"""Visualization of robustness results across different extractors."""
from pathlib import Path
from typing import Any
import matplotlib.pyplot as plt
import pandas as pd
import seaborn as sns
sns.set_style("darkgrid")
pd.set_option("future.no_silent_downcasting", True)
# Please leave those 2 dictionnaries as is.
EXTRACTOR_LABELS_DICT = {
"conch": "CONCH",
"gpfm": "GPFM",
"hibou_vit_base": "Hibou Base",
"hibou_vit_large": "Hibou Large",
"h0_mini": "H0-Mini",
"hoptimus0": "H-Optimus-0",
"kaiko_vit_base_8": "Kaiko ViT-B/8",
"kaiko_vit_large_14": "Kaiko ViT-L/14",
"phikon": "Phikon",
"phikon_v2": "Phikon v2",
"plip": "PLIP",
"provgigapath": "Prov-GigaPath",
"uni": "UNI",
"uni2h": "UNI2-h",
"virchow": "Virchow",
"virchow2": "Virchow2",
}
EXTRACTOR_PARAMETERS_DICT = {
"conch": 86_000_000,
"gpfm": 307_000_000,
"hibou_vit_base": 86_000_000,
"hibou_vit_large": 307_000_000,
"h0_mini": 86_000_000,
"hoptimus0": 1_100_000_000,
"kaiko_vit_base_8": 86_000_000,
"kaiko_vit_large_14": 307_000_000,
"phikon": 86_000_000,
"phikon_v2": 307_000_000,
"plip": 86_000_000,
"provgigapath": 1_100_000_000,
"uni": 307_000_000,
"uni2h": 681_000_000,
"virchow": 632_000_000,
"virchow2": 632_000_000,
}
[docs]
def expand_columns(raw_results: pd.DataFrame) -> pd.DataFrame:
"""Expand columns so as to have one column per metric and robustness type."""
output = []
# Robustness types are "all", "inter-scanner", "inter-staining",
# "inter-scanner, inter-staining"
for robustness_type in raw_results["robustness_type"].unique():
subset = raw_results[
raw_results["robustness_type"] == robustness_type
].sort_values("extractor")
subset = subset.set_index("extractor").iloc[:, 1:]
subset.columns = [f"{c}__{robustness_type}" for c in subset.columns]
output.append(subset)
output_df = pd.concat(output, axis=1)
output_df.insert(
0, "extractor", output_df.index.to_series().replace(EXTRACTOR_LABELS_DICT)
)
output_df.insert(
1, "Parameters", output_df.index.to_series().replace(EXTRACTOR_PARAMETERS_DICT)
)
return output_df
[docs]
def display_plism_metrics(
raw_results: pd.DataFrame,
metric_x: str = "cosine_similarity_median",
metric_y: str = "top_1_accuracy_median",
robustness_x: str = "all",
robustness_y: str = "all",
label_x: str = "Median Cosine Similarity",
label_y: str = "Median Top-1 Accuracy",
fig_save_path: str | Path | None = None,
xlim: tuple[float, float] | None = None,
ylim: tuple[float, float] | None = None,
palette: Any | None = None,
):
"""Display PLISM robustness metrics.
Parameters
----------
raw_results: pd.DataFrame
Raw results as computed by ``plismbench.utils.metrics.format_results``.
metric_x: str = "cosine_similarity_median"
Metric to display for x-axis. Should be of type 'metric_aggregation'.
Supported metrics depends on the columns of ``raw_results`` but are
by "cosine_similarity", "top_1_accuracy", "top_3_accuracy",
"top_5_accuracy" and "top_10_accuracy". Supported aggregation types
are either "mean" or "median".
metric_y: str = "top_1_accuracy_median"
Metric to display for y-axis.
robustness_x: str = "all"
Type of robustness for ``metric_x``.
Supported types are "all", "inter-scanner", "inter-staining",
"inter-scanner" and "inter-staining".
robustness_y: str = "all"
Type of robustness for ``metric_y``.
label_x: str = "Median Cosine Similarity"
Label for x-axis (can be anything).
label_y: str = "Median Top-1 Accuracy"
Label for y-axis (can be anything).
xlim: tuple[float, float] | None = None
Limits for x-axis.
ylim: tuple[float, float] | None = None
Limits for y-axis.
palette = None
Color palette.
fig_save_path: str | Path | None = None
Figure save path.
"""
# Set figure
fig, ax = plt.subplots(1, 1, figsize=(10, 6), dpi=200)
# Default limits for axes
xlim = (0, 1) if xlim is None else xlim
ylim = (0, 1) if xlim is None else ylim
# Define x and y column in ``results_df``
col_x = f"{metric_x}__{robustness_x}"
col_y = f"{metric_y}__{robustness_y}"
results_df = expand_columns(raw_results)
results_df = results_df[["extractor", "Parameters", col_x, col_y]]
results_df[[col_x, col_y]] = results_df[[col_x, col_y]].astype(float)
# Default color palette
if palette is None:
palette = sns.color_palette("tab20")[: results_df.shape[0]]
# Display metrics for each extractor
sns.scatterplot(
data=results_df,
x=col_x,
y=col_y,
hue="extractor",
size="Parameters",
sizes=(50, 2000),
palette=palette,
edgecolor="black",
alpha=0.9,
legend=True,
ax=ax,
)
# Display number of parameters for each extractor
sns.scatterplot(
data=results_df,
x=col_x,
y=col_y,
s=5,
color="black",
marker="+",
facecolor="black",
alpha=0.7,
legend=True,
ax=ax,
)
# Set labels and limits
ax.set_xlabel(label_x)
ax.set_ylabel(label_y)
ax.set_xlim(xlim)
ax.set_ylim(ylim)
# Please leave these values as-is
for _, row in results_df.iterrows():
plt.text(
row[col_x] + 0.001,
row[col_y] + 0.001,
row["extractor"],
fontsize=10,
ha="left",
va="bottom",
fontweight="bold",
)
# Please leave these values as-is: adding circles in the legend
# proportionnal to the model size.
scatter_handles = [
plt.scatter(
[], [], s=50, edgecolor="black", color="gray", alpha=0.5, label=" 22M"
),
plt.scatter(
[], [], s=155, edgecolor="black", color="gray", alpha=0.5, label=" 86M"
),
plt.scatter(
[], [], s=555, edgecolor="black", color="gray", alpha=0.5, label=" 307M"
),
plt.scatter(
[],
[],
s=1143,
edgecolor="black",
color="gray",
alpha=0.5,
label=" 632M",
),
plt.scatter(
[],
[],
s=2000,
edgecolor="black",
color="gray",
alpha=0.5,
label=" 1,100M",
),
]
ax.legend(
handles=scatter_handles,
title="No. parameters",
loc="center left",
bbox_to_anchor=(1, 0.6),
fontsize=12,
labelspacing=0.2,
handleheight=2,
)
# Export figure
if fig_save_path is not None:
fig.savefig(fig_save_path, dpi=300, bbox_inches="tight")
plt.show()
