Spaces:
Running
on
Zero
Running
on
Zero
DongfuJiang
commited on
Commit
•
6ff1b6e
1
Parent(s):
583f0ee
cutoff 6.04
Browse files- arena_elo/elo_rating/elo_analysis.py +14 -14
- arena_elo/results/20240602/elo_results_image_editing.pkl +2 -2
- arena_elo/results/20240604/elo_results_t2i_generation.pkl +2 -2
- arena_elo/results/20240604/elo_results_video_generation.pkl +2 -2
- arena_elo/results/latest/elo_results_image_editing.pkl +2 -2
- arena_elo/results/latest/elo_results_t2i_generation.pkl +2 -2
- arena_elo/results/latest/elo_results_video_generation.pkl +2 -2
arena_elo/elo_rating/elo_analysis.py
CHANGED
@@ -171,20 +171,20 @@ def visualize_pairwise_win_fraction(battles, model_order):
|
|
171 |
xaxis_side="top",
|
172 |
title_y=0.07,
|
173 |
title_x=0.5,
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
)
|
183 |
fig.update_traces(
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
hovertemplate="Model A: %{y}<br>Model B: %{x}<br>Fraction of A Wins: %{z}<extra></extra>"
|
189 |
)
|
190 |
|
@@ -379,11 +379,11 @@ if __name__ == "__main__":
|
|
379 |
|
380 |
|
381 |
# # save heatmap results in the same directory of the cleaned battle file
|
382 |
-
|
383 |
# battle_count_heatmap_file = args.clean_battle_file.replace(".json", "_battle_count_heatmap.jpg")
|
384 |
# average_win_rate_bar_file = args.clean_battle_file.replace(".json", "_average_win_rate_bar.jpg")
|
385 |
# bootstrap_elo_rating_file = args.clean_battle_file.replace(".json", "_bootstrap_elo_rating.jpg")
|
386 |
-
|
387 |
# anony_results["battle_count_heatmap"].write_image(battle_count_heatmap_file)
|
388 |
# anony_results["average_win_rate_bar"].write_image(average_win_rate_bar_file)
|
389 |
# anony_results["bootstrap_elo_rating"].write_image(bootstrap_elo_rating_file)
|
|
|
171 |
xaxis_side="top",
|
172 |
title_y=0.07,
|
173 |
title_x=0.5,
|
174 |
+
xaxis=dict(
|
175 |
+
tickfont=dict(size=16),
|
176 |
+
title=dict(font=dict(size=16)),
|
177 |
+
),
|
178 |
+
yaxis=dict(
|
179 |
+
tickfont=dict(size=16),
|
180 |
+
title=dict(font=dict(size=16)),
|
181 |
+
),
|
182 |
)
|
183 |
fig.update_traces(
|
184 |
+
textfont=dict(size=16),
|
185 |
+
colorbar=dict(
|
186 |
+
title=dict(font=dict(size=16))
|
187 |
+
),
|
188 |
hovertemplate="Model A: %{y}<br>Model B: %{x}<br>Fraction of A Wins: %{z}<extra></extra>"
|
189 |
)
|
190 |
|
|
|
379 |
|
380 |
|
381 |
# # save heatmap results in the same directory of the cleaned battle file
|
382 |
+
win_fraction_heatmap_file = args.clean_battle_file.replace(".json", "_win_fraction_heatmap.jpg")
|
383 |
# battle_count_heatmap_file = args.clean_battle_file.replace(".json", "_battle_count_heatmap.jpg")
|
384 |
# average_win_rate_bar_file = args.clean_battle_file.replace(".json", "_average_win_rate_bar.jpg")
|
385 |
# bootstrap_elo_rating_file = args.clean_battle_file.replace(".json", "_bootstrap_elo_rating.jpg")
|
386 |
+
anony_results["win_fraction_heatmap"].write_image(win_fraction_heatmap_file)
|
387 |
# anony_results["battle_count_heatmap"].write_image(battle_count_heatmap_file)
|
388 |
# anony_results["average_win_rate_bar"].write_image(average_win_rate_bar_file)
|
389 |
# anony_results["bootstrap_elo_rating"].write_image(bootstrap_elo_rating_file)
|
arena_elo/results/20240602/elo_results_image_editing.pkl
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:51cebad92fed9a2c15871c3dae4ef1ffbbaea02b5906f3100e5e9f1dc25e97da
|
3 |
+
size 62732
|
arena_elo/results/20240604/elo_results_t2i_generation.pkl
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:787326f9d085b668bdf7d2f24439f7080a402c7333423f4856c86ef1c11288b8
|
3 |
+
size 68373
|
arena_elo/results/20240604/elo_results_video_generation.pkl
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e67b05d84d6bccdaefaba734220e3a87d1ef897a4969f415a8a94831a2a9d9a1
|
3 |
+
size 60162
|
arena_elo/results/latest/elo_results_image_editing.pkl
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:51cebad92fed9a2c15871c3dae4ef1ffbbaea02b5906f3100e5e9f1dc25e97da
|
3 |
+
size 62732
|
arena_elo/results/latest/elo_results_t2i_generation.pkl
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:787326f9d085b668bdf7d2f24439f7080a402c7333423f4856c86ef1c11288b8
|
3 |
+
size 68373
|
arena_elo/results/latest/elo_results_video_generation.pkl
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e67b05d84d6bccdaefaba734220e3a87d1ef897a4969f415a8a94831a2a9d9a1
|
3 |
+
size 60162
|