quantization error between SiLU and ReLU, please check this Gist by @motokimura who helped us in our research. Thanks Motoki!

Gist: Quantization error simulation of SiLU (Swish) activation

The accuracy error rates after quantization for different activation functions are shown in the figure below. The graph plots the distribution of absolute error, so a position with a higher value on the horizontal axis indicates a larger error. The vertical axis is the number of samples. SiLU (Swish) produces catastrophic errors after INT8 quantization.

image

  • e.g. YOLOX-Nano

    • https://github.com/motokimura/yolox-ti-lite_tflite

    • https://github.com/TexasInstruments/edgeai-yolox

      Before After
      Swish/SiLU
      image
      ReLU
      image
      DepthwiseConv2D
      image
      Conv2D
      image
      MaxPool, kernel_size=5x5,9x9,13x13
      image
      MaxPool, kernel_size=3x3
      image
      ### Float32 - YOLOX-Nano
      (1, 52, 52, 85)
      array([[[
          [ 0.971787,  0.811184,  0.550566, ..., -5.962632, -7.403673, -6.735206],
          [ 0.858804,  1.351296,  1.231673, ..., -6.479690, -8.277064, -7.664936],
          [ 0.214827,  1.035119,  1.458006, ..., -6.291425, -8.229385, -7.761562],
              ...,
          [ 0.450116,  1.391900,  1.533354, ..., -5.672194, -7.121591, -6.880231],
          [ 0.593133,  2.112723,  0.968755, ..., -6.150078, -7.370633, -6.874294],
          [ 0.088263,  1.985220,  0.619998, ..., -5.507928, -6.914980, -6.234259]]]]),
      
      ### INT8 - YOLOX-Nano
      (1, 52, 52, 85)
      array([[[
          [ 0.941908,  0.770652,  0.513768, ..., -5.993958, -7.449634, -6.850238],
          [ 0.856280,  1.284420,  1.198792, ..., -6.507727, -8.391542, -7.792146],
          [ 0.256884,  0.941908,  1.455676, ..., -6.336471, -8.305914, -7.877774],
              ...,
          [ 0.342512,  1.370048,  1.541304, ..., -5.737075, -7.192750, -7.107122],
          [ 0.513768,  2.226327,  1.027536, ..., -6.165215, -7.449634, -7.021494],
          [ 0.085628,  2.055072,  0.685024, ..., -5.480191, -7.021494, -6.422099]]]]),
      
  • Other recommended replacement OP

    Before After
    HardSwish
    image
    ReLU
    image
    ReLU6
    image
    Paper: A Quantization-Friendly Separable Convolution for MobileNets https://arxiv.org/pdf/1803.08607.pdf
    ReLU
    image
  • Quantization range collapse due to non-zero constant padding

    If padding is performed with a constant other than zero, the padding value may destroy the quantization range of the input tensor. For example, the pattern is shown in the figure below. The MaxPool2D is done after padding the 4 sides of the input tensor with the minimum value of Float32. It seems that if INT8 quantization is performed with this structure, the quantization range is determined by MaxPool2D during quantization, including the values padded to the tensor. See: #444 image

    Therefore, the following two similar examples are equally likely to result in divergent output values for the model after INT8 quantization, with all output values being Nan or zero.

    1. Pattern with fixed value -255.0 padded on 4 sides of tensor image

    2. Pattern with fixed value -128.0 padded on 4 sides of tensor image

8. Calibration data creation for INT8 quantization

Click to expand

Calibration data (.npy) for INT8 quantization (-cind) is generated as follows. This is a sample when the data used for training is image data. See: #222

https://www.tensorflow.org/lite/performance/post_training_quantization

import cv2
import glob
import numpy as np

# Not used during data generation ################################
# You will need to do the calculations yourself using the test data
MEAN = np.asarray([[[[0.485, 0.456, 0.406]]]], dtype=np.float32) # [1,1,1,3]
STD = np.asarray([[[[0.229, 0.224, 0.225]]]], dtype=np.float32) # [1,1,1,3]
# Not used during data generation ################################

files = glob.glob("data/*.png")
img_datas = []
for idx, file in enumerate(files):
    bgr_img = cv2.imread(file)
    rgb_img = cv2.cvtColor(bgr_img, cv2.COLOR_BGR2RGB)
    resized_img = cv2.resize(rgb_img, dsize=(200,112))
    extend_batch_size_img = resized_img[np.newaxis, :]
    normalized_img = extend_batch_size_img / 255.0 # 0.0 - 1.0
    print(
        f'{str(idx+1).zfill(2)}. extend_batch_size_img.shape: {extend_batch_size_img.shape}'
    ) # [1,112,200,3]
    img_datas.append(extend_batch_size_img)
calib_datas = np.vstack(img_datas)
print(f'calib_datas.shape: {calib_datas.shape}') # [10,112,200,3]
np.save(file='data/calibdata.npy', arr=calib_datas)

loaded_data = np.load('data/calibdata.npy')
print(f'loaded_data.shape: {loaded_data.shape}') # [10,112,200,3]

"""
-cind INPUT_NAME NUMPY_FILE_PATH MEAN STD
int8_calib_datas = (loaded_data - MEAN) / STD # -1.0 - 1.0

e.g. How to specify calibration data in CLI or Script respectively.
1. CLI
  -cind "pc_dep" "data/calibdata.npy" "[[[[0.485,0.456,0.406]]]]" "[[[[0.229,0.224,0.225]]]]"
  -cind "feat" "data/calibdata2.npy" "[[[[0.123,...,0.321]]]]" "[[[[0.112,...,0.451]]]]"

2. Script
  custom_input_op_name_np_data_path=[
      ["pc_dep", "data/calibdata.npy", [[[[0.485,0.456,0.406]]]], [[[[0.229,0.224,0.225]]]]],
      ["feat", "data/calibdata2.npy", [[[[0.123,...,0.321]]]], [[[[0.112,...,0.451]]]],
  ]
"""

9. INT8 quantization of models with multiple inputs requiring non-image data

Click to expand

If you do not need to perform INT8 quantization with this tool alone, the following method is the easiest.

The -osd option will output a saved_model.pb in the saved_model folder with the full size required for quantization. That is, a default signature named serving_default is embedded in .pb. The -b option is used to convert the batch size by rewriting it as a static integer.

Note: INT8 TFLite generated by following this procedure as is will result in a model with significantly degraded accuracy. This tutorial only demonstrates the INT8 quantization procedure; if you wish to correct for accuracy, please refer to Parameter replacement to correct for transposition errors in the operation.

# Ref: https://github.com/onnx/models/tree/main/text/machine_comprehension/bert-squad
wget https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_248/bertsquad-12.onnx

onnx2tf -i bertsquad-12.onnx -b 1 -osd -cotof

image

Use the saved_model_cli command to check the saved_model signature. INT8 quantization calibration using signatures allows correct control of the input order of data for calibration. Therefore, calibration with signatures is recommended for INT8 quantization of models with multiple inputs.

saved_model_cli show --dir saved_model/ --tag_set serve --signature_def serving_default

The given SavedModel SignatureDef contains the following input(s):
  inputs['input_ids_0'] tensor_info:
      dtype: DT_INT64
      shape: (1, 256)
      name: serving_default_input_ids_0:0
  inputs['input_mask_0'] tensor_info:
      dtype: DT_INT64
      shape: (1, 256)
      name: serving_default_input_mask_0:0
  inputs['segment_ids_0'] tensor_info:
      dtype: DT_INT64
      shape: (1, 256)
      name: serving_default_segment_ids_0:0
  inputs['unique_ids_raw_output___9_0'] tensor_info:
      dtype: DT_INT64
      shape: (1)
      name: serving_default_unique_ids_raw_output___9_0:0

Calibrate by specifying the input OP name displayed in inputs. The np.ones([xxx], dtype=np.int64) part must be replaced with the correct calibration test data. In practice, several pieces of data used for training are extracted and used.

import tensorflow as tf
import numpy as np

def representative_dataset():
    unique_ids = np.ones([10, 256], dtype=np.int64)
    segment_ids = np.ones([10, 256], dtype=np.int64)
    input_masks = np.ones([10, 256], dtype=np.int64)
    input_ids = np.ones([10], dtype=np.int64)

    for unique_id, segment_id, input_mask, input_id \
        in zip(unique_ids, segment_ids, input_masks, input_ids):

        yield {
            "unique_ids_raw_output___9_0": unique_id,
            "segment_ids_0": segment_id,
            "input_mask_0": input_mask,
            "input_ids_0": input_id,
        }

converter = tf.lite.TFLiteConverter.from_saved_model('saved_model')
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = representative_dataset
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
converter.inference_input_type = tf.int8  # or tf.uint8
converter.inference_output_type = tf.int8  # or tf.uint8
tflite_quant_model = converter.convert()

with open('saved_model/int8_model.tflite', 'wb') as w:
    w.write(tflite_quant_model)

image

https://www.tensorflow.org/lite/performance/post_training_quantization

See: #248

10. Fixing the output of NonMaxSuppression (NMS)

Click to expand

PyTorch's NonMaxSuppression (torchvision.ops.nms) and ONNX's NonMaxSuppression are not fully compatible. TorchVision's NMS is very inefficient. Therefore, it is inevitable that converting ONNX using NMS in object detection models and other models will be very redundant and will be converted with a structure that is difficult for TensorFlow.js and TFLite models to take advantage of in devices. This is due to the indefinite number of tensors output by the NMS. In this chapter, I share how to easily tune the ONNX generated using TorchVision's redundant NMS to generate an optimized NMS.

  1. There are multiple issues with TorchVision's NMS. First, the batch size specification is not supported; second, the max_output_boxes_per_class parameter cannot be specified. Please see the NMS sample ONNX part I generated. The max_output_boxes_per_class has been changed to 896 instead of -Infinity. The biggest problem with TorchVision NMS is that it generates ONNX with max_output_boxes_per_class set to -Infinity or 9223372036854775807 (Maximum value of INT64), resulting in a variable number of NMS outputs from zero to infinite. Thus, by rewriting -Infinity or 9223372036854775807 (Maximum value of INT64) to a constant value, it is possible to output an NMS that can be effortlessly inferred by TFJS or TFLite. image

    Here you will find committed ONNX components optimized for various devices. https://github.com/PINTO0309/components_of_onnx/tree/main/components_of_onnx/ops

  2. In the following example, the max_output_boxes_per_class of NMS in the post-processing generated by YOLOv7 is changed from -Infinity or 9223372036854775807 (Maximum value of INT64) to 20, as shown in the figure below. The name main01_max_output_boxes_per_class has been rewritten by me for clarity, but it originally appears as max_output_boxes_per_class. image

    Simply execute the following command. The command rewrites the specified attribute value of the OP specified by ONNX.

    pip install sam4onnx
    
    sam4onnx \
    --op_name main01_nonmaxsuppression11 \
    --input_onnx_file_path yolov7.onnx \
    --output_onnx_file_path nms_yolov7_update.onnx \
    --input_constants main01_max_output_boxes_per_class int64 [20]

    A tutorial on one of my ONNX modification tools, sam4onnx, can be found here.

    https://github.com/PINTO0309/sam4onnx

    Many detailed tutorials are provided below, so if you are interested, please play with them.

    https://github.com/PINTO0309/PINTO_model_zoo/tree/main/307_YOLOv7/post_process_gen_tools

  3. Finally, simply convert ONNX to TFLite or saved_model or TFJS using onnx2tf. onnx2tf performs an internal operation to automatically optimize the NMS output to a fixed shape if max_output_boxes_per_class is set to a value other than -Infinity and 9223372036854775807 (Maximum value of INT64). Specify --output_nms_with_dynamic_tensor or -onwdt if you do not want to optimize for a fixed shape.

    onnx2tf -i nms_yolov7_update.onnx -osd -cotof
    

    I would be happy if this is a reference for Android + Java or TFJS implementations. There are tons more tricky model optimization techniques described in my blog posts, so you'll have to find them yourself. I don't dare to list the URL here because it is annoying to see so many issues being posted. And unfortunately, all articles are in Japanese. image

11. RNN (RNN, GRU, LSTM) Inference Acceleration

Click to expand

TensorFlow's RNN has a speedup option called unroll. The network will be unrolled, else a symbolic loop will be used. Unrolling can speed-up a RNN, although it tends to be more memory-intensive. Unrolling is only suitable for short sequences. onnx2tf allows you to deploy RNNs into memory-intensive operations by specifying the --enable_rnn_unroll or -eru options. The --enable_rnn_unroll option is available for RNN, GRU, and LSTM.

An example of BidirectionalLSTM conversion with the --enable_rnn_unroll option is shown below. Please ignore that the shapes of the input and output tensors do not match, since the samples are shown by picking up separate models.

