ONNX to TensorFlow/TFLite Converter

Warnings

• breaking Starting with v2.4.0, the default TFLite backend for `onnx2tf.convert()` and the CLI has switched from `tf_converter` to `flatbuffer_direct`. Code relying on the implicit `tf_converter` behavior for TFLite conversion will now use `flatbuffer_direct`.
  Fix: If you require the `tf_converter` backend for TFLite, explicitly set `tflite_backend='tf_converter'` in your `onnx2tf.convert()` call or use the `--tflite_backend tf_converter` CLI option.
• deprecated The `tf_converter` TFLite backend is deprecated starting from v2.4.0. While still available as an explicit option, users are encouraged to migrate to `flatbuffer_direct` as it may be removed in future versions.
  Fix: Update your conversion scripts to use the `flatbuffer_direct` backend, which is now the default, or keep an eye on release notes for its eventual removal if `tf_converter` is critical for your workflow.
• gotcha onnx2tf requires Python 3.12 or newer. Using older Python versions will lead to installation failures or runtime errors.
  Fix: Ensure your development environment uses Python 3.12 or a later version. You can manage Python versions using `pyenv` or `conda`.
• gotcha Implicit network downloads for validation sample data have been removed since v2.3.17. For integer quantization (`-oiqt`), explicit calibration input mapping is now required.
  Fix: When performing integer quantization, ensure you provide calibration inputs explicitly rather than relying on the library to download them. Refer to the documentation for the `--calibration_data_path` or equivalent API parameters.

Install

• pip install onnx2tf Default Install

Imports

• convert

  import onnx2tf
  
  # ... then onnx2tf.convert(...)

Quickstart

This quickstart demonstrates the end-to-end process of defining a simple PyTorch model, exporting it to ONNX format, and then using `onnx2tf` to convert the ONNX model into a TensorFlow SavedModel. It highlights the primary `onnx2tf.convert()` function and the necessary input/output paths.

import onnx2tf
import torch
import torch.nn as nn
import os

# 1. Define a simple PyTorch model
class SimpleModel(nn.Module):
    def __init__(self):
        super(SimpleModel, self).__init__()
        self.conv = nn.Conv2d(3, 16, 3, 1, 1)

    def forward(self, x):
        return self.conv(x)

# 2. Instantiate and export to ONNX
model = SimpleModel()
dummy_input = torch.randn(1, 3, 224, 224)
onnx_file_path = "simple_model.onnx"
torch.onnx.export(
    model,
    dummy_input,
    onnx_file_path,
    opset_version=17,
    input_names=["input"],
    output_names=["output"],
    dynamic_axes={"input": {0: "batch_size"}, "output": {0: "batch_size"}}
)
print(f"PyTorch model exported to {onnx_file_path}")

# 3. Convert ONNX to TensorFlow SavedModel
output_folder = "./converted_tf_model"
os.makedirs(output_folder, exist_ok=True)

onnx2tf.convert(
    input_onnx_file_path=onnx_file_path,
    output_folder_path=output_folder,
    # For TFLite conversion, you might add:
    # tflite_output_file_path="./converted_tf_model/model.tflite"
)
print(f"ONNX model converted to TensorFlow SavedModel at {output_folder}")

# Clean up generated ONNX file
os.remove(onnx_file_path)

# To use a specific backend for TFLite (e.g., the deprecated tf_converter):
# onnx2tf.convert(
#     input_onnx_file_path=onnx_file_path,
#     output_folder_path="./converted_tf_model_tfconv",
#     tflite_output_file_path="./converted_tf_model_tfconv/model.tflite",
#     tflite_backend='tf_converter'
# )