Organic Analyzer Model

English Українська

A bilingual Streamlit application for biological-species classification with dedicated ResNet-50 models trained on iNaturalist 2021 Full and the image-only labeled subset of BIOSCAN-5M. The interface supports animals, insects, plants and fungi, top predictions, confidence indicators, input-quality warnings, Google search links, and Grad-CAM visualization.

Model output is intended for research and education and requires expert review.

Current datasets

Status	Dataset	Purpose	Scale
Current	iNaturalist 2021 Full	Classification of animals, plants, fungi, and other organisms	10,000 classes; 2,686,843 training images
Current	BIOSCAN-5M image-only labeled subset	Insect species classification without DNA input	11,846 classes; 289,203 training images

System requirements

Component	Minimum	Recommended
Operating system	64-bit Windows 10/11 or a current 64-bit Linux distribution	64-bit Windows 11
Python	3.12 or 3.13, 64-bit	Python 3.12, 64-bit
Processor	4-core x64 processor	6 or more modern cores
Memory	8 GB RAM	16 GB RAM or more
Storage	9 GB of free space for the application, environment, and checkpoints	12 GB or more
Graphics	CPU inference is supported	NVIDIA GPU with at least 6 GB VRAM and a compatible CUDA-enabled PyTorch build
Network	Required to install Python packages	Not required for local image analysis after installation

Training datasets are not required to run image analysis. GPU acceleration is optional; without a compatible NVIDIA GPU, the application automatically uses the CPU and processes images more slowly.

Run the application

Windows: installation without terminal commands

1. Install 64-bit Python 3.12 and enable Add Python to PATH in the installer.
2. Download or clone the complete repository.
3. Double-click install.bat. It creates .venv and installs all packages automatically. Run it only once, or again when requirements.txt changes.
4. Add the inference files using the structure below.
5. Double-click start_app.bat whenever you want to start the application. Keep its console window open while the application is in use; closing it stops the local server.

checkpoints/
├── inaturalist/
│   └── resnet50_inaturalist_best.pt
└── bioscan/
    ├── bioscan_species_resnet50_best.pt
    └── classes.json

The application normally opens http://localhost:8501 automatically. Other devices on the same local network can use http://COMPUTER_IP:8501 if Windows Firewall permits the connection.

Manual or Linux launch

Create a virtual environment, install requirements.txt, and run python -m streamlit run app.py. The BAT files are intended for Windows only.

The training datasets are not needed for inference. Only the three files listed above are required.

Connection log

The application records the connection time and client IP address once per browser session in logs/access.log. Localhost connections are recorded as local. The log rotates automatically at 5 MB and retains up to three older files. IP addresses can be personal data, so disclose this logging to users and define an appropriate retention period before making the application publicly accessible.

Project structure

app.py                     Streamlit interface
install.bat                One-click Windows environment setup
start_app.bat              One-click Windows application launcher
models/inaturalist/        Training configuration and entry points
models/bioscan/            BIOSCAN configuration, data loader, training and evaluation
src/                       Model, data, inference, taxonomy, and Grad-CAM code
checkpoints/inaturalist/   Best and latest iNaturalist checkpoints
checkpoints/bioscan/       Best/latest BIOSCAN checkpoints and class mapping
results/inaturalist/       Training and evaluation outputs
tests/                     Automated tests