Skip to content

rework some pair quality metrics #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
amj wants to merge 2 commits into
base: cross_eval_pairer
Choose a base branch
from
Open

rework some pair quality metrics #4

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d367b33
rework some pair quality metrics
amj Feb 25, 2019
11e41a4
remove stray debug call
amj Feb 25, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
33 changes: 19 additions & 14 deletions cluster/evaluator/launch_eval.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -161,11 +161,12 @@ def get_cross_eval_pairs():
print ("Found", sum(game_counts.values()), "games")

model_game_counts = Counter()
for (winner, losser), count in game_counts.items():
for (winner, loser), count in game_counts.items():
model_game_counts[winner] += count
model_game_counts[losser] += count
model_game_counts[loser] += count

existing_pairs, previous_pairs = restore_pairs()
max_uncertainty = float(max([r[1] for r in rs.values()]))

# priority is roughly related to expected gain of information
pairs = []
Expand Down Expand Up @@ -207,32 +208,35 @@ def get_cross_eval_pairs():

# priority based on being highly ranked
# Higher = better
rank_num = max(ranks.get(model_a, 0), ranks.get(model_b, 0))
rank_adjustment = (1 + rank_num / len(rs)) ** 2 / 4
rank_num = min(ranks.get(model_a, 0), ranks.get(model_b, 0))

# Do this for a while
if (rank_num + 25 < len(rs)):
continue
if (rank_num + 25) < len(rs):
continue

rank_adjustment = (1 + rank_num / len(rs)) ** 2 / 4

# priority based on model variances
joint_uncertainty = (r_a[1] ** 2 + r_b[1] ** 2) ** 0.5
uncertainty_priority = joint_uncertainty / uncertainty_const
avg_u = (r_a[1] + r_b[1]) / 2.0
uncertainty_priority = avg_u / max_uncertainty

# priority based on information gained by playing this pairing
win_prob = 1 / (1 + 10 ** (-(r_a[0] - r_b[0])/400))
variance = win_prob * (1 - win_prob)
pairing_priority = equality_const * variance / (1 + games) ** games_power
win_prob = abs(1 / (1 + 10 ** (-(r_a[0] - r_b[0])/400)))
game_quality = win_prob * (1 - win_prob) + rank_adjustment

# priority based on playing a game with this model
model_priority = (1 / (1 + model_a_games) ** games_power +
1 / (1 + model_b_games) ** games_power)

priority = pairing_priority + model_priority + uncertainty_priority
priority = game_quality + model_priority + uncertainty_priority

# print("{}:{} vs {}:{}"
# "\t{:0.2f} {:0.2f} {:0.2f} = {:0.2f}".format(
# run_a, model_a, run_b, model_b,
# game_quality, model_priority, uncertainty_priority, priority))

pairs.append((
[rank_adjustment * priority,
rank_adjustment, win_prob, joint_uncertainty,
rank_adjustment, win_prob, uncertainty_priority,
games, model_a_games, model_b_games],
pair))

Expand Down Expand Up @@ -414,6 +418,7 @@ def make_pairs_for_model(model_num=0):
'same_run_eval': same_run_eval,
'cross_run_eval': cross_run_eval,
'cleanup': cleanup,
'get_cross_eval_pairs': get_cross_eval_pairs,
'add_top_pairs': add_top_pairs,
'launch_eval_job': launch_eval_job,
}, remaining_argv[1:])