-
Notifications
You must be signed in to change notification settings - Fork 14
/
train.py
203 lines (176 loc) · 9.84 KB
/
train.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
import torchvision.transforms as ttf
from src.factory import *
import torch
from torch.optim.lr_scheduler import StepLR
from torch import optim
import shutil
from extract_predictions import extract_msls_top_k
from src.validate import validate
from src.criteria import *
import argparse
def extract_features(dl, net, f_length, feats_file):
feats = np.zeros((len(dl.dataset), f_length))
for i, batch in tqdm(enumerate(dl), desc="Extracting features"):
if torch.cuda.is_available():
x = net.forward_single(batch.cuda())
else:
x = net.forward_single(batch)
feats[i * dl.batch_size:i * dl.batch_size + dl.batch_size] = x.cpu().detach().squeeze(0)
np.save(feats_file, feats)
class TrainParser():
def __init__(self):
self.parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
self.parser.add_argument('--root_dir', required=True, help='Root directory of the dataset')
self.parser.add_argument('--cities', required=False, default='train', help='Subset of MSLS')
self.parser.add_argument('--batch_size', type=int, default=32, help='input batch size')
self.parser.add_argument('--name', type=str, required=False, help='name of the experiment', default='testexp')
self.parser.add_argument('--backbone', type=str, default='vgg16', help='which architecture to use. [resnet50, resnet152, resnext, vgg16]')
self.parser.add_argument('--snapshot_dir', type=str, default='./snapshots', help='models are saved here')
self.parser.add_argument('--result_dir', type=str, default='./results', help='predictions and results are saved here')
self.parser.add_argument('--use_gpu', help='Use GPU mode',action='store_true')
self.parser.add_argument('--save_freq', type=int, default=1, help='save frequency in steps')
self.parser.add_argument('--dataset', type=str, default='soft_MSLS', help='[binary_MSLS|soft_MSLS]')
self.parser.add_argument('--pool', type=str, default='GeM', help='Global pool layer max|avg|GeM')
self.parser.add_argument('--p', required=False, type=int, default=3, help='P parameter for GeM pool')
self.parser.add_argument('--norm', type=str, default="L2", help='Norm layer')
self.parser.add_argument('--image_size', type=str, default="480,640", help='Input size, separated by commas')
self.parser.add_argument('--last_layer', type=int, default=None, help='Last layer to keep')
self.parser.add_argument('--display_freq', type=int, default=10, help='frequency of showing training results on screen')
self.parser.add_argument('--steps', type=int, default=52, help='Number of training steps. 52= 1epoch')
self.parser.add_argument('--margin', type=float, default='.5', help='margin parameter for the contrastive loss')
self.parser.add_argument('--learning_rate', type=float, default='.1', help='learning rate')
self.parser.add_argument('--lr_gamma', type=float, default='.1', help='learning rate decay')
self.parser.add_argument('--step_size', type=float, default='25', help='Learning rate update frequency (in steps)')
def parse(self):
self.opt = self.parser.parse_args()
def val(params, model, image_t, best_metric, reference_metric="recall@5", metric="EuclideanDistance"):
print("Validating...")
