File size: 7,507 Bytes
205a7af |
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 |
"""Light HamHead Decoder.
Adapted from:
https://github.com/Visual-Attention-Network/SegNeXt/blob/main/mmseg/models/decode_heads/ham_head.py
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from siclib.models import BaseModel
from siclib.models.utils.modules import ConvModule, FeatureFusionBlock
# flake8: noqa
# mypy: ignore-errors
class _MatrixDecomposition2DBase(nn.Module):
def __init__(self):
super().__init__()
self.spatial = True
self.S = 1
self.D = 512
self.R = 64
self.train_steps = 6
self.eval_steps = 7
self.inv_t = 100
self.eta = 0.9
self.rand_init = True
def _build_bases(self, B, S, D, R, device="cpu"):
raise NotImplementedError
def local_step(self, x, bases, coef):
raise NotImplementedError
# @torch.no_grad()
def local_inference(self, x, bases):
# (B * S, D, N)^T @ (B * S, D, R) -> (B * S, N, R)
coef = torch.bmm(x.transpose(1, 2), bases)
coef = F.softmax(self.inv_t * coef, dim=-1)
steps = self.train_steps if self.training else self.eval_steps
for _ in range(steps):
bases, coef = self.local_step(x, bases, coef)
return bases, coef
def compute_coef(self, x, bases, coef):
raise NotImplementedError
def forward(self, x, return_bases=False):
B, C, H, W = x.shape
# (B, C, H, W) -> (B * S, D, N)
if self.spatial:
D = C // self.S
N = H * W
x = x.view(B * self.S, D, N)
else:
D = H * W
N = C // self.S
x = x.view(B * self.S, N, D).transpose(1, 2)
if not self.rand_init and not hasattr(self, "bases"):
bases = self._build_bases(1, self.S, D, self.R, device=x.device)
self.register_buffer("bases", bases)
# (S, D, R) -> (B * S, D, R)
if self.rand_init:
bases = self._build_bases(B, self.S, D, self.R, device=x.device)
else:
bases = self.bases.repeat(B, 1, 1)
bases, coef = self.local_inference(x, bases)
# (B * S, N, R)
coef = self.compute_coef(x, bases, coef)
# (B * S, D, R) @ (B * S, N, R)^T -> (B * S, D, N)
x = torch.bmm(bases, coef.transpose(1, 2))
# (B * S, D, N) -> (B, C, H, W)
x = x.view(B, C, H, W) if self.spatial else x.transpose(1, 2).view(B, C, H, W)
# (B * H, D, R) -> (B, H, N, D)
bases = bases.view(B, self.S, D, self.R)
return x
class NMF2D(_MatrixDecomposition2DBase):
def __init__(self):
super().__init__()
self.inv_t = 1
def _build_bases(self, B, S, D, R, device="cpu"):
bases = torch.rand((B * S, D, R)).to(device)
bases = F.normalize(bases, dim=1)
return bases
# @torch.no_grad()
def local_step(self, x, bases, coef):
# (B * S, D, N)^T @ (B * S, D, R) -> (B * S, N, R)
numerator = torch.bmm(x.transpose(1, 2), bases)
# (B * S, N, R) @ [(B * S, D, R)^T @ (B * S, D, R)] -> (B * S, N, R)
denominator = coef.bmm(bases.transpose(1, 2).bmm(bases))
# Multiplicative Update
coef = coef * numerator / (denominator + 1e-6)
# (B * S, D, N) @ (B * S, N, R) -> (B * S, D, R)
numerator = torch.bmm(x, coef)
# (B * S, D, R) @ [(B * S, N, R)^T @ (B * S, N, R)] -> (B * S, D, R)
denominator = bases.bmm(coef.transpose(1, 2).bmm(coef))
# Multiplicative Update
bases = bases * numerator / (denominator + 1e-6)
return bases, coef
def compute_coef(self, x, bases, coef):
# (B * S, D, N)^T @ (B * S, D, R) -> (B * S, N, R)
numerator = torch.bmm(x.transpose(1, 2), bases)
# (B * S, N, R) @ (B * S, D, R)^T @ (B * S, D, R) -> (B * S, N, R)
denominator = coef.bmm(bases.transpose(1, 2).bmm(bases))
# multiplication update
coef = coef * numerator / (denominator + 1e-6)
return coef
class Hamburger(nn.Module):
def __init__(self, ham_channels=512, norm_cfg=None, **kwargs):
super().__init__()
self.ham_in = ConvModule(ham_channels, ham_channels, 1)
self.ham = NMF2D()
self.ham_out = ConvModule(ham_channels, ham_channels, 1)
def forward(self, x):
enjoy = self.ham_in(x)
enjoy = F.relu(enjoy, inplace=False)
enjoy = self.ham(enjoy)
enjoy = self.ham_out(enjoy)
ham = F.relu(x + enjoy, inplace=False)
return ham
class LightHamHead(BaseModel):
"""Is Attention Better Than Matrix Decomposition?
This head is the implementation of `HamNet
<https://arxiv.org/abs/2109.04553>`_.
Args:
ham_channels (int): input channels for Hamburger.
ham_kwargs (int): kwagrs for Ham.
"""
default_conf = {
"predict_uncertainty": True,
"out_channels": 64,
"in_channels": [64, 128, 320, 512],
"in_index": [0, 1, 2, 3],
"ham_channels": 512,
"with_low_level": True,
}
def _init(self, conf):
self.in_index = conf.in_index
self.in_channels = conf.in_channels
self.out_channels = conf.out_channels
self.ham_channels = conf.ham_channels
self.align_corners = False
self.predict_uncertainty = conf.predict_uncertainty
self.squeeze = ConvModule(sum(self.in_channels), self.ham_channels, 1)
self.hamburger = Hamburger(self.ham_channels)
self.align = ConvModule(self.ham_channels, self.out_channels, 1)
if self.predict_uncertainty:
self.linear_pred_uncertainty = nn.Sequential(
ConvModule(
in_channels=self.out_channels,
out_channels=self.out_channels,
kernel_size=3,
padding=1,
bias=False,
),
nn.Conv2d(in_channels=self.out_channels, out_channels=1, kernel_size=1),
)
self.with_ll = conf.with_low_level
if self.with_ll:
self.out_conv = ConvModule(
self.out_channels, self.out_channels, 3, padding=1, bias=False
)
self.ll_fusion = FeatureFusionBlock(self.out_channels, upsample=False)
def _forward(self, features):
"""Forward function."""
# inputs = self._transform_inputs(inputs)
inputs = [features["hl"][i] for i in self.in_index]
inputs = [
F.interpolate(
level, size=inputs[0].shape[2:], mode="bilinear", align_corners=self.align_corners
)
for level in inputs
]
inputs = torch.cat(inputs, dim=1)
x = self.squeeze(inputs)
x = self.hamburger(x)
feats = self.align(x)
if self.with_ll:
assert "ll" in features, "Low-level features are required for this model"
feats = F.interpolate(feats, scale_factor=2, mode="bilinear", align_corners=False)
feats = self.out_conv(feats)
feats = F.interpolate(feats, scale_factor=2, mode="bilinear", align_corners=False)
feats_ll = features["ll"].clone()
feats = self.ll_fusion(feats, feats_ll)
uncertainty = (
self.linear_pred_uncertainty(feats).squeeze(1) if self.predict_uncertainty else None
)
return feats, uncertainty
def loss(self, pred, data):
raise NotImplementedError
|