-
Notifications
You must be signed in to change notification settings - Fork 9
/
main_fcn.py
328 lines (268 loc) · 13.8 KB
/
main_fcn.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
import torch
import torch.nn as nn
import random
import ipdb
import os
from tqdm import tqdm
from functools import partial
import numpy as np
np.warnings.filterwarnings('ignore')
import torch.utils.data as data
import torchvision.transforms as transforms
from torch import default_generator
from PIL import Image
from datasets import datasets
import datasets.transforms as ext_transforms
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
import argparse
import matplotlib.pyplot as plt
from conformal import helper
from conformal.fcn_model import make_fcn, FCN, enet_weighing
from conformal.icp import IcpRegressor, RegressorNc, FeatRegressorNc
from conformal.icp import AbsErrorErrFunc, FeatErrorErrFunc
from conformal.utils import compute_coverage, WeightedMSE, seed_torch
torch.backends.cuda.matmul.allow_tf32 = False
torch.backends.cudnn.allow_tf32 = False
def makedirs(path):
if not os.path.exists(path):
print('creating dir: {}'.format(path))
os.makedirs(path)
else:
print(path, "already exist!")
def visualize(image, height, width, save_dir):
fig, ax = plt.subplots()
ax.imshow(image)
plt.axis("off")
fig.set_size_inches(width / 100.0 / 3.0, height / 100.0 / 3.0)
plt.gca().xaxis.set_major_locator(plt.NullLocator())
plt.gca().yaxis.set_major_locator(plt.NullLocator())
plt.subplots_adjust(top=1, bottom=0, left=0, right=1, hspace=0, wspace=0)
plt.margins(0, 0)
plt.xticks([])
plt.yticks([])
plt.savefig(save_dir, pad_inches=0)
plt.close()
def random_split(dataset, lengths, generator=default_generator, cal_test_generator=None):
"""
Refactor the random_split function of pytorch.
Randomly split a dataset into non-overlapping new datasets of given lengths.
Optionally fix the generator for reproducible results, e.g.:
>>> random_split(range(10), [3, 7], generator=torch.Generator().manual_seed(42))
Arguments:
dataset (Dataset): Dataset to be split
lengths (sequence): lengths of splits to be produced
generator (Generator): Generator used for the random permutation.
"""
# Cannot verify that dataset is Sized
if sum(lengths) != len(dataset): # type: ignore
raise ValueError("Sum of input lengths does not equal the length of the input dataset!")
indices = torch.randperm(sum(lengths), generator=generator)
if cal_test_generator is None:
indices = indices.tolist()
return [data.Subset(dataset, indices[offset - length: offset]) for offset, length in zip(torch._utils._accumulate(lengths), lengths)]
else:
# only shuffle the calibration and test set
assert len(lengths) == 4
train_val_indices = indices[:sum(lengths[:2])]
cal_test_indices = indices[-sum(lengths[-2:]):]
assert len(train_val_indices) + len(cal_test_indices) == len(indices)
cal_test_indices = cal_test_indices[torch.randperm(len(cal_test_indices), generator=cal_test_generator)]
final_indices = train_val_indices.tolist() + cal_test_indices.tolist()
assert len(final_indices) == len(indices)
return [data.Subset(dataset, final_indices[offset - length: offset]) for offset, length in zip(torch._utils._accumulate(lengths), lengths)]
def load_dataset(dataset, seed):
print("\nLoading dataset...\n")
print("Selected dataset:", args.data)
print("Dataset directory:", args.dataset_dir)
image_transform = transforms.Compose(
[transforms.Resize((args.height, args.width)),
transforms.ToTensor()])
label_transform = transforms.Compose([
transforms.Resize((args.height, args.width), Image.NEAREST),
ext_transforms.PILToLongTensor(),
ext_transforms.ToOnehotGaussianBlur(kernel_size=7, num_classes=dataset.num_classes)
])
whole_train_set = dataset(
args.dataset_dir,
transform=image_transform,
label_transform=label_transform)
if args.data_seed is None:
cal_test_generator = torch.Generator().manual_seed(seed)
else:
cal_test_generator = torch.Generator().manual_seed(args.data_seed)
train_set, val_set, cal_set, test_set = random_split(whole_train_set, [1800, 200, 600, len(whole_train_set) - 2600],
generator=torch.Generator().manual_seed(0),
cal_test_generator=cal_test_generator)
train_loader = data.DataLoader(
train_set,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
cal_loader = data.DataLoader(
cal_set,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True
)
test_loader = data.DataLoader(
test_set,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
# Get encoding between pixel valus in label images and RGB colors
class_encoding = whole_train_set.color_encoding
# Get number of classes to predict
num_classes = len(class_encoding)
# Print information for debugging
print("Number of classes to predict:", num_classes)
print("Train dataset size:", len(train_set))
print("Calibration dataset size:", len(cal_set))
print("test dataset size:", len(test_set))
# Get a batch of samples to display
images, _, labels = iter(train_loader).next()
print("Image size:", images.size())
print("Label size:", labels.size())
print("Class-color encoding:", class_encoding)
print("Computing class weights...")