val_cities = ["cph", "sf"]
save_dir = params.result_dir
if not os.path.exists(save_dir):
os.makedirs(save_dir)
is_best = False
model.eval()
ret_file = save_dir+ "/" + params.name + "_retrieved.csv"
with open(ret_file, "w"):
pass
for c in val_cities:
print(c)
m_raw_file = params.root_dir+"train_val/"+c+"/database/raw.csv"
q_idx_file = params.root_dir+"train_val/"+c+"/query.json"
m_idx_file = params.root_dir+"train_val/"+c+"/database.json"
gt_file = params.root_dir+"train_val/"+c+"_gt.h5"
q_dl = create_dataloader("test", params.root_dir, q_idx_file, None, image_t, 2)
m_dl = create_dataloader("test", params.root_dir, m_idx_file, None, image_t, 2)
q_feats_file = save_dir + "/" + params.name + "_"+c+ "_query_features.npy"
q_feats = extract_features(q_dl, model, model.feature_length,q_feats_file)
m_feats_file = save_dir + "/" + params.name + "_"+c+ "_database_features.npy"
m_feats = extract_features(m_dl, model, model.feature_length,m_feats_file)
extract_msls_top_k(m_feats_file,q_feats_file, m_idx_file, q_idx_file, ret_file, 25, m_raw_file)
res_file = save_dir + "/" + params.name + "_val_results.txt"
metrics = validate(ret_file, params.root_dir, res_file)
if metrics[reference_metric] > best_metric:
shutil.copy(ret_file, ret_file.replace(".csv", "_best.csv"))
shutil.copy(res_file, res_file.replace(".txt", "_best.txt"))
for c in val_cities:
q_feats_file = save_dir + "/" + params.name + "_"+c+ "_query_features.npy"
shutil.copy(q_feats_file, q_feats_file.replace(".npy", "_best.npy"))
m_feats_file = save_dir + "/" + params.name + "_"+c+ "_database_features.npy"
shutil.copy(m_feats_file, m_feats_file.replace(".npy", "_best.npy"))
is_best = True
model.train()
return metrics, is_best
def train(params):
image_size = [int(x) for x in (params.image_size).split(",")]
best_metric = 0
ref_metric = "recall@5"
print("training with images of size",image_size[0],image_size[1])
if image_size[0] == image_size[1]: #If we want to resize to square, we do resize+crop
image_t = ttf.Compose([ttf.Resize(size=(image_size[0])),
ttf.CenterCrop(size=(image_size[0])),
ttf.ToTensor(),
ttf.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
else:
image_t = ttf.Compose([ttf.Resize(size=(image_size[0], image_size[1])),
ttf.ToTensor(),
ttf.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
# writer = SummaryWriter('runs/'+params.name+"_"+datetime.now().isoformat("-").split(".")[0].replace(":","-"))
model = create_model(params.backbone, params.pool, last_layer=params.last_layer, norm=params.norm, p_gem=params.p)
if torch.cuda.is_available():
model = model.cuda()
loss = ContrastiveLoss(params.margin)
print(params.dataset)
dataloader = create_msls_dataloader(params.dataset, params.root_dir, params.cities,
transform=image_t,
batch_size=params.batch_size)
# if params.use_gpu:
if torch.cuda.is_available():
model = model.cuda()
loss = loss.cuda()
total_iterations = 0
optimizer = optim.SGD(model.parameters(), lr=params.learning_rate)
scheduler = StepLR(optimizer, step_size=params.step_size, gamma=params.lr_gamma)
init_step = 0
optimizer.zero_grad()
# metrics, is_best = val(params, model, image_t, best_metric, reference_metric=ref_metric)
# best_metric = metrics[ref_metric]
best_metric = 0
# for step in tqdm(range(init_step, params.steps), desc="Steps"):
for step in range(init_step, params.steps):
# step
e_iteration = 0
for i, data in enumerate(dataloader):
e_iteration += params.batch_size
mini_batch_size = 2
accum_iterations = int(data["im0"].shape[0]/mini_batch_size)
for j in range(accum_iterations):
a = j * mini_batch_size
b = a + mini_batch_size
if torch.cuda.is_available(): # params.use_gpu:
x0, x1 = model(data["im0"][a:b,:].cuda(), data["im1"][a:b,:].cuda())
error = loss(x0, x1, (data["label"][a:b]).cuda())
else:
x0, x1 = model(data["im0"][a:b,:], data["im1"][a:b,:])
error = loss(x0, x1, data["label"][a:b])
null_losses = torch.sum(error==0).item()/len(error)
error = torch.mean(error) / accum_iterations
# writer.add_scalar('Debug/null_losses', null_losses, total_iterations)
error.backward()
# writer.add_scalar('Loss/train', error.cpu(), total_iterations)
total_iterations += mini_batch_size
# Visualize
if i % params.display_freq == 0:
print("Step %d, Iteration %d, Loss %.4f, Null loss %.4f" % (step, e_iteration, error, null_losses))
optimizer.step()
optimizer.zero_grad()
metrics, is_best = val(params, model, image_t, best_metric, reference_metric=ref_metric)
# Save
if step % params.save_freq == 0:
save_path = params.snapshot_dir + "/" + params.name + ".pth"
torch.save({'step': step,
'model_state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict()
}, save_path)
if is_best:
best_metric = metrics[ref_metric]
best_metrics = metrics
save_path = params.snapshot_dir + "/" + params.name +"_best.pth"
torch.save({'step': step,
'model_state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict()
}, save_path)
scheduler.step()
dataloader.dataset.load_pairs()
# writer.close()
print("Done. Best results on val:")
print(best_metrics)
if __name__ == "__main__":
p = TrainParser()
p.parse()
params = p.opt
print(params)
train(params)