-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathffm2.py
408 lines (336 loc) · 17.1 KB
/
ffm2.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
import os
import math
from shutil import copyfile
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
from torch.autograd import Variable
from torchvision import datasets, transforms
import torch.backends.cudnn as cudnn
from torchvision.transforms import InterpolationMode
from torchvision import models
import numpy as np
import random
from utils.common import setup_seed
# from utils.loader import environments, init_dataset_train
from torch import nn
from tqdm import tqdm
from torch.utils.data import Subset, DataLoader
from torch.nn import functional as F
from MLPN.loader import init_dataset_train
from utils.loader import init_dataset_test, environments, tensor2label, label2tensor, init_dataset_test
from utils.metrics import metrics
from FFM.model import CSWinTransv2_threeIn
from FFM.utils import extract_feature, get_id, extract_feature, SAM, SupConLoss, one_LPN_output
from tqdm.contrib import tzip
import tarfile
import zipfile
from LPN.image_folder_ import CustomData160k_drone, CustomData160k_sat
from utils.competition import get_result_rank10, get_SatId_160k
class FFilter(nn.Module):
def __init__(self, domin_num=10, h=256, w=256, channel=3) -> None:
super(FFilter, self).__init__()
self.learnable_filter = nn.Parameter(torch.ones(channel, h, w))
def forward(self, x):
return self.learnable_filter * x
class FFM(nn.Module):
def __init__(self, domin_num=10, h=256, w=256, channel=3) -> None:
super(FFM, self).__init__()
self.filter_invariant = FFilter(domin_num=domin_num, h=h, w=w, channel=channel)
def forward(self, img):
fft_img = torch.fft.fftn(img, dim=(-2,-1))
amplitude = torch.abs(fft_img)
phase = torch.angle(fft_img)
spec_invariant = self.filter_invariant(amplitude)
component_invariant = torch.real(torch.fft.ifftn(torch.polar(spec_invariant, phase), dim=(-2,-1)))
return component_invariant
def get_fixed_dataloader(dataloaders):
dataloader_drone = list()
dataloader_sat = list()
for data, data3 in tzip(dataloaders['drone'], dataloaders['satellite']):
dataloader_drone.append(data)
dataloader_sat.append(data3)
print('done synthesis {} style image'.format(style))
return dataloader_drone, dataloader_sat
class Ours:
def __init__(self,
use_wandb=True,
wandb_key = '16c9a3f92163ef4df08841029e02fded0cd0cfed'
) -> None:
self.seed = 2024
self.use_wandb = use_wandb # use wandb to monitor training instead of CLI
self.wandb_key = wandb_key
# init
setup_seed(self.seed)
self.model_dir = os.path.join(os.getcwd(), 'model', 'FFMv2')
os.environ['TORCH_HOME']='./'
if not os.path.isdir(self.model_dir):
os.makedirs(self.model_dir)
#
# self.domain_classifer = DomainClassifier()
def train(self,
data_dir=None,
style='mixed',
model_name1='FFMv2',
model_name2='MLPN',
num_epochs=210,
lr = 0.005,
ffm_lr = 0.01,
batchsize = 8,
block = 4,
checkpoint_interval = 10,
checkpoint_start = 40,
droprate = 0.75,
update_aug_img = [40, 80, 120, 160, 180, 200],
num_worker_imgaug = 16,
opt_iter_epoch = 32
):
# use wandb to log
if self.use_wandb:
import wandb
os.environ["WANDB_API_KEY"] = self.wandb_key
wandb.init(project="ACMMMW24", name=model_name1)
if data_dir==None:
data_dir = os.path.join(os.getcwd(), 'University-Release', 'train')
image_datasets, dataloaders, dataset_sizes = init_dataset_train(data_dir, batchsize=batchsize, style=style, num_worker_imgaug=num_worker_imgaug)
