Implemenetation of 2016 paper "Show, Attend and Tell: Neural Image Caption Generation with Visual Attention" on Flickr8k dataset.

1. Overview

   • Image Captioning using high level VGG19 features
   • Soft and Hard Attention
   • Focus will be on decoder

2. Encoder

   • VGG19 will take 224 * 224 * 3 image and generate 512 feature maps of size 14 * 14
   • Each pixel of 14x14 feature map represents a region of image, so we have 512 features for a region of image.
   • We will represent 196 regions with 512 features for each region.

3. Attention

   • Soft attention is when we calculate the context vector as a weighted sum of the encoder features.
   • Hard attention is when, instead of weighted average of all encoder features, we use attention scores to select a single hidden state.
   • Here we will use Soft Attention.