-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_zero.py
378 lines (298 loc) · 14.2 KB
/
train_zero.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
"""
Training a CLIP model
"""
import argparse
import os
import time
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as T
from datetime import timedelta
import torchvision
from torchvision import models
from torch.nn.parallel import DataParallel
import utils
import logging
import clip
import pdb
def parse_args(argv):
"""
Parsing input arguments.
"""
# Arguments
parser = argparse.ArgumentParser(description='Training on RSI datasets.')
parser.add_argument('--work-dir', type=str, default='./work_dir/', required=False,
help='Model directory to use to save models (default=%(default)s)')
# miscellaneous args
parser.add_argument('--seed', type=int, default=0,
help='Random seed (default=%(default)s)')
# data args
parser.add_argument('--dataset', type=str, default='UCMerced_LandUse',
help='Dataset (default=%(default)s)')
parser.add_argument('--season', type=str, default=None, required=False,
help='only used for sen12ms dataset')
parser.add_argument('--n-shots', default=10, type=int, required=False,
help='Number of classes to train with')
parser.add_argument('--num-workers', default=4, type=int, required=False,
help='Number of subprocesses to use for the dataloader (default=%(default)s)')
parser.add_argument('--pseudo-version', default='1', type=str, required=False,
help='Number ofs classes to train with')
# model args
parser.add_argument('--vis-model', type=str, required=False, default='ViT-L/14',
help='Vision Model')
parser.add_argument('--gpu', type=str, required=False, default='0',
help='GPU assigned')
# training args
parser.add_argument('--nbatches', default=300, type=int, required=False,
help='Maximum number of batches to see per session (default=%(default)s)')
parser.add_argument('--batch-size', default=24, type=int, required=False,
help='Batch size to use (default=%(default)s)')
parser.add_argument('--decay-step', default=20, type=int, required=False,
help='decay-step to use (default=%(default)s)')
parser.add_argument('--decay', default=0.7, type=float, required=False,
help='decay rate to use (default=%(default)s)')
parser.add_argument('--weight-decay', default=0.0005, type=float, required=False,
help='Weight decay (default=%(default)s)')
parser.add_argument('--momentum', default=0.9, type=float, required=False,
help='Momentum (default=%(default)s)')
parser.add_argument('--patience', type=int, default=3, required=False,
help='Use patience while training (default=%(default)s)')
parser.add_argument('--load-from', type=str, default='',
help='Resume training with the last saved model (default=%(default)s)')
args = parser.parse_args(argv)
return args
# prompt = "This is a photo of a {}."
# prompt = "This is a satellite image of a {}."
# prompt = "This is a land use image of a {}."
prompt = "This is an aerial image of a {}."
def evaluate(model, test_loader, device):
"""
Evaluating the resulting classifier using a given test set loader.
"""
model.eval()
model.float()
text_descriptions = [prompt.format(label) for label in test_loader.dataset.classes]
text_tokens = clip.tokenize(text_descriptions).cuda()
text_features = model.encode_text(text_tokens).float()
text_features /= text_features.norm(dim=-1, keepdim=True)
total_top1_hits, total_top5_hits, N = 0, 0, 0
top1_avg, top5_avg = 0, 0
with torch.no_grad():
cnt = 0
for data in test_loader:
images, targets, _ = data
images = images.to(device)
targets = targets.to(device).squeeze()
image_features = model.encode_image(images)
similarity = (100.0 * image_features @ text_features.T).softmax(dim=-1)
top5_hits, top1_hits = utils.calculate_metrics(similarity, targets)
total_top1_hits += top1_hits
total_top5_hits += top5_hits
N += targets.shape[0]
top1_avg = 100 * (float(total_top1_hits) / N)
top5_avg = 100 * (float(total_top5_hits) / N)
return top1_avg, top5_avg
def convert_models_to_fp32(model):
for p in model.parameters():
p.data = p.data.float()
p.grad.data = p.grad.data.float()
def run_training(args):
"""
Main training routine.