# fix setting that
# ====================
# LPN: true
# SAM: 1
# balance: true
# infonce: 1
# decouple: false
# only_decouple: false
# moving_avg: 1.0
# warm_epoch: 0 --> No Warmup
# extra_Google: false
# select_id: false
# normal: false --> dataloader give couple
#### MLPN
model = CSWinTransv2_threeIn(701, droprate=droprate, decouple=False, infonce=1)
model = model.cuda()
#### OUR FFM Module
ffm_sat = FFM()
ffm_sat = ffm_sat.cuda()
ffm_drone = FFM()
ffm_drone = ffm_drone.cuda()
ignored_params = list()
for i in range(block):
cls_name = 'classifier' + str(i)
c = getattr(model, cls_name)
ignored_params += list(map(id, c.parameters()))
base_params = filter(lambda p: id(p) not in ignored_params, model.parameters())
optim_params = [{'params': base_params, 'lr': 0.1 * lr}]
for i in range(block):
cls_name = 'classifier' + str(i)
c = getattr(model, cls_name)
optim_params.append({'params': c.parameters(), 'lr': lr})
optim_params.append({'params': ffm_drone.parameters(), 'lr': ffm_lr})
optim_params.append({'params': ffm_sat.parameters(), 'lr': ffm_lr})
infonce = SupConLoss(temperature=0.1)
# SAM = 1
base_optimizer = torch.optim.SGD
optimizer_ft = SAM(optim_params, base_optimizer, lr=lr, weight_decay=5e-4, momentum=0.9, nesterov=True)
exp_lr_scheduler = lr_scheduler.MultiStepLR(optimizer_ft, milestones=[120, 180, 210], gamma=0.1)
criterion = nn.CrossEntropyLoss()
bestAcc, bestAp, bestEp = 0, 0, 0
dataloader_drone, dataloader_sat = get_fixed_dataloader(dataloaders)
optimizer = optimizer_ft
for epoch in range(1, num_epochs+1):
running_loss, running_corrects, running_corrects3 = 0.0, 0.0, 0.0
ins_loss, dec_loss, on_loss, off_loss = 0.0, 0.0, 0.0, 0.0
lossinfo1, lossinfo2 = 0.0, 0.0
if epoch in update_aug_img:
dataloader_drone, dataloader_sat = get_fixed_dataloader(dataloaders)
# Iterate over data.
for data, data3 in tzip(dataloader_sat, dataloader_drone):
# get the inputs
inputs, inputs_d, labels, weather = data
inputs3, inputs3_s, labels3, weather3 = data3
now_batch_size, c, h, w = inputs.shape
if now_batch_size < batchsize: # skip the last batch
continue
inputs = Variable(inputs.cuda().detach())
inputs_d = Variable(inputs_d.cuda().detach())
inputs3 = Variable(inputs3.cuda().detach())
inputs3_s = Variable(inputs3_s.cuda().detach())
labels = Variable(labels.cuda().detach())
labels3 = Variable(labels3.cuda().detach())
# zero the parameter gradients
optimizer.zero_grad()
inputs = ffm_sat(inputs)
inputs3_s = ffm_sat(inputs3_s)
inputs_d = ffm_drone(inputs_d)
inputs3 = ffm_drone(inputs3)
# forward (invariant part and satellite img)
outs_c, outs_info = model(inputs) # decouple: false
outs_d_c, outs_d_info = model(inputs_d) # balance: true and infonce: 1
outd_c, outd_info = model(inputs3)
outs3_s_c, outs3_s_info = model(inputs3_s)
# calculate loss
preds, loss = one_LPN_output(outs_c, labels, criterion, block)
_, loss_d = one_LPN_output(outs_d_c, labels, criterion, block)
loss = loss + loss_d
preds3, loss3 = one_LPN_output(outd_c, labels3, criterion, block)
_, loss3_s = one_LPN_output(outs3_s_c, labels3, criterion, block)
loss3 = loss3 + loss3_s
loss = (loss + loss3) / 2
sate = F.normalize(outs_info, dim=1)
drone = F.normalize(outd_info, dim=1)
sate_ = F.normalize(outs_d_info, dim=1)
drone_ = F.normalize(outs3_s_info, dim=1)
features1 = torch.cat([sate.unsqueeze(1), sate_.unsqueeze(1)], dim=1)
features2 = torch.cat([drone.unsqueeze(1), drone_.unsqueeze(1)], dim=1)
