TFLite export fails when onnxsim is not on PATH #1365

Search before asking

• I searched the RF-DETR issues and found no similar bug report.

#1035 concerns TFLite support in 1.6.5, not this PATH failure. Searches in PINTO0309/onnx2tf also found no matching issue.

Bug

model.export(format="tflite") fails when RF-DETR runs through <venv>/bin/python without activating the virtual environment.

onnx2tf 2.4.3 calls the bare onnxsim console script. Invoking Python by absolute path does not add <venv>/bin to PATH, so the onnxsim installed by rfdetr[tflite] cannot be found.

onnx2tf contains this comment above the first call:

To fully optimize the model, run onnxsim three times in a row.
Due to unstable script execution of onnxsim in v0.4.8,
I have no choice but to use subprocesses that we do not want to use.

A clean pip install "rfdetr[tflite]==1.9.3" with stock RFDETRSmall() produced:

FileNotFoundError: [Errno 2] No such file or directory: 'onnxsim'
WARNING: Failed to optimize the onnx file.
RuntimeError: onnx2tf conversion failed: Output tensors of a Functional model must be the output of a TensorFlow Layer (thus holding past layer metadata).

No .tflite file is written. The final RuntimeError above was observed on stock RFDETRSmall(); it is not claimed for other variants.

Adding <venv>/bin to PATH makes the same export finish. onnxsim still returns exit status 1 after rewriting the graph.

Failed to optimize the onnx file appears in both runs. FileNotFoundError: ... 'onnxsim' identifies this failure.

Linux, Windows, and onnx2tf versions after 2.4.3 were not runtime-tested. Source inspection found the same bare onnxsim invocation in onnx2tf 2.6.8. The reason onnxsim returns exit status 1 when found is unknown.

Environment

• macOS 26.5.2 (25F84) arm64, Apple M4, Python 3.12.12, no CUDA
• RF-DETR 1.9.3 from PyPI (sha256:3f19602ee7487dd80d8d2774e8dcd9380aba1eae9f86a8ba1c5d558b143236ef)
• onnx2tf 2.4.3, onnxsim 0.7.3, TensorFlow 2.19.1, torch 2.13.0, onnx 1.20.1

Minimal Reproducible Example

REPRO_ROOT="$(mktemp -d /tmp/rfdetr-onnxsim-clean.XXXXXX)"
REPRO_VENV="$REPRO_ROOT/venv"

python3.12 -m venv "$REPRO_VENV"
"$REPRO_VENV/bin/python" -m pip install --upgrade pip
"$REPRO_VENV/bin/python" -m pip install --no-cache-dir "rfdetr[tflite]==1.9.3"

cat > "$REPRO_ROOT/repro.py" <<'PY'
import os
import shutil
import sys

from rfdetr import RFDETRSmall

print("python:", sys.executable)
print("PATH:", os.environ["PATH"])
print("onnxsim:", shutil.which("onnxsim"))

RFDETRSmall().export(
    format="tflite",
    output_dir=sys.argv[1],
    shape=(512, 512),
    batch_size=1,
)
PY

env -u PYTHONPATH -u VIRTUAL_ENV \
  PATH="/usr/bin:/bin:/usr/sbin:/sbin" \
  "$REPRO_VENV/bin/python" "$REPRO_ROOT/repro.py" "$REPRO_ROOT/no-path"

The same command succeeds when PATH starts with "$REPRO_VENV/bin".

Are you willing to submit a PR?

• Yes, I'd like to help by submitting a PR!