"""
def log_string(str):
logger.info(str)
print(str)
# Fixing random seed
utils.seed_everything(args.seed)
os.environ['CUDA_DEVICE_ORDER']='PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES']=str(args.gpu)
# Checking directory folder
if not os.path.exists(args.work_dir):
print("Output model directory [%s] not found. Creating..." % args.work_dir)
os.makedirs(args.work_dir)
os.system('cp train_zero.py {}'.format(args.work_dir))
logger = logging.getLogger("Model")
logger.setLevel(logging.INFO)
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
file_handler = logging.FileHandler('%s/%s.txt' % (args.work_dir, 'train'))
file_handler.setLevel(logging.INFO)
file_handler.setFormatter(formatter)
logger.addHandler(file_handler)
log_string('PARAMETER ...')
log_string(args)
'''DATA LOADING'''
# Setting up dataset
log_string("Loading data : [%s]" % args.dataset)
device = torch.device("cuda:0")
# device = torch.device("cpu")
model, preprocess = clip.load(args.vis_model,device=device,jit=False) #Must set jit=False for training
# Optionally resume training
if os.path.isfile(args.load_from):
log_string("Loading pretrained model : [%s]" % args.load_from)
checkpoint = torch.load(open(args.load_from, 'rb'), map_location="cpu")
model.load_state_dict(checkpoint['state_dict'])
if device == torch.device("cpu"):
model.float()
else :
clip.model.convert_weights(model) # Actually this line is unnecessary since clip by default already on float16
model.to(device)
# model = DataParallel(model)
# Defining Loss and Optimizer
loss_img = nn.CrossEntropyLoss()
loss_txt = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=1e-5,betas=(0.9,0.98),eps=1e-6,weight_decay=0.2) #Params used from paper, the lr is smaller, more safe for fine tuning to new dataset
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.decay_step, gamma=args.decay)
# Instantiating data loaders
# train_transforms = T.Compose([
# T.RandomHorizontalFlip(),
# T.CenterCrop(224),
# T.ToTensor(),
# T.Normalize((0.48422758, 0.49005175, 0.45050276), (0.17348297, 0.16352356, 0.15547496)), #recompute
# ])
train_transforms = T.Compose([
T.Resize(256),
T.CenterCrop(224),
T.RandomHorizontalFlip(),
# T.ColorJitter(0.05, 0.05, 0.05),
# T.RandomRotation(10),
T.ToTensor(),
T.Normalize(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711))
])
test_transforms = T.Compose([
T.Resize(256),
T.CenterCrop(224),
T.ToTensor(),
T.Normalize(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711))
])
if 'sen12ms' in args.dataset.lower():
bands_mean = {'s1_mean': [-11.76858, -18.294598],
's2_mean': [1226.4215, 1137.3799, 1139.6792, 1350.9973, 1932.9058,
2211.1584, 2154.9846, 2409.1128, 2001.8622, 1356.0801]}
bands_std = {'s1_std': [4.525339, 4.3586307],
's2_std': [741.6254, 740.883, 960.1045, 946.76056, 985.52747,
1082.4341, 1057.7628, 1136.1942, 1132.7898, 991.48016]}
# data path
data_root = './data/SEN12MS/'
list_dir = "./data/SEN12MS/SEN12MS/splits/" # split lists/ label dirs
# define image transform
from data.sen12ms_dataset import SEN12MS, ToTensor, Normalize, CenterCrop, SEN12_Pseudo_Dataset
imgTransform = T.Compose([
ToTensor(),
Normalize(bands_mean, bands_std),
CenterCrop(224),
])
# test multi_label part with normalization
print("\n\nSEN12MS train")
if args.season is None:
season = 'all_seasons'
else:
season = args.season
assert season in ['sprint', 'summer', 'fall', 'winter']
train_dataset = SEN12_Pseudo_Dataset(data_root, transform=imgTransform,
txt_path='pseudo_v{}_train_{}shot.txt'.format(args.pseudo_version, args.n_shots))
test_dataset = SEN12MS(data_root, list_dir, imgTransform, season=season,
label_type="single_label", threshold=0.1, subset="test",
use_s1=False, use_s2=False, use_RGB=True, IGBP_s=True,
txt_path='test_{}.txt'.format(season))
else:
if 'ucm' in args.dataset.lower():
data_root='./data/UCMerced_LandUse/'
elif 'whurs' in args.dataset.lower():
data_root='./data/WHU-RS19/'
elif 'nwpu' in args.dataset.lower():
data_root='./data/NWPU-RESISC45/'
elif 'aid' in args.dataset.lower():
data_root='./data/AID/'
elif 'eurosat' in args.dataset.lower():
data_root = './data/EuroSAT/'
else:
print('dataset not defined')
return
from data.ucm_dataset import RSDataset
train_dataset = RSDataset(data_root=data_root, txt_path='pseudo_v{}_train_{}shot.txt'.format(args.pseudo_version, args.n_shots), transform=train_transforms)
test_dataset = RSDataset(data_root=data_root, txt_path='all.txt', transform=test_transforms)
log_string('train size: {}/test size: {}'.format(len(train_dataset), len(test_dataset)))
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=args.batch_size, shuffle=True,
num_workers=args.num_workers, pin_memory=True)
test_loader = torch.utils.data.DataLoader(
test_dataset,
batch_size=24, shuffle=False,
num_workers=args.num_workers, pin_memory=True)
# Training
starting_time = time.time()
# Main loop
tstart = time.time()
log_string("Starting training...")