  • ONNX LSTM (Bidirectional)

    image

  • BidirectionalLSTM with --enable_rnn_unroll option unspecified

    Recurrent layer is implemented from scratch.

    image

  • BidirectionalLSTM with --enable_rnn_unroll option

    image

12. If the accuracy of the Float32 model degrades significantly

Click to expand

The pattern of accuracy degradation of the converted model does not only occur when INT8 quantization is performed. A special edge case is when there is a problem with the implementation of a particular OP on the TFLite runtime side. Below, I will reproduce the problem by means of a very simple CNN model and further explain its workaround. Here is the issue that prompted me to add this explanation. [Conv-TasNet] Facing issue in converting Conv-TasNet model #447

Download a sample model for validation.

curl \
-L https://github.com/PINTO0309/onnx2tf/files/12367312/prelu_check.onnx.zip \
-o prelu_check.onnx.zip

unzip prelu_check.onnx.zip

The part of the downloaded model where the problem occurs is the PRelu part in the figure below.

  • ONNX

    image

Reproduce the problem. The following command converts an ONNX file to a TFLite file.

onnx2tf -i prelu_check.onnx -cotof

The conversion was successful and, as shown in the figure below, the inference test results from ONNX and the inference results for the Float32 model in TensorFlow (Keras) match perfectly. It is important to note that the comparison of inference results between ONNX and TensorFlow transformed models is comparing ONNX models with TensorFlow (Keras) models, not ONNX models with TFLite models.

  • Conversion results

    20230817175146

  • tflite

    20230817175530

Now, let's try inference with the TFLite runtime instead of the TensorFlow runtime.

  • test.py
    import time
    import numpy as np
    np.random.seed(0)
    import tensorflow as tf
    
    # Load TFLite model
    interpreter = tf.lite.Interpreter(model_path="./saved_model/prelu_check_float32.tflite")
    interpreter.allocate_tensors()
    tensor_shape = (256, 20)
    input_data = {'waveform': np.random.randn(*tensor_shape).astype(np.float32)}
    
    # Load and preprocess
    input_details = interpreter.get_input_details()
    input_shape = input_details[0]['shape']
    print(input_shape)
    
    # Run inference
    interpreter.set_tensor(input_details[0]['index'], input_data["waveform"])
    separate_time = time.time()
    interpreter.invoke()
    print("Done! {:.3f} s".format(time.time() - separate_time))
    output_details = interpreter.get_output_details()
    output_data = interpreter.get_tensor(output_details[0]['index'])
    
    output_data = []
    for output_detail in output_details:
        output_data.append(interpreter.get_tensor(output_detail['index']))
    
    print(output_data)

Oddly enough, the output value of PReLU contains multiple nan. However, as can be seen by converting the ONNX model to the middle of the model using the -onimc option, nan does not occur until just before PReLU. Thus, it is clear that the PReLU OP in the TFLite runtime has a problem with divergent inference results.

  • TFLite inference results

    20230817175454

The following is a work-around to avoid this problem. Use the -rtpo option to replace PReLU with a similar primitive operation when transforming a model, and then perform the model transformation.

onnx2tf -i prelu_check.onnx -cotof -rtpo PReLU

As before, the inference results from ONNX and TensorFlow (Keras) match perfectly.

  • Conversion results

    20230817175614

However, -rtpo PReLU will generate a .tflite file with the PRelu OP replaced by a primitive OP combination.

  • tflite

    20230817175724

Again, run the test code to check the inference results. The figure below shows that no nan occurs when inference is performed by replacing the PReLU OP with only combinations of primitive operations. In other words, it is important to know that large arithmetic errors are not only due to the broken structure of the model, but can also be caused by internal implementations such as the TFLite runtime. I have implemented the -rtpo option to replace operators as a work-around to avoid such runtime problems.

  • TFLite inference results

    20230817175701

13. Problem of extremely large calculation error in InstanceNormalization

Click to expand

Even if the conversion is successful, InstanceNormalization tends to have very large errors. This is an ONNX specification.

I verified this with a very simple sample model. There are more than 8 million elements, and the calculation error reached 1e-2.

image

image

14. Inference with dynamic tensors in TFLite

Click to expand

For some time now, TFLite runtime has supported inference by dynamic tensors. However, the existence of this important function is not widely recognized. In this chapter, I will show how I can convert an ONNX file that contains dynamic geometry in batch size directly into a TFLite file that contains dynamic geometry and then further infer it in variable batch conditions. The issue that inspired me to add this tutorial is here. [Dynamic batch / Dynamic shape] onnx model with dynamic input is converted to tflite with static input 1 #441, or Cannot use converted model with dynamic input shape #521

First, download the sample ONNX file.

wget https://s3.ap-northeast-2.wasabisys.com/temp-models/onnx2tf_441/osnet_x0_25_msmt17.onnx

This model calculates the similarity of features by cosine similarity. The batch size dimension of the input tensor is batch, allowing various numbers of images to be input simultaneously. This is often used, for example, to achieve tracking by calculating the similarity of people or objects reflected between successive video frames. However, the total number of objects to be tracked changes rapidly with each video frame because the number of people and objects in the image constantly increases and decreases. Therefore, there is a very significant use case for generating models with variable settings for the number of input images (batch size) of the model.

image

Convert the downloaded OSNet to tflite and saved_model as a variable batch. If you do not specify the -b or -ois options, onnx2tf does not change the batch size as N. The only important point is to convert the model with the -osd and -coion options. Note that if you use the -coion option, you must install flatbuffers-compiler with apt-get install, run the commands for building the environment described first in this README, or use a Docker container.

onnx2tf -i osnet_x0_25_msmt17.onnx -osd -coion
  • .tflite

    When viewing tflite in Netron, the batch size appears to be fixed at 1. image

  • saved_model

    However, checking the structure of saved_model, the batch size is correctly set to -1.

    saved_model_cli show --dir saved_model/ --all
    
    MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:
    
    signature_def['__saved_model_init_op']:
      The given SavedModel SignatureDef contains the following input(s):
      The given SavedModel SignatureDef contains the following output(s):
        outputs['__saved_model_init_op'] tensor_info:
            dtype: DT_INVALID
            shape: unknown_rank
            name: NoOp
      Method name is:
    
    signature_def['serving_default']:
      The given SavedModel SignatureDef contains the following input(s):
        inputs['images'] tensor_info:
            dtype: DT_FLOAT
            shape: (-1, 256, 128, 3)
            name: serving_default_images:0
      The given SavedModel SignatureDef contains the following output(s):
        outputs['output'] tensor_info:
            dtype: DT_FLOAT
            shape: (-1, 512)
            name: PartitionedCall:0
      Method name is: tensorflow/serving/predict

To prove that the tflite structure has been converted correctly, I will convert the tflite to JSON and look at the structure.

docker run --rm -it \
-v `pwd`:/home/user/workdir \
ghcr.io/pinto0309/tflite2json2tflite:latest

./flatc -t \
--strict-json \
--defaults-json \
-o workdir \
./schema.fbs -- workdir/saved_model/osnet_x0_25_msmt17_float32.tflite

ls -l workdir

-rw-rw-r-- 1 user user   921564 Aug  4 10:24 osnet_x0_25_msmt17.onnx
-rw-r--r-- 1 user user 10369524 Aug  4 10:30 osnet_x0_25_msmt17_float32.json
drwxrwxr-x 4 user user     4096 Aug  4 10:26 saved_model

image

  • osnet_x0_25_msmt17_float32.json

    "shape_signature" is correctly set to -1. However, "shape" is set to 1. This could be a problem with TFLiteConverter, or it could be a problem with Netron's graphical display capabilities. image

In other words, although onnx2tf converts TFLiteConverer as specified, with the batch size of -1 without any model processing, only Netron's display is broken. This is a problem I have known for quite some time. However, the inference itself does not cause the problem.

If you want to infer in variable batches, you need to infer using signature. In such cases, the -coion option must be specified when converting the model. Note that I have identified a problem with quantization with the -coion option, which can corrupt tflite files. #429

https://github.com/PINTO0309/onnx2tf#4-match-tflite-inputoutput-names-and-inputoutput-order-to-onnx

You can use signature_runner to handle dynamic input tensors by performing inference using signature. Below I show that both batch_size=5 and batch_size=3 tensors can be inferred with the same model.

  • test.py - Batch size: 5
    import numpy as np
    import tensorflow as tf
    from pprint import pprint
    
    interpreter = tf.lite.Interpreter(model_path="saved_model/osnet_x0_25_msmt17_float32.tflite")
    tf_lite_model = interpreter.get_signature_runner()
    inputs = {
        'images': np.ones([5,256,128,3], dtype=np.float32),
    }
    tf_lite_output = tf_lite_model(**inputs)
    print(f"[TFLite] Model Predictions shape: {tf_lite_output['output'].shape}")
    print(f"[TFLite] Model Predictions:")
    pprint(tf_lite_output)
  • Results
    [TFLite] Model Predictions shape: (5, 512)
    [TFLite] Model Predictions:
    {'output': array([[0.0000000e+00, 2.4730086e-04, 0.0000000e+00, ..., 1.0528549e+00,
            3.7874988e-01, 0.0000000e+00],
           [0.0000000e+00, 2.4730086e-04, 0.0000000e+00, ..., 1.0528549e+00,
            3.7874988e-01, 0.0000000e+00],
           [0.0000000e+00, 2.4730086e-04, 0.0000000e+00, ..., 1.0528549e+00,
            3.7874988e-01, 0.0000000e+00],
           [0.0000000e+00, 2.4730086e-04, 0.0000000e+00, ..., 1.0528549e+00,
            3.7874988e-01, 0.0000000e+00],
           [0.0000000e+00, 2.4730084e-04, 0.0000000e+00, ..., 1.0528525e+00,
            3.7874976e-01, 0.0000000e+00]], dtype=float32)}
    
  • test.py - Batch size: 3
    import numpy as np
    import tensorflow as tf
    from pprint import pprint
    
    interpreter = tf.lite.Interpreter(model_path="saved_model/osnet_x0_25_msmt17_float32.tflite")
    tf_lite_model = interpreter.get_signature_runner()
    inputs = {
        'images': np.ones([3,256,128,3], dtype=np.float32),
    }
    tf_lite_output = tf_lite_model(**inputs)
    print(f"[TFLite] Model Predictions shape: {tf_lite_output['output'].shape}")
    print(f"[TFLite] Model Predictions:")
    pprint(tf_lite_output)
  • Results
    [TFLite] Model Predictions shape: (3, 512)
    [TFLite] Model Predictions:
    {'output': array([[0.0000000e+00, 2.4730084e-04, 0.0000000e+00, ..., 1.0528525e+00,
            3.7874976e-01, 0.0000000e+00],
           [0.0000000e+00, 2.4730084e-04, 0.0000000e+00, ..., 1.0528525e+00,
            3.7874976e-01, 0.0000000e+00],
           [0.0000000e+00, 2.4730084e-04, 0.0000000e+00, ..., 1.0528525e+00,
            3.7874976e-01, 0.0000000e+00]], dtype=float32)}
    

15. Significant optimization of the entire model through Einsum and OneHot optimizations

Click to expand

Einsum and OneHot are not optimized to the maximum by the standard behavior of onnx-optimizer. Therefore, pre-optimizing the Einsum OP and OneHot OP using my original method can significantly improve the success rate of model conversion, and the input ONNX model itself can be significantly optimized compared to when onnxsim alone is optimized. See: #569

For example

python export.py \
--img_size 512 512 \
--lightglue_path weights/sjy_fused_static.onnx \
--end2end

pip install -U spo4onnx onnx2tf

cd weights
spo4onnx -if sjy_fused_static.onnx -of sjy_fused_static_spo.onnx

onnx2tf -i sjy_fused_static_spo.onnx

image

16. Add constant outputs to the model that are not connected to the model body

Click to expand

Sometimes you want to always output constants that are not connected to the model body. See: #627. For example, in the case of ONNX as shown in the figure below. You may want to keep scaling parameters and other parameters as fixed values inside the model and always include the same value in the output.

image

In such cases, the process of optimizing the ONNX file in onnxsim must be bypassed and not executed. You can bypass the execution of onnxsim by specifying -nuo or --not_use_onnxsim as a conversion option. Running onnxsim will remove constants from the model definition that are not connected to the body of the model in the process of optimizing the model structure.

wget https://github.com/PINTO0309/onnx2tf/files/15292126/toy_with_constant.onnx.zip
unzip toy_with_constant.onnx.zip
onnx2tf -i toy_with_constant.onnx -nuo -cotof

The relationship between the ONNX before conversion and the TFLite file after conversion is shown in the figure below.

ONNX TFLite
image image

image

Use the generated TFLite file to inference and ensure that it always contains fixed value output.

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)

17. Conversion of models that use variable length tokens and embedding, such as LLM and sound models

Click to expand

This refers to a model with undefined dimensions, either all dimensions or multiple dimensions including batch size, as shown in the figure below.

