Model Studio Workflow

This repository contains a YOLO workflow with:

• Inference GUI for images/videos.
• Polygon labeling GUI and YOLO-seg dataset export.
• Pseudo-label + train pipeline (Optimize tab).
• Direct training (detect or obb) in Training tab.
• .pt -> .onnx conversion in Converter tab.

Project Layout

• scripts/python/: GUI and Python automation scripts.
• scripts/bash/: shell scripts (including AMD OBB training launcher).
• runs/: training outputs.
• models/: model artifacts (ONNX, checkpoints, exports).
• pt/: local .pt models.
• Backup/: archived datasets/images/models.
• requirements.rocm.txt: ROCm stack Python requirements.
• requirements.nvidia.txt: NVIDIA/CUDA stack Python requirements.

Environment Setup

AMD ROCm

python -m venv .venv-rocm --system-site-packages
source .venv-rocm/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.rocm.txt

NVIDIA CUDA

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.nvidia.txt

Launch GUI

.venv-rocm/bin/python scripts/python/model_viewer_gui.py

Use .venv/bin/python on NVIDIA if needed.

Build AppImage

1. Install build prerequisites in your active environment:

python -m pip install pyinstaller

2. Install appimagetool on your system, or place appimagetool-*.AppImage in tools/.

3. Build the AppImage:

./scripts/bash/build_appimage.sh .venv-rocm/bin/python

Output will be generated in dist/ as ModelStudio-<version>-<arch>.AppImage.

GUI Tabs

Inference

• Load model (.pt, .onnx, etc.), image folder, or video.
• Adjustable confidence threshold.

Labeling

• Draw polygon annotations per class.
• Export YOLO-seg dataset with train/test/valid splits.

Optimize

• Pseudo-label unlabeled images, then train.
• Platform profiles: AMD / NVIDIA / CPU.
• Uses scripts/python/pseudo_label_and_train.py.

Training

• Task selector: detect or obb.
• Platform selector: AMD (ROCm) or NVIDIA (CUDA).
• Runs task-aware training command and output paths (runs/detect or runs/obb).

Converter

• Convert .pt to .onnx with task/imgsz/opset/runtime controls.
• Output folder must already exist (converter does not create folders).

Dataset Path Behavior

The GUI now resolves dataset paths from your current structure and falls back across common locations:

• data/...
• Backup/Data/...

When switching training task (obb <-> detect), dataset YAML defaults are updated automatically if current value is old/invalid.

CLI: AMD OBB Training

./scripts/bash/train_obb_amd.sh --device 0 --batch 8

The script now auto-resolves default OBB dataset YAML from available data/ or Backup/Data/ locations.

CLI: Direct Ultralytics Training

Detect

.venv-rocm/bin/yolo detect train \
  data=/absolute/path/to/data.yaml \
  model=yolo11n.pt \
  epochs=100 imgsz=640 batch=8 device=0 workers=8 \
  project=runs/detect name=my_detect_run

OBB

.venv-rocm/bin/yolo obb train \
  data=/absolute/path/to/data.yaml \
  model=yolo11n-obb.pt \
  epochs=100 imgsz=640 batch=8 device=0 workers=8 \
  project=runs/obb name=my_obb_run

Notes

• ONNX inference models should be loaded with the correct task (detect, obb, etc.).
• For AMD/ROCm setups, validation/AMP settings may need conservative values depending on torchvision/ROCm compatibility.