-
Notifications
You must be signed in to change notification settings - Fork 0
/
generate_plots.py
415 lines (342 loc) · 17.7 KB
/
generate_plots.py
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
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
from functools import cache
import matplotlib.pyplot as plt
import matplotlib.ticker as ticker
import numpy as np
import torch
import seaborn as sns
import pandas as pd
plt.rcParams.update({'font.size': 21})
SEEDS = list(range(100))
DESIRED_GAP = 0.1
DPI = 300
DATASET_TO_MAX_LEN = {
'Tiselac': 23,
'ElectricDevices': 96,
'PenDigits': 8,
'Crop': 46,
'WalkingSittingStanding': 206,
'quality': 10,
}
@cache
def load_res(path: str):
return torch.load(path)
@cache
def get_accuracy_gap_df(dataset: str, cal_type: str, accuracy_gap: float):
accuracy_list = []
late_accuracy_list = []
t_accuracy_list_list = []
t_late_accuracy_list_list = []
t_num_correct_list_list = []
t_late_num_correct_list_list = []
t_num_samples_list_list = []
t_gap_list_list = []
halt_timesteps_list_list = []
for seed in SEEDS:
res_path = f'results/dataset={dataset}_seed={seed}_cal_type={cal_type}_accuracy_gap={accuracy_gap}.pt'
res = load_res(res_path)
accuracy_list.append(res['accuracy'])
late_accuracy_list.append(res['late_accuracy'])
t_accuracy_list_list.append(res['t_accuracy_list'])
t_late_accuracy_list_list.append(res['t_late_accuracy_list'])
t_num_correct_list_list.append(res['t_num_correct_list'])
t_late_num_correct_list_list.append(res['t_late_num_correct_list'])
t_num_samples_list_list.append(res['t_num_samples_list'])
t_gap_list_list.append(res['t_gap_list'])
halt_timesteps_list_list.append(res['halt_timesteps'])
t_gap_list_list = np.array(t_gap_list_list)
halt_timesteps_list_list = np.array(halt_timesteps_list_list)
t_accuracy_list_list = np.array(t_accuracy_list_list)
t_late_accuracy_list_list = np.array(t_late_accuracy_list_list)
t_num_samples_list_list = np.array(t_num_samples_list_list)
t_num_correct_list_list = np.array(t_num_correct_list_list)
t_late_num_correct_list_list = np.array(t_late_num_correct_list_list)
num_samples_until_t = t_num_samples_list_list.cumsum(axis=1)
t_gap_until_t = t_gap_list_list.cumsum(axis=1)
num_accuracy_gap_until_t = t_gap_until_t / num_samples_until_t
num_timesteps = DATASET_TO_MAX_LEN[dataset]
timesteps = np.tile(np.arange(1, num_timesteps+1), len(SEEDS))
seeds = np.repeat(np.arange(1, len(SEEDS)+1), num_timesteps)
values = num_accuracy_gap_until_t.flatten()
data = {
'TimeStep': timesteps,
'Seed': seeds,
'Value': values,
'NumHalted': t_num_samples_list_list.flatten(),
'NumHaltedUntilT': num_samples_until_t.flatten(),
}
df = pd.DataFrame(data)
return df
def plot_marginal_vs_conditional(dataset: str):
marginal_df = get_accuracy_gap_df(dataset, 'marginal_accuracy_gap', accuracy_gap=DESIRED_GAP)
conditional_df = get_accuracy_gap_df(dataset, 'conditional_accuracy_gap', accuracy_gap=DESIRED_GAP)
marginal_df['Experiment'] = 'Marginal'
conditional_df['Experiment'] = 'Conditional'
combined_df = pd.concat([marginal_df, conditional_df], ignore_index=True)
max_timestep = DATASET_TO_MAX_LEN[dataset]
first_tick = 1
tick_spacing = max_timestep // 5
xticks = [first_tick] + list(range(first_tick + tick_spacing, max_timestep+1, tick_spacing))
xticks[-1] = max_timestep