If such a model is converted without any options, TensorFlow/Keras will abort. This is an internal TensorFlow/Keras implementation issue rather than an onnx2tf issue. TensorFlow/Keras does not allow more than two undefined dimensions in the shape attribute of Reshape due to the specification, so an error occurs during the internal transformation operation of the Reshape OP as shown below. This has been an inherent problem in TensorFlow/Keras since long ago and has not been resolved to this day. See: RuntimeError: tensorflow/lite/kernels/range.cc:39 (start > limit && delta < 0) || (start < limit && delta > 0) was not true.Node number 3 (RANGE) failed to invoke. Node number 393 (WHILE) failed to invoke. current error :RuntimeError: tensorflow/lite/kernels/reshape.cc:55 stretch_dim != -1 (0 != -1)Node number 83 (RESHAPE) failed to prepare. #40504

  • OP where the problem occurs

    image

  • Error message

    error: 'tf.Reshape' op requires 'shape' to have at most one dynamic dimension, but got multiple dynamic dimensions at indices 0 and 3
    

Thus, for models such as this, where all dimensions, including batch size, are dynamic shapes, it is often possible to convert by fixing the batch size to 1 with the -b 1 or --batch_size 1 option.

onnx2tf -i model.onnx -b 1 -osd
  • Results

    image

    When the converted tflite is displayed in Netron, all the dimensions of the dynamic shape are displayed as 1, but this is a display problem in Netron, and the shape is actually converted to -1 or None.

    image

Click here to see how to perform inference using the dynamic shape tensor.

https://github.com/PINTO0309/onnx2tf/tree/main?tab=readme-ov-file#14-inference-with-dynamic-tensors-in-tflite

18. Convert only the intermediate structural part of the ONNX model

Click to expand

By specifying ONNX input or output names, only the middle part of the model can be converted. This is useful when you want to see what output is obtained in what part of the model after conversion, or when debugging the model conversion operation itself.

For example, take a model with multiple inputs and multiple outputs as shown in the figure below to try a partial transformation.

image

  • To convert by specifying only the input name to start the conversion

    wget https://github.com/PINTO0309/onnx2tf/releases/download/1.25.0/cf_fus.onnx
    onnx2tf -i cf_fus.onnx -inimc 448 -coion
    

    image

  • To convert by specifying only the output name to end the conversion

    wget https://github.com/PINTO0309/onnx2tf/releases/download/1.25.0/cf_fus.onnx
    onnx2tf -i cf_fus.onnx -onimc dep_sec -coion
    

    image

  • To perform a conversion by specifying the input name to start the conversion and the output name to end the conversion

    wget https://github.com/PINTO0309/onnx2tf/releases/download/1.25.0/cf_fus.onnx
    onnx2tf -i cf_fus.onnx -inimc 448 -onimc velocity -coion
    

    image

19. Conversion to TensorFlow.js

Click to expand

When converting to TensorFlow.js, process as follows.

pip install -U --no-deps \
tensorflowjs \
tensorflow_decision_forests \
ydf \
tensorflow_hub

onnx2tf -i mobilenetv2-12.onnx -ois input:1,3,224,224 -osd -dgc

tensorflowjs_converter \
--input_format tf_saved_model \
--output_format tfjs_graph_model \
saved_model \
tfjs_model

See: https://github.com/tensorflow/tfjs/tree/master/tfjs-converter

image

20. Conversion to CoreML

Click to expand

When converting to CoreML, process as follows. The -k option is for conversion while maintaining the input channel order in ONNX's NCHW format.

pip install coremltools

onnx2tf -i mobilenetv2-12.onnx -k input -ois input:1,3,224,224 -osd
import coremltools as ct

FOLDER_PATH = 'saved_model'

model = ct.convert(
    model=FOLDER_PATH,
    source='tensorflow',
)
model.save(f'{FOLDER_PATH}/model.mlmodel')

See: https://github.com/apple/coremltools

image

CLI Parameter

Click to expand
onnx2tf -h

usage: onnx2tf
[-h]
(-i INPUT_ONNX_FILE_PATH | -V)
[-o OUTPUT_FOLDER_PATH]
[-osd]
[-oh5]
[-okv3]
[-otfv1pb]
[-ow]
[-coion]
[-odrqt]
[-oiqt]
[-qt {per-channel,per-tensor}]
[-cind INPUT_NAME NUMPY_FILE_PATH MEAN STD]
[-iqd {int8,uint8,float32}]
[-oqd {int8,uint8,float32}]
[-nuo]
[-nuonag]
[-b BATCH_SIZE]
[-ois OVERWRITE_INPUT_SHAPE [OVERWRITE_INPUT_SHAPE ...]]
[-nlt]
[-onwdt]
[-k KEEP_NCW_OR_NCHW_OR_NCDHW_INPUT_NAMES [KEEP_NCW_OR_NCHW_OR_NCDHW_INPUT_NAMES ...]]
[-kt KEEP_NWC_OR_NHWC_OR_NDHWC_INPUT_NAMES [KEEP_NWC_OR_NHWC_OR_NDHWC_INPUT_NAMES ...]]
[-kat KEEP_SHAPE_ABSOLUTELY_INPUT_NAMES [KEEP_SHAPE_ABSOLUTELY_INPUT_NAMES ...]]
[-inimc INPUT_NAMES [INPUT_NAMES ...]]
[-onimc OUTPUT_NAMES [OUTPUT_NAMES ...]]
[-dgc]
[-eatfp16]
[-ebu]
[-eru]
[-dsft]
[-nodaftc]
[-dsfs]
[-dsm]
[-nodafsc]
[-ofgd]
[-rari64 | -rarf32 | -rafi64 | -raff32]
[-fasr FUSED_ARGMAX_SCALE_RATIO]
[-rtpo REPLACE_TO_PSEUDO_OPERATORS [REPLACE_TO_PSEUDO_OPERATORS ...]]
[-me MVN_EPSILON]
[-prf PARAM_REPLACEMENT_FILE]
[-cgdc]
[-coto | -cotof]
[-coton]
[-cotor CHECK_ONNX_TF_OUTPUTS_ELEMENTWISE_CLOSE_RTOL]
[-cotoa CHECK_ONNX_TF_OUTPUTS_ELEMENTWISE_CLOSE_ATOL]
[-dms]
[-uc]
[-n]
[-v]

optional arguments:
  -h, --help
    show this help message and exit

  -i INPUT_ONNX_FILE_PATH, --input_onnx_file_path INPUT_ONNX_FILE_PATH
    Input onnx file path.

  -V, --version
    Show version and exit.

  -o OUTPUT_FOLDER_PATH, --output_folder_path OUTPUT_FOLDER_PATH
    Output folder path. Default: "saved_model"

  -osd, --output_signaturedefs
    Signature is added to the output for serving or for conversion
    to other model formats. However, this can significantly reduce the speed
    of model conversion and significant increase the size of the model.

  -oh5, --output_h5
    Output model in Keras (hdf5) format.

  -okv3, --output_keras_v3
    Output model in Keras (keras_v3) format.

  -otfv1pb, --output_tfv1_pb
    Output model in TF v1 (.pb) format.

  -ow, --output_weights
    Output weights in hdf5 format.

  -coion, --copy_onnx_input_output_names_to_tflite
    Copy the input/output OP name of ONNX to the input/output OP name of tflite.
    Due to Tensorflow internal operating specifications,
    the input/output order of ONNX does not necessarily match
    the input/output order of tflite.
    Be sure to check that the input/output OP names in the generated
    tflite file have been converted as expected.
    Also, this option generates a huge JSON file as a temporary file for processing.
    Therefore, it is strongly discouraged to use it on large models of hundreds
    of megabytes or more.

  -odrqt, --output_dynamic_range_quantized_tflite
    Output of dynamic range quantized tflite.

  -oiqt, --output_integer_quantized_tflite
    Output of integer quantized tflite.

  -qt {per-channel,per-tensor}, --quant_type {per-channel,per-tensor}
    Selects whether "per-channel" or "per-tensor" quantization is used.
    Default: "per-channel"

  -cind INPUT_NAME NUMPY_FILE_PATH MEAN STD, \
    --custom_input_op_name_np_data_path INPUT_NAME NUMPY_FILE_PATH MEAN STD
    Input name of OP and path of data file (Numpy) for custom input for -cotof or -oiqt,
    and mean (optional) and std (optional).

    <Usage in -cotof>
      When using -cotof, custom input defined by the user, instead of dummy data, is used.
      In this case, mean and std are omitted from the input.
      -cind {input_op_name} {numpy_file_path}
      e.g. -cind onnx::Equal_0 test_cind/x_1.npy -cind onnx::Add_1 test_cind/x_2.npy -cotof
      The input_op_name must be the same as in ONNX,
      and it may not work if the input format is different between ONNX and TF.

    <Usage in -oiqt>
      INPUT Name of OP and path of calibration data file (Numpy) for quantization
      and mean and std.
      The specification can be omitted only when the input OP is a single 4D tensor image data.
      If omitted, it is automatically calibrated using 20 normalized MS-COCO images.
      The type of the input OP must be Float32.
      Data for calibration must be pre-normalized to a range of 0 to 1.
      -cind {input_op_name} {numpy_file_path} {mean} {std}
      Numpy file paths must be specified the same number of times as the number of input OPs.
      Normalize the value of the input OP based on the tensor specified in mean and std.
      (input_value - mean) / std
      Tensors in Numpy file format must be in dimension order after conversion to TF.
      Note that this is intended for deployment on low-resource devices,
      so the batch size is limited to 1 only.

      e.g.
      The example below shows a case where there are three input OPs.
      Assume input0 is 128x128 RGB image data.
      In addition, input0 should be a value that has been divided by 255
      in the preprocessing and normalized to a range between 0 and 1.
      input1 and input2 assume the input of something that is not an image.
      Because input1 and input2 assume something that is not an image,
      the divisor is not 255 when normalizing from 0 to 1.
      "n" is the number of calibration data.

      ONNX INPUT shapes:
        input0: [n,3,128,128]
            mean: [1,3,1,1] -> [[[[0.485]],[[0.456]],[[0.406]]]]
            std:  [1,3,1,1] -> [[[[0.229]],[[0.224]],[[0.225]]]]
        input1: [n,64,64]
            mean: [1,64] -> [0.1, ..., 0.64]
            std:  [1,64] -> [0.05, ..., 0.08]
        input2: [n,5]
            mean: [1] -> [0.3]
            std:  [1] -> [0.07]
      TensorFlow INPUT shapes (Numpy file ndarray shapes):
        input0: [n,128,128,3]
            mean: [1,1,1,3] -> [[[[0.485, 0.456, 0.406]]]]
            std:  [1,1,1,3] -> [[[[0.229, 0.224, 0.225]]]]
        input1: [n,64,64]
            mean: [1,64] -> [0.1, ..., 0.64]
            std:  [1,64] -> [0.05, ..., 0.08]
        input2: [n,5]
            mean: [1] -> [0.3]
            std:  [1] -> [0.07]
      -cind "input0" "../input0.npy" "[[[[0.485,0.456,0.406]]]]" "[[[[0.229,0.224,0.225]]]]"
      -cind "input1" "./input1.npy" "[0.1,...,0.64]" "[0.05,...,0.08]"
      -cind "input2" "input2.npy" "[0.3]" "[0.07]"

    <Using -cotof and -oiqt at the same time>
      To use -cotof and -oiqt simultaneously,
      you need to enter the Input name of OP, path of data file, mean, and std all together.
      And the data file must be in Float32 format,
      and {input_op_name}, {numpy_file_path}, {mean}, and {std} must all be entered.
      Otherwise, an error will occur during the -oiqt stage.

  -iqd {int8,uint8,float32}, --input_quant_dtype {int8,uint8,float32}
    Input dtypes when doing Full INT8 Quantization.
    "int8"(default) or "uint8" or "float32"

  -oqd {int8,uint8,float32}, --output_quant_dtype {int8,uint8,float32}
    Output dtypes when doing Full INT8 Quantization.
    "int8"(default) or "uint8" or "float32"

  -nuo, --not_use_onnxsim
    No optimization by onnx-simplifier is performed.
    If this option is used, the probability of a conversion error is very high.

  -nuonag, --not_use_opname_auto_generate
    Automatic generation of each OP name in the old format ONNX file
    and assignment of OP name are not performed.

  -b BATCH_SIZE, --batch_size BATCH_SIZE
    Fixes the dynamic batch size to the specified numeric batch size.
    A value of 1 or more must be specified.

