Trolo

A framework for harnessing the power of transformers with YOLO models and other single-shot detectors!

Note: This is an early release. The package is under active development. Please report any issues and I'll try to fix them ASAP.

Quickstart on Colab:

🔥 NEW 🔥 D-Fine models are now available. Inspired by RT-DETR outperform all real-time detectors including YOLO-series models

Installation

pip install trolo

Features

• 🔥 Transformer-enhanced object detection
• 🎯 Single-shot detection capabilities
• ⚡ High performance inference
• 🛠️ Easy to use CLI interface
• 🚀 Fast video stream inference
• 🧠 Automatic DDP handling

Available Models

D-FINE

The D-FINE model redefines regression tasks in DETR-based detectors using Fine-grained Distribution Refinement (FDR). Official Paper | Official Repo

( All models will be automatically downloaded when you pass the name for any task)

Model	Dataset	APval	#Params	Latency	GFLOPs
dfine-n	COCO	42.8	4M	2.12ms	7
dfine-s	COCO	48.5	10M	3.49ms	25
dfine-m	COCO	52.3	19M	5.62ms	57
dfine-l	COCO	54.0	31M	8.07ms	91
dfine-x	COCO	55.8	62M	12.89ms	202

RT-DETR v3 (Coming Soon)

RT-DETR v2 (Coming Soon)

Trolo-2024 (WIP)

Quick Start

The CLI command structure is:

trolo [command] [options]

For detailed help:

trolo --help # for general help
trolo [command] --help # for command-specific help

Inference

Example inference command:

trolo predict --model dfine-n # automatically downloads model from trolo model hub

Support for single image, image folder, and video input

trolo predict --model dfine-n.pth --input img.jpg # folder/ or video.mp4 or 0 (for webcam)

🔥 Smart Video stream inference - infers on videos in streaming mode so you never have to worry about memory issues!

Python API:

from trolo.inference import DetectionPredictor

predictor = DetectionPredictor(model="dfine-n")
predictions = predictor.predict() # get predictions
poltted_preds = predictor.visualize(show=True, save=True) # or get visualized outputs

Visit Inference Docs for more details

Model export

Example export command:

trolo export --model dfine-n --export_format onnx --input_size 640  

Python API:

python from trolo.inference import ModelExporter model_path = "/path/to/model"  
input_size = 640 # Inference resolution  
export_format = "onnx"  
exporter = ModelExporter(model=model_path) exporter.export(input_size=input_size, export_format=export_format)  

Visit Export Docs for more details. Please check deployment for inference script for various deployment.

Training

Example training command:

trolo train --config dfine_n # automatically find the config file

🔥 Automatically handle DDP by simply passing the GPUs to the CLI

trolo train --config dfine_n --device 0,1,2,3 

That's it!

Python API

from trolo.trainers import DetectionTrainer

trainer = DetectionTrainer(config="dfine_n") # or pass custom config path
trainer.train() # pass device = 0,1,2,3 to automatically handle DDP 

Visit Training Docs for more details

Docker 🐳 Usage Guide

Build and Run Options

Build Commands

# Standard build
docker build -t trolo .

# Build with a specific tag
docker build -t trolo:v1 .

# Build with build arguments (if needed)
docker build --build-arg SOME_ARG=value -t trolo .

Run Commands

1. Basic Run (No GPU, No Volume)

docker run -it --name containaer_name  trolo

2. Run with GPU Mounting

docker run -it --gpus  --name containaer_name  all trolo

3. Run with Volume Mounting

docker run -it -v /local/path/on/host:/workspace/app/trolo --name containaer_name trolo

4. Comprehensive Run (GPU and Volume)

docker run -it --gpus all -v /local/path/on/host:/workspace/app/trolo --name containaer_name  trolo

5. Additional Run Options

# Run with custom entrypoint
docker run -it --entrypoint /bin/bash trolo

# Run with environment variables
docker run -it -e CUSTOM_ENV=value trolo

# Run in detached mode
docker run -d trolo

Notes:

• Replace /local/path/on/host with your actual host path
• --gpus all requires NVIDIA Container Toolkit
• Volume mounting allows persistent data and code modifications

Totally open source and free

TLDR: This is a non-profit project.Use it, modify it, copy it, do whatever you want with it. And if something doesn't allow you to do that, please open an issue.

More details

• Apache 2.0

• The license has simply been copied from official apache repo. Please open an issue if something doesn't allow you to use it.

• This project is built on top of open licensed projects as mentioned below.

• I intend to keep this project free and open source FOREVER. There are no plans for direct/indirect monetization of this project. I only accept sponsorships for compute resources to train models and perform independent research.

Credits

This project builds upon several excellent open source projects:

• D-FINE: Original D-FINE model implementation
• RT-DETR: Real-time DETR architecture
• PaddlePaddle: Detection framework

More details

- The original trainer is based on D-fine with major modifications for handling pre-trained weights, DDP, and other features. - The architecture is for D-fine is same as the original paper and repo.

Contributing

Contributions are most welcome! Please feel free to submit a Pull Request.

---

Note: This is an early work in progress. Many features are still under development.

Immidiate TODOs

• Docusaurus documentation