conditional_accuracy_gap_fig, axs = plt.subplots(1, 2, figsize=(10, 4), gridspec_kw={'width_ratios': [1, 1]}, constrained_layout=True)
pos1 = axs[0].get_position() # Get the original position
axs[0].set_position([pos1.x0, pos1.y0, pos1.width, pos1.height])
pos2 = axs[1].get_position() # Get the original position
axs[1].set_position([pos2.x0+0.1, pos2.y0, pos2.width, pos2.height])
ax = axs[0]
hue_order = ['Marginal', 'Conditional']
df_dashed = combined_df[combined_df['Experiment'] == 'Marginal']
df_solid = combined_df[combined_df['Experiment'] != 'Marginal']
sns.lineplot(data=df_dashed, x='TimeStep', y='Value', hue='Experiment', units='Seed', estimator=None, lw=1, alpha=0.2, ax=ax, linestyle='--', hue_order=hue_order)
sns.lineplot(data=df_solid, x='TimeStep', y='Value', hue='Experiment', units='Seed', estimator=None, lw=1, alpha=0.2, ax=ax, linestyle='-', hue_order=hue_order)
ax.set_xlim([1-max_timestep*(1/20), max_timestep*(21/20)])
ax.set_xticks(xticks)
desired_accuracy_gap = DESIRED_GAP
if desired_accuracy_gap is not None:
ax.axhline(y=DESIRED_GAP, color='r', linestyle=':')
# make a legend for the red line
ax.plot([], [], color='r', linestyle=':', label=f'$\\alpha$ level')
ax.legend(borderaxespad=0.2, borderpad=0.25)
ax.set_xlabel('Timestep $t$')
ax.set_ylabel('Conditional accuracy gap')
if dataset == 'quality':
ax.set_ylim([0, 0.2])
else:
ax.set_ylim([0, 0.35])
last_timestep = DATASET_TO_MAX_LEN[dataset]
# Adding a vertical line on the last timestep
ax.axvline(x=last_timestep, color='black', linestyle='--')
text_y_position = ax.get_ylim()[1] * 0.87
if dataset == 'quality':
text_x_position = last_timestep*0.71
elif dataset == 'Tiselac':
text_x_position = last_timestep*0.70
text_y_position = ax.get_ylim()[1] * 0.44
elif dataset in {'Crop', 'PhonemeSpectra', 'ElectricDevices'}:
text_x_position = last_timestep*0.69
elif dataset == 'PenDigits':
text_x_position = last_timestep*0.72
elif dataset == 'WalkingSittingStanding':
text_x_position = last_timestep*0.69
ax.text(text_x_position, text_y_position, 'Marginal risk', color='black', ha='center', va='bottom')
if dataset in {'quality', 'Tiselac'}:
handles, labels = ax.get_legend_handles_labels()
if dataset == 'quality':
handles = handles[4:]
labels = labels[4:]
else:
handles = handles[2:]
labels = labels[2:]
handles[0].set_linestyle('--')
# Put the legend in the bottom left
if dataset == 'quality':
ax.legend(handles, labels, loc='lower left', borderaxespad=0.2, borderpad=0.25)
else:
loc = 'upper left'
ax.legend(handles, labels, loc='upper left', borderaxespad=0.2, borderpad=0.25)
else:
ax.get_legend().remove()
pos = ax.yaxis.label.get_position()
ax.yaxis.label.set_position((pos[0], pos[1]-0.065))
ax = axs[1]
sns.lineplot(data=df_dashed, x='TimeStep', y='NumHaltedUntilT', hue='Experiment', errorbar=('ci', 95), ax=ax, hue_order=hue_order, linewidth=3, linestyle='--')
sns.lineplot(data=df_solid, x='TimeStep', y='NumHaltedUntilT', hue='Experiment', errorbar=('ci', 95), ax=ax, hue_order=hue_order, linewidth=3, linestyle='-')
ax.set_xlim([1-max_timestep*(1/20), max_timestep*(21/20)])
ax.set_xticks(xticks)
ax.set_xlabel('Timestep $t$')
ax.set_ylabel('#Samples with\nhalt time $\\hat{\\tau}(X) \leq t$')
# Get current handles and labels
handles, labels = ax.get_legend_handles_labels()
# Remove the first handle and label (which is typically the hue name)
handles = handles[2:]