  -ois OVERWRITE_INPUT_SHAPE [OVERWRITE_INPUT_SHAPE ...], \
      --overwrite_input_shape OVERWRITE_INPUT_SHAPE [OVERWRITE_INPUT_SHAPE ...]
    Overwrite the input shape.
    The format is
    "i1:dim0,...,dimN" "i2:dim0,...,dimN" "i3:dim0,...,dimN"
    When there is only one input, for example,
    "data:1,3,224,224"
    When there are multiple inputs, for example,
    "data1:1,3,224,224" "data2:1,3,112" "data3:5"
    A value of 1 or more must be specified.
    Numerical values other than dynamic dimensions are ignored.
    Ignores --batch_size if specified at the same time as --batch_size.

  -nlt, --no_large_tensor
    Suppresses constant bloat caused by Tile OP when optimizing models in onnxsim.
    See: https://github.com/daquexian/onnx-simplifier/issues/178

  -onwdt, --output_nms_with_dynamic_tensor
    The number of bounding boxes in the NMS output results is
    not fixed at the maximum number of max_output_boxes_per_class,
    but rather at the smallest possible number of dynamic tensors.
    If this option is disabled, NMS output is padded to the number
    set in the max_output_boxes_per_class attribute.
    e.g.
    disable --output_nms_with_dynamic_tensor:
        output_tensor_shape: [100, 7]
    enable --output_nms_with_dynamic_tensor:
        output_tensor_shape: [N, 7]

  -k KEEP_NCW_OR_NCHW_OR_NCDHW_INPUT_NAMES [KEEP_NCW_OR_NCHW_OR_NCDHW_INPUT_NAMES ...], \
      --keep_ncw_or_nchw_or_ncdhw_input_names KEEP_NCW_OR_NCHW_OR_NCDHW_INPUT_NAMES \
          [KEEP_NCW_OR_NCHW_OR_NCDHW_INPUT_NAMES ...]
    Holds the NCW or NCHW or NCDHW of the input shape for the specified INPUT OP names.
    If a nonexistent INPUT OP name is specified, it is ignored.
    Valid only for 3D, 4D and 5D input tensors.
    e.g. --keep_ncw_or_nchw_or_ncdhw_input_names "input0" "input1" "input2"

  -kt KEEP_NWC_OR_NHWC_OR_NDHWC_INPUT_NAMES [KEEP_NWC_OR_NHWC_OR_NDHWC_INPUT_NAMES ...], \
      --keep_nwc_or_nhwc_or_ndhwc_input_names KEEP_NWC_OR_NHWC_OR_NDHWC_INPUT_NAMES \
          [KEEP_NWC_OR_NHWC_OR_NDHWC_INPUT_NAMES ...]
    Holds the NWC or NHWC or NDHWC of the input shape for the specified INPUT OP names.
    If a nonexistent INPUT OP name is specified, it is ignored.
    If the input OP name is the same as the input OP name specified
    in the keep_ncw_or_nchw_or_ncdhw_input_names option, it is ignored.
    Valid only for 3D, 4D and 5D input tensors.
    e.g. --keep_nwc_or_nhwc_or_ndhwc_input_names "input0" "input1" "input2"

  -kat KEEP_SHAPE_ABSOLUTELY_INPUT_NAMES [KEEP_SHAPE_ABSOLUTELY_INPUT_NAMES ...], \
      --keep_shape_absolutely_input_names KEEP_SHAPE_ABSOLUTELY_INPUT_NAMES \
        [KEEP_SHAPE_ABSOLUTELY_INPUT_NAMES ...]
    Name of the INPUT that unconditionally maintains its shape.
    If a nonexistent INPUT OP name is specified, it is ignored.
    e.g. --keep_shape_absolutely_input_names "input0" "input1" "input2"

  -inimc INPUT_NAMES [INPUT_NAMES ...], \
      --input_names_to_interrupt_model_conversion INPUT_NAMES [INPUT_NAMES ...]
    Input names of ONNX that interrupt model conversion.
    Interrupts model transformation at the specified input name and inputs the
    model partitioned into subgraphs.
    e.g. --input_names_to_interrupt_model_conversion "input0" "input1" "input2"

  -onimc OUTPUT_NAMES [OUTPUT_NAMES ...], \
      --output_names_to_interrupt_model_conversion OUTPUT_NAMES [OUTPUT_NAMES ...]
    Output names of ONNX that interrupt model conversion.
    Interrupts model transformation at the specified output name and outputs the
    model partitioned into subgraphs.
    e.g. --output_names_to_interrupt_model_conversion "output0" "output1" "output2"

  -dgc, --disable_group_convolution
    Disable GroupConvolution and replace it with SeparableConvolution for
    output to saved_model format.

  -eatfp16, --enable_accumulation_type_float16 ENABLE_ACCUMULATION_TYPE_FLOAT16
    Hint for XNNPACK fp16 inference on float16 tflite model.
    XNNPACK float16 inference on certain ARM64 cores is 2x faster.
    Float16 inference doubling on devices with ARM64 ARMv8.2 or higher instruction set.
    See: https://github.com/PINTO0309/onnx2tf/pull/553

  -ebu, --enable_batchmatmul_unfold
    BatchMatMul is separated batch by batch to generate a primitive MatMul.

  -eru, --enable_rnn_unroll
    Instead of increasing inference speed by expanding all symbolic loops of
    the RNN (LSTM, GRU, RNN), RAM consumption will increase because all tensors
    are expanded and embedded in the model.
    https://keras.io/api/layers/recurrent_layers/

  -dsft, --disable_suppression_flextranspose
    Disables FlexTranspose generation suppression.

  -nodaftc, --number_of_dimensions_after_flextranspose_compression
    Number of Transpose OP dimensions generated after avoiding FlexTranspose generation.
    Also suppress the creation of the Transpose itself by specifying 2.
    Default: 6

  -dsfs, --disable_suppression_flexstridedslice
    Disables FlexStridedSlice generation suppression.

  -dsm, --disable_strict_mode
    If specified, the conversion speed is greatly accelerated because the strict accuracy
    correction process is skipped, but the frequency of transposition errors increases
    and accuracy errors are more likely to occur. Strict mode is enabled by default.
    As of 2023.05.07, this is a work in progress and is an experimental feature.
    Therefore, only some OPs are converted in strict mode for accuracy correction.

  -nodafsc, --number_of_dimensions_after_flexstridedslice_compression
    Number of StridedSlice OP dimensions generated after avoiding FlexStridedSlice generation.
    Default: 5

  -ofgd, --optimization_for_gpu_delegate
    Replace operations that do not support gpu delegate with those
    that do as much as possible.

  -rari64, --replace_argmax_to_reducemax_and_indices_is_int64
    Replace ArgMax with a ReduceMax. The returned indices are int64.
    Only one of replace_argmax_to_reducemax_and_indices_is_int64
    and replace_argmax_to_reducemax_and_indices_is_float32
    and replace_argmax_to_fused_argmax_and_indices_is_int64
    and replace_argmax_to_fused_argmax_and_indices_is_float32 can be specified.

  -rarf32, --replace_argmax_to_reducemax_and_indices_is_float32
    Replace ArgMax with a ReduceMax. The returned indices are float32.
    Only one of replace_argmax_to_reducemax_and_indices_is_int64
    and replace_argmax_to_reducemax_and_indices_is_float32
    and replace_argmax_to_fused_argmax_and_indices_is_int64
    and replace_argmax_to_fused_argmax_and_indices_is_float32 can be specified.

  -rafi64, --replace_argmax_to_fused_argmax_and_indices_is_int64
    Replace ArgMax with a Fused_ArgMax. The returned indices are int64.
    It improves inference speed at the cost of a small sacrifice in accuracy.
    See. https://github.com/tensorflow/models/tree/master/official/projects/edgetpu/vision#argmax-fusion-to-improve-segmentation-model-latency
    Currently, only 4D tensors are supported.
    Only one of replace_argmax_to_reducemax_and_indices_is_int64
    and replace_argmax_to_reducemax_and_indices_is_float32
    and replace_argmax_to_fused_argmax_and_indices_is_int64
    and replace_argmax_to_fused_argmax_and_indices_is_float32 can be specified.

  -raff32, --replace_argmax_to_fused_argmax_and_indices_is_float32
    Replace ArgMax with a Fused_ArgMax. The returned indices are float32.
    It improves inference speed at the cost of a small sacrifice in accuracy.
    See. https://github.com/tensorflow/models/tree/master/official/projects/edgetpu/vision#argmax-fusion-to-improve-segmentation-model-latency
    Currently, only 4D tensors are supported.
    Only one of replace_argmax_to_reducemax_and_indices_is_int64
    and replace_argmax_to_reducemax_and_indices_is_float32
    and replace_argmax_to_fused_argmax_and_indices_is_int64
    and replace_argmax_to_fused_argmax_and_indices_is_float32 can be specified.

  -fasr FUSED_ARGMAX_SCALE_RATIO, --fused_argmax_scale_ratio FUSED_ARGMAX_SCALE_RATIO
    For Fused ArgMax.
    Scale ratio when generating Fused ArgMax.
    0.0 < fused_argmax_scale_ratio <= 1.0
    Default: 0.5

  -rtpo, --replace_to_pseudo_operators
    Replace list of operators to pseudo operators.
    Full name of the target operators should be given.
    Currently supported operators :
    Asin, Acos, Atan, Abs, PReLU, LeakyReLU, Power, GatherND, Neg, HardSwish, Erf, GeLU, MatMulInteger

  -me, --mvn_epsilon
    For MeanVarianceNormalization.
    The number to be added to the variance to avoid division by zero
    when normalizing the value.
    (input_tensor - mean) / tf.sqrt(variance + mvn_epsilon)
    Default: 0.0000000001

  -prf PARAM_REPLACEMENT_FILE, --param_replacement_file PARAM_REPLACEMENT_FILE
    Parameter replacement file path. (.json)

  -cgdc, --check_gpu_delegate_compatibility
    Run TFLite ModelAnalyzer on the generated Float16 tflite model
    to check if the model can be supported by GPU Delegate.
    e.g.
    """
    === TFLite ModelAnalyzer ===

    Your TFLite model has '1' subgraph(s). In the subgraph description below,
    T# represents the Tensor numbers. For example, in Subgraph#0, the RESHAPE op takes
    tensor #0 and tensor #6 as input and produces tensor #7 as output.

    Subgraph#0 main(T#0) -> [T#17]
      Op#0 RESHAPE(T#0, T#6[2, 8, 8, 3, 2, ...]) -> [T#7]
      Op#1 SPLIT(T#5[0], T#7) -> [T#8, T#9]
      Op#2 RESHAPE(T#8, T#1[8, 8, 3, 2, 2]) -> [T#10]
      Op#3 TRANSPOSE(T#10, T#4[0, 3, 1, 4, 2]) -> [T#11]
      Op#4 RESHAPE(T#11, T#2[1, 8, 2, 8, 2, ...]) -> [T#12]
      Op#5 RESHAPE(T#9, T#1[8, 8, 3, 2, 2]) -> [T#13]
      Op#6 TRANSPOSE(T#13, T#4[0, 3, 1, 4, 2]) -> [T#14]
      Op#7 RESHAPE(T#14, T#2[1, 8, 2, 8, 2, ...]) -> [T#15]
      Op#8 CONCATENATION(T#12, T#15) -> [T#16]
      Op#9 RESHAPE(T#16, T#3[2, 16, 16, 3]) -> [T#17]

    Tensors of Subgraph#0
      T#0(inputs_0) shape:[2, 8, 8, 12], type:FLOAT32
      T#1(model/tf.compat.v1.squeeze_2/Squeeze) shape:[5], type:INT32 RO 20 bytes, data:[8, 8, 3, 2, 2]
      T#2(model/tf.expand_dims_1/ExpandDims) shape:[6], type:INT32 RO 24 bytes, data:[1, 8, 2, 8, 2, ...]
      T#3(model/tf.reshape_1/Reshape/shape) shape:[4], type:INT32 RO 16 bytes, data:[2, 16, 16, 3]
      T#4(model/tf.compat.v1.transpose/transpose/perm) shape:[5], type:INT32 RO 20 bytes, data:[0, 3, 1, 4, 2]
      T#5(model/tf.concat/concat/axis) shape:[], type:INT32 RO 4 bytes, data:[0]
      T#6(model/tf.reshape/Reshape/shape) shape:[6], type:INT32 RO 24 bytes, data:[2, 8, 8, 3, 2, ...]
      T#7(model/tf.reshape/Reshape) shape:[2, 8, 8, 3, 2, 2], type:FLOAT32
      T#8(model/tf.split/split) shape:[1, 8, 8, 3, 2, 2], type:FLOAT32
      T#9(model/tf.split/split1) shape:[1, 8, 8, 3, 2, 2], type:FLOAT32
      T#10(model/tf.compat.v1.squeeze_1/Squeeze) shape:[8, 8, 3, 2, 2], type:FLOAT32
      T#11(model/tf.compat.v1.transpose/transpose) shape:[8, 2, 8, 2, 3], type:FLOAT32
      T#12(model/tf.expand_dims/ExpandDims) shape:[1, 8, 2, 8, 2, 3], type:FLOAT32
      T#13(model/tf.compat.v1.squeeze_2/Squeeze1) shape:[8, 8, 3, 2, 2], type:FLOAT32
      T#14(model/tf.compat.v1.transpose_1/transpose) shape:[8, 2, 8, 2, 3], type:FLOAT32
      T#15(model/tf.expand_dims_1/ExpandDims1) shape:[1, 8, 2, 8, 2, 3], type:FLOAT32
      T#16(model/tf.concat/concat) shape:[2, 8, 2, 8, 2, 3], type:FLOAT32
      T#17(Identity) shape:[2, 16, 16, 3], type:FLOAT32

