File size: 8,767 Bytes
485127c | 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 | import os
import pandas as pd
import matplotlib.pyplot as plt
from collections import defaultdict
import numpy as np
import json
from scipy import stats
from scipy.stats import norm
CASES = ['exact', 'inexact']
plot_path = 'visualization/normal_results'
include_legend = False
def get_results():
datasets=["xsum", "squad", "writing"]
source_models=["gpt2-xl", "gpt-neo-2.7B"]
methods2 = {
'classification': 'AdaDetectGPT',
}
result_dir_template = 'exp_normal/results_{}'
def get_stats(result_file):
with open(result_file, 'r') as fin:
res = json.load(fin)
return res['predictions']['samples']
def _get_method_stats(dataset, model, method, cases, filter=''):
res_path = result_dir_template.format(cases)
result_file = f'{res_path}/{dataset}_{model}.{method}.json'
if os.path.exists(result_file):
stats = np.array(get_stats(result_file))
else:
stats = np.array([0.0])
return stats
result_list = []
for cases in CASES:
for dataset in datasets:
for model in source_models:
for method in methods2:
results = {'datasets': dataset, 'models': model, 'cases': cases}
method_name = methods2[method]
results['methods'] = method_name
cols = _get_method_stats(dataset, model, method, cases)
results['values'] = cols
# print(f"{method_name} with mean {np.mean(cols)} and std {np.std(cols)} on {dataset} with cases {cases}")
result_list.append(results)
def merge_dicts_of_lists(dataset_list: list[dict]) -> dict:
"""
将一系列 dict(键→list) 合并为一个 dict,
同一个键对应的 list 会被 extend 到一起。
"""
merged = defaultdict(list)
for d in dataset_list:
for key, value in d.items():
# 如果 value 本身是 list,则 extend;否则 append
if isinstance(value, list):
merged[key].extend(value)
else:
merged[key].append(value)
return dict(merged)
result_list = merge_dicts_of_lists(result_list)
df = pd.DataFrame(result_list)
df['values'] = df['values'].apply(lambda arr: arr.tolist())
df = df.explode('values').reset_index(drop=True)
# print(df)
return df
def plot_hist_by_dataset_and_method(df):
datasets = sorted(df['datasets'].unique())
models = sorted(df['models'].unique())
# datasets.remove('xsum')
methods = sorted(df['methods'].unique())
# pick a distinct color for each cases
N = len(CASES)
cmap = plt.get_cmap('viridis')
palette = cmap(np.linspace(1, 0, N))
fig, axes = plt.subplots(
nrows=len(methods),
ncols=len(datasets),
figsize=(3.9*len(datasets), 3.3*len(methods)),
sharey=False,
)
# if only one row or one col, axes may be 1-D
if len(methods) == 1 and len(datasets) == 1:
axes = [[axes]]
elif len(methods) == 1:
axes = [axes]
elif len(datasets) == 1:
axes = [[ax] for ax in axes]
for i, method in enumerate(methods):
for j, dataset in enumerate(datasets):
ax = axes[i][j]
sub = df[(df['methods']==method) & (df['datasets']==dataset)]
value_list = []
model_list = []
for l, model in enumerate(models):
sub2 = sub[sub['models']==model]
# draw one histogram per cases
for k, n in enumerate(CASES):
vals = sub2[sub2['cases']==n]['values']
value_list.append(vals)
model_list.append(model)
print(f"------------ Test on {model} at {dataset} ------------")
test_result = stats.kstest(vals.to_numpy().astype(np.float16), stats.norm.cdf)
print("KS test: ", test_result.pvalue)
test_result = stats.shapiro(vals.to_numpy().astype(np.float16))
print("SW test: ", test_result.pvalue)