loss_info = infonce(features1, labels)
loss = loss + loss_info
loss_info1 = infonce(features2, labels3)
loss = loss + loss_info1
# backward
loss.backward(retain_graph=True)
optimizer.first_step(zero_grad=True)
# again
# forward ffm
inputs = ffm_sat(inputs)
inputs3_s = ffm_sat(inputs3_s)
inputs_d = ffm_drone(inputs_d)
inputs3 = ffm_drone(inputs3)
# forward (invariant part and satellite img)
outs_c, outs_info = model(inputs) # decouple: false
outs_d_c, outs_d_info = model(inputs_d) # balance: true and infonce: 1
outd_c, outd_info = model(inputs3)
outs3_s_c, outs3_s_info = model(inputs3_s)
# calculate loss
preds, loss = one_LPN_output(outs_c, labels, criterion, block)
_, loss_d = one_LPN_output(outs_d_c, labels, criterion, block)
loss = loss + loss_d
preds3, loss3 = one_LPN_output(outd_c, labels3, criterion, block)
_, loss3_s = one_LPN_output(outs3_s_c, labels3, criterion, block)
loss3 = loss3 + loss3_s
loss = (loss + loss3) / 2
sate = F.normalize(outs_info, dim=1)
drone = F.normalize(outd_info, dim=1)
sate_ = F.normalize(outs_d_info, dim=1)
drone_ = F.normalize(outs3_s_info, dim=1)
features1 = torch.cat([sate.unsqueeze(1), sate_.unsqueeze(1)], dim=1)
features2 = torch.cat([drone.unsqueeze(1), drone_.unsqueeze(1)], dim=1)
loss_info = infonce(features1, labels)
loss = loss + loss_info
loss_info1 = infonce(features2, labels3)
loss = loss + loss_info1
# backward
loss.backward()
optimizer.second_step(zero_grad=True)
# statistics
running_loss += loss.item() * now_batch_size
lossinfo1 += loss_info.item() * now_batch_size
running_corrects += float(torch.sum(preds == labels.data))
running_corrects3 += float(torch.sum(preds3 == labels3.data))
epoch_loss = running_loss / dataset_sizes['satellite']
epoch_acc = running_corrects / dataset_sizes['satellite']
epoch_acc3 = running_corrects3 / dataset_sizes['satellite']
epoch_loss_info1 = lossinfo1 / dataset_sizes['satellite']
print('{} Loss: {:.4f} Satellite_Acc: {:.4f} Drone_Acc: {:.4f} infoloss1: {:.4f} '.format(
epoch, epoch_loss, epoch_acc,
epoch_acc3, epoch_loss_info1))
if self.use_wandb:
wandb.log({'Loss': epoch_loss,
'Satellite_Acc': epoch_acc,
'Drone_Acc': epoch_acc3,
'infoloss1': epoch_loss_info1
})
exp_lr_scheduler.step()
if epoch % checkpoint_interval == 0 and epoch > checkpoint_start:
save_filename = '{}_drone_{:03d}.pth'.format(model_name1, epoch)
save_path = os.path.join(os.path.join(self.model_dir, save_filename))
torch.save(ffm_drone.cpu().state_dict(), save_path)
ffm_drone.cuda()
save_filename = '{}_sat_{:03d}.pth'.format(model_name1, epoch)
save_path = os.path.join(os.path.join(self.model_dir, save_filename))
torch.save(ffm_sat.cpu().state_dict(), save_path)
ffm_sat.cuda()
save_filename = '{}_{:03d}.pth'.format(model_name2, epoch)
save_path = os.path.join(os.path.join(self.model_dir, save_filename))
torch.save(model.cpu().state_dict(), save_path)
model.cuda() # essential!
def get_competition_submit(self, data160k_dir='D://dataset/university-160k-wx', save_file = 'answer.txt', pth=None, multiple_scale=[1], batchsize=128, block=4, MLPN_file='MLPN_200.pth', DC_file='DomainClassifier_060.pth', FFM_file='FFM_200.pth'):
# a part of test set of UAVM'24 competition is provided as tar file
if not os.path.exists(os.path.join(data160k_dir, 'gallery_satellite_160k')):
tar_file = os.path.join(data160k_dir, 'gallery_satellite_160k.tar.gz')
if os.path.isfile(tar_file):
print('Found dataset tar file. Extracting...')