    Your model looks compatibile with GPU delegate with TFLite runtime version 2.10.0.
    But it doesn't guarantee that your model works well with GPU delegate.
    There could be some runtime incompatibililty happen.
    ---------------------------------------------------------------
                  Model size:       2988 bytes
        Non-data buffer size:       2757 bytes (92.27 %)
      Total data buffer size:        231 bytes (07.73 %)
        (Zero value buffers):          4 bytes (00.13 %)

    * Buffers of TFLite model are mostly used for constant tensors.
      And zero value buffers are buffers filled with zeros.
      Non-data buffers area are used to store operators, subgraphs and etc.
      You can find more details from https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/schema/schema.fbs
    """

  -coto, --check_onnx_tf_outputs_elementwise_close
    Returns "Matches" if the output of onnx and the output of TF are
    within acceptable proximity element by element.
    Returns "Unmatched" if the output of onnx and the output of TF are
    not within acceptable proximity element by element.
    If the output of onnx is 1D, it returns "Skipped" and skips the comparison
    between the output of onnx and that of TF. This is because when undefined
    dimensions are present, a situation often arises where very large index
    values are compared, causing OutOfMemory.
    Only the output content of the models final output OP is checked.

  -cotof, --check_onnx_tf_outputs_elementwise_close_full
    Returns "Matches" if the output of onnx and the output of TF are
    within acceptable proximity element by element.
    Check the output of all OPs in sequence from the beginning,
    including all but the final output OP of the model.
    Returns "Unmatched" if the output of onnx and the output of TF are
    not within acceptable proximity element by element.
    If the output of onnx is 1D, it returns "Skipped" and skips the comparison
    between the output of onnx and that of TF. This is because when undefined
    dimensions are present, a situation often arises where very large index
    values are compared, causing OutOfMemory.
    It is very time consuming because it performs as many inferences as
    there are operations.

  -coton, --check_onnx_tf_outputs_sample_data_normalization
    norm: Validate using random data normalized to the range 0.0 to 1.0
    denorm: Validate using random data in the range 0.0 to 255.0
    If there is a normalization layer at the model's entry point, or
    if the model was trained on denormalized data, "denorm" must be specified.
    Default: "norm"

  -cotor CHECK_ONNX_TF_OUTPUTS_ELEMENTWISE_CLOSE_RTOL,\
    --check_onnx_tf_outputs_elementwise_close_rtol CHECK_ONNX_TF_OUTPUTS_ELEMENTWISE_CLOSE_RTOL
    The relative tolerance parameter.
    Default: 0.0

  -cotoa CHECK_ONNX_TF_OUTPUTS_ELEMENTWISE_CLOSE_ATOL,\
    --check_onnx_tf_outputs_elementwise_close_atol CHECK_ONNX_TF_OUTPUTS_ELEMENTWISE_CLOSE_ATOL
    The absolute tolerance parameter.
    Default: 1e-4

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

  -n, --non_verbose
    Shorthand to specify a verbosity of "error".

  -v, --verbosity
    Change the level of information printed.
    Values are "debug", "info", "warn", and "error".
    Default: "debug" (for backwards compatability)

In-script Usage

Click to expand
>>> from onnx2tf import convert
>>> help(convert)

Help on function convert in module onnx2tf:

convert(
  input_onnx_file_path: Union[str, NoneType] = '',
  onnx_graph: Union[onnx.onnx_ml_pb2.ModelProto, NoneType] = None,
  output_folder_path: Union[str, NoneType] = 'saved_model',
  output_signaturedefs: Optional[bool] = False,
  output_h5: Optional[bool] = False,
  output_keras_v3: Optional[bool] = False,
  output_tfv1_pb: Optional[bool] = False,
  output_weights: Optional[bool] = False,
  copy_onnx_input_output_names_to_tflite: Optional[bool] = False,
  output_integer_quantized_tflite: Optional[bool] = False,
  quant_type: Optional[str] = 'per-channel',
  custom_input_op_name_np_data_path: Optional[List] = None,
  input_quant_dtype: Optional[str] = 'int8',
  output_quant_dtype: Optional[str] = 'int8',
  not_use_onnxsim: Optional[bool] = False,
  not_use_opname_auto_generate: Optional[bool] = False,
  batch_size: Union[int, NoneType] = None,
  overwrite_input_shape: Union[List[str], NoneType] = None,
  no_large_tensor: Optional[bool] = False,
  output_nms_with_dynamic_tensor: Optional[bool] = False,
  keep_ncw_or_nchw_or_ncdhw_input_names: Union[List[str], NoneType] = None,
  keep_nwc_or_nhwc_or_ndhwc_input_names: Union[List[str], NoneType] = None,
  keep_shape_absolutely_input_names: Optional[List[str]] = None,
  input_names_to_interrupt_model_conversion: Union[List[str], NoneType] = None,
  output_names_to_interrupt_model_conversion: Union[List[str], NoneType] = None,
  disable_group_convolution: Union[bool, NoneType] = False,
  enable_batchmatmul_unfold: Optional[bool] = False,
  enable_rnn_unroll: Optional[bool] = False,
  disable_suppression_flextranspose: Optional[bool] = False,
  number_of_dimensions_after_flextranspose_compression: Optional[int] = 6,
  disable_suppression_flexstridedslice: Optional[bool] = False,
  disable_strict_mode: Optional[bool] = False,
  number_of_dimensions_after_flexstridedslice_compression: Optional[int] = 5,
  optimization_for_gpu_delegate: Optional[bool] = False,
  replace_argmax_to_reducemax_and_indices_is_int64: Union[bool, NoneType] = False,
  replace_argmax_to_reducemax_and_indices_is_float32: Union[bool, NoneType] = False,
  replace_argmax_to_fused_argmax_and_indices_is_int64: Union[bool, NoneType] = False,
  replace_argmax_to_fused_argmax_and_indices_is_float32: Union[bool, NoneType] = False,
  fused_argmax_scale_ratio: Union[float, NoneType] = 0.5,
  replace_to_pseudo_operators: List[str] = None,
  mvn_epsilon: Union[float, NoneType] = 0.0000000001,
  param_replacement_file: Optional[str] = '',
  check_gpu_delegate_compatibility: Optional[bool] = False,
  check_onnx_tf_outputs_elementwise_close: Optional[bool] = False,
  check_onnx_tf_outputs_elementwise_close_full: Optional[bool] = False,
  check_onnx_tf_outputs_sample_data_normalization: Optional[str] = 'norm',
  check_onnx_tf_outputs_elementwise_close_rtol: Optional[float] = 0.0,
  check_onnx_tf_outputs_elementwise_close_atol: Optional[float] = 1e-4,
  disable_model_save: Union[bool, NoneType] = False,
  non_verbose: Union[bool, NoneType] = False,
  verbosity: Optional[str] = 'debug'
) -> keras.engine.training.Model

    Convert ONNX to TensorFlow models.

    Parameters
    ----------
    input_onnx_file_path: Optional[str]
      Input onnx file path.
      Either input_onnx_file_path or onnx_graph must be specified.

    onnx_graph: Optional[onnx.ModelProto]
      onnx.ModelProto.
      Either input_onnx_file_path or onnx_graph must be specified.
      onnx_graph If specified, ignore input_onnx_file_path and process onnx_graph.

    output_folder_path: Optional[str]
      Output tensorflow model folder path.
      Default: "saved_model"

    output_signaturedefs: Optional[bool]
      Signature is added to the output for serving or for conversion
      to other model formats. However, this can significantly reduce the speed
      of model conversion and significant increase the size of the model.

    output_h5: Optional[bool]
      Output model in Keras H5 format.

    output_keras_v3: Optional[bool]
      Output model in Keras (keras_v3) format.

    output_tfv1_pb: Optional[bool]
      Output model in TF v1 (.pb) format.

    output_weights: Optional[bool]
      Output weights in hdf5 format.

    copy_onnx_input_output_names_to_tflite: Optional[bool]
      Copy the input/output OP name of ONNX to the input/output OP name of tflite.
      Due to Tensorflow internal operating specifications,
      the input/output order of ONNX does not necessarily match
      the input/output order of tflite.
      Be sure to check that the input/output OP names in the generated
      tflite file have been converted as expected.
      Also, this option generates a huge JSON file as a temporary file for processing.
      Therefore, it is strongly discouraged to use it on large models of hundreds
      of megabytes or more.

    output_integer_quantized_tflite: Optional[bool]
      Output of integer quantized tflite.

    quant_type: Optional[str]
      Selects whether "per-channel" or "per-tensor" quantization is used.
      Default: "per-channel"

    custom_input_op_name_np_data_path: Optional[List]
      --custom_input_op_name_np_data_path INPUT_NAME NUMPY_FILE_PATH MEAN STD
      Input name of OP and path of data file (Numpy) for custom input for -cotof or -oiqt,
      and mean (optional) and std (optional).

      <Usage in -cotof>
        When using -cotof, custom input defined by the user, instead of dummy data, is used.
        In this case, mean and std are omitted from the input.
        -cind {input_op_name} {numpy_file_path}
        e.g. -cind onnx::Equal_0 test_cind/x_1.npy -cind onnx::Add_1 test_cind/x_2.npy -cotof
        The input_op_name must be the same as in ONNX,
        and it may not work if the input format is different between ONNX and TF.

      <Usage in -oiqt>
        INPUT Name of OP and path of calibration data file (Numpy) for quantization
        and mean and std.
        The specification can be omitted only when the input OP is a single 4D tensor image data.
        If omitted, it is automatically calibrated using 20 normalized MS-COCO images.
        The type of the input OP must be Float32.
        Data for calibration must be pre-normalized to a range of 0 to 1.
        -cind {input_op_name} {numpy_file_path} {mean} {std}
        Numpy file paths must be specified the same number of times as the number of input OPs.
        Normalize the value of the input OP based on the tensor specified in mean and std.
        (input_value - mean) / std
        Tensors in Numpy file format must be in dimension order after conversion to TF.
        Note that this is intended for deployment on low-resource devices,
        so the batch size is limited to 1 only.
        e.g.
        The example below shows a case where there are three input OPs.
        Assume input0 is 128x128 RGB image data.
        In addition, input0 should be a value that has been divided by 255
        in the preprocessing and normalized to a range between 0 and 1.
        input1 and input2 assume the input of something that is not an image.
        Because input1 and input2 assume something that is not an image,
        the divisor is not 255 when normalizing from 0 to 1.
        "n" is the number of calibration data.

