conditioning-transformer

Design of a transformer-based architecture for object detection conditioned by metadata:

• DEtection TRanformer (DETR)
• You Only Look at One Sequence (YOLOS)

Installation

To install the project, simply clone the repository and get the necessary dependencies. Then, create a new project on Weights & Biases. Log in and paste your API key when prompted.

# clone repo
git clone https://github.com/MarcoParola/conditioning-transformer.git
cd conditioning-transformer
mkdir models data

# Create virtual environment and install dependencies 
python -m venv env
. env/bin/activate
python -m pip install -r requirements.txt 

# Weights&Biases login 
wandb login 

Usage

To perform a training run by setting model parameter that can assume the following value detr, early-sum-detr, early-concat-detr, yolos, early-sum-yolos, early-concat-yolos

python train.py model=detr

The command could also be run specifying the cropBackground option by setting it at true or false resulting on the following training image.

Whole image	Cropped image

To run inference on test set to compute some metrics, specify the weight model path by setting weight parameter (I ususally download it from wandb and I copy it in checkpoint folder).

python test.py model=detr weight=checkpoint/best.pt

Training params

Training hyperparams can be edited in the config file or ovewrite by shell

Params	Value
batchSize	16
lr	1e-6

Acknowledgement

Special thanks to @clive819 for making an implementation of DETR public here. Special thanks to @hustvl for YOLOS original implementation