kevinhuads/deepvision-workflow

Deepvision Workflow · Food-101 Image Classification

This Space hosts the Streamlit demo for the Deepvision Workflow project, a complete computer vision workflow built around the Food-101 dataset.

The application allows interactive exploration of a fine-tuned vision transformer model that predicts the dish category from a single food image.

What this demo does

• —Accepts a single food image upload (JPEG or PNG)
• —Runs inference with a fine-tuned backbone on the Food-101 dataset (101 food classes)
• —Displays the top-k predicted classes with their associated probabilities
• —Provides a short textual interpretation of the model’s confidence
• —Summarises the model and dataset used in the underlying project

All preprocessing and inference logic is shared with the training and evaluation code in the main repository to ensure consistent behaviour across experimentation and the demo.

How to use

1. 1.Upload a food image in the sidebar.
2. 2.Optionally adjust the number of top-k predictions to display.
3. 3.Inspect:
4. 4.The predicted classes and their probabilities
5. 5.The confidence commentary
6. 6.The accompanying description of the model and dataset

Close-up images of a single dish generally produce the most informative predictions.

Project background

This demo is part of a broader project that covers:

• —Exploratory analysis of Food-101 and pretrained visual embeddings
• —Benchmarking of modern CNN and transformer architectures
• —Detailed evaluation (per-class metrics, calibration, error analysis)
• —MLOps components (MLflow tracking, tests, Docker, CI/CD, Streamlit demo)

Full source code, documentation and notebooks are available in the main repository:

GitHub: Deepvision Workflow