File size: 14,172 Bytes
12f8999 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 | import torch
def get_activation(actv_config):
actv_cls = getattr(torch.nn, actv_config.name, None)
assert actv_cls is not None, "No activation function"
if actv_config.params:
return (
actv_cls(**actv_config.params)
if isinstance(actv_config.params, dict)
else actv_cls(**actv_config.params.model_dump())
)
else:
return actv_cls()
"""
ECGRecoverRandomMaskWithRS4 와 차이: lead 내에서 VCP 만 적용 + lead II 를 k/v 로 사용해서 다른 lead 로 rhythm 정보 전달
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
class Convolution1D_layer(nn.Module):
def __init__(
self, in_channels, out_channels, kernel_size, padding, leaky_relu, dropout
):
super(Convolution1D_layer, self).__init__()
self.out_channels = out_channels
self.kernel_size = kernel_size
self.padding = padding
self.conv = nn.Sequential(
nn.Conv1d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=2,
padding=padding,
),
nn.BatchNorm1d(num_features=out_channels),
nn.LeakyReLU(leaky_relu),
nn.Dropout(dropout),
)
def forward(self, x: torch.Tensor):
out_size = (x.shape[-1] + 2 * self.padding - self.kernel_size) // 2 + 1
new_x = torch.zeros(
(len(x), self.out_channels, 12, out_size),
dtype=x.dtype,
device=x.device,
)
for i in range(12):
new_x[:, :, i, :] = self.conv(x[:, :, i, :])
return new_x
class Convolution2D_layer(nn.Module):
def __init__(
self, in_channels, out_channels, kernel_size, padding, leaky_relu, dropout
):
super(Convolution2D_layer, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=(1, 2),
padding=padding,
),
nn.BatchNorm2d(num_features=out_channels),
nn.LeakyReLU(leaky_relu),
# nn.Dropout(dropout)
)
def forward(self, x):
return self.conv(x)
class Deconvolution2D_layer(nn.Module):
def __init__(
self, in_channels, out_channels, kernel_size, padding, leaky_relu, dropout
):
super(Deconvolution2D_layer, self).__init__()
self.deconv = nn.Sequential(
nn.ConvTranspose2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=(1, 2),
padding=padding,
),
nn.BatchNorm2d(num_features=out_channels),
nn.LeakyReLU(leaky_relu),
# nn.Dropout(dropout)
)
def forward(self, x):
return self.deconv(x)
class VCBlock(nn.Module):
"""
enc: (B, C, 12, D)
1) lead-wise self-attention (mask 를 이용한 VCP 방식):
- q: full lead (B, D, C)
- k, v: visible 구간 from mask
2) lead II -> others cross-attention:
- q: full lead (B, D, C)
- k, v: lead ii 의 visible 구간 from mask
3) residual: enc + 1) + 2)
"""
def __init__(self, channels: int, num_heads: int = 4):
super().__init__()
self.self_attn = nn.MultiheadAttention(
embed_dim=channels, num_heads=num_heads, batch_first=True
)
self.cross_attn = nn.MultiheadAttention(
embed_dim=channels, num_heads=num_heads, batch_first=True
)
def forward(self, enc: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
_, _, L, _ = enc.shape # (B, C, 12, D)
enc_r = enc.permute(0, 2, 3, 1) # (B, 12, D, C)
attn_self = torch.zeros_like(enc_r, dtype=enc_r.dtype, device=enc_r.device)
attn_lead2 = torch.zeros_like(enc_r, dtype=enc_r.dtype, device=enc_r.device)
# print(f"in refineblock 1: {enc_r.shape}")
# Lead II K/V (모든 lead에 공통)
k2 = v2 = enc_r[:, 1, :, :] # (B, D, C)
key_padding_mask2 = mask[:, 1, :].bool() # (B, D)
for lead in range(L):
# lead 내에서 self-attention (mask 를 활용한 VCP 방식)
q = enc_r[:, lead, :, :] # (B, D, C)
