A YOLOv5 model trained on modified TACO dataset to perform object detection on waste and rubbish. Receiving an accuracy as high as 87.6%, it can confidently predict a wide variety of recyclable rubbish lying around us.
Modified official YOLOv5 training script was used to train the model, for 100 Epochs. (Best results are at 200)
In the repository, I have attached the trained weights in the form of “best.pt” file, which was later used to perform inference via torch.load.
And a number of different flavours of saved versions of the model was attached with this repository as well.
We used a modified TACO dataset. This version of the dataset was created by @manaporkun. A special thanks to him for making it open source.
Inference has always been a headache, ugh! Even with Gradio and Streamlit, it doesn’t get better. Which is why, we resorted to load and make a web application (flask as backend) to perform inference.
I always prefer clear code than shabby coding where authors try to make everything complex and stranded me with only !python command. Like W!! Which is why, I used Flask as a backend to perform inference while using honey sweet pytorch (Love!)
(Jupyter Notebooks are provided for learners to learn) (Experts run app.py) :P (CPU is fine, I love hot GPU air)
Quite simple! I have stored all the notebooks, I used both for training and interference.
- 'notebooks/hufload_pytorch.ipynb' -> inteference notebook
notebooks/yolo5_trash.ipynb-> training notebook
- Open
terminal - Navigate to the repository
- Inside the repo run
python app.py
- Labelling TrashNet dataset for Object Detection. (Stanford Version)
- Releasing TrashNet weights
- Finding waste food costs :P
- DarkNet (haha! _)
Special thanks to Roboflow for such wonderful free support! Couldn’t make the dataset without you.
A project developed by me under my team Alpha Tauri. Licenced under Open Source GPL.
We believe in developing human friendly Software!