File size: 13,702 Bytes
e7eaeed |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 |
from collections import defaultdict
from typing import Any, Dict, List, Optional, Tuple
import pandas as pd
from pytorch_ie import Document, DocumentMetric
class ScoreDistribution(DocumentMetric):
"""Computes the distribution of prediction scores for annotations in a layer. The scores are
separated into true positives (TP) and false positives (FP) based on the gold annotations.
Args:
layer: The name of the annotation layer to analyze.
per_label: If True, the scores are separated per label. Default is False.
label_field: The field name of the label to use for separating the scores per label. Default is "label".
equal_sample_size_binning: If True, the scores are binned into equal sample sizes. If False,
the scores are binned into equal width. The former is useful when the distribution of scores is skewed.
Default is True.
show_plot: If True, a plot of the score distribution is shown. Default is False.
plotting_backend: The plotting backend to use. Default is "plotly".
plotting_caption_mapping: A mapping to rename any caption entries for plotting, i.e., the layer name,
labels, or TP/FP. Default is None.
plotting_colors: A dictionary mapping from gold scores to colors for plotting. Default is None.
"""
def __init__(
self,
layer: str,
label_field: str = "label",
per_label: bool = False,
show_plot: bool = False,
equal_sample_size_binning: bool = True,
plotting_backend: str = "plotly",
plotting_caption_mapping: Optional[Dict[str, str]] = None,
plotting_colors: Optional[Dict[str, str]] = None,
plotly_use_create_distplot: bool = True,
plotly_barmode: Optional[str] = None,
plotly_marginal: Optional[str] = "violin",
plotly_font_size: int = 18,
plotly_font_family: Optional[str] = None,
plotly_background_color: Optional[str] = None,
):
super().__init__()
self.layer = layer
self.label_field = label_field
self.per_label = per_label
self.equal_sample_size_binning = equal_sample_size_binning
self.plotting_backend = plotting_backend
self.show_plot = show_plot
self.plotting_caption_mapping = plotting_caption_mapping or {}
self.plotting_colors = plotting_colors
self.plotly_use_create_distplot = plotly_use_create_distplot
self.plotly_barmode = plotly_barmode
self.plotly_marginal = plotly_marginal
self.plotly_font_size = plotly_font_size
self.plotly_font_family = plotly_font_family
self.plotly_background_color = plotly_background_color
self.scores: Dict[str, Dict[str, List[float]]] = defaultdict(lambda: defaultdict(list))
def reset(self):
self.scores = defaultdict(lambda: defaultdict(list))
def _update(self, document: Document):
gold_annotations = set(document[self.layer])
for ann in document[self.layer].predictions:
if self.per_label:
label = getattr(ann, self.label_field)
else:
label = "ALL"
if ann in gold_annotations:
self.scores[label]["TP"].append(ann.score)
else:
self.scores[label]["FP"].append(ann.score)
def _combine_scores(
self,
scores_tp: List[float],
score_fp: List[float],
col_name_pred: str = "prediction",
col_name_gold: str = "gold",
) -> pd.DataFrame:
scores_tp_df = pd.DataFrame(scores_tp, columns=[col_name_pred])
scores_tp_df[col_name_gold] = 1.0
scores_fp_df = pd.DataFrame(score_fp, columns=[col_name_pred])
scores_fp_df[col_name_gold] = 0.0
scores_df = pd.concat([scores_tp_df, scores_fp_df])
return scores_df
def _get_calibration_data_and_metrics(
self, scores: pd.DataFrame, q: int = 20
) -> Tuple[pd.DataFrame, pd.Series]:
from sklearn.metrics import brier_score_loss
if self.equal_sample_size_binning:
# Create bins with equal number of samples.
scores["bin"] = pd.qcut(scores["prediction"], q=q, labels=False)
else:
# Create bins with equal width.
scores["bin"] = pd.cut(
scores["prediction"],
bins=q,
include_lowest=True,
right=True,
labels=False,
)
calibration_data = (
scores.groupby("bin")
.apply(
lambda x: pd.Series(
{
"avg_score": x["prediction"].mean(),
"fraction_positive": x["gold"].mean(),
"count": len(x),
}
)
)
.reset_index()
)
total_count = scores.shape[0]
calibration_data["bin_weight"] = calibration_data["count"] / total_count
# Calculate the absolute differences and squared differences.
calibration_data["abs_diff"] = abs(
calibration_data["avg_score"] - calibration_data["fraction_positive"]
)
calibration_data["squared_diff"] = (
calibration_data["avg_score"] - calibration_data["fraction_positive"]
) ** 2
# Compute Expected Calibration Error (ECE): weighted average of absolute differences.
ece = (calibration_data["abs_diff"] * calibration_data["bin_weight"]).sum()
# Compute Maximum Calibration Error (MCE): maximum absolute difference.
mce = calibration_data["abs_diff"].max()
# Compute Mean Squared Error (MSE): weighted average of squared differences.
mse = (calibration_data["squared_diff"] * calibration_data["bin_weight"]).sum()
