Argus: Visual Intelligence at your hand

A from-scratch implementation of vision models for image classification, served via a Flask API and deployed at argus.iashyam.in.

What it does

Upload any image, get a classification label from 1000 ImageNet categories. The backend runs custom-trained vision models built entirely in PyTorch — no pretrained backbone dependencies at inference time.

Models

Model	Paper	Weights
SimpleCNN	Custom architecture	Trained on CIFAR-100
MobileNet V2	MobileNetV2: Inverted Residuals and Linear Bottlenecks	burrah_mobilenet_v1.pth

Both models are implemented from scratch in src/models/.

Stack

• Framework: PyTorch + Flask
• ETL: Custom ImageDataset pipeline (src/ETL/)
• Training: Trainer class with train/eval loops, tqdm progress (src/Train/)
• Experiment tracking: MLFlow
• CI: GitHub Actions (lint + pytest on push to main)
• Deployment: Docker + Cloudflare Tunnel

Project structure

src/
  models/     # SimpleCNN and MobileNet V2 implementations
  ETL/        # Dataset loading and preprocessing
  Train/      # Training loop (Trainer class)
  utils/      # ImageNet 1000-class label map
  weights/    # Saved model weights
Data/         # CIFAR-10 and CIFAR-100 datasets
tests/        # pytest test suite

Run locally

Docker (recommended)

cp .env.example .env
# Add your CLOUDFLARE_TUNNEL_TOKEN to .env (optional, only needed for tunnel)

docker compose up

App runs at localhost:5911.

From source

pip install .
# run the Flask app
python -m app

Train a model

from src.Train.train import Trainer
import torch

trainer = Trainer(model, criterion, optimizer, device=torch.device("cuda"))
history = trainer.train_loop(n_epochs=20, train_dataloader=..., test_dataloader=...)

Data

CIFAR-10 and CIFAR-100 batches are in Data/. To download:

python Data/Download.py

Roadmap

• Containerize the model with Docker.
• Put automated testing with github actions.
• Deploy the model to server.
• Write a ETL pipeline.
• Write training loop.
• Write documentation for ETL.
• Use ML Flow for experiment tracking
• User DVC pipelines.
• Run some basic experiments.

Contributing

Personal project, not accepting contributions right now. EfficientNet is next.