with tarfile.open(tar_file, 'r:gz') as tar:
tar.extractall(path=data160k_dir)
print('Extract done')
else:
print('Found dataset')
query_name = os.path.join(data160k_dir, 'query_drone_name.txt')
if os.path.isfile(save_file):
os.remove(save_file)
os.remove('answer.zip')
results_rank10 = []
data_transforms = transforms.Compose([
transforms.Resize((self.h, self.w), interpolation=InterpolationMode.BICUBIC),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
image_datasets = {}
image_datasets['gallery_satellite_160k'] = CustomData160k_sat(os.path.join(data160k_dir, 'gallery_satellite_160k'), data_transforms)
image_datasets['query_drone_160k'] = CustomData160k_drone( os.path.join(data160k_dir, 'query_drone160k_wx') ,data_transforms, query_name = query_name)
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=batchsize,
shuffle=False, num_workers=16) for x in
['gallery_satellite_160k','query_drone_160k']}
gallery_path = image_datasets['gallery_satellite_160k'].imgs
gallery_label, gallery_path = get_SatId_160k(gallery_path)
# load model
## load MLPN model
print("load MLPN model: {}".format(MLPN_file))
MLPN_file = os.path.join(self.model_dir, MLPN_file)
MLPN_model = CSWinTransv2_threeIn(701, droprate=0.75, decouple=False, infonce=1)
MLPN_model.load_state_dict(torch.load(MLPN_file))
# LPN: true
for i in range(block):
cls_name = 'classifier'+str(i)
c = getattr(MLPN_model, cls_name)
c.classifier = nn.Sequential()
MLPN_model = MLPN_model.cuda()
MLPN_model.train(False)
## load Domain Classifier model
DC_file = os.path.join(self.model_dir, DC_file)
domain_classifier = DomainClassifier()
domain_classifier.load_state_dict(torch.load(DC_file))
domain_classifier = domain_classifier.cuda()
domain_classifier.train(False)
## load FFM model
FFM_file = os.path.join(self.model_dir, FFM_file)
ffm = FFM()
ffm.load_state_dict(torch.load(FFM_file))
ffm = ffm.cuda()
ffm.train(False)
# Extract features
with torch.no_grad():
query_feature = extract_feature(domain_classifier,ffm, MLPN_model, dataloaders['query_drone_160k'], view='drone', ms=multiple_scale, testing=True)
gallery_feature = extract_feature(domain_classifier,ffm, MLPN_model, dataloaders['gallery_satellite_160k'], view='satellite', ms=multiple_scale)
query_feature = query_feature.cuda()
gallery_feature = gallery_feature.cuda()
gallery_label = np.array(gallery_label)
for i in tqdm(range(len(query_feature)), desc='Evaluate Rank10 results'):
result_rank10 = get_result_rank10(query_feature[i], gallery_feature, gallery_label)
results_rank10.append(result_rank10)
results_rank10 = np.row_stack(results_rank10)
with open(save_file, 'w') as f:
for row in results_rank10:
f.write('\t'.join(map(str, row)) + '\n')
# zip
zip_name = os.path.join(os.getcwd(), 'answer.zip')
with zipfile.ZipFile(zip_name, 'w', zipfile.ZIP_DEFLATED) as zipf:
zipf.write(save_file, save_file)
if __name__ == '__main__':
m = Ours()
# m.train_multi_MLPNs(data_dir='/dataset/University-Release/train')
#style='normal'
#m.train(data_dir='/dataset/University-Release/train', style=style, model_name=style, num_epochs=20, checkpoint_interval=5, num_worker_imgaug=32, fix_img=True)
#style='rain'
#m.train(data_dir='/dataset/University-Release/train', style=style, model_name=style, num_epochs=20, checkpoint_interval=5, num_worker_imgaug=32, fix_img=True)
style='mixed'
m.train(data_dir='/dataset/University-Release/train', update_aug_img=list(range(10,200)), checkpoint_interval=10, checkpoint_start=0)