-
Notifications
You must be signed in to change notification settings - Fork 0
/
CSQ_HashUQ.py
391 lines (330 loc) · 16.8 KB
/
CSQ_HashUQ.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
from utils.tools import *
from network import *
import os
import torch
import torch.optim as optim
import time
import numpy as np
from scipy.linalg import hadamard # direct import hadamrd matrix from scipy
import argparse
import random
import pandas as pd
from tqdm import tqdm
from functools import partialmethod
import math
torch.multiprocessing.set_sharing_strategy('file_system')
# CSQ(CVPR2020)
# paper [Central Similarity Quantization for Efficient Image and Video Retrieval](https://openaccess.thecvf.com/content_CVPR_2020/papers/Yuan_Central_Similarity_Quantization_for_Efficient_Image_and_Video_Retrieval_CVPR_2020_paper.pdf)
# code [CSQ-pytorch](https://github.com/yuanli2333/Hadamard-Matrix-for-hashing)
# AlexNet
# [CSQ] epoch:65, bit:64, dataset:cifar10-1, MAP:0.787, Best MAP: 0.790
# [CSQ] epoch:90, bit:16, dataset:imagenet, MAP:0.593, Best MAP: 0.596, paper:0.601
# [CSQ] epoch:150, bit:64, dataset:imagenet, MAP:0.698, Best MAP: 0.706, paper:0.695
# [CSQ] epoch:40, bit:16, dataset:nuswide_21, MAP:0.784, Best MAP: 0.789
# [CSQ] epoch:40, bit:32, dataset:nuswide_21, MAP:0.821, Best MAP: 0.821
# [CSQ] epoch:40, bit:64, dataset:nuswide_21, MAP:0.834, Best MAP: 0.834
# ResNet50
# [CSQ] epoch:20, bit:64, dataset:imagenet, MAP:0.881, Best MAP: 0.881, paper:0.873
# [CSQ] epoch:10, bit:64, dataset:nuswide_21_m, MAP:0.844, Best MAP: 0.844, paper:0.839
# [CSQ] epoch:40, bit:64, dataset:coco, MAP:0.870, Best MAP: 0.883, paper:0.861
eps = 1e-8
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--grad_est", help = 'Gradient Estimator for Discrete Latent', type = str, default = 'st')
parser.add_argument("--u2g_samples", help = 'Number of u2g samples for CLS loss head', type = int, default = 1)
#Assume u2g_recon < u2g_samples
# parser.add_argument("--u2g_recon", help = 'Number of u2g samples for reconstruction head', type = int, default = 1)
parser.add_argument("--KL_regularization", help = 'scalar to tradeoff likelihood term and KL term', type = float, default = 0)
parser.add_argument("--prior", help = 'prior Bernoulli success probability', type = float, default = 0.5)
# parser.add_argument("--st_pretrain", help = 'Number of epochs pretrain with ST (Valid for U2G for now)', type = int, default = 0)
#Todo: inplement pairwise dataset class for CIFAR-10 dataset
parser.add_argument('--pairwise', action='store_true')
parser.add_argument('--no-pairwise', dest='pairwise', action='store_false')
parser.set_defaults(pairwise=False)
parser.add_argument('--pair_weight', help = 'scalar to tradeoff pairwise term', type = float, default = 0)
parser.add_argument('--tqdm', action='store_true')
parser.add_argument('--no-tqdm', dest='tqdm', action='store_false')
parser.set_defaults(tqdm=False)
parser.add_argument('--dataset', type = str, default = "imagenet", help = " ")
parser.add_argument('--rt_st', type = str, default = "ham", help = "Retrieval Stratege, \"ham\" or \"hamuct\"")
parser.add_argument('--uct_quant_level', type = int, default = 0, help = "Quantization level of uncertainty, 0=No Quant")