        ONNX INPUT shapes:
          input0: [n,3,128,128]
            mean: [1,3,1,1] -> [[[[0.485]],[[0.456]],[[0.406]]]]
            std : [1,3,1,1] -> [[[[0.229]],[[0.224]],[[0.225]]]]
          input1: [n,64,64]
            mean: [1,64] -> [[0.1, ..., 0.64]]
            std : [1,64] -> [[0.05, ..., 0.08]]
          input2: [n,5]
            mean: [1] -> [0.3]
            std : [1] -> [0.07]

        TensorFlow INPUT shapes (Numpy file ndarray shapes):
          input0: [n,128,128,3]
            mean: [1,1,1,3] -> [[[[0.485, 0.456, 0.406]]]]
            std : [1,1,1,3] -> [[[[0.229, 0.224, 0.225]]]]
          input1: [n,64,64]
            mean: [1,64] -> [[0.1, ..., 0.64]]
            std : [1,64] -> [[0.05, ..., 0.08]]
          input2: [n,5]
            mean: [1] -> [0.3]
            std : [1] -> [0.07]

          cind=[
              ["input0","../input0.npy",[[[[0.485, 0.456, 0.406]]]],[[[[0.229, 0.224, 0.225]]]]],
              ["input1","./input1.npy",[0.1, ..., 0.64],[0.05, ..., 0.08]],
              ["input2","input2.npy",[0.3],[0.07]],
          ]

      <Using -cotof and -oiqt at the same time>
        To use -cotof and -oiqt simultaneously,
        you need to enter the Input name of OP, path of data file, mean, and std all together.
        And the data file must be in Float32 format,
        and {input_op_name}, {numpy_file_path}, {mean}, and {std} must all be entered.
        Otherwise, an error will occur during the -oiqt stage.

    input_quant_dtype: Optional[str]
      Input dtypes when doing Full INT8 Quantization.
      "int8"(default) or "uint8" or "float32"

    output_quant_dtype: Optional[str]
      Output dtypes when doing Full INT8 Quantization.
      "int8"(default) or "uint8" or "float32"

    not_use_onnxsim: Optional[bool]
      No optimization by onnx-simplifier is performed.
      If this option is used, the probability of a conversion error is very high.

    not_use_opname_auto_generate: Optional[bool]
      Automatic generation of each OP name in the old format ONNX file
      and assignment of OP name are not performed.

    batch_size: Optional[int]
      Fixes the dynamic batch size to the specified numeric batch size.
      A value of 1 or more must be specified.

    overwrite_input_shape: Optional[List[str]]
      Overwrite the input shape.
      The format is
      ['i1:dim0,dim1,...,dimN', 'i2:dim0,dim1,...,dimN', 'i3:dim0,dim1,...,dimN']
      When there is only one input, for example,
      ['data:1,3,224,224']
      When there are multiple inputs, for example,
      ['data1:1,3,224,224','data2:1,3,112','data3:5']
      A value of 1 or more must be specified.
      Numerical values other than dynamic dimensions are ignored.
      Ignores batch_size if specified at the same time as batch_size.

    no_large_tensor: Optional[bool]
      Suppresses constant bloat caused by Tile OP when optimizing models in onnxsim.
      See: https://github.com/daquexian/onnx-simplifier/issues/178

    output_nms_with_dynamic_tensor: Optional[bool]
      The number of bounding boxes in the NMS output results is
      not fixed at the maximum number of max_output_boxes_per_class,
      but rather at the smallest possible number of dynamic tensors.
      If this option is disabled, NMS output is padded to the number
      set in the max_output_boxes_per_class attribute.
      e.g.
      disable --output_nms_with_dynamic_tensor:
          output_tensor_shape: [100, 7]
      enable --output_nms_with_dynamic_tensor:
          output_tensor_shape: [N, 7]

    keep_ncw_or_nchw_or_ncdhw_input_names: Optional[List[str]]
      Holds the NCW or NCHW or NCDHW of the input shape for the specified INPUT OP names.
      If a nonexistent INPUT OP name is specified, it is ignored.
      Valid only for 3D, 4D and 5D input tensors.
      e.g.
      keep_ncw_or_nchw_or_ncdhw_input_names=['input0','input1','input2']

    keep_nwc_or_nhwc_or_ndhwc_input_names: Optional[List[str]]
      Holds the NWC or NHWC or NDHWC of the input shape for the specified INPUT OP names.
      If a nonexistent INPUT OP name is specified, it is ignored.
      If the input OP name is the same as the input OP name specified
      in the keep_ncw_or_nchw_or_ncdhw_input_names option, it is ignored.
      Valid only for 3D, 4D and 5D input tensors.
      e.g.
      keep_nwc_or_nhwc_or_ndhwc_input_names=['input0','input1','input2']

    keep_shape_absolutely_input_names: Optional[List[str]]
      Name of the INPUT that unconditionally maintains its shape.
      If a nonexistent INPUT OP name is specified, it is ignored.
      e.g.
      keep_shape_absolutely_input_names=['input0','input1','input2']

    input_names_to_interrupt_model_conversion: Optional[List[str]]
      Input names of ONNX that interrupt model conversion.
      Interrupts model transformation at the specified input name
      and inputs the model partitioned into subgraphs.
      e.g.
      input_names_to_interrupt_model_conversion=['input0','input1','input2']

    output_names_to_interrupt_model_conversion: Optional[List[str]]
      Output names of ONNX that interrupt model conversion.
      Interrupts model transformation at the specified output name
      and outputs the model partitioned into subgraphs.
      e.g.
      output_names_to_interrupt_model_conversion=['output0','output1','output2']

    disable_group_convolution: Optional[bool]
      Disable GroupConvolution and replace it with SeparableConvolution for
      output to saved_model format.

    enable_accumulation_type_float16: Optional[bool]
      Hint for XNNPack fp16 inference on float16 tflite model.
      XNNPACK float16 inference on certain ARM64 cores is 2x faster.
      Float16 inference doubling on devices with ARM64 ARMv8.2 or higher instruction set.
      https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/delegates/xnnpack/README.md#floating-point-ieee-fp16-operators

    enable_batchmatmul_unfold: Optional[bool]
      BatchMatMul is separated batch by batch to generate a primitive MatMul.

    enable_rnn_unroll: Optional[bool]
      Instead of increasing inference speed by expanding all symbolic loops of
      the RNN (LSTM, GRU, RNN), RAM consumption will increase because all tensors
      are expanded and embedded in the model.
      https://keras.io/api/layers/recurrent_layers/

    disable_suppression_flextranspose: Optional[bool]
      Disables FlexTranspose generation suppression.

    number_of_dimensions_after_flextranspose_compression: Optional[int]
      Number of Transpose OP dimensions generated after avoiding FlexTranspose generation.
      Also suppress the creation of the Transpose itself by specifying 2.
      Default: 6

    disable_suppression_flexstridedslice: Optional[bool]
      Disables FlexStridedSlice generation suppression.

    disable_strict_mode: Optional[bool]
      If specified, the conversion speed is greatly accelerated because the strict accuracy
      correction process is skipped, but the frequency of transposition errors increases
      and accuracy errors are more likely to occur. Strict mode is enabled by default.
      As of 2023.05.07, this is a work in progress and is an experimental feature.
      Therefore, only some OPs are converted in strict mode for accuracy correction.

    number_of_dimensions_after_flexstridedslice_compression: Optional[int]
      Number of StridedSlice OP dimensions generated after avoiding FlexStridedSlice generation.
      Default: 5

    optimization_for_gpu_delegate: Optional[bool]
      Replace operations that do not support gpu delegate with those
      that do as much as possible.

    replace_argmax_to_reducemax_and_indices_is_int64: Optional[bool]
      Replace ArgMax with a ReduceMax. The returned indices are int64.
      Only one of replace_argmax_to_reducemax_and_indices_is_int64 and
      replace_argmax_to_reducemax_and_indices_is_float32 and
      replace_argmax_to_fused_argmax_and_indices_is_int64 and
      replace_argmax_to_fused_argmax_and_indices_is_float32 can be specified.
      Default: False

    replace_argmax_to_reducemax_and_indices_is_float32: Optional[bool]
      Replace ArgMax with a ReduceMax. The returned indices are float32.
      Only one of replace_argmax_to_reducemax_and_indices_is_int64 and
      replace_argmax_to_reducemax_and_indices_is_float32 and
      replace_argmax_to_fused_argmax_and_indices_is_int64 and
      replace_argmax_to_fused_argmax_and_indices_is_float32 can be specified.
      Default: False

    replace_argmax_to_fused_argmax_and_indices_is_int64: Optional[bool]
      Replace ArgMax with a ReduceMax. The returned indices are int64.
      It improves inference speed at the cost of a small sacrifice in accuracy.
      See. https://github.com/tensorflow/models/tree/master/official/projects/edgetpu/vision#argmax-fusion-to-improve-segmentation-model-latency
      Currently, only 4D tensors are supported.
      Only one of replace_argmax_to_reducemax_and_indices_is_int64 and
      replace_argmax_to_reducemax_and_indices_is_float32 and
      replace_argmax_to_fused_argmax_and_indices_is_int64 and
      replace_argmax_to_fused_argmax_and_indices_is_float32 can be specified.
      Default: False

    replace_argmax_to_fused_argmax_and_indices_is_float32: Optional[bool]
      Replace ArgMax with a ReduceMax. The returned indices are float32.
      It improves inference speed at the cost of a small sacrifice in accuracy.
      See. https://github.com/tensorflow/models/tree/master/official/projects/edgetpu/vision#argmax-fusion-to-improve-segmentation-model-latency
      Currently, only 4D tensors are supported.
      Only one of replace_argmax_to_reducemax_and_indices_is_int64 and
      replace_argmax_to_reducemax_and_indices_is_float32 and
      replace_argmax_to_fused_argmax_and_indices_is_int64 and
      replace_argmax_to_fused_argmax_and_indices_is_float32 can be specified.
      Default: False

    fused_argmax_scale_ratio: Optional[float]
      For Fused ArgMax.
      Scale ratio when generating Fused ArgMax.
      0.0 < fused_argmax_scale_ratio <= 1.0
      Default: 0.5

    replace_to_pseudo_operators: List[str]
      Replace list of operators to pseudo operators.
      Full name of the target operators should be given.
      Currently supported operators :
      Asin, Acos, Atan, Abs, PReLU, LeakyReLU, Power, GatherND, Neg, HardSwish, Erf, GeLU, MatMulInteger

    mvn_epsilon: Optional[float]
      For MeanVarianceNormalization.
      The number to be added to the variance to avoid division by zero
      when normalizing the value.
      (input_tensor - mean) / tf.sqrt(variance + mvn_epsilon)
      Default: 0.0000000001

    param_replacement_file: Optional[str]
      Parameter replacement file path. (.json)

    check_gpu_delegate_compatibility: Optional[bool]
      Run TFLite ModelAnalyzer on the generated Float16 tflite model
      to check if the model can be supported by GPU Delegate.
      e.g.
      """
      === TFLite ModelAnalyzer ===

      Your TFLite model has '1' subgraph(s). In the subgraph description below,
      T# represents the Tensor numbers. For example, in Subgraph#0, the RESHAPE op takes
      tensor #0 and tensor #6 as input and produces tensor #7 as output.

      Subgraph#0 main(T#0) -> [T#17]
        Op#0 RESHAPE(T#0, T#6[2, 8, 8, 3, 2, ...]) -> [T#7]
        Op#1 SPLIT(T#5[0], T#7) -> [T#8, T#9]
        Op#2 RESHAPE(T#8, T#1[8, 8, 3, 2, 2]) -> [T#10]
        Op#3 TRANSPOSE(T#10, T#4[0, 3, 1, 4, 2]) -> [T#11]
        Op#4 RESHAPE(T#11, T#2[1, 8, 2, 8, 2, ...]) -> [T#12]
        Op#5 RESHAPE(T#9, T#1[8, 8, 3, 2, 2]) -> [T#13]
        Op#6 TRANSPOSE(T#13, T#4[0, 3, 1, 4, 2]) -> [T#14]
        Op#7 RESHAPE(T#14, T#2[1, 8, 2, 8, 2, ...]) -> [T#15]
        Op#8 CONCATENATION(T#12, T#15) -> [T#16]
        Op#9 RESHAPE(T#16, T#3[2, 16, 16, 3]) -> [T#17]

      Tensors of Subgraph#0
        T#0(inputs_0) shape:[2, 8, 8, 12], type:FLOAT32
        T#1(model/tf.compat.v1.squeeze_2/Squeeze) shape:[5], type:INT32 RO 20 bytes, data:[8, 8, 3, 2, 2]
        T#2(model/tf.expand_dims_1/ExpandDims) shape:[6], type:INT32 RO 24 bytes, data:[1, 8, 2, 8, 2, ...]
        T#3(model/tf.reshape_1/Reshape/shape) shape:[4], type:INT32 RO 16 bytes, data:[2, 16, 16, 3]
        T#4(model/tf.compat.v1.transpose/transpose/perm) shape:[5], type:INT32 RO 20 bytes, data:[0, 3, 1, 4, 2]
        T#5(model/tf.concat/concat/axis) shape:[], type:INT32 RO 4 bytes, data:[0]
        T#6(model/tf.reshape/Reshape/shape) shape:[6], type:INT32 RO 24 bytes, data:[2, 8, 8, 3, 2, ...]
        T#7(model/tf.reshape/Reshape) shape:[2, 8, 8, 3, 2, 2], type:FLOAT32
        T#8(model/tf.split/split) shape:[1, 8, 8, 3, 2, 2], type:FLOAT32
        T#9(model/tf.split/split1) shape:[1, 8, 8, 3, 2, 2], type:FLOAT32
        T#10(model/tf.compat.v1.squeeze_1/Squeeze) shape:[8, 8, 3, 2, 2], type:FLOAT32
        T#11(model/tf.compat.v1.transpose/transpose) shape:[8, 2, 8, 2, 3], type:FLOAT32
        T#12(model/tf.expand_dims/ExpandDims) shape:[1, 8, 2, 8, 2, 3], type:FLOAT32
        T#13(model/tf.compat.v1.squeeze_2/Squeeze1) shape:[8, 8, 3, 2, 2], type:FLOAT32
        T#14(model/tf.compat.v1.transpose_1/transpose) shape:[8, 2, 8, 2, 3], type:FLOAT32
        T#15(model/tf.expand_dims_1/ExpandDims1) shape:[1, 8, 2, 8, 2, 3], type:FLOAT32
        T#16(model/tf.concat/concat) shape:[2, 8, 2, 8, 2, 3], type:FLOAT32
        T#17(Identity) shape:[2, 16, 16, 3], type:FLOAT32

