Anas1010/flickr8k-image-caption-generator-attn

Flickr8k Image Caption Generator (ResNet50 Spatial + Bahdanau Attention + LSTM)

This repo contains an attention-based captioning model trained on Flickr8k.

Architecture

• Encoder: ResNet50 backbone feature map (2048x7x7), cached
• Projected encoder dim: 512
• Attention: Bahdanau (additive) over 49 spatial locations
• Decoder: LSTMCell + greedy decoding
• Vocab: train-only, min_freq=2, max_len=20

Test Metrics (Flickr8k)

• BLEU-1: 0.5895
• BLEU-2: 0.4030
• BLEU-3: 0.2651
• BLEU-4: 0.1820
• ROUGE-L(F1): 0.4597
• METEOR: 0.4033

Files

• best_model_attn.pt
• vocab.json
• test_metrics_attn.json