parser.add_argument('--val', action='store_true')
parser.add_argument('--no-val', dest='val', action='store_false')
parser.add_argument("--train", action = 'store_true')
parser.add_argument('--no-train', dest='train', action='store_false')
parser.set_defaults(train=True)
parser.add_argument('--sample', type = int, default = 1)
parser.add_argument('--sample_method', type = str, default = "MCD") #MCD: MC Dropout, IGN: Input Gaussian Noise
parser.add_argument('--eps', type = float, default = 1e-100)
parser.set_defaults(val=False)
args = parser.parse_args()
print('Arguments:', args)
return args
def get_config():
config = {
"lambda": 0.0001,
"optimizer": {"type": optim.RMSprop, "optim_params": {"lr": 1e-5, "weight_decay": 10 ** -5}},
"info": "[CSQ]",
"resize_size": 256,
"crop_size": 224,
"batch_size": 64,
"net": AlexNet,
# "net": ResNet,
# "dataset": "cifar10-1",
# "dataset": "imagenet",
# "dataset": "coco",
# "dataset": "nuswide_21",
# "dataset": "nuswide_21_m",
"epoch": 100,
"test_map": 10,
# "device":torch.device("cpu"),
"device": torch.device("cuda:0"),
"bit_list": [16],
"seed": 1,
}
# config = config_dataset(config)
return config
class CSQLoss(torch.nn.Module):
def __init__(self, config, bit):
super(CSQLoss, self).__init__()
self.is_single_label = config["dataset"] not in {"nuswide_21", "nuswide_21_m", "coco"}
self.hash_targets = self.get_hash_targets(config["n_class"], bit).to(config["device"])
self.multi_label_random_center = torch.randint(2, (bit,)).float().to(config["device"])
self.criterion = torch.nn.BCELoss().to(config["device"])
def forward(self, u, y, ind, config):
# u = u.tanh()
prob = torch.sigmoid(u)
hash_center = self.label2center(y)
# print("hash centre:", hash_center)
# Hamming_loss = torch.sum(torch.abs(hash_center - b)/2, dim = (0, 1))/(config['batch_size']*bit)
if config["grad_est"] == 'cf':
Hamming_loss = torch.sum(prob + (hash_center+1)/2 - 2*prob*(hash_center+1)/2)/(config['batch_size']*bit)
elif config["grad_est"] == 'cf2':
Hamming_loss = torch.sum(prob*(1 - (hash_center+1)) + ((hash_center+1)/2)**2)/(config['batch_size']*bit)
else:
if config["grad_est"] == 'st':
b = 2 * self.Sample_Bernoulli.apply(prob) - 1
elif config["grad_est"] == 'gs':
b = 2 * self.Sample_GS(prob, config) - 1
elif config["grad_est"] == 'u2g':
b = 2 * self.Sample_Bernoulli.apply(prob) - 1
Hamming_loss = torch.sum(torch.abs(hash_center - b)/2, dim = (0, 1))/(config['batch_size']*bit)
return Hamming_loss
def label2center(self, y):
if self.is_single_label:
hash_center = self.hash_targets[y.argmax(axis=1)]
else:
# to get sign no need to use mean, use sum here
center_sum = y @ self.hash_targets
random_center = self.multi_label_random_center.repeat(center_sum.shape[0], 1)
center_sum[center_sum == 0] = random_center[center_sum == 0]
hash_center = 2 * (center_sum > 0).float() - 1
return hash_center
def is_power_of_two(self, n):
if n <= 0:
return False
return (n & (n - 1)) == 0
# use algorithm 1 to generate hash centers
def get_hash_targets(self, n_class, bit):
if self.is_power_of_two(bit):
#If bit = 2^n, ELSE:
H_K = hadamard(bit)
else:
# ceil_bit = int(math.pow(2,math.ceil(math.log2(bit))))
# print("ceil bit", ceil_bit)
# H_K = hadamard(ceil_bit)[:, random.sample(range(ceil_bit), bit)]
floor_bit = int(math.pow(2,math.floor(math.log2(bit))))
H_K = np.concatenate((hadamard(floor_bit), 2*np.random.randint(2, size = (floor_bit, bit - floor_bit)) - 1), axis = 1)
H_2K = np.concatenate((H_K, -H_K), 0)
hash_targets = torch.from_numpy(H_2K[:n_class]).float()
if H_2K.shape[0] < n_class:
hash_targets.resize_(n_class, bit)
for k in range(20):
for index in range(H_2K.shape[0], n_class):
ones = torch.ones(bit)
# Bernouli distribution
sa = random.sample(list(range(bit)), bit // 2)
ones[sa] = -1
hash_targets[index] = ones
# to find average/min pairwise distance
c = []
for i in range(n_class):
for j in range(n_class):
if i < j:
TF = sum(hash_targets[i] != hash_targets[j])
c.append(TF)
c = np.array(c)
# choose min(c) in the range of K/4 to K/3
# see in https://github.com/yuanli2333/Hadamard-Matrix-for-hashing/issues/1
# but it is hard when bit is small
if c.min() > bit / 4 and c.mean() >= bit / 2:
print(c.min(), c.mean())
break
return hash_targets
def f_function_cls(self, b, y, config):
hash_center = self.label2center(y)
# center_loss = self.criterion(0.5 * (u + 1), 0.5 * (hash_center + 1))
Hamming_loss = torch.sum(torch.abs(hash_center - b)/2, dim = (0, 1))/(config['batch_size']*bit)
return Hamming_loss
def estimate_grad(self, u, y, ind, config):
batch_size, bit_length = u.size()
sample_u = torch.rand((config["u2g_samples"], batch_size, bit_length)).to(config["device"])
Score2 = torch.sigmoid(u)
Score1 = 1 - Score2
Indicator1 = (sample_u>Score1).float()
Indicator2 = (sample_u<Score2).float()
loss_1 = []
loss_2 = []
with torch.no_grad():
for i in range(config["u2g_samples"]):
loss_1_im = self.f_function_cls(2 * Indicator1[i] - 1, y, config)
loss_2_im = self.f_function_cls(2 * Indicator2[i] - 1, y, config)
loss_1.append(loss_1_im.item())
loss_2.append(loss_2_im.item())
del loss_1_im, loss_2_im
torch.cuda.empty_cache()
loss_1 = torch.as_tensor(loss_1).to(config["device"])
loss_2 = torch.as_tensor(loss_2).to(config["device"])
grad_u2g = torch.mean(torch.sigmoid(torch.abs(u))*(loss_1 - loss_2).unsqueeze(1).unsqueeze(2).repeat(1, batch_size, bit_length)*(Indicator1 - Indicator2)/2, dim = 0)
return grad_u2g.detach()
def estimate_loss(self, u, y, ind, config):
prob = torch.sigmoid(u)
b = 2 * self.Sample_Bernoulli.apply(prob) - 1
hash_center = self.label2center(y)
Hamming_loss = torch.sum(torch.abs(hash_center - b)/2, dim = (0, 1))/(config['batch_size']*bit)
return Hamming_loss
def Sample_GS(self, prob, config, lam = 0.5):
size = prob.size()
u = torch.rand(size).to(config['device'])
alpha = (prob + eps)/(1 - prob + eps)
L = torch.log(u) - torch.log(1 - u)
bb = torch.sigmoid((torch.log(alpha) + L)/lam)
return bb
class Sample_Bernoulli(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
code = torch.bernoulli(input)
return code
@staticmethod
def backward(ctx, grad_output):
return grad_output
def train_val(config, bit):
device = config["device"]
random.seed(config["seed"])
np.random.seed(config["seed"])
torch.manual_seed(config["seed"])
if not config["tqdm"]:
tqdm.__init__ = partialmethod(tqdm.__init__, disable=True)
#Need to check whether we use save_path somewhere else.
