Spaces:
Running on Zero
Running on Zero
| title: Image Caption Generator | |
| emoji: 👁 | |
| colorFrom: red | |
| colorTo: yellow | |
| sdk: gradio | |
| sdk_version: 6.26.0 | |
| python_version: '3.12' | |
| app_file: app.py | |
| pinned: false | |
| # Image Caption Generator | |
| Upload an image and get an automatically generated caption. Built with a **ResNet50 (frozen, transfer learning) + LSTM decoder**, trained on the [Flickr8k dataset](https://www.kaggle.com/datasets/adityajn105/flickr8k). | |
| This Space runs the **regularized model checkpoint with beam search decoding** (beam width 3) — the best-performing configuration from the project's optimization experiments (BLEU-4 = 0.1557 on the held-out test set, a ~27% relative improvement over the initial baseline). | |
| - **Full project, code, and documentation:** https://github.com/adhamashraf7788/Image-Caption-Generator | |
| - **Model weights:** https://huggingface.co/AdhamAshraf/image-caption-generator | |
| ## How it works | |
| 1. Image is resized/cropped/normalized and passed through a frozen, ImageNet-pretrained ResNet50 to extract a 2048-d feature vector. | |
| 2. The feature vector is projected into a 256-d embedding and fed as the first input to an LSTM decoder. | |
| 3. The LSTM generates the caption word-by-word using beam search (keeping the top-3 candidate sequences at each step, rather than greedily committing to a single best word). | |
| See the [main repo's README](https://github.com/adhamashraf7788/Image-Caption-Generator) for full details on the dataset, architecture, training process, evaluation metrics, and optimization experiments (beam search + regularization comparisons). | |
| Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference |