Moving the description generation to its own function and expanding it.

This commit is contained in:
2026-05-07 17:13:47 +01:00
parent 92db500bd0
commit 7c72462e53

View File

@@ -31,12 +31,138 @@ def _to_age_bracket(row):
return "{}-{}".format(d, d + 9)
def generate_description(dataset: pd.DataFrame, name: str | None = None) -> tuple[str]:
"""Generates a textual description of `dataset`, its variables and values.
For all supported variables in `dataset`, this function lists their names,
types, value ranges and descriptive statistics. It also groups each
variable, comparing it to the others, thus creating a subgroup description
as well.
Note that we only support the variables defined by the `Capability` enum.
Parameters
----------
dataset: pd.DataFrame
name: str, optional
The dataset name. Only ever used in the first line of the returned
description, for pretty printing.
Returns
-------
desc: str
The textual description as a human-readable string.
See Also
--------
`Capability`
"""
outstr = ""
if name is not None:
outstr = f"The dataset {name}"
else:
outstr = f"The dataset"
outstr += f" has a total of {len(dataset)} {dataset.index.name}s, with the following supported features/capabilities:"
caps = []
for c in Capability:
if c.value in dataset:
caps.append(c)
outstr += f"\n- {c.value}"
outstr += "\n\nEach feature/capability has the following types and values:"
for c in caps:
if c == Capability.AGE:
outstr += f"\n{c.value}: numeric"
else:
outstr += f"\n{c.value}: categorical"
outstr += f"\n - {sorted(dataset[c].unique())}"
outstr += "\n\nData distribution statistics."
for c in caps:
outstr += (
f'\nThe feature/capability "{c}" has the following distribution of values:'
)
if c == Capability.AGE:
m1 = dataset[c].min()
m2 = dataset[c].max()
mean = dataset[c].mean()
std = dataset[c].std()
p25 = dataset[c].quantile(0.25)
median = dataset[c].median()
p75 = dataset[c].quantile(0.75)
outstr += f"\n - min = {m1}"
outstr += f"\n - max = {m2}"
outstr += f"\n - mean = {mean:.2f}"
outstr += f"\n - std = {std:.2f}"
outstr += f"\n - p25 = {p25}"
outstr += f"\n - p50 = {median}"
outstr += f"\n - p75 = {p75}"
outstr += "\n Interqualtile ranges:"
outstr += f"\n - p25-min = {p25 - m1}"
outstr += f"\n - p50-p25 = {median - p25}"
outstr += f"\n - p75-p50 = {p75 - median}"
outstr += f"\n - max-p75 = {m2 - p75}"
else:
series = dataset[c].value_counts().sort_index()
for s in series.index:
outstr += f"\n - {s}: {series[s]}"
outstr += "\n\nPer capability/class data distribution statistics."
if Capability.AGE in caps and caps[-1] != Capability.AGE:
# Rotating AGE to be the last in the list, as it cannot be the grouping
# variable, since it is numeric.
idx = caps.index(Capability.AGE)
del caps[idx]
caps.append(Capability.AGE)
# caps = caps[1:] + caps[:1]
for c1, c2 in combinations(caps, 2):
# Here we test for age, but this should really be a test for any
# numerical variable.
if c1 == Capability.AGE:
continue
gb = dataset.groupby(c1)[[c2]]
tmpdf = None
tmpstr = '\nGrouping by "{}", the dataset has the following {} for each value of "{}"'
if c2 != Capability.AGE:
outstr += tmpstr.format(c1, "number of elements", c2)
tmpdf = gb.value_counts().sort_index().unstack(level=c1)
tmpdf = tmpdf.fillna(0).astype(int)
# tmpdf = pd.concat([tmpdf, gb.apply(gini)], axis=1)
elif c1 != Capability.AGEGROUP:
outstr += tmpstr.format(c1, "statistics", c2)
tmpdf = pd.concat(
[
gb.min(),
gb.quantile(0.25),
gb.median(),
gb.quantile(0.75),
gb.max(),
gb.apply(gini),
],
axis=1,
)
tmpdf.columns = ["min", "p25", "p50", "p75", "max", "gini_impurity"]
else:
continue
outstr += f"\n{tmpdf}\n"
return outstr
if __name__ == "__main__":
import os
logger.info(os.getcwd())
DATASET_PATH = Path("../../data/facing2-train/")
DATASET_PATH = Path("data/facing2-train/")
METADATA_PATH = DATASET_PATH / "meta-w-age.csv"
meta = pd.read_csv(METADATA_PATH, sep=",", index_col="image")
@@ -49,9 +175,8 @@ if __name__ == "__main__":
# GINI IMPURITY
# Lower values means a concentration of values around a single class, i.e. bias.
age_gini = gini(meta["age"])
age_gini = gini(meta[Capability.AGE])
# gt_age_group_ord = meta["age_group"].apply(lambda x: _agegroup_int_map[x])
agegroup_gini = gini(meta[Capability.AGEGROUP + "_cat"])
# Should be close to 0.5, indicating a 50/50 split of males and females,
@@ -78,64 +203,12 @@ if __name__ == "__main__":
gini_per_sex = sex_gb.apply(gini)
gini_per_agegroup = agegroup_gb.apply(gini)
# Prototype textual description of the dataset. To be incorporated into a
# "generate_report" function.
print(
f'The dataset "{DATASET_PATH.name}" has a total of {len(meta)} {meta.index.name}s,'
" with the following features/capabilities:"
)
caps = []
for c in Capability:
if c.value in meta:
caps.append(c)
print(f"- {c.value}")
# Prototype textual description of the dataset.
s = generate_description(meta, name=DATASET_PATH.name)
print(s)
print("\nEach feature/capability has the following types and values:")
for c in caps:
if c == Capability.AGE:
print(f"{c.value}: numeric")
else:
print(f"{c.value}: categorical")
print(f" - {sorted(meta[c].unique())}")
print("\nData distribution statistics.")
for c in caps:
print(f'The feature/capability "{c}" has the following distribution of values:')
if c == Capability.AGE:
m1 = meta[c].min()
m2 = meta[c].max()
mean = meta[c].mean()
std = meta[c].std()
p25 = meta[c].quantile(0.25)
median = meta[c].median()
p75 = meta[c].quantile(0.75)
print(
f" - min = {m1}, max = {m2}, mean = {mean:.2f}, std = {std:.2f}"
f" p25 = {p25}, p50 = {median}, p75 = {p75}"
)
print(" Interqualtile ranges:")
print(f" - p25-min = {p25 - m1}")
print(f" - p50-p25 = {median - p25}")
print(f" - p75-p50 = {p75 - median}")
print(f" - max-p75 = {m2 - p75}")
else:
series = meta[c].value_counts().sort_index()
for s in series.index:
print(f" - {s}: {series[s]}")
print("\nPer capability/class data distribution statistics.")
for c1, c2 in combinations(caps, 2):
if c1 == Capability.AGE:
continue
if c2 != Capability.AGE:
gb = meta.groupby(c1)[[c2]]
print(
f'Grouping by "{c1}", the dataset has the following data distribution for "{c2}"'
)
print(gb.value_counts().sort_index().unstack(level=c1).fillna(0))
# Diagnostics of biases in the dataset. To be incorporated into a
# "generate_diagnostics" function later on.
# Diagnostics of biases in the dataset. To be incorporated into a
# "generate_diagnostics" function later on.
def generate_bias_diagnostics():
pass