# Compute the Brier Score on the raw predictions.
brier = brier_score_loss(scores["gold"], scores["prediction"])
values = {
"ece": ece,
"mce": mce,
"mse": mse,
"brier": brier,
}
return calibration_data, pd.Series(values)
def calculate_calibration_metrics(self, scores_combined: pd.DataFrame) -> pd.DataFrame:
calibration_data_dict = {}
calibration_metrics_dict = {}
for label, current_scores in scores_combined.groupby("label"):
calibration_data, calibration_metrics = self._get_calibration_data_and_metrics(
current_scores, q=20
)
calibration_data_dict[label] = calibration_data
calibration_metrics_dict[label] = calibration_metrics
all_calibration_data = pd.concat(
calibration_data_dict, names=["label", "idx"]
).reset_index(level=0)
all_calibration_metrics = pd.concat(calibration_metrics_dict, axis=1).T
if self.show_plot:
self.plot_calibration_data(calibration_data=all_calibration_data)
return all_calibration_metrics
def calculate_correlation(self, scores: pd.DataFrame) -> pd.Series:
result_dict = {}
for label, current_scores in scores.groupby("label"):
result_dict[label] = current_scores.drop("label", axis=1).corr()["prediction"]["gold"]
return pd.Series(result_dict, name="correlation")
@property
def mapped_layer(self):
return self.plotting_caption_mapping.get(self.layer, self.layer)
def plot_score_distribution(self, scores: pd.DataFrame):
if self.plotting_backend == "plotly":
for label in scores["label"].unique():
description = f"Distribution of Predicted Scores for {self.mapped_layer}"
if self.per_label:
label_mapped = self.plotting_caption_mapping.get(label, label)
description += f" ({label_mapped})"
if self.plotly_use_create_distplot:
import plotly.figure_factory as ff
current_scores = scores[scores["label"] == label]
# group by gold score
scores_dict = (
current_scores.groupby("gold")["prediction"].apply(list).to_dict()
)
group_labels, hist_data = zip(*scores_dict.items())
group_labels_renamed = [
self.plotting_caption_mapping.get(label, label) for label in group_labels
]
if self.plotting_colors is not None:
colors = [
self.plotting_colors[group_label] for group_label in group_labels
]
else:
colors = None
fig = ff.create_distplot(
hist_data,
group_labels=group_labels_renamed,
show_hist=True,
colors=colors,
bin_size=0.025,
)
else:
import plotly.express as px
fig = px.histogram(
scores,
x="prediction",
color="gold",
marginal=self.plotly_marginal, # "violin", # or box, violin, rug
hover_data=scores.columns,
color_discrete_map=self.plotting_colors,
nbins=50,
)
fig.update_layout(
height=600,
width=800,
title_text=description,
title_x=0.5,
font=dict(size=self.plotly_font_size),
legend=dict(yanchor="top", y=0.99, xanchor="left", x=0.01),
)
if self.plotly_barmode is not None:
fig.update_layout(barmode=self.plotly_barmode)
if self.plotly_font_family is not None:
fig.update_layout(font_family=self.plotly_font_family)
if self.plotly_background_color is not None:
fig.update_layout(
plot_bgcolor=self.plotly_background_color,
paper_bgcolor=self.plotly_background_color,
)
fig.show()
else:
raise NotImplementedError(f"Plotting backend {self.plotting_backend} not implemented")
def plot_calibration_data(self, calibration_data: pd.DataFrame):
import plotly.express as px
import plotly.graph_objects as go
color = "label" if self.per_label else None
x_col = "avg_score"
y_col = "fraction_positive"
fig = px.scatter(
calibration_data,
x=x_col,
y=y_col,
color=color,
trendline="ols",
labels=self.plotting_caption_mapping,
)
if not self.per_label:
fig["data"][1]["name"] = "prediction vs. gold"
# show legend only for trendlines
for idx, trace_data in enumerate(fig["data"]):
if idx % 2 == 0:
trace_data["showlegend"] = False
else:
trace_data["showlegend"] = True
# add the optimal line
minimum = calibration_data[x_col].min()
maximum = calibration_data[x_col].max()
fig.add_trace(
go.Scatter(
x=[minimum, maximum],
y=[minimum, maximum],
mode="lines",
name="optimal",
line=dict(color="black", dash="dash"),
)
)
fig.update_layout(
height=600,
width=800,
title_text=f"Mean Binned Scores for {self.mapped_layer}",
title_x=0.5,
font=dict(size=self.plotly_font_size),
)
fig.update_layout(
legend=dict(
yanchor="top",
y=0.99,
xanchor="left",
x=0.01,
title="OLS trendline" + ("s" if self.per_label else ""),
),
)
if self.plotly_background_color is not None:
fig.update_layout(
plot_bgcolor=self.plotly_background_color,
paper_bgcolor=self.plotly_background_color,
)
if self.plotly_font_family is not None:
fig.update_layout(font_family=self.plotly_font_family)
fig.show()
def _compute(self) -> Dict[str, Dict[str, Any]]:
scores_combined = pd.concat(
{
label: self._combine_scores(scores["TP"], scores["FP"])
for label, scores in self.scores.items()
},
names=["label", "idx"],
).reset_index(level=0)
result_df = scores_combined.groupby("label")["prediction"].agg(["mean", "std", "count"])
if self.show_plot:
self.plot_score_distribution(scores=scores_combined)
calibration_metrics = self.calculate_calibration_metrics(scores_combined)
calibration_metrics["correlation"] = self.calculate_correlation(scores_combined)
result_df = pd.concat(
{"prediction": result_df, "prediction vs. gold": calibration_metrics}, axis=1
)
if not self.per_label:
result = result_df.xs("ALL")
else:
result = result_df.T.stack().unstack()
result_dict = {
main_key: result.xs(main_key).T.to_dict()
for main_key in result.index.get_level_values(0).unique()
}
return result_dict
|