      Your model looks compatibile with GPU delegate with TFLite runtime version 2.10.0.
      But it doesn't guarantee that your model works well with GPU delegate.
      There could be some runtime incompatibililty happen.
      ---------------------------------------------------------------
                    Model size:       2988 bytes
          Non-data buffer size:       2757 bytes (92.27 %)
        Total data buffer size:        231 bytes (07.73 %)
          (Zero value buffers):          4 bytes (00.13 %)

      * Buffers of TFLite model are mostly used for constant tensors.
        And zero value buffers are buffers filled with zeros.
        Non-data buffers area are used to store operators, subgraphs and etc.
        You can find more details from https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/schema/schema.fbs
      """

    check_onnx_tf_outputs_elementwise_close: Optional[bool]
      Returns "Matches" if the output of onnx and the output of TF are
      within acceptable proximity element by element.
      Returns "Unmatched" if the output of onnx and the output of TF are
      not within acceptable proximity element by element.
      If the output of onnx is 1D, it returns "Skipped" and skips the comparison
      between the output of onnx and that of TF. This is because when undefined
      dimensions are present, a situation often arises where very large index
      values are compared, causing OutOfMemory.
      Only the output content of the models final output OP is checked.

    check_onnx_tf_outputs_elementwise_close_full: Optional[bool]
      Returns "Matches" if the output of onnx and the output of TF are
      within acceptable proximity element by element.
      Check the output of all OPs in sequence from the beginning,
      including all but the final output OP of the model.
      Returns "Unmatched" if the output of onnx and the output of TF are
      not within acceptable proximity element by element.
      If the output of onnx is 1D, it returns "Skipped" and skips the comparison
      between the output of onnx and that of TF. This is because when undefined
      dimensions are present, a situation often arises where very large index
      values are compared, causing OutOfMemory.
      It is very time consuming because it performs as many inferences as
      there are operations.

    check_onnx_tf_outputs_sample_data_normalization: Optional[str]
      norm: Validate using random data normalized to the range 0.0 to 1.0
      denorm: Validate using random data in the range 0.0 to 255.0
      If there is a normalization layer at the models entry point, or
      if the model was trained on denormalized data, "denorm" must be specified.
      Default: "norm"

    check_onnx_tf_outputs_elementwise_close_rtol: Optional[float]
      The relative tolerance parameter.
      Default: 0.0

    check_onnx_tf_outputs_elementwise_close_atol: Optional[float]
      The absolute tolerance parameter.
      Default: 1e-4

    disable_model_save: Optional[bool]
      Does not save the converted model. For CIs RAM savings.
      Default: False

    non_verbose: Optional[bool]
      Shorthand to specify a verbosity of "error".
      Default: False

    verbosity: Optional[str]
      Change the level of information printed.
      Values are "debug", "info", "warn", and "error".
      Default: "debug" (for backwards compatability)

    Returns
    ----------
    model: tf_keras.Model
      Model

Parameter replacement

This tool is used to convert NCW to NWC, NCHW to NHWC, NCDHW to NDHWC, NCDDHW to NDDHWC, NCDDDDDDHW to NDDDDDDHWC. Therefore, as stated in the Key Concepts, the conversion will inevitably break down at some point in the model. You need to look at the entire conversion log to see which OP transpositions are failing and correct them yourself. I dare to explain very little because I know that no matter how much detail I put in the README, you guys will not read it at all. attribute or INPUT constant or INPUT Initializer can be replaced with the specified value.

Click to expand

Starting from v1.3.0, almost all OPs except for some special OPs support pre- and post-transposition by pre_process_transpose and post_process_transpose.

  1. "A conversion error occurs."
  2. "Output results are wrong."

Do not submit an issue that only contains an amount of information that cannot be reproduced.

  • convert option

    --param_replacement_file param_replacement.json
    
    or
    
    -prf param_replacement.json
    
  • param_replacement.json

    See a sample of replacement JSON
    {
      "format_version": 1,
      "operations": [
        {
          "op_name": "StatefulPartitionedCall/Tile_4",
          "param_target": "inputs", # attributes or inputs
          "param_name": "const_fold_opt__677",
          "values": [1,1,17] # Disable parameter transposition or overwrite parameters
        },
        {
          "op_name": "StatefulPartitionedCall/Cast_3",
          "param_target": "attributes", # attributes or inputs
          "param_name": "to",
          "values": 1 # Disable parameter transposition or overwrite "to" parameters
        },
        {
          "op_name": "Resize__697",
          "param_target": "inputs",
          "param_name": "Concat__696:0",
          "values": [26,26] # Replacement of unk__x (Resize OP, sizes height/width parameter)
        },
        {
          "op_name": "Transpose__927",
          "param_target": "attributes",
          "param_name": "perm",
          "values": [0,1,2,3] # Disable parameter transposition or overwrite "perm" parameters
        },
        {
          "op_name": "StatefulPartitionedCall/functional_1/max_unpooling2d_2/Reshape_1",
          "param_target": "inputs",
          "param_name": "const_fold_opt__911",
          "values": [4,131072] # Overwrite "shape" parameters
        },
        {
          "op_name": "Reshape_25",
          "param_target": "outputs",
          "param_name": "onnx::InstanceNormalization_270",
          "post_process_transpose_perm": [0,2,1] # Extrapolate 3D Transpose after Reshape
        },
        {
          "op_name": "Reshape_30",
          "param_target": "outputs",
          "param_name": "onnx::Mul_275",
          "post_process_transpose_perm": [0,2,3,1] # Extrapolate 4D Transpose after Reshape
        },
        {
          "op_name": "flatten_1127",
          "param_target": "inputs",
          "param_name": "dropout0",
          "pre_process_transpose_perm": [0,3,1,2]
        },
        {
          "op_name": "/Slice",
          "param_target": "op",
          "begin": [0,0,1,0],
          "end": [0,0,0,0],
          "end_mask": 15
        },
        {
          "op_name": "/Slice_1",
          "param_target": "op",
          "begin": [0,0,0,0],
          "end": [0,0,39,0],
          "end_mask": 11
        },
        {
          "op_name": "/backbone/backbone.1/Unsqueeze_1",
          "param_target": "op",
          "new_shape": [1,15,15,1]
        }
      ]
    }
  • Replacement Supported OPs

    See list of replacement specifications
    No. OP type Remarks
    1 Add 1. "param_target": "inputs"
    pre_process_transpose_perm: Transpose is applied to the tensor before the Add operation with the perm specified as pre-processing.
    2. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Add operation with the perm specified as post-processing.
    2 Cast
    TypeValuesTypeValues
    float1610int83
    float321int165
    float6411int326
    bool9int647
    uint82
    uint164
    uint3212
    uint6413
    3 Concat 1. "param_target": "attributes"
    axis: Value of axis
    2. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Concat operation with the perm specified as post-processing.
    4 ConvTranspose ConvTranspose implements special replacements separately ignore all automatic conversions and generate tf.nn.conv1d_transpose or tf.nn.conv2d_transpose or tf.nn.conv3d_transpose directly by specifying all parameters.
    https://www.tensorflow.org/api_docs/python/tf/nn/conv1d_transpose
    https://www.tensorflow.org/api_docs/python/tf/nn/conv2d_transpose
    https://www.tensorflow.org/api_docs/python/tf/nn/conv3d_transpose
    1. "param_target": "op"
    output_shape: Value of output_shape
    strides: Value of strides
    padding: Value of padding
    dilations: Value of dilations
    5 Div 1. "param_target": "inputs"
    values: Value of input
    pre_process_transpose_perm: Transpose is applied to the tensor before the Div operation with the perm specified as pre-processing.
    2. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Div operation with the perm specified as post-processing.
    6 Expand 1. "param_target": "inputs"
    values: Value of shape
    pre_process_transpose_perm: Transpose is applied to the tensor before the Expand operation with the perm specified as pre-processing.
    2. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Expand operation with the perm specified as post-processing.
    7 Flatten 1. "param_target": "attributes"
    axis: Value of axis
    2. "param_target": "inputs"
    pre_process_transpose_perm: Transpose is applied to the tensor before the Flatten operation with the perm specified as pre-processing.
    3. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Flatten operation with the perm specified as post-processing.
    8 Gemm
    9 Gather 1. "param_target": "attributes"
    axis: Value of axis
    2. "param_target": "inputs"
    values: Value of indices
    pre_process_transpose_perm: Transpose is applied to the tensor before the Gather operation with the perm specified as pre-processing.
    3. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Gather operation with the perm specified as post-processing.
    10 MatMul 1. "param_target": "inputs"
    pre_process_transpose_perm: Transpose is applied to the tensor before the MatMul operation with the perm specified as pre-processing.
    2. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the MatMul operation with the perm specified as post-processing.
    11 Mul 1. "param_target": "inputs"
    values: Value of input
    pre_process_transpose_perm: Transpose is applied to the tensor before the Mul operation with the perm specified as pre-processing.
    2. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Mul operation with the perm specified as post-processing.
    12 NonMaxSuppression
    13 ReduceL1
    ReduceL2
    ReduceLogSum
    ReduceLogSumExp
    ReduceMax
    ReduceMean
    ReduceMin
    ReduceProd
    ReduceSum
    ReduceSumSquare
    1. "param_target": "attributes"
    axes: Value of axes
    keepdims: Value of keepdims
    2. "param_target": "inputs"
    pre_process_transpose_perm: Transpose is applied to the tensor before the ReduceXX operation with the perm specified as pre-processing.
    3. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the ReduceXX operation with the perm specified as post-processing.
    14 Unsqueeze 1. "param_target": "inputs"
    pre_process_transpose_perm: Transpose is applied to the tensor before the Unsqueeze operation with the perm specified as pre-processing.
    2. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Unsqueeze operation with the perm specified as post-processing.
    3. "param_target": "op"
    new_shape: Specifies directly the shape after Unsqueeze processing.
    {
      "op_name": "/backbone/backbone.1/Unsqueeze_1",
      "param_target": "op",
      "new_shape": [1,15,15,1]
    }
    15 Reshape 1. "param_target": "inputs"
    values: Value of shape
    pre_process_transpose_perm: Transpose is applied to the tensor before the Reshape operation with the perm specified as pre-processing.
    2. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Reshape operation with the perm specified as post-processing.
    16 Resize 1. "param_target": "attributes"
    coordinate_transformation_mode: Value of coordinate_transformation_mode
    extrapolation_value: Value of extrapolation_value
    mode: Value of mode
    2. "param_target": "inputs"
    values: Value of roi or scales or sizes. scales=[scale_h,scale_w],sizes=[h,w]
    pre_process_transpose_perm: Transpose is applied to the tensor before the Resize operation with the perm specified as pre-processing.
    3. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Resize operation with the perm specified as post-processing.
    17 Slice Slice implements special replacements separately ignore all automatic conversions and generate tf.strided_slice directly by specifying all parameters of tf.strided_slice directly.
    https://www.tensorflow.org/api_docs/python/tf/strided_slice
    See json_samples/replace_slice.json for a sample description.
    20221221222956
    1. "param_target": "op"
    begin: Value of begin
    end: Value of end
    strides: Value of strides
    begin_mask: Value of begin_mask
    end_mask: Value of end_mask
    ellipsis_mask: Value of ellipsis_mask
    new_axis_mask: Value of new_axis_mask
    shrink_axis_mask: Value of shrink_axis_mask
    {
      "op_name": "/Slice",
      "param_target": "op",
      "begin": [0,0,1,0],
      "end": [0,0,0,0],
      "end_mask": 15
    }
    18 Softmax 1. "param_target": "attributes"
    axis: Value of axis. The transpositions corresponding to the specified axis are extrapolated before and after Softmax.
    2. "param_target": "inputs"
    values: Value of tensor
    19 Split 1. "param_target": "inputs"
    values: Value of split
    2. "param_target": "attributes"
    axis: Value of axis.
    num_outputs: Value of num_outputs.
    20 Sub 1. "param_target": "inputs"
    values: Value of input
    pre_process_transpose_perm: Transpose is applied to the tensor before the Sub operation with the perm specified as pre-processing.
    2. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Sub operation with the perm specified as post-processing.
    21 Tile 1. "param_target": "inputs"
    values: Value of input
    pre_process_transpose_perm: Transpose is applied to the tensor before the Tile operation with the perm specified as pre-processing.
    2. "param_target": "outputs"
    post_process_transpose_perm: Transpose is applied to the tensor after the Tile operation with the perm specified as post-processing.
    22 Transpose 1. "param_target": "attributes"
    perm: Value of perm
    2. "param_target": "inputs"
    values: Value of tensor

