Testing/inference
At the inference stage, the image representation is provided to the first time step of the decoder. Set x1 =<START> vector and compute the distribution over the first word, y1. One could either sample a set of words from the distribution, or greedily choose the argmax over the computed probability of the next word, or sample the top ones from the returned candidates (Beam Search). If using LSTM, both the current state and the output will be used in the next time step as input information. The system repeats this process until it hits an <EOS> token.
During testing, the output of the decoder at time t is fed back and becomes the input of the decoder at time t + 1 (RNN/LSTM unrolling):
The preceding figure shows a selection ...
Become an O’Reilly member and get unlimited access to this title plus top books and audiobooks from O’Reilly and nearly 200 top publishers, thousands of courses curated by job role, 150+ live events each month,
and much more.
Read now
Unlock full access