test_result = stats.anderson(vals.to_numpy().astype(np.float16), dist='norm')
print("Anderson test: ", test_result.statistic, test_result.critical_values, test_result.significance_level)
print(f"------------------------------------------------------")
_, bins, _ = ax.hist(
value_list,
density=True,
bins=12,
alpha=0.6,
range=(-3, 3),
histtype='stepfilled',
color=['#0571b0', '#f4a582'],
label=model_list,
)
# overlay the standard normal curve
x = np.linspace(bins[0], bins[-1], 200)
y = norm.pdf(x, loc=0, scale=1) # standard normal
ax.plot(
x, y,
color='darkred',
linestyle='--',
linewidth=2,
)
ax.set_ylabel('density', fontsize=12, fontweight='bold')
# titles & labels
if i == 0:
pass
if j == 0:
pass
if i == len(methods)-1:
ax.set_xlabel('statistics', fontsize=13, fontweight='bold')
# common legend on the right
handles, labels = axes[0][-1].get_legend_handles_labels()
if include_legend:
fig.legend(
handles, labels,
# title="cases",
loc='lower center',
bbox_to_anchor=(0.5, -0.017),
ncol=5,
)
plt.tight_layout()
plt.subplots_adjust(bottom=0.24)
plt.savefig(f'{plot_path}/normal_histogram.pdf', dpi=300, bbox_inches='tight')
plt.show()
def plot_box_by_dataset_and_method(df):
"""
df must have columns:
- 'datasets' (categorical)
- 'methods' (categorical)
- 'cases' (numeric, e.g. 2,4,8,16)
- 'values' (each entry is a 1D array or list of numbers)
"""
datasets = sorted(df['datasets'].unique())
methods = sorted(df['methods'].unique())
n_rows, n_cols = len(methods), len(datasets)
fig, axes = plt.subplots(
n_rows, n_cols,
figsize=(4*n_cols, 4*n_rows),
sharey=False
)
# In case of single row/col, normalize axes to 2D list
if n_rows == 1 and n_cols == 1:
axes = [[axes]]
elif n_rows == 1:
axes = [axes]
elif n_cols == 1:
axes = [[ax] for ax in axes]
# pick a color for each cases
N = len(CASES)
cmap = plt.get_cmap('viridis')
palette = cmap(np.linspace(1, 0, N))
for i, method in enumerate(methods):
for j, dataset in enumerate(datasets):
ax = axes[i][j]
sub = df[(df['datasets']==dataset) & (df['methods']==method)]
# build list-of-lists for each prompt
data_list = []
for n in CASES:
block = sub[sub['cases']==n]
print("KS test for ", n, f"at {dataset} {method}")
print(stats.kstest(block['values'], stats.norm.cdf))
data_list.append(block['values'])
# draw the boxplots at positions 0,1,2,3
bp = ax.boxplot(
data_list,
positions=range(len(CASES)),
patch_artist=True,
widths=0.6,
medianprops=dict(color="black")
)
# color them
for patch, color in zip(bp['boxes'], palette):
patch.set_facecolor(color)
patch.set_alpha(0.7)
# x‐ticks & labels
ax.set_xticks(list(range(len(CASES))))
ax.set_xticklabels([str(n) for n in CASES])
ax.set_xlabel("cases")
if j == 0:
ax.set_ylabel(method, fontweight="bold")
if i == 0:
pass
ax.set_title(dataset, fontweight="bold")
# shared legend in upper right
handles = [
plt.Line2D([0],[0], color=palette[k], marker='s', linestyle='', alpha=0.7)
for k in range(len(CASES))
]
labels = [f"cases={n}" for n in CASES]
fig.legend(
handles, labels,
# title="cases",
loc='lower center',
bbox_to_anchor=(0.5, 0.0),
ncol=5,
)
plt.tight_layout()
plt.subplots_adjust(bottom=0.2) # leave extra space at the bottom
plt.savefig(f'{plot_path}/normal_boxplot.pdf', dpi=300, bbox_inches='tight')
plt.show()
if __name__ == '__main__':
df = get_results()
plot_hist_by_dataset_and_method(df)
|