-
Notifications
You must be signed in to change notification settings - Fork 13
/
db_quant_utils.py
157 lines (141 loc) · 5.97 KB
/
db_quant_utils.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
import numpy as np
import torch
import torch.nn as nn
from skimage import measure
from PIL import Image
import os
from data import make_planeloader
from evaluation import decision_boundary
##Margin stuff
def delta_counter(max_steps):
delta_list = [0]
delta = 0
step = 0.01
for count in range(max_steps-1):
# print(count,step)
if count > 10 and count < 20:
step = 0.1
elif count >= 20 and count <= 50:
step = 0.5
elif count > 50 and count <= 75:
step = 1
elif count > 75 and count <= 100:
step = 10
elif count > 100:
step = 20
delta+=step
delta_list.append(delta)
# print(delta_list)
return np.array(delta_list)
def avg_margin(dlist1,dlist2,loader,num_iters, num_samples,net,device,max_steps=200):
print(f'In the case of list lengths: {len(dlist1)},{len(dlist2)}')
baselist = np.random.choice(dlist1, num_samples)
deltas = torch.from_numpy(delta_counter(max_steps).reshape(max_steps,1,1,1))
deltas = deltas.to(device)
deltas_temp = torch.squeeze(deltas)
mean_margins = []
for i in baselist:
# print(f'The image: {i}')
dirlist = np.random.choice(dlist2, num_iters+10)
margins = []
iter_count = 0
while len(margins) < num_iters and iter_count < num_iters+10:
j = dirlist[iter_count]
iter_count+=1
img1 = loader.dataset[i][0].unsqueeze(0)
img2 = loader.dataset[j][0].unsqueeze(0)
a = img2 - img1
a_norm = torch.dot(a.flatten(), a.flatten()).sqrt()
a = a / a_norm
img1 = loader.dataset[i][0].unsqueeze(0).expand(max_steps,-1,-1,-1)
a = a.expand(max_steps,-1,-1,-1)
img1 = img1.to(device)
a = a.to(device)
img_batch = img1 + deltas*a
img_batch = img_batch.to(device=device, dtype=torch.float)
preds = torch.argmax(net(img_batch),dim=1).cpu().numpy()
where_db = np.where(np.diff(preds) != 0)[0]
if where_db.size !=0:
delta = deltas_temp[where_db[0]].item()
margins.append(delta)
if len(margins) >= num_iters//2:
mean_margins.append(np.mean(margins))
pdone = len(mean_margins)
if pdone%1000 ==0:
print(f'At {pdone}th point')
return mean_margins
## Class Skips
def rel_index(class_index,trainset,testset,noise_rate):
l_all_train = np.where(np.array(trainset.targets) == class_index)[0]
l_mis_cls = 0
l_fromcls_mislab = 0
l_corr_cls = l_all_train
if noise_rate > 0:
l_mis = np.where(np.array(trainset.targets) != np.array(trainset.true_targets))[0] #mislabeled images
l_corr = np.where(np.array(trainset.targets) == np.array(trainset.true_targets))[0] #correctly labeled images
l_mis_cls = np.intersect1d(l_all_train,l_mis)
l_corr_cls = np.intersect1d(l_all_train,l_corr)
l_fromcls_mislab = np.intersect1d(np.where(np.array(trainset.true_targets) == class_index)[0],l_mis)
l_all_test = np.where(np.array(testset.targets) == class_index)[0]
return l_all_train, l_mis_cls, l_corr_cls,l_all_test,l_fromcls_mislab
def count_skips(dlist1,dlist2,loader1,loader2,num_steps,net,device):
skips_list = []
for i in range(len(dlist1)):
img1 = loader1.dataset[dlist1[i]][0].unsqueeze(0).expand(num_steps,-1,-1,-1)
img2 = loader2.dataset[dlist2[i]][0].unsqueeze(0).expand(num_steps,-1,-1,-1)
img1 = img1.to(device)
img2 = img2.to(device)
alphas = torch.from_numpy(np.linspace(0,1,num_steps)).reshape(num_steps,1,1,1)
alphas = alphas.to(device)
img_batch = img1 + alphas*(img2-img1)
img_batch = img_batch.to(device=device, dtype=torch.float)
preds = torch.argmax(net(img_batch),dim=1).cpu().numpy()
skip = (np.diff(preds)!=0).sum()
skips_list.append(skip)
return skips_list
def connected_components(preds,args):
preds = torch.stack((preds))
preds = nn.Softmax(dim=1)(preds)
class_pred = torch.argmax(preds, dim=1).cpu().numpy()
resolution = int(args.resolution)
img = np.zeros((resolution, resolution)).astype(np.uint8)
img[np.arange(resolution).repeat(resolution), np.tile(np.arange(resolution), resolution)] = class_pred
unique_classes = np.unique(img)
cc_counts = []
for lbl in unique_classes:
_, count = measure.label((img==lbl).astype(np.uint8), background=0, return_num=True, connectivity=2)
cc_counts.append(count)
if args.plot_path is not None:
path = args.plot_path
img_dir = '/'.join([p for p in (path.split('/'))[:-1]])
os.makedirs(img_dir, exist_ok=True)
colors = np.array([[ 3, 41, 8],
[231, 50, 144],
[144, 96, 19],
[141, 21, 179],
[ 26, 42, 130],
[215, 93, 7],
[ 85, 88, 251],
[137, 112, 156],
[167, 245, 192],
[243, 20, 230]])
Image.fromarray(colors[img].astype(np.uint8)).save(f'{path}_fragmentation.png')
return np.sum(cc_counts)
import wandb
def num_connected_components(dlist1,dlist2, loader1,loader2, num_samples,net,device,args):
cc_list = []
for i in range(num_samples):
# import ipdb; ipdb.set_trace()
dirlist = np.random.choice(dlist1, 2)
dirlist2 = np.random.choice(dlist2, 1)
images = [loader1.dataset[j][0] for j in dirlist]
images.append(loader2.dataset[dirlist2[0]][0])
planeloader = make_planeloader(images, args)
preds = decision_boundary(args, net, planeloader, device)
ncc = connected_components(preds,args)
cc_list.append(ncc)
# print(dirlist,dirlist2,ncc)
if i%100==0:
if args.active_log:
wandb.log({'iteration':i})
return cc_list