print("(this can take a while depending on the dataset size)")
class_weights = enet_weighing(train_loader, num_classes)
if class_weights is not None:
class_weights = torch.from_numpy(class_weights).float().to(device)
# Set the weight of the unlabeled class to 0
ignore_index = list(class_encoding).index('unlabeled')
class_weights[ignore_index] = 0
print("Class weights:", class_weights)
return (train_loader, cal_loader, test_loader), class_weights, class_encoding
def main(train_loader, cal_loader, test_loader, args):
dir = f"ckpt/{args.data}"
if os.path.exists(os.path.join(dir, "model.pt")):
model = make_fcn(pretrained=True, img_shape=image_shape, num_classes=20)
print(f"==> Load model from {dir}")
model.load_state_dict(torch.load(os.path.join(dir, "model.pt"), map_location=device)['state_dict'])
else:
raise ValueError("does not find the checkpoint in {}".format(dir))
mean_estimator = helper.MSENet_RegressorAdapter(model=model, device=device, fit_params=None,
in_shape=image_shape, out_shape=args.height*args.width*num_classes,
hidden_size=args.hidden_size, learn_func=nn_learn_func, epochs=args.epochs,
batch_size=args.batch_size, dropout=args.dropout, lr=args.lr, wd=args.wd,
test_ratio=cv_test_ratio, random_state=cv_random_state, )
if float(args.feat_norm) <= 0 or args.feat_norm == "inf":
args.feat_norm = "inf"
print("Use inf as feature norm")
else:
args.feat_norm = float(args.feat_norm)
criterion = WeightedMSE(out_shape=(args.height, args.width), weight=class_weights)
# FeatureCP
nc = FeatRegressorNc(mean_estimator, criterion=criterion, inv_lr=args.feat_lr, inv_step=args.feat_step,
feat_norm=args.feat_norm, certification_method=args.cert_method,
g_out_process=partial(model.output_post_process, img_shape=image_shape))
icp = IcpRegressor(nc)
icp.calibrate_batch(cal_loader)
# calculating the coverage of FCP
in_coverage = icp.if_in_coverage_batch(test_loader, significance=alpha)
coverage_fcp = np.sum(in_coverage) * 100 / len(in_coverage)
test_intervals = []
all_y_test = []
img_idx = 0
for x_test, discrete_y_test, y_test in tqdm(test_loader):
intervals = icp.predict(x_test, significance=alpha)
test_intervals.append(intervals)
all_y_test.append(y_test.cpu().numpy())
# this is used for visualization
if args.visualize:
loglog_interval = np.exp(-np.exp(intervals))
show_interval = np.abs(loglog_interval[..., 1] - loglog_interval[..., 0])
for img_interval in show_interval:
img_interval = img_interval.reshape(19, args.height, args.width).mean(axis=0)
img_name = os.path.basename(test_loader.dataset.dataset.train_data[test_loader.dataset.indices[img_idx]])
visualize(img_interval, height=args.height, width=args.width, save_dir=os.path.join(f'visualization/seed{seed}', img_name))
img_idx += 1
test_intervals = np.concatenate(test_intervals, axis=0)
all_y_test = np.concatenate(all_y_test, axis=0)
# estimating the length of FCP
y_lower, y_upper = test_intervals[..., 0], test_intervals[..., 1]
_, length_fcp = compute_coverage(all_y_test, y_lower, y_upper, alpha, "FeatRegressorNc")
# Vanilla CP
icp2 = IcpRegressor(RegressorNc(mean_estimator))
icp2.calibrate_batch(cal_loader)
test_intervals = []
all_y_test = []
for x_test, _, y_test in tqdm(test_loader):
intervals = icp2.predict(x_test, significance=alpha)
test_intervals.append(intervals)
all_y_test.append(y_test.cpu().numpy())
test_intervals = np.concatenate(test_intervals, axis=0)
all_y_test = np.concatenate(all_y_test, axis=0)
y_lower, y_upper = test_intervals[..., 0], test_intervals[..., 1]
coverage_cp, length_cp = compute_coverage(all_y_test, y_lower, y_upper, alpha, "RegressorNc")
return coverage_fcp, length_fcp, coverage_cp, length_cp
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--device", "--d", default=-1, type=int)
parser.add_argument('--seed', type=int, nargs='+', default=[0])
parser.add_argument("--data", type=str, default="cityscapes", help="only support cityscapes now", choices=['cityscapes'])
parser.add_argument("--data-seed", type=int, default=None, help="the random seed to split the calibration and test sets")
parser.add_argument("--dataset-dir", type=str, default=None,
help="Path to the root directory of the selected dataset.")