log_string("Number of batches: [%d]" % (len(train_loader.dataset)/args.batch_size))
n_batch = 0
loss = None
while n_batch <= args.nbatches:
# Training the model
total_top1_hits, total_top5_hits, N = 0, 0, 0
top1_avg, top5_avg = 0, 0
for images, targets, _ in train_loader:
## Making a checkpoint
if n_batch % 100 == 0:
# Measuring model test-accuracy
top1_test_acc, top5_test_acc = evaluate(model, test_loader, device)
log_string('Evaluattion {:03}/{:03}, top1_test_acc: {:.3f}, top5_test_acc: {:.3f}'.format(n_batch, args.nbatches, top1_test_acc, top5_test_acc))
if n_batch == 300:
saved_model = os.path.join(args.work_dir, "batch_%d.ckpt" % (n_batch))
log_string("Saved model at: [" + saved_model + "]")
state = {
"top1_train_acc": top1_avg,
"top5_train_acc": top5_avg,
"top1_test_acc": top1_test_acc,
"top5_test_acc": top5_test_acc,
"state_dict": model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}
torch.save(state, saved_model)
# pdb.set_trace()
model.train()
optimizer.zero_grad()
images, targets = images.to(device), targets.to(device)
texts = [prompt + test_loader.dataset.classes[t] for t in targets]
texts = torch.stack([clip.tokenize(t) for t in texts])
texts = texts.squeeze().to(device)
## Forward + backward + optimize
logits_per_image, logits_per_text = model(images, texts) #logits_per_image_orig
ground_truth = torch.arange(len(images),dtype=torch.long,device=device)
loss = (loss_img(logits_per_image,ground_truth) + loss_txt(logits_per_text,ground_truth))/2
loss.backward()
if device == torch.device("cpu"):
optimizer.step()
else:
convert_models_to_fp32(model)
optimizer.step()
clip.model.convert_weights(model)
## Logging results
outputs = logits_per_image.softmax(dim=-1)
top5_hits, top1_hits = utils.calculate_metrics(outputs, ground_truth)
total_top1_hits += top1_hits
total_top5_hits += top5_hits
N += images.shape[0]
top1_avg = 100 * (float(total_top1_hits) / N)
top5_avg = 100 * (float(total_top5_hits) / N)
log_string('Training {:03}/{:03} | loss = {:.4f} | top-1 acc = {:.3f} | top-5 acc = {:.3f}'.format(n_batch, args.nbatches, loss.item(), top1_avg, top5_avg))
n_batch += 1
scheduler.step()
if (n_batch >= args.nbatches): break
# Logging results
current_elapsed_time = time.time() - starting_time
log_string('{:03}/{:03} | {} | Train : loss = {:.4f} | top-1 acc = {:.3f} | top-5 acc = {:.3f}'.
format(n_batch, args.nbatches,
timedelta(seconds=round(current_elapsed_time)),
loss, top1_avg, top5_avg))
# Final output
log_string('[Elapsed time = {:.1f} mn]'.format((time.time() - tstart) / 60))
log_string('Done!')
log_string('-' * 108)
def main(argv=None):
args = parse_args(argv)
run_training(args)
if __name__ == "__main__":
main()