-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathkmeans_choose_prompt.py
195 lines (164 loc) · 7.07 KB
/
kmeans_choose_prompt.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
import numpy as np
import os
import time
import torch
from torch.autograd import Variable
from torchvision import transforms
from utils.misc import check_mkdir
from model.prompt_features import convnext_fea
from utils.dataset_rgb_strategy2 import test_get_loader
import cv2
import ttach as tta
from sklearn.cluster import KMeans
torch.manual_seed(2018)
torch.cuda.set_device(0)
to_pil = transforms.ToPILImage()
def Resize(image,H, W):
image = cv2.resize(image, dsize=(W, H), interpolation=cv2.INTER_LINEAR)
return image
img_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
transforms = tta.Compose(
[
tta.HorizontalFlip(),
tta.Scale(scales=[0.75, 1, 1.25], interpolation='bilinear', align_corners=False),
# tta.Scale(scales=[1], interpolation='bilinear', align_corners=False),
]
)
def genearte_prompt_features(image_shadow_root,ckpt_path,task_list):
t0 = time.time()
net = convnext_fea().cuda()
net = torch.nn.DataParallel(net)
net.eval()
with torch.no_grad():
test_loader = test_get_loader(image_shadow_root, batchsize=64, trainsize=384)
for i, (img, gt, img_name, w_, h_) in enumerate(test_loader):
img = Variable(img)
img_var = img.cuda()
n, c, h, w = img_var.size()
assert not torch.isnan(img_var).any()
model_output = net(img_var)
check_mkdir(os.path.join(ckpt_path, task_list))
for j in range(n):
# print(model_output.shape)
result = model_output[j,:,:,:].unsqueeze(0)
res = result.data.cpu().numpy()
# assert np.any(np.isnan(res))
print(i*n+j)
np.save(os.path.join(ckpt_path,task_list, img_name[j][:-4] + '.npy'), res)
# cv2.imwrite(os.path.join(ckpt_path, exp_name ,args['snapshot']+'epoch',task,name,task_list[3], img_name[j][:-4] + '.png'), res)
class KMEANS:
def __init__(self, n_clusters=64, max_iter=None, verbose=True,device = torch.device("cpu")):
# self.n_cluster = n_clusters
self.n_clusters = n_clusters
self.labels = None
self.dists = None # shape: [x.shape[0],n_cluster]
self.centers = None
self.variation = torch.Tensor([float("Inf")]).to(device)
self.verbose = verbose
self.started = False
self.representative_samples = None
self.max_iter = max_iter
self.count = 0
self.device = device
def fit(self, x):
# 随机选择初始中心点，想更快的收敛速度可以借鉴sklearn中的kmeans++初始化方法
init_row = torch.randint(0, x.shape[0], (self.n_clusters,)).to(self.device)
self.centers = x[init_row]
assert not torch.isnan(self.centers).any()
while True:
print("Iteration: ", self.count)
# 聚类标记
self.nearest_center(x)
# 更新中心点
self.update_center(x)
if self.verbose:
print(self.variation, torch.argmin(self.dists, (0)))
if torch.abs(self.variation) < 1e-3 and self.max_iter is None:
break
elif self.max_iter is not None and self.count == self.max_iter:
break
self.count += 1
print(self.dists.shape)
print(self.dists)
return self.representative_sample()
def nearest_center(self, x):
labels = torch.empty((x.shape[0],)).long().to(self.device)
dists = torch.empty((0, self.n_clusters)).to(self.device)
for i, sample in enumerate(x):
assert not torch.isnan(sample).any()
assert not torch.isnan(self.centers).any()
# dist = torch.sum(torch.mul(sample - self.centers, sample - self.centers), (1))
# print(sample.shape, self.centers.shape) # torch.Size([1024]) torch.Size([64, 1024])
dist = (sample - self.centers).pow(2).sum(dim=1)
labels[i] = torch.argmin(dist)
assert not torch.isnan(dist).any()
dists = torch.cat([dists, dist.unsqueeze(0)], (0))
self.labels = labels
if self.started:
self.variation = torch.sum(self.dists - dists)
self.dists = dists
self.started = True
def update_center(self, x):
centers = torch.empty((0, x.shape[1])).to(self.device)
for i in range(self.n_clusters):
mask = self.labels == i
cluster_samples = x[mask]
centers = torch.cat([centers, torch.mean(cluster_samples, (0)).unsqueeze(0)], (0))
self.centers = centers
def representative_sample(self):
# 查找距离中心点最近的样本，作为聚类的代表样本，更加直观
self.representative_samples = torch.argmin(self.dists, (0))
def time_clock(matrix,device):
a = time.time()
k = KMEANS(max_iter=10,verbose=False,device=device)
k.fit(matrix)
b = time.time()
return (b-a)/k.count
def choose_device(cuda=False):
if cuda:
device = torch.device("cuda:0")
else:
device = torch.device("cpu")
return device
def kemans_choose_prompt(ckpt_path,task_list):
device = choose_device(True)
# gpu_speeds = []
root1 = os.path.join(ckpt_path, task_list)
img_list = [os.path.splitext(f) for f in os.listdir(root1)]
for idx, img_name in enumerate(img_list):
print(img_name)
fea = np.load(os.path.join(ckpt_path,task_list, img_name[0] + '.npy'))
fea = (fea.mean(axis=(2,3)))
fea = fea/np.linalg.norm(fea,axis=1,keepdims=True)
assert not np.any(np.isnan(fea))
# print(fea.max(),fea.min())
if idx==0:
fea_total = fea
else:
fea_total = np.concatenate((fea_total,fea),0)
# print(fea_total.shape)
print(fea_total.shape)
k = KMeans(n_clusters=64)
dis = k.fit_transform(fea_total)
prompt_index = np.argmin(dis,axis=0).tolist()
prompt_list = []
for i in prompt_index:
prompt_list.append(img_list[i])
print(prompt_list)
if __name__ == '__main__':
ckpt_path = '/root/autodl-tmp/coding/DBS_code/prompt_features'
image_sod_root = "/root/autodl-tmp/datasets/dbs/DUTS/DUTS-TR"
image_cod_root = "/root/autodl-tmp/datasets/dbs/cod/TrainDataset/TrainDataset"
image_shadow_root = "/root/autodl-tmp/datasets/dbs/shadow_detection/SBU/SBU-shadow/SBUTrain4KRecoveredSmall"
image_transparent_root = "/root/autodl-tmp/datasets/dbs/transparent/train"
image_polyp_root = "/root/autodl-tmp/datasets/dbs/polyp/TrainDataset/TrainDataset"
image_covid_root = "/root/autodl-tmp/datasets/dbs/COVID-19_Lung_Infection_train"
image_breast_root = "/root/autodl-tmp/datasets/dbs/breast"
image_skin_root = "/root/autodl-tmp/datasets/dbs/isic2018/train"
task_list = ['SOD', 'COD', 'Shadow', 'Transparent', 'Polyp', 'COVID', 'Breast', 'Skin']
##############first run the genearte_prompt_features(), then run the kemans_choose_prompt
# genearte_prompt_features(image_breast_root,ckpt_path,task_list[6])
kemans_choose_prompt(ckpt_path,task_list[6])