A team has 1,000 product reviews to classify.
| review length | how many reviews |
|---|
| 5 tokens | 600 |
| 10 tokens | 300 |
| 50 tokens | 100 |
There are two ways to run them.
- Padded — the fixed-input approach from Chapter 1. Take the longest review, 50 tokens, and pad every shorter review with blanks up to that length. Every review then occupies 50 slots, and every slot is processed.
- Unrolled — an RNN, unrolled to each review's own length. A five-token review runs the cell five times and stops.
Call one run of the cell on one slot a token-step.
Across all 1,000 reviews, what percentage of the padded approach's token-steps is spent on blanks?
Round to the nearest whole percent.