-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathSettingB-SO2-Spread-B.py
184 lines (165 loc) · 6.25 KB
/
SettingB-SO2-Spread-B.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
import os
from argparse import ArgumentParser
import torch
from torch.optim import AdamW
from torch.optim.lr_scheduler import CosineAnnealingLR
from torch.utils.data import ConcatDataset
import torch.nn as nn
from DeDoDe.datasets.megadepth import MegadepthBuilder
from DeDoDe.descriptors.dedode_descriptor import DeDoDeDescriptor
from DeDoDe.encoder import VGG
from DeDoDe.decoder import ConvRefiner, Decoder
from DeDoDe import dedode_detector_L
from DeDoDe.benchmarks import MegadepthNLLBenchmark
from rotation_steerers.train import train_k_steps
from rotation_steerers.descriptor_loss import DescriptorLoss
from rotation_steerers.checkpoint import CheckPoint
from rotation_steerers.steerers import ContinuousSteerer
def train(detector_weights):
NUM_PROTOTYPES = 256 # == descriptor size
residual = True
hidden_blocks = 5
amp_dtype = torch.float16#torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
amp = True
conv_refiner = nn.ModuleDict(
{
"8": ConvRefiner(
512,
512,
256 + NUM_PROTOTYPES,
hidden_blocks = hidden_blocks,
residual = residual,
amp = amp,
amp_dtype = amp_dtype,
),
"4": ConvRefiner(
256+256,
256,
128 + NUM_PROTOTYPES,
hidden_blocks = hidden_blocks,
residual = residual,
amp = amp,
amp_dtype = amp_dtype,
),
"2": ConvRefiner(
128+128,
64,
32 + NUM_PROTOTYPES,
hidden_blocks = hidden_blocks,
residual = residual,
amp = amp,
amp_dtype = amp_dtype,
),
"1": ConvRefiner(
64 + 32,
32,
1 + NUM_PROTOTYPES,
hidden_blocks = hidden_blocks,
residual = residual,
amp = amp,
amp_dtype = amp_dtype,
),
}
)
import os
experiment_name = os.path.splitext(os.path.basename(__file__))[0]
encoder = VGG(size = "19", pretrained = True, amp = amp, amp_dtype = amp_dtype)
decoder = Decoder(conv_refiner, num_prototypes=NUM_PROTOTYPES)
model = DeDoDeDescriptor(encoder = encoder, decoder = decoder).cuda()
generator = torch.block_diag(
torch.zeros(
[NUM_PROTOTYPES // 2,
NUM_PROTOTYPES // 2],
device='cuda',
),
*(
torch.tensor([[0., 1],
[-1, 0]],
device='cuda')
for _ in range(NUM_PROTOTYPES // 8)
),
*(
torch.tensor([[0., 2],
[-2, 0]],
device='cuda')
for _ in range(NUM_PROTOTYPES // 16)
),
*(
torch.tensor([[0., 3],
[-3, 0]],
device='cuda')
for _ in range(NUM_PROTOTYPES // 32)
),
*(
torch.tensor([[0., 4],
[-4, 0]],
device='cuda')
for _ in range(NUM_PROTOTYPES // 64)
),
*(
torch.tensor([[0., 5],
[-5, 0]],
device='cuda')
for _ in range(NUM_PROTOTYPES // 128)
),
*(
torch.tensor([[0., 6],
[-6, 0]],
device='cuda')
for _ in range(NUM_PROTOTYPES // 128)
),
)
steerer = ContinuousSteerer(generator)
params = [
{"params": model.encoder.parameters(), "lr": 1e-5},
{"params": model.decoder.parameters(), "lr": 2e-4},
{"params": steerer.parameters(), "lr": 2e-4},
]
optim = AdamW(params, weight_decay = 1e-5)
n0, N, k = 0, 100_000, 1000
lr_scheduler = CosineAnnealingLR(optim, T_max = N)
checkpointer = CheckPoint("workspace/", name = experiment_name)
model, optim, lr_scheduler, n0 = checkpointer.load(model, optim, lr_scheduler, n0, steerer=steerer)
detector = dedode_detector_L(weights = detector_weights)
loss = DescriptorLoss(detector=detector, normalize_descriptions = True, inv_temp = 20)
H, W = 512, 512
mega = MegadepthBuilder(data_root="data/megadepth", loftr_ignore=True, imc21_ignore = True, use_detections=False)
use_horizontal_flip_aug = False
megadepth_train1 = mega.build_scenes(
split="train_loftr", min_overlap=0.01, ht=H, wt=W, shake_t=32, use_horizontal_flip_aug = use_horizontal_flip_aug,
)
megadepth_train2 = mega.build_scenes(
split="train_loftr", min_overlap=0.35, ht=H, wt=W, shake_t=32, use_horizontal_flip_aug = use_horizontal_flip_aug,
)
megadepth_train = ConcatDataset(megadepth_train1 + megadepth_train2)
mega_ws = mega.weight_scenes(megadepth_train, alpha=0.75)
megadepth_test = mega.build_scenes(
split="test_loftr", min_overlap=0.01, ht=H, wt=W, shake_t=32, use_horizontal_flip_aug = use_horizontal_flip_aug,
)
mega_test = MegadepthNLLBenchmark(ConcatDataset(megadepth_test))
grad_scaler = torch.cuda.amp.GradScaler()
for n in range(n0, N, k):
mega_sampler = torch.utils.data.WeightedRandomSampler(
mega_ws, num_samples = 8 * k, replacement=False
)
mega_dataloader = iter(
torch.utils.data.DataLoader(
megadepth_train,
batch_size = 8,
sampler = mega_sampler,
num_workers = 8,
)
)
train_k_steps(
n, k, mega_dataloader, model, loss, optim, lr_scheduler, grad_scaler = grad_scaler, rot_cont=True, steerer=steerer,
)
checkpointer.save(model, optim, lr_scheduler, n, steerer=steerer)
mega_test.benchmark(detector = detector, descriptor = model)
if __name__ == "__main__":
os.environ["TORCH_CUDNN_V8_API_ENABLED"] = "1" # For BF16 computations
os.environ["OMP_NUM_THREADS"] = "16"
parser = ArgumentParser()
parser.add_argument("--detector_path", default="dedode_detector_SO2.pth")
args = parser.parse_args()
weights = torch.load(args.detector_path)
train(weights)