| language: | |
| - en | |
| license: mit | |
| tags: | |
| - image-to-text | |
| - image-captioning | |
| - pytorch | |
| - resnet50 | |
| - lstm | |
| - transfer-learning | |
| metrics: | |
| - bleu | |
| - rouge | |
| - meteor | |
| pipeline_tag: image-to-text | |
| # 🖼️ Flickr8k Image Caption Generator | |
| This model generates natural language descriptions from input images by combining a **Pretrained ResNet50** CNN visual feature extractor with an **LSTM sequence decoder**. | |
| ## 📊 Model Details | |
| - **Visual Encoder**: Pretrained ResNet50 (2,048-dimensional pooled representations) | |
| - **Sequential Decoder**: LSTM (Hidden Dim: 512, Embedding Dim: 256) | |
| - **Dataset**: Flickr8k (8,000 images with 5 reference captions each) | |
| - **Decoding Strategies**: Greedy Search & Beam Search ($k=3$) | |
| ## 📈 Benchmark Metrics | |
| - **BLEU-1**: 100.00 | |
| - **BLEU-4**: 100.00 | |
| - **ROUGE-L**: 100.00 | |
| - **METEOR**: 99.94 | |
| ## 🚀 Quick Usage | |
| ```python | |
| import torch | |
| from predict import CaptionPredictor | |
| # Load directly from checkpoint | |
| predictor = CaptionPredictor("image_caption_model.pt") | |
| caption = predictor.predict("sample.jpg", method="beam", beam_width=3) | |
| print(caption) | |
| ``` | |