Introduce the Skip Links to the Hypotheses Graph, that jump over i>2 frames.
Because we want the normal links to be favored though, we impose a greater energy for the SkipLinks than for the normal ones.
Solution now: works without any extra multiplicative parameter from structured learning (in weights.json), but has a bias parameter --skip-links-bias (see /scripts/hypotheses_graph_to_json.py).
Commented out solution: extra parameter in weights.json. In contrast to the bias parameter, that was easily found, did not manage to find a sensible value for the multiplicative extra parameter (yet?).
Oh and one more thing: could you please add a test to tests/core/test_hypothesesgraph.py for the skip links? Then we'll know whether it still works after merging it with master.
Introduce the Skip Links to the Hypotheses Graph, that jump over i>2 frames.
Because we want the normal links to be favored though, we impose a greater energy for the SkipLinks than for the normal ones.
Solution now: works without any extra multiplicative parameter from structured learning (in weights.json), but has a bias parameter --skip-links-bias (see /scripts/hypotheses_graph_to_json.py).
Commented out solution: extra parameter in weights.json. In contrast to the bias parameter, that was easily found, did not manage to find a sensible value for the multiplicative extra parameter (yet?).