Anas1010/flickr8k-image-caption-generator

Flickr8k Image Caption Generator (ResNet50 + LSTM)

This repository contains a trained image captioning model.

Architecture

• Image encoder: ResNet50 (ImageNet pretrained) feature extractor (2048-d)
• Decoder: LSTM (hidden=256) with greedy decoding
• Vocabulary: built from train captions only (min_freq=2), max_len=20

Test Metrics (Flickr8k)

• BLEU-1: 0.5647
• BLEU-2: 0.3731
• BLEU-3: 0.2453
• BLEU-4: 0.1685
• ROUGE-L(F1): 0.4350
• METEOR: 0.3678

Files

• best_model.pt: model weights
• vocab.json: vocabulary + max_len + special tokens
• test_metrics.json: evaluation metrics
• train_history.json: training curves data