Spaces:
Running
Running
File size: 4,597 Bytes
98847a8 a1f1bf8 98847a8 a1f1bf8 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 |
import collections
import pandas as pd
def get_leaderboard_filters(df, categories) -> dict[str, list[str]]:
# Create groups based on categories
groups = collections.OrderedDict({"Overall": set()})
for k in categories.values():
groups[k] = set()
default_selection = set()
for k, v in categories.items():
if v not in default_selection:
for k in list(df.columns):
if k.startswith(v):
groups["Overall"].add(k)
default_selection.add(k)
for col in list(df.columns):
for k in categories.keys():
if col.startswith(k):
cat = categories[k]
groups[cat].add(col)
break
return groups, default_selection
def add_avg_as_columns(
benchmark_df: pd.DataFrame, attack_scores: list[str]
) -> pd.DataFrame:
# average over the attack variants (inequal number of attack variants)
attack_avg_df = (
benchmark_df[["model", "attack_name"] + attack_scores]
.groupby(["model", "attack_name"])
.agg(lambda x: x.mean(skipna=False))
.reset_index(drop=False)
)
avg_df = (
attack_avg_df[["model"] + attack_scores]
.groupby(["model"])
.agg(lambda x: x.mean(skipna=False))
.reset_index(drop=False)
)
return avg_df.rename(columns={s: f"avg_{s}" for s in attack_scores})
def add_attack_variants_as_columns(
df: pd.DataFrame, first_cols: list[str], attack_scores: list[str]
) -> pd.DataFrame:
model_dfs = []
for model in df.model.unique():
model_view = df[df.model == model]
attack_dfs = []
for i, row in model_view.iterrows():
attack_name = row["attack_name"]
attack_variant = row["attack_variant"]
g_df = model_view[
(model_view.attack_name == attack_name)
& (model_view.attack_variant == attack_variant)
]
if (
attack_name == "none"
or attack_name == "identity"
or attack_name == "Identity"
):
g_df = g_df[["model"] + first_cols + attack_scores]
else:
g_df = g_df[attack_scores]
if attack_variant == "default":
prefix = attack_name
else:
prefix = f"{attack_name}_{attack_variant}"
g_df = g_df.rename(columns={s: f"{prefix}_{s}" for s in attack_scores})
attack_dfs.append(g_df.reset_index(drop=True))
model_df = pd.concat(attack_dfs, axis=1)
model_dfs.append(model_df)
final_df = pd.concat(model_dfs, axis=0, ignore_index=True)
first_cols_ = ["model"] + first_cols
reordered_cols = first_cols_ + list(
set(final_df.columns.tolist()) - set(first_cols_)
)
return final_df[reordered_cols]
def add_attack_categories_as_columns(
benchmark_df: pd.DataFrame, attack_scores: list[str]
) -> pd.DataFrame:
# average over the attack variants (inequal number of attack variants)
attack_avg_df = (
benchmark_df[["model", "attack_name", "cat"] + attack_scores]
.groupby(["model", "attack_name", "cat"])
.agg(lambda x: x.mean(skipna=False))
.reset_index(drop=False)
)
df = (
attack_avg_df.groupby(["model", "cat"])[attack_scores]
.agg(lambda x: x.mean(skipna=False))
.reset_index()
)
model_dfs = []
for model in df.model.unique():
cat_dfs = []
for cat in df.cat.unique():
if cat == "None":
continue
cat_df = df[(df.model == model) & (df.cat == cat)]
cat_df = cat_df[attack_scores]
cat_df = cat_df.rename(columns={s: f"{cat}_{s}" for s in attack_scores})
cat_dfs.append(cat_df.reset_index(drop=True))
model_dfs.append(pd.concat(cat_dfs, axis=1))
return pd.concat(model_dfs, axis=0, ignore_index=True)
def get_old_format_dataframe(
benchmark_df: pd.DataFrame, first_cols: list[str], attack_scores: list[str]
) -> pd.DataFrame:
benchmark_df = benchmark_df.fillna("None")
avg_df = add_avg_as_columns(benchmark_df, attack_scores)
attack_variants_df = add_attack_variants_as_columns(
benchmark_df, first_cols, attack_scores
)
categories_df = add_attack_categories_as_columns(benchmark_df, attack_scores)
final_df = pd.concat([attack_variants_df, categories_df, avg_df], axis=1)
final_df = final_df.loc[:, ~final_df.columns.duplicated()].copy()
return final_df
|