parser.add_argument("--workers", type=int, default=4, help="Number of subprocesses to use for data loading. Default: 4")
parser.add_argument("--alpha", type=float, default=0.1, help="miscoverage error")
parser.add_argument("--height", type=int, default=256, help="The image height. Default: 512")
parser.add_argument("--width", type=int, default=512, help="The image width. Default: 1024")
parser.add_argument("--lr", type=float, default=5e-4)
parser.add_argument("--epochs", type=int, default=10)
parser.add_argument("--batch_size", "--bs", type=int, default=64)
parser.add_argument("--hidden_size", "--hs", type=int, default=64)
parser.add_argument("--dropout", "--do", type=float, default=0.1)
parser.add_argument("--wd", type=float, default=1e-6)
parser.add_argument("--no-resume", action="store_true", default=False)
parser.add_argument("--feat_opt", "--fo", type=str, default="sgd", choices=["sgd", "adam"])
parser.add_argument("--feat_lr", "--fl", type=float, default=1e-2)
parser.add_argument("--feat_step", "--fs", type=int, default=None)
parser.add_argument("--feat_norm", "--fn", default=-1)
parser.add_argument("--cert_method", "--cm", type=int, default=0, choices=[0, 1, 2, 3])
parser.add_argument("--visualize", action="store_true", default=False, help="visualize the length in the image")
args = parser.parse_args()
fcp_coverage_list, fcp_length_list, cp_coverage_list, cp_length_list = [], [], [], []
for seed in args.seed:
seed_torch(seed)
os.environ['CUDA_VISIBLE_DEVICES'] = "{:}".format(args.device)
device = torch.device("cpu") if args.device < 0 else torch.device("cuda")
if args.visualize:
makedirs(f"./visualization/seed{seed}")
nn_learn_func = torch.optim.Adam
# ratio of held-out data, used in cross-validation
cv_test_ratio = 0.05
# desired miscoverage error
# alpha = 0.1
alpha = args.alpha
# used to determine the size of test set
test_ratio = 0.2
# seed for splitting the data in cross-validation.
cv_random_state = 1
if args.data.lower() == 'cityscapes':
from datasets.cityscapes import Cityscapes as dataset
else:
# Should never happen...but just in case it does
raise RuntimeError("\"{0}\" is not a supported dataset.".format(
args.dataset))
(train_loader, cal_loader, test_loader), class_weights, class_encoding = load_dataset(dataset, seed)
# remove the first dimension "unlabeled"
class_weights = class_weights[1:]
image_shape = (args.height, args.width)
n_train = len(train_loader.dataset)
num_classes = len(class_encoding)
print("Dataset: %s" % (args.data))
coverage_fcp, length_fcp, coverage_cp, length_cp = \
main(train_loader, cal_loader, test_loader, args)
fcp_coverage_list.append(coverage_fcp)
fcp_length_list.append(length_fcp)
cp_coverage_list.append(coverage_cp)
cp_length_list.append(length_cp)
print(f'FeatureCP coverage: {np.mean(fcp_coverage_list)} \pm {np.std(fcp_coverage_list)}',
f'FeatureCP estimated length: {np.mean(fcp_length_list)} \pm {np.std(fcp_length_list)}')
print(f'VanillaCP coverage: {np.mean(cp_coverage_list)} \pm {np.std(cp_coverage_list)}',
f'VanillaCP length: {np.mean(cp_length_list)} \pm {np.std(cp_length_list)}')