VisualStorytelling

My bachelor's thesis -- Visual Storytelling -- an encoder-decoder model consisting of both convolutional and attentional recurrent networks that transform a sequence of 5 images into a context-generative story in natural language (english).

Base model architecture

Results

From the above-illustrated architecture, two models were trained and compared:

• Encoding the whole album's into one embedding vector (CS)
• Encoding each photo separately to its own embedding vector and transmitting them to the decoder sequentially (CC)

How overfitting, even though it increases METEOR and ROGUE score, harm the results