save_path = os.path.join(".", "saved_model", "BerVAE_bit_%i_optimizer_%s_KL_%.2f_dataset_%s" %(bit, config["grad_est"], config["KL_regularization"], config["dataset"]))
print("save_path", save_path)
config["save_path"] = save_path
if not os.path.exists(os.path.join(".", "saved_model")):
os.mkdir(os.path.join(".", "saved_model"))
if not os.path.exists(save_path):
os.mkdir(save_path)
training_loss_list = []
val_loss_list = []
map_list = []
kl_list = []
pw_list = []
torch.set_printoptions(sci_mode = True)
if config["val"] == True:
train_loader, val_loader, test_loader, dataset_loader, num_train, num_val, num_test, num_dataset = get_data(config)
else:
train_loader, test_loader, dataset_loader, num_train, num_test, num_dataset = get_data(config)
config["num_train"] = num_train
net = config["net"](bit).to(device)
# optimizer = config["optimizer"]["type"](net.parameters(), **(config["optimizer"]["optim_params"]))
criterion = CSQLoss(config, bit)
optimizer_net = config["optimizer"]["type"](list(net.parameters()), **(config["optimizer"]["optim_params"]))
# optimizer_criterion = config["optimizer"]["type"](list(criterion.parameters()), **(config["optimizer"]["optim_params"]))
Best_mAP = 0
save_hash_center(criterion, config)
if config["train"] == True:
for epoch in range(config["epoch"]):
current_time = time.strftime('%H:%M:%S', time.localtime(time.time()))
print("%s[%2d/%2d][%s] bit:%d, dataset:%s, training...." % (
config["info"], epoch + 1, config["epoch"], current_time, bit, config["dataset"]), end="")
net.train()
train_loss = 0
for image, label, ind in train_loader:
image = image.to(device)
label = label.to(device)
# optimizer.zero_grad()
optimizer_net.zero_grad()
u = net(image)
# print('u', u)
prob = torch.sigmoid(u)
loss = criterion(u, label.float(), ind, config)
KL = torch.sum((prob*torch.log(torch.div(prob, config["prior"]) + eps) + (1 - prob)*torch.log(torch.div(1- prob, 1 - config["prior"]) + eps)))/(config['batch_size']*bit)
KL_loss = config["KL_regularization"]*KL
train_loss += loss.item()
if config["grad_est"] in ['st', 'gs', 'cf', 'cf2']:
loss.backward(retain_graph = True)
KL_loss.backward(retain_graph = True)
optimizer_net.step()
elif config["grad_est"] == 'u2g':
u_grad = criterion.estimate_grad(u, label.float(), ind, config)
u.backward(u_grad.detach(), retain_graph = True)
KL_loss.backward(retain_graph = True)
optimizer_net.step()
train_loss = train_loss / len(train_loader)
print("\b\b\b\b\b\b\b loss:%.3f" % (train_loss))
if (epoch + 1) % config["test_map"] == 0:
if config["val"] == True:
Best_mAP, mAP = self_validate(config, Best_mAP, val_loader, net, bit, epoch, num_val, return_map = True)
else:
Best_mAP, mAP = validate(config, Best_mAP, test_loader, dataset_loader, net, bit, epoch, num_dataset, return_map = True)
net.eval()
criterion.eval()
train_loss = 0
with torch.no_grad():
for data in train_loader:
(image, label, ind) = data
image = image.to(device)
label = label.to(device)
u = net(image)
loss = criterion.estimate_loss(u, label.float(), ind, config)
train_loss += loss.item()
train_loss = train_loss / len(train_loader)
val_loss = 0
if config["val"] == True:
with torch.no_grad():
for data in val_loader:
(image, label, ind) = data
image = image.to(device)
label = label.to(device)
u = net(image)
loss = criterion.estimate_loss(u, label.float(), ind, config)
val_loss += loss.item()
val_loss = val_loss / len(val_loader)
torch.save(net.state_dict(), os.path.join(save_path, "epoch_%i.pth" %(epoch)))
training_loss_list.append(train_loss)
val_loss_list.append(val_loss)
map_list.append(mAP)
kl_list.append(KL.item())
else:
map_list = pd.read_csv(os.path.join(save_path, "map.csv"))["0"]
#Training Done
Best_Epoch = (np.argmax(map_list) + 1)*config["test_map"]
Best_Model_Path = os.path.join(save_path, "epoch_%i.pth" %(Best_Epoch - 1))
net.load_state_dict(torch.load(Best_Model_Path))
Best_mAP = 0
Best_mAP, mAP = validate(config, Best_mAP, test_loader, dataset_loader, net, bit, Best_Epoch - 1, num_dataset, return_map = True)
if config["train"] == True:
pd.DataFrame(training_loss_list).to_csv(os.path.join(save_path, "train_loss.csv"))
pd.DataFrame(training_loss_list).to_csv(os.path.join(save_path, "val_loss.csv"))
pd.DataFrame(map_list).to_csv(os.path.join(save_path, "map.csv"))
pd.DataFrame(kl_list).to_csv(os.path.join(save_path, "kl.csv"))
if __name__ == "__main__":
config = get_config()
print(config)
args = get_args()
config.update(vars(args))
config = config_dataset(config)
# assert config["u2g_samples"]>=config["u2g_recon"]
for bit in config["bit_list"]:
if config["rt_st"] == 'ham':
config["pr_curve_path"] = f"log/alexnet/CSQ_{config['dataset']}_{bit}.json"
train_val(config, bit)