File size: 5,428 Bytes
fcaa164 |
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 |
import json
import os
from collections import defaultdict
from contextlib import contextmanager
from glob import glob
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
from scipy.stats import pearsonr, spearmanr
@contextmanager
def science_plot(font_size=16):
import scienceplots
with plt.style.context(["ieee", "grid", "no-latex", "light"]):
plt.rcParams.update({"font.size": font_size})
yield
def statistic_humaneval(eval_file: str, print_diff: bool = False):
llm_eval = json.load(open(eval_file))
llm_data = []
for dimension, files in llm_eval.items():
for filename, values in files.items():
try:
setting, basename = filename.split("/", 2)[1:]
basename = basename.split("/")[0]
if not isinstance(values["score"], int):
raise ValueError(f"score is not int: {values['score']}")
llm_data.append(
{
"setting": setting,
"sample": basename,
"dimension": dimension,
"score": values["score"],
}
)
except:
continue
file_path = "human_eval/human_scores_2024-12-13.xlsx"
human_eval = pd.read_excel(file_path).to_dict("records")
human_data = []
for record in human_eval:
setting = record.get("setting")
score = record[dimension]
basename = record["PPT"]
try:
score = int(score)
except:
continue
human_data.append(
{
"setting": setting,
"sample": basename,
"dimension": dimension,
"score": score,
}
)
# Compare and output differences between human and llm evaluations
llm_df = pd.DataFrame(llm_data)
human_df = pd.DataFrame(human_data)
merged = pd.merge(
llm_df,
human_df,
on=["setting", "sample", "dimension"],
suffixes=("_llm", "_human"),
how="outer",
indicator=True,
)
# Calculate and print correlation coefficients for common records
common_records = merged[merged["_merge"] == "both"].drop(columns=["_merge"])
dimensions = common_records["dimension"].unique()
for dimension in dimensions:
scores_human = common_records[common_records["dimension"] == dimension][
"score_human"
]
scores_llm = common_records[common_records["dimension"] == dimension][
"score_llm"
]
pearson_correlation = pearsonr(scores_human, scores_llm)
spearman_correlation = spearmanr(scores_human, scores_llm)
print(
f"{dimension}, pearson: {pearson_correlation}, spearman: {spearman_correlation}"
)
if print_diff:
difference_df = common_records[
common_records["score_human"] != common_records["score_llm"]
]
for _, row in difference_df.iterrows():
print(row)
def statistic_ppteval():
data = []
eval_files = glob("./data/evals/PPTCrew*")
for eval_file in eval_files:
setting = eval_file.split("/")[-1].removesuffix(".json")
eval_stats = json.load(open(eval_file))
for dimension, files in eval_stats.items():
if dimension == "vision":
dimension = "design"
for filename, score in files.items():
domain = filename.split("/")[1]
if isinstance(score, dict):
score = score["score"]
if isinstance(score, str):
continue
if score > 5000 or score < 0:
continue
data.append(
{
"setting": setting,
"dimension": dimension,
"sample": filename,
"score": score,
"domain": domain,
}
)
return pd.DataFrame(data)
def setting_perfomance(df: pd.DataFrame):
df = df.drop(columns=["domain"])
for setting, dimension in df[["setting", "dimension"]].drop_duplicates().values:
avg_score = df[(df["setting"] == setting) & (df["dimension"] == dimension)][
"score"
].mean()
print(f"{setting}, {dimension}, {avg_score}")
def plot_correlation(df: pd.DataFrame):
df = df.drop(columns=["domain"])
correlation_matrix = df[df["setting"] == "PPTCrew-gpt-4o+gpt-4o+gpt-4o"][
["ppl", "fid", "content", "design"]
].corr()
# Plot the heatmap with axis limits set from -1 to 1
plt.figure(figsize=(10, 8))
sns.heatmap(
correlation_matrix,
annot=True,
cmap="coolwarm",
fmt=".2f",
linewidths=0.5,
vmin=-1,
vmax=1,
annot_kws={"size": 15},
)
plt.xticks(fontsize=16)
plt.yticks(fontsize=16)
plt.savefig("correlation.pdf", bbox_inches="tight")
plt.show()
def domain_perfomance(df: pd.DataFrame):
df = df[df["setting"] == "PPTCrew-gpt-4o+gpt-4o+gpt-4o"]
for domain, scores in df.groupby("domain")["score"]:
print(f"{domain}, {scores.mean()}")
if __name__ == "__main__":
statistic_humaneval("./test-logic.json")
|