ece / app.py
jordyvl's picture
Fix to reliability diagram - correct with test
897eeff
import evaluate
import json
import sys
from pathlib import Path
import gradio as gr
import numpy as np
import pandas as pd
import ast
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
plt.rcParams["figure.dpi"] = 300
plt.switch_backend(
"agg"
) # ; https://stackoverflow.com/questions/14694408/runtimeerror-main-thread-is-not-in-main-loop
def default_plot():
fig = plt.figure()
ax1 = plt.subplot2grid((3, 1), (0, 0), rowspan=2)
ax2 = plt.subplot2grid((3, 1), (2, 0))
ranged = np.linspace(0, 1, 10)
ax1.plot(
ranged,
ranged,
color="darkgreen",
ls="dotted",
label="Perfect",
)
# Bin differences
ax1.set_ylabel("Conditional Expectation")
ax1.set_ylim([0, 1.05])
ax1.set_title("Reliability Diagram")
ax1.set_xlim([-0.05, 1.05]) # respective to bin range
# Bin frequencies
ax2.set_xlabel("Confidence")
ax2.set_ylabel("Count")
ax2.set_xlim([-0.05, 1.05]) # respective to bin range
return fig, ax1, ax2
def reliability_plot(results):
# DEV: might still need to write tests in case of equal mass binning
# DEV: nicer would be to plot like a polygon
# see: https://github.com/markus93/fit-on-the-test/blob/main/Experiments_Synthetic/binnings.py
fig, ax1, ax2 = default_plot()
# Bin differences
bins_with_left_edge = np.insert(results["y_bar"], 0, 0, axis=0)
bins_with_right_edge = np.insert(results["y_bar"], -1, 1.0, axis=0)
bins_with_leftright_edge = np.insert(bins_with_left_edge, -1, 1.0, axis=0)
weights = np.nan_to_num(results["p_bar"], copy=True, nan=0)
# NOTE: the histogram API is strange
_, _, patches = ax1.hist(
bins_with_left_edge,
weights=weights,
bins=bins_with_leftright_edge,
)
for b in range(len(patches)):
perfect = bins_with_right_edge[b] # if b != n_bins else
empirical = weights[b] # patches[b]._height
bin_color = (
"limegreen"
if perfect == empirical
else "dodgerblue"
if empirical < perfect
else "orangered"
)
patches[b].set_facecolor(bin_color) # color based on over/underconfidence
ax1handles = [
mpatches.Patch(color="orangered", label="Overconfident"),
mpatches.Patch(color="limegreen", label="Perfect", linestyle="dotted"),
mpatches.Patch(color="dodgerblue", label="Underconfident"),
]
# Bin frequencies
anindices = np.where(~np.isnan(results["p_bar"]))[0]
bin_freqs = np.zeros(len(results["p_bar"]))
bin_freqs[anindices] = results["bin_freq"]
ax2.hist(
bins_with_left_edge, weights=bin_freqs, color="midnightblue", bins=bins_with_leftright_edge
)
acc_plt = ax2.axvline(x=results["accuracy"], ls="solid", lw=3, c="black", label="Accuracy")
conf_plt = ax2.axvline(
x=results["p_bar_cont"], ls="dotted", lw=3, c="#444", label="Avg. confidence"
)
ax1.legend(loc="lower right", handles=ax1handles)
ax2.legend(handles=[acc_plt, conf_plt])
ax1.set_xticks(bins_with_left_edge)
ax2.set_xticks(bins_with_left_edge)
plt.tight_layout()
return fig
def compute_and_plot(data, n_bins, bin_range, scheme, proxy, p):
# DEV: check on invalid datatypes with better warnings
if isinstance(data, pd.DataFrame):
data.dropna(inplace=True)
predictions = [
ast.literal_eval(prediction) if not isinstance(prediction, list) else prediction
for prediction in data["predictions"]
]
references = [reference for reference in data["references"]]
results = metric._compute(
predictions,
references,
n_bins=n_bins,
scheme=scheme,
proxy=proxy,
p=p,
detail=True,
)
plot = reliability_plot(results)
return results["ECE"], plot
sliders = [
gr.Slider(0, 100, value=10, label="n_bins"),
gr.Slider(
0, 100, value=None, label="bin_range", visible=False
), # DEV: need to have a double slider
gr.Dropdown(choices=["equal-range", "equal-mass"], value="equal-range", label="scheme"),
gr.Dropdown(choices=["upper-edge", "center"], value="upper-edge", label="proxy"),
gr.Dropdown(choices=[1, 2, np.inf], value=1, label="p"),
]
slider_defaults = [slider.value for slider in sliders]
# example data
component = gr.inputs.Dataframe(
headers=["predictions", "references"], col_count=2, datatype="number", type="pandas"
)
component.value = [
[[0.6, 0.2, 0.2], 0],
[[0.7, 0.1, 0.2], 2],
[[0, 0.95, 0.05], 1],
]
sample_data = [[component] + slider_defaults]
local_path = Path(sys.path[0])
metric = evaluate.load("jordyvl/ece")
outputs = [gr.outputs.Textbox(label="ECE"), gr.Plot(label="Reliability diagram")]
# outputs[1].value = default_plot().__dict__ #DEV: Does not work in gradio; needs to be JSON encoded
iface = gr.Interface(
fn=compute_and_plot,
inputs=[component] + sliders,
outputs=outputs,
description=metric.info.description,
article=evaluate.utils.parse_readme(local_path / "README.md"),
title=f"Metric: {metric.name}",
# examples=sample_data; #DEV: ValueError: Examples argument must either be a directory or a nested list, where each sublist represents a set of inputs.
).launch()