Generated Model

Validated models (without replacement.json)

ONNX file for testing. https://github.com/PINTO0309/onnx2tf/releases/tag/1.1.28

See a list of verified models
No. Model Pass
1 age_googlenet.onnx ✔️
2 alike_t_opset11_192x320.onnx ✔️
3 arcfaceresnet100-8.onnx ✔️
4 baseline_simplified.onnx ✔️
5 big_slice_11.onnx ✔️
6 bvlcalexnet-12.onnx ✔️
7 caffenet-12.onnx ✔️
8 convtranspose_3_1_5_2.onnx ✔️
9 convtranspose_4_5_2_2.onnx ✔️
10 convtranspose_5_5_6_1.onnx ✔️
11 convtranspose_6_5_5_8.onnx ✔️
12 convtranspose_7_1_3_4.onnx ✔️
13 damoyolo_tinynasL20_T_192x192_post.onnx ✔️
14 deeplabv3_mobilenet_v3_large.onnx ✔️
15 densenet-12.onnx ✔️
16 depth_to_spase_17.onnx ✔️
17 double_gru.onnx ✔️
18 digits.onnx ✔️
19 detr_demo.onnx ✔️
20 efficientformer_l1.onnx ✔️
21 efficientdet_lite2_detection_1.onnx ✔️
22 efficientnet-lite4-11_nchw.onnx ✔️
23 effnet_opset11_dynamic_axis.onnx ✔️
24 emotion-ferplus-8_rename.onnx ✔️
25 face_detection_yunet_2022mar.onnx ✔️
26 face_recognition_sface_2021dec-act_int8-wt_int8-quantized.onnx ✔️
27 face_recognition_sface_2021dec.onnx ✔️
28 faster_rcnn-10.onnx ✔️
29 fastestdet.onnx ✔️
30 fused_conv_clip.onnx ✔️
31 fused_conv_hardsigmoid.onnx ✔️
32 fused_conv_leakyrelu.onnx ✔️
33 fused_conv_relu.onnx ✔️
34 fused_conv_sigmoid.onnx ✔️
35 fused_conv_tanh.onnx ✔️
36 gender_googlenet.onnx ✔️
37 gmflow-scale1-mixdata-train320x576-4c3a6e9a_1x3x480x640_bidir_flow_sim.onnx ✔️
38 handpose_estimation_mediapipe_2022may.onnx ✔️
39 htnet_1x17x2_without_norm.onnx ✔️
40 iat_llie_180x320.onnx ✔️
41 if_p1_11.onnx ✔️
42 if_p2_11.onnx ✔️
43 if_p3_11.onnx ✔️
44 imageclassifier.onnx ✔️
45 inception-v2-9.onnx ✔️
46 inverse11.onnx ✔️
47 mhformer_NxFxKxXY_1x27x17x2.onnx ✔️
48 mnist.onnx ✔️
49 mnist-12.onnx ✔️
50 mobilenetv2-12.onnx ✔️
51 mosaic_11.onnx ✔️
52 mosaic-9.onnx ✔️
53 movenet_multipose_lightning_192x256_p6.onnx ✔️
54 nanodet-plus-m_416.onnx ✔️
55 object_tracking_dasiamrpn_kernel_cls1_2021nov.onnx ✔️
56 object_tracking_dasiamrpn_kernel_r1_2021nov.onnx ✔️
57 object_tracking_dasiamrpn_model_2021nov.onnx ✔️
58 pidnet_S_cityscapes_192x320.onnx ✔️
59 ppmattingv2_stdc1_human_480x640.onnx ✔️
60 qlinear_conv_tensor_test.onnx ✔️
61 rcnn-ilsvrc13-9.onnx ✔️
62 regnet_x_400mf.onnx ✔️
63 ResNet101-DUC-12.onnx ✔️
64 resnet18-v1-7.onnx ✔️
65 resnet50-v1-12.onnx ✔️
66 resnet50-v2-7.onnx ✔️
67 retinanet-9.onnx ✔️
68 sinet_320_op.onnx ✔️
69 squeezenet1.0-12.onnx ✔️
70 super-resolution-10.onnx ✔️
71 swinir-m_64x64_12.onnx ✔️
72 text_recognition_CRNN_EN_2021sep.onnx ✔️
73 tinyyolov2-8.onnx ✔️
74 version-RFB-640.onnx ✔️
75 vit-b-32_textual.onnx ✔️
76 vit-b-32_visual.onnx ✔️
77 yolact_edge_mobilenetv2_550x550.onnx ✔️
78 yolact_regnetx_600mf_d2s_31classes_512x512.onnx ✔️
79 yolact_regnetx_800mf_20classes_512x512.onnx ✔️
80 yolo_free_nano_crowdhuman_192x320_post.onnx ✔️
81 yolov7_tiny_head_0.768_post_480x640.onnx ✔️
82 yolox_nano_192x192.onnx ✔️
83 yolox_nano_416x416.onnx ✔️
84 yolox_s.onnx ✔️
85 yolox_x_crowdhuman_mot17_bytetrack.onnx ✔️
86 zero_dce_640_dele.onnx ✔️
87 zfnet512-12.onnx ✔️

Key concept

List of Key concept
  • onnx-tensorflow is a very useful tool, but the performance of the generated TensorFlow models is significantly degraded due to the extrapolation of a large number of Transpose OPs before and after each OP during the format conversion from NCHW to NHWC. Therefore, I will make this tool myself as a derivative tool of onnx-tensorflow without extrapolating Transpose.

  • Most of the internal processing of the tool is full-scratch, but some of the more complex OPs have been adapted from onnx-tensorflow. I am very grateful to the engineers at International Business Machines Corporation / LeapMind / Microsoft / IBM for developing onnx-tensorflow.

  • I have incorporated all my knowledge of model optimization to other models such as TFLite, EdgeTPU, TensorFlow.js and Myriad based on my years of experience implementing openvino2tensorflow and tflite2tensorflow. It probably has the best model optimization performance and conversion efficiency of any tool I have created in the past, and the lowest rate of conversion errors.

  • Supported layers list. Supported layers

  • If you are having trouble with conversion errors, searching for resolved or open issues will almost always solve your problems. Issues are knowledge for engineers around the world.

  • Contributors to this repository should first read Contribution Guide.

    Kazam_screencast_00065_.mp4
  • All OPs are decomposed into primitive operations as much as possible. This is beneficial for lateral deployment of models to frameworks other than TFLite. Therefore, OPs belonging to tf_keras.layers are almost never used, and the tool consists only of tf.xxx. (except for a very few OPs)

  • As I do not want to add more dependent packages, I do not use tensorflow_addons (tfa), but replace it with the standard OP of tensorflow.

  • Not only does it handle conversions of 4-dimensional inputs, such as NCHW to NHWC, but also the number of input dimensions in 3, 5, or even more dimensions. For example, NCDHW to NDHWC, etc. However, since 1-D, 2-D, 3-D and 6-D input may produce patterns that are mechanically difficult to convert, it should be possible to give parameters to externally modify the tool's behavior. See Parameter replacement

  • If there are undefined dimensions in the input OP, the model structure is not fully optimized and conversion errors are very likely to occur.

  • Immediately following a Reshape OP with dimensional compression and dimensional decompression, there is a 95% probability that the model transformation operation will be disrupted and errors will occur. For example, patterns such as [1,200,200,5] -> [1,200,-1] or [10,20,30,40,50] -> [10,2,10,30,10,4,50] or Flatten. See #8 Not able to reshape input in replace.json, or #15 Conv layer shape wrong, or #18 Question about channel_transpose in common_functions.py, or #105 [MobileFormer]Converted model outputs values mismatch with original ones., or #133 When Onnx Matmul inputs have different dimension.

  • TensorFlow's Convolution does not have an equivalent operation to ONNX's Padding operation. Therefore, a Pad OP is inserted immediately before a Convolution with Padding of size greater than 1.

  • Support conversion to TensorFlow saved model and TFLite (Float32/Float16/INT8).

  • Files exceeding the Protocol Buffers file size limit of 2GB are not supported. Therefore, the external format is not supported at the initial stage of tool creation.

  • If there are ONNX OPs that are not supported by TensorFlow, use simple-onnx-processing-tools to replace them with harmless OPs in advance and then use this tool to convert them. In other words, you can convert any model with your efforts.

  • ONNX splitting, merging, generating OPs, rewriting OP attributes, BGR<->RGB conversion, converting to JSON and editing in the IDE, batch size changes for undefined dimensions, and various other processing can be done with the simple-onnx-processing-tools. Therefore, it is recommended that models with very complex structures be converted to TFLite after modifying the structure beforehand.

  • BatchNormalization supports only inference mode.

  • LayerNormalization supports only inference mode.

  • Only for opset=11 or higher

  • If you do not like the generated TFLite OP name, edit it using tflite2json2tflite.

  • The generated Keras models cannot be used for retraining. If you want to train, you must build your own model.

  • When converting to TensorFlow.js, CoreML, etc., please generate saved_model with the --output_signaturedefs option and use the generated saved_model to convert with various converters. tensorflowjs_converter, coremltools, edgetpu_compilier, etc... If this option is not enabled, saved_model records only the minimum necessary information and its size is minimized. When this option is enabled, saved_model records the maximum amount of information, and instead of being maximized in size, the output is in a format that supports conversion to other frameworks. It can also be used for serving.

  • There are many OPs on ONNX that do not support TFLite/EdgeTPU/TFJS/CoreML/TensorRT. Therefore, if you need to generate an EdgeTPU model, please specify --replace_to_pseudo_operators to convert your model. onnx2tf will attempt to replace the OP with an TFLite/EdgeTPU/TFJS/CoreML/TensorRT-compatible OP whenever possible.

  • The main factors that cause accuracy degradation after model conversion are as follows

  1. differences in Padding specifications
  2. difference in Python division specification in the process of model transformation (error due to even rounding)
  3. Divide epsilon without consideration
  4. deprecated TrueDivision
  5. support difference of powers
  6. differences in interpolation operation specifications during resizing
  7. Difference in arithmetic precision supported by each operation
  8. Calculation error due to scaling up or down by specifying a scale when resizing images

The above differences often cannot be dealt with by simply converting the model in a straightforward manner. Therefore, you need to replace the model yourself in advance with an operation that is less prone to errors.

  • Support for INT8 Quantization, Full INT8 Quantization, INT8 Quantization with INT16 activation, Full INT8 Quantization with INT16 activation and Dynamic Range Quantization.
  • Support for Per-Channel Quantization and Per-Tensor Quantization.
  • Support for GroupConvolution.
  • TFLite does not support TrueDiv(INT), so TrueDiv is avoided if possible.
  • Implement the Resize process for the 5D tensor.
  • Add process to replace Asin with pseudo-Asin.
  • Add process to replace Acos with pseudo-Acos.
  • Add process to replace Atan with pseudo-Atan.
  • Add process to replace Abs with pseudo-Abs.
  • Add process to replace GatherND with pseudo-GatherND.
  • Add process to replace HardSwish with pseudo-HardSwish.
  • Add process to replace GridSample with pseudo-GridSample.
  • Add process to replace PRelu with pseudo-PRelu.
  • Add process to replace LeakyRelu with pseudo-LeakyRelu.
  • Add process to replace Power with pseudo-Power.
  • Add process to replace Neg with pseudo-Neg.
  • Add process to replace ArgMax with pseudo-ArgMax.
  • Add process to replace Erf with pseudo-Erf.
  • Add process to replace GeLU with pseudo-GeLU.
  • Added option to fix dynamic batch size N to a specified number.
  • Added option to overwrite dynamic shape input OPs with static shape. --overwrite_input_shape
  • Output in Keras H5 format.
  • Automatically run onnx-simplifier (onnxsim) backend and optimize onnx files before model transformation.
  • Added the ability to automatically generate each OP name and assign OP names to ONNX files in the old format.
  • Supports model splitting. Interrupts model transformation at the specified output name and outputs the model partitioned into subgraphs.

Related tools

  1. tflite2tensorflow
  2. openvino2tensorflow
  3. tflite2json2tflite
  4. tensorflowjs_converter
  5. coremltools
  6. simple-onnx-processing-tools
  7. tflite-input-output-rewriter
  8. onnx-simplifier
  9. onnx_graphsurgeon
  10. onnx
  11. onnx-tensorflow
  12. onnx2keras
  13. TinyNeuralNetwork
  14. nobuco
  15. onnx2torch
  16. ai-edge-torch

Acknowledgement

  1. https://github.com/onnx/models
  2. https://github.com/opencv/opencv_zoo
  3. https://pytorch.org/vision/stable/models.html
  4. https://tfhub.dev/
  5. https://www.kaggle.com/models
  6. https://github.com/TexasInstruments/edgeai-modelzoo

Contributors

Made with contrib.rocks.