🧪 VibeLab Coding Playground (vibelab)

Welcome to VibeLab's playground! This repository is a sandbox for experimenting with ideas, learning new tech, and storing useful code snippets.

---

🎯 Purpose

• Prototype new concepts quickly
• Explore different tool-chains and libraries
• Practise coding techniques
• Keep reference examples for future projects

---

🚀 Getting Started

1 Create a virtual environment (conda recommended)

conda create -n vbl python=3.12
conda activate vbl

2 Clone and install

git clone --recursive https://github.com/wvibe/vibe_coding.git
cd vibe_coding

# Installs every dependency *and* sets up editable packages
pip install -r requirements.txt

requirements.txt contains

-e .                   # editable install of src/vibelab
-e ./ext/ultralytics   # editable install of our forked Ultralytics submodule
-e ./ext/qwen2.5-vl/qwen-vl-utils  # editable install of Qwen VL utilities
<all third-party packages>

---

🗂 Project Structure

vibe_coding/
├─ configs/            # dataset / training / model YAML
├─ docs/               # design docs & guides
├─ ext/
│   ├─ ultralytics/    # YOLO models (fork)
│   └─ qwen2.5-vl/     # Qwen VLM repository (submodule)
├─ notebooks/          # Jupyter experiments
│   └─ vlm/            # Vision Language Model experiments
├─ ref/                # read-only reference code
├─ scripts/            # CLI utilities (not part of the package)
├─ src/
│   ├─ vibelab/        # ← top-level Python package (new)
│   │   ├─ models/
│   │   └─ utils/
│   └─ …legacy code    # to be moved into `vibelab`
├─ tests/              # mirrors src/ hierarchy
├─ requirements.txt
└─ setup.py

New code should live under src/vibelab/… and be imported as, for example:

from vibelab.utils.geometry import mask_to_yolo_polygons

Legacy from src.… imports will be migrated gradually.

---

⚙️ Environment Variables

Create a .env file in the repo root (example):

VIBE_ROOT=/abs/path/to/vibe
VHUB_ROOT=${VIBE_ROOT}/vhub
DATA_ROOT=${VHUB_ROOT}/data

---

🤖 Vision Language Models (VLM)

This repository includes support for Vision Language Models, particularly Qwen2.5-VL:

Features

• Multi-modal understanding: Process images, videos, and text together
• Fine-tuning capabilities: Train VLMs on custom datasets using TRL/PEFT
• Video processing: Support for long videos with temporal understanding
• Document parsing: Advanced OCR and document understanding

Quick Start

from qwen_vl_utils import process_vision_info
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor

# Load model and processor
model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")

# Process multimodal input
messages = [{"role": "user", "content": [
    {"type": "image", "image": "path/to/image.jpg"},
    {"type": "text", "text": "Describe this image."}
]}]

Notebooks

Check out notebooks/vlm/ for example notebooks including:

• Fine-tuning VLMs with TRL
• Multi-modal data processing
• Video understanding examples

---

🖥️ Cursor IDE Setup

For an enhanced experience with the Cursor IDE, follow the guide:

➡️ docs/cursor/setup-guide.md

---

👥 Contributing

• Work on a feature branch
• Keep commits focused and documented
• Feel free to add new folders for distinct experiments
• Update docs / notebooks with your findings

---

YOLO Training Scripts

The training scripts now live under vibelab/models/ext/…. Update your command lines accordingly, e.g.:

python src/vibelab/models/ext/yolov8/train_yolov8.py \
  --config src/vibelab/models/ext/yolov8/configs/voc_finetune_config.yaml \
  --name yolov8l_voc_finetune_run1

(Full instructions remain the same as before.)

---

📝 Note

This is an experimental playground—break things, fix them, and learn!

Happy coding! ✨