labels = labels[2:]
handles[0].set_linestyle('--')
# Recreate the legend without the hue name
ax.legend(handles, labels, borderaxespad=0.2, borderpad=0.25)
# remove the legend
if dataset not in {'quality'}:
ax.get_legend().remove()
return conditional_accuracy_gap_fig
def plot_stage1_vs_stage2(dataset: str):
stage1_df = get_accuracy_gap_df(dataset, 'conditional_without_stage2', accuracy_gap=DESIRED_GAP)
stage2_df = get_accuracy_gap_df(dataset, 'conditional_accuracy_gap', accuracy_gap=DESIRED_GAP)
stage1_df['Experiment'] = 'Only Stage 1'
stage2_df['Experiment'] = 'Conditional Method: Stage 1+2'
combined_df = pd.concat([stage1_df, stage2_df], ignore_index=True)
# plot accuracy gap
conditional_accuracy_gap_fig, ax = plt.subplots(1, 1, figsize=(10, 5))
hue_order = ['Only Stage 1', 'Conditional Method: Stage 1+2']
default_palette = sns.color_palette()
orange_color = default_palette[1]
palette = ['#404040', orange_color]
df_dashed = combined_df[combined_df['Experiment'] == 'Only Stage 1']
df_solid = combined_df[combined_df['Experiment'] != 'Only Stage 1']
sns.lineplot(data=df_dashed, x='TimeStep', y='Value', hue='Experiment', units='Seed', estimator=None, lw=1, alpha=0.2, ax=ax, linestyle='--', hue_order=hue_order, palette=[palette[0]])
sns.lineplot(data=df_solid, x='TimeStep', y='Value', hue='Experiment', units='Seed', estimator=None, lw=1, alpha=0.2, ax=ax, linestyle='-', hue_order=hue_order, palette=[palette[1]])
desired_accuracy_gap = DESIRED_GAP
if desired_accuracy_gap is not None:
ax.axhline(y=DESIRED_GAP, color='r', linestyle=':')
# make a legend for the red line
ax.plot([], [], color='r', linestyle=':', label=f'$\\alpha$ level')
ax.set_xlabel('Timestep $t$')
ax.set_ylabel('Conditional accuracy gap')
if dataset == 'quality':
ax.set_ylim([0, 0.2])
else:
ax.set_ylim([0, 0.35])
# Get current handles and labels
handles, labels = ax.get_legend_handles_labels()
# Remove the first handle and label (which is typically the hue name)
handles = [handles[0], handles[3], handles[-1]]
labels = [labels[0], labels[3], '$\\alpha$ level']
handles[0].set_linestyle('--')
# Recreate the legend without the hue name
plt.legend(handles, labels, borderaxespad=0.1, borderpad=0.25)
return conditional_accuracy_gap_fig
def calc_cond_acc_gap_until_quantile_of_halt_time(dataset: str, cal_type: str, accuracy_gap: float, quantile: float):
is_correct_list_list = []
late_is_correct_list_list = []
halt_timesteps_list_list = []
for seed in SEEDS:
res_path = f'results/dataset={dataset}_seed={seed}_cal_type={cal_type}_accuracy_gap={accuracy_gap}.pt'
res_path2 = f'results/dataset={dataset}_seed={seed}_cal_type={cal_type}_accuracy_gap={accuracy_gap}.pt.2'
res = load_res(res_path)
res2 = load_res(res_path2)
is_correct_list_list.append(res2['is_correct'])
late_is_correct_list_list.append(res2['late_is_correct'])
halt_timesteps_list_list.append(res['halt_timesteps'])
is_correct_list_list = np.array(is_correct_list_list)
late_is_correct_list_list = np.array(late_is_correct_list_list)
halt_timesteps_list_list = np.array(halt_timesteps_list_list)
# sort indices by halt time
indices_by_halt_time = np.argsort(halt_timesteps_list_list, axis=1)
indices_before_quantile = indices_by_halt_time[:, :int(len(indices_by_halt_time[0]) * quantile)]
