"FrontEnd API failed with InitializationFailure" when converting YOLOv8 from Ultralytics

[msmiatac]

When running a script to convert the YoloV8 model from Ultralytics (https://huggingface.co/Ultralytics/YOLOv8/resolve/main/yolov8n.pt), I encounter the following error on one of the machines:

OpenVINO: export failure ❌ 10.8s: Check 'false' failed at src/frontends/common/src/manager.cpp:64:
FrontEnd API failed with InitializationFailure:
FrontEnd for Framework pytorch is not found

However, on other machines with seemingly identical configurations, the conversion completes successfully. Has anyone encountered this issue or have any suggestions on how to resolve it?

Here is the conversion script:

#!/bin/bash
#Set the name of the virtual environment directory
VENV_DIR="yolo_venv"

#Create a Python virtual environment if it doesn't exist
if [ ! -d "$VENV_DIR" ]; then
echo "Creating virtual environment in $VENV_DIR..."
python3 -m venv "$VENV_DIR"
fi

#Activate the virtual environment
source "$VENV_DIR/bin/activate"

#Upgrade pip in the virtual environment
pip install --upgrade pip

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

pip install openvino
pip install ultralytics

#Function to export YOLO model
export_yolo_model() {
local weights_path=$1 # User-provided weights path

if [ ! -f "$weights_path" ]; then
  echo "Weights file not found: ${weights_path}"
  return 1
fi

echo "Using weights from: ${weights_path}"

python3 - <<EOF "$weights_path"

import sys
from ultralytics import YOLO
import openvino

weights_path = sys.argv[1]
model_type = "YOLOv8"

try:
model = YOLO(weights_path)
model.info()
converted_path = model.export(format='openvino')

except Exception as e:
print(f"An error occurred during model conversion: {e}")
EOF
}

#Check if a weights path is provided as an argument
if [ $# -eq 0 ]; then
echo "Usage: $0 <path_to_weights_file>"
exit 1
fi

#Call the function with the user-provided weights path
export_yolo_model "$1"

[Ganesh-403]

@msmiatac The key part of the error is:

FrontEnd API failed with InitializationFailure:
FrontEnd for Framework pytorch is not found

Since the same script works on other machines, I'd first suspect an environment difference rather than an issue with the YOLO model itself.

A few things to check:

1. Verify the OpenVINO installation

   The PyTorch frontend is shipped as part of the OpenVINO installation. You can check which frontends are available with:

   from openvino.frontend import FrontEndManager
   
   fem = FrontEndManager()
   print(fem.get_available_front_ends())

   If pytorch is missing from the list on the failing machine, that would explain the error immediately. OpenVINO's FrontEndManager only loads frontends that are available in the installation. :contentReference[oaicite:0]{index=0}

2. Check the installed package versions

   On both the working and failing machines, compare:

   pip show openvino
   pip show ultralytics
   pip show torch

   Recent Ultralytics versions also require PyTorch ≥ 2.1 for OpenVINO export, so it's worth confirming that both environments satisfy the same requirements. :contentReference[oaicite:1]{index=1}

3. Make sure only one OpenVINO installation is being used

   Sometimes an older system-wide OpenVINO installation (or leftover runtime libraries) gets picked up instead of the one installed in the virtual environment.

   You can verify:

   import openvino
   print(openvino.__file__)
   print(openvino.__version__)

   on both machines to ensure they're loading the same package.

4. Try a clean virtual environment

   Since your script installs everything from scratch, I'd also try:

   python -m venv clean_env
   source clean_env/bin/activate
   
   pip install --upgrade pip
   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
   pip install openvino ultralytics

   and then rerun the export.

One question that would help narrow it down:

• What is the output of:

  from openvino.frontend import FrontEndManager
  print(FrontEndManager().get_available_front_ends())

  on both the working and failing machines?

If the failing machine doesn't list the PyTorch frontend, then the issue is likely an incomplete or mismatched OpenVINO installation rather than anything specific to Ultralytics or your conversion script.

If this solves your problem, feel free to mark it as the accepted answer so others can find it easily.