Spaces:
Running
Running
File size: 4,740 Bytes
c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 |
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 |
import torch
from torch import nn
from ..dkm import *
from ..encoders import *
def DKMv3(
weights,
h,
w,
symmetric=True,
sample_mode="threshold_balanced",
device=None,
**kwargs
):
if device is None:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
gp_dim = 256
dfn_dim = 384
feat_dim = 256
coordinate_decoder = DFN(
internal_dim=dfn_dim,
feat_input_modules=nn.ModuleDict(
{
"32": nn.Conv2d(512, feat_dim, 1, 1),
"16": nn.Conv2d(512, feat_dim, 1, 1),
}
),
pred_input_modules=nn.ModuleDict(
{
"32": nn.Identity(),
"16": nn.Identity(),
}
),
rrb_d_dict=nn.ModuleDict(
{
"32": RRB(gp_dim + feat_dim, dfn_dim),
"16": RRB(gp_dim + feat_dim, dfn_dim),
}
),
cab_dict=nn.ModuleDict(
{
"32": CAB(2 * dfn_dim, dfn_dim),
"16": CAB(2 * dfn_dim, dfn_dim),
}
),
rrb_u_dict=nn.ModuleDict(
{
"32": RRB(dfn_dim, dfn_dim),
"16": RRB(dfn_dim, dfn_dim),
}
),
terminal_module=nn.ModuleDict(
{
"32": nn.Conv2d(dfn_dim, 3, 1, 1, 0),
"16": nn.Conv2d(dfn_dim, 3, 1, 1, 0),
}
),
)
dw = True
hidden_blocks = 8
kernel_size = 5
displacement_emb = "linear"
conv_refiner = nn.ModuleDict(
{
"16": ConvRefiner(
2 * 512 + 128 + (2 * 7 + 1) ** 2,
2 * 512 + 128 + (2 * 7 + 1) ** 2,
3,
kernel_size=kernel_size,
dw=dw,
hidden_blocks=hidden_blocks,
displacement_emb=displacement_emb,
displacement_emb_dim=128,
local_corr_radius=7,
corr_in_other=True,
),
"8": ConvRefiner(
2 * 512 + 64 + (2 * 3 + 1) ** 2,
2 * 512 + 64 + (2 * 3 + 1) ** 2,
3,
kernel_size=kernel_size,
dw=dw,
hidden_blocks=hidden_blocks,
displacement_emb=displacement_emb,
displacement_emb_dim=64,
local_corr_radius=3,
corr_in_other=True,
),
"4": ConvRefiner(
2 * 256 + 32 + (2 * 2 + 1) ** 2,
2 * 256 + 32 + (2 * 2 + 1) ** 2,
3,
kernel_size=kernel_size,
dw=dw,
hidden_blocks=hidden_blocks,
displacement_emb=displacement_emb,
displacement_emb_dim=32,
local_corr_radius=2,
corr_in_other=True,
),
"2": ConvRefiner(
2 * 64 + 16,
128 + 16,
3,
kernel_size=kernel_size,
dw=dw,
hidden_blocks=hidden_blocks,
displacement_emb=displacement_emb,
displacement_emb_dim=16,
),
"1": ConvRefiner(
2 * 3 + 6,
24,
3,
kernel_size=kernel_size,
dw=dw,
hidden_blocks=hidden_blocks,
displacement_emb=displacement_emb,
displacement_emb_dim=6,
),
}
)
kernel_temperature = 0.2
learn_temperature = False
no_cov = True
kernel = CosKernel
only_attention = False
basis = "fourier"
gp32 = GP(
kernel,
T=kernel_temperature,
learn_temperature=learn_temperature,
only_attention=only_attention,
gp_dim=gp_dim,
basis=basis,
no_cov=no_cov,
)
gp16 = GP(
kernel,
T=kernel_temperature,
learn_temperature=learn_temperature,
only_attention=only_attention,
gp_dim=gp_dim,
basis=basis,
no_cov=no_cov,
)
gps = nn.ModuleDict({"32": gp32, "16": gp16})
proj = nn.ModuleDict(
{"16": nn.Conv2d(1024, 512, 1, 1), "32": nn.Conv2d(2048, 512, 1, 1)}
)
decoder = Decoder(coordinate_decoder, gps, proj, conv_refiner, detach=True)
encoder = ResNet50(pretrained=False, high_res=False, freeze_bn=False)
matcher = RegressionMatcher(
encoder,
decoder,
h=h,
w=w,
name="DKMv3",
sample_mode=sample_mode,
symmetric=symmetric,
**kwargs
).to(device)
res = matcher.load_state_dict(weights)
return matcher
|