Implementation of of the paper Show-Atten-Tell and use of beam search.
Bleu score of 22 with beam size of 3.
Flickr8k dataset, 8000 images 64x64 pixels.
Execute training:
python train.py
You can choose to use pre-trained embedding.
Change the WANDB_TRACK = False to WANDB_TRACK = True
The full training and evaluation of metrics is shown in this notebook