is_correct_before_quantile = is_correct_list_list[np.arange(len(indices_before_quantile))[:, None], indices_before_quantile]
late_is_correct_before_quantile = late_is_correct_list_list[np.arange(len(indices_before_quantile))[:, None], indices_before_quantile]
before_agg = np.maximum(late_is_correct_before_quantile.astype(float) - is_correct_before_quantile.astype(float), 0).mean(axis=1)
accuracy_gap_before_quantile = before_agg.mean()
accuracy_gap_before_quantile_std = before_agg.std() / np.sqrt(len(before_agg))
return accuracy_gap_before_quantile, accuracy_gap_before_quantile_std
def calc_time_used_and_accuracy(dataset: str, cal_type: str, accuracy_gap: float):
mean_halt_time_list = []
accuracy_list = []
late_accuracy_list = []
for seed in SEEDS:
res_path = f'results/dataset={dataset}_seed={seed}_cal_type={cal_type}_accuracy_gap={accuracy_gap}.pt'
res = load_res(res_path)
mean_halt_time_list.append(res['mean_halt_timesteps'])
accuracy_list.append(res['accuracy'])
late_accuracy_list.append(res['late_accuracy'])
mean_halt_time_array = (np.array(mean_halt_time_list) + 1)/DATASET_TO_MAX_LEN[dataset]
accuracy_array = np.array(accuracy_list)
late_accuracy_array = np.array(late_accuracy_list)
time_used_mean = np.mean(mean_halt_time_array)
time_used_std_err = np.std(mean_halt_time_array) / np.sqrt(len(mean_halt_time_array))
accuracy_mean = np.mean(accuracy_array)
accuracy_std_err = np.std(accuracy_array) / np.sqrt(len(accuracy_array))
late_accuracy_mean = np.mean(late_accuracy_array)
late_accuracy_std_err = np.std(late_accuracy_array) / np.sqrt(len(late_accuracy_array))
return time_used_mean, time_used_std_err, accuracy_mean, accuracy_std_err, late_accuracy_mean, late_accuracy_std_err
def get_t_avg_of_accuracy_gap(dataset: str, cal_type: str, accuracy_gap: float):
mean_halt_time_list = []
for seed in SEEDS:
res_path = f'results/dataset={dataset}_seed={seed}_cal_type={cal_type}_accuracy_gap={accuracy_gap}.pt'
res = load_res(res_path)
mean_halt_time_list.append(res['mean_halt_timesteps'])
mean_halt_time_array = (np.array(mean_halt_time_list) + 1)/DATASET_TO_MAX_LEN[dataset]
return mean_halt_time_array
def plot_t_avg_of_accuracy_gap_tiselac():
accuracy_gaps = [0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.1, 0.11, 0.12, 0.13, 0.14, 0.15, 0.16, 0.17, 0.18, 0.19, 0.2]
t_avg_array_list = []
for accuracy_gap in accuracy_gaps:
t_avg_array = get_t_avg_of_accuracy_gap('Tiselac', 'conditional_accuracy_gap', accuracy_gap)
t_avg_array_list.append(t_avg_array)
t_avg_array_array = np.array(t_avg_array_list)
# t_avg_array_array is of size (len(accuracy_gaps), len(SEEDS))
data = {
'Accuracy Gap': np.repeat(accuracy_gaps, len(SEEDS)),
'Seed': np.tile(SEEDS, len(accuracy_gaps)),
't_avg': t_avg_array_array.flatten()
}
df = pd.DataFrame(data)
plt.rcParams.update({'font.size': 12})
t_avg_of_accuracy_gap_fig, ax = plt.subplots(1, 1, figsize=(6, 3))
sns.lineplot(x='Accuracy Gap', y='t_avg', data=df, marker='o', ax=ax, errorbar='se', err_style='bars')
ax.set_xlabel('$\\alpha$')
ax.set_ylabel('$T_{\\text{avg}}$')
# change the font size of the x and y titles
ax.xaxis.label.set_size(14)
ax.yaxis.label.set_size(14)
ax.set_ylim([0, 1])
tick_spacing = 0.02
ax.xaxis.set_major_locator(ticker.MultipleLocator(tick_spacing))
return t_avg_of_accuracy_gap_fig
def get_latex_table():
dataset_list = []
cal_type_list = []
quantile_20_list = []
quantile_50_list = []
quantile_100_list = []
quantile_20_std_err_list = []
quantile_50_std_err_list = []
quantile_100_std_err_list = []
time_used_list = []
time_used_std_err_list = []
accuracy_list = []
accuracy_std_err_list = []
late_accuracy_list = []
late_accuracy_std_err_list = []
for dataset in DATASET_TO_MAX_LEN:
for cal_type in ['marginal_accuracy_gap', 'conditional_accuracy_gap']:
accuracy_gap_before_quantile20, accuracy_gap_before_quantile20_std = calc_cond_acc_gap_until_quantile_of_halt_time(dataset, cal_type, DESIRED_GAP, 0.2)
accuracy_gap_before_quantile50, accuracy_gap_before_quantile50_std = calc_cond_acc_gap_until_quantile_of_halt_time(dataset, cal_type, DESIRED_GAP, 0.5)
accuracy_gap_before_quantile100, accuracy_gap_before_quantile100_std = calc_cond_acc_gap_until_quantile_of_halt_time(dataset, cal_type, DESIRED_GAP, 1.0)
dataset_list.append(dataset)
cal_type_list.append(cal_type)
quantile_20_list.append(accuracy_gap_before_quantile20)
quantile_50_list.append(accuracy_gap_before_quantile50)
quantile_100_list.append(accuracy_gap_before_quantile100)
quantile_20_std_err_list.append(accuracy_gap_before_quantile20_std)
quantile_50_std_err_list.append(accuracy_gap_before_quantile50_std)
quantile_100_std_err_list.append(accuracy_gap_before_quantile100_std)
time_used, time_used_std_err, accuracy_mean, accuracy_std_err, late_accuracy_mean, late_accuracy_std_err = calc_time_used_and_accuracy(dataset, cal_type, DESIRED_GAP)
time_used_list.append(time_used)
time_used_std_err_list.append(time_used_std_err)
accuracy_list.append(accuracy_mean)
accuracy_std_err_list.append(accuracy_std_err)
late_accuracy_list.append(late_accuracy_mean)
late_accuracy_std_err_list.append(late_accuracy_std_err)
df = pd.DataFrame({
'Dataset': dataset_list,
'Calibration Type': cal_type_list,
'Accuracy Gap 20% first': quantile_20_list,
'Accuracy Gap 20% first Std Err': quantile_20_std_err_list,
'Accuracy Gap 50% first': quantile_50_list,
'Accuracy Gap 50% first Std Err': quantile_50_std_err_list,
'Accuracy Gap 100% first': quantile_100_list,
'Accuracy Gap 100% first Std Err': quantile_100_std_err_list,
'Time Used': time_used_list,
'Time Used Err': time_used_std_err_list,
'Accuracy': accuracy_list,
'Accuracy Err': accuracy_std_err_list,
'Late Accuracy': late_accuracy_list,
'Late Accuracy Err': late_accuracy_std_err_list,
})
# print latex table, each cell has the accuracy gap
latex = df.to_latex(index=False, float_format='%.3f')
return latex
def main():
print('Writing latex table')
latex = get_latex_table()
with open('figures/accuracy_gap.tex', 'w') as f:
f.write(latex)
print('Plotting marginal_vs_conditional')
for dataset in DATASET_TO_MAX_LEN:
conditional_accuracy_gap_fig = plot_marginal_vs_conditional(dataset)
conditional_accuracy_gap_fig.savefig(f'figures/{dataset}_marginal_vs_conditional_accuracy_gap.png', bbox_inches='tight', dpi=DPI)
print('Plotting stage1_vs_stage2')
stage1_vs_stage2_fig = plot_stage1_vs_stage2('quality')
stage1_vs_stage2_fig.savefig(f'figures/quality_stage1_vs_stage2.png', bbox_inches='tight', dpi=DPI)
print('Plotting t_avg_of_accuracy_gap_tiselac')
plot_t_avg_of_accuracy_gap_tiselac().savefig(f'figures/t_avg_of_accuracy_gap_tiselac.png', bbox_inches='tight', dpi=DPI)
if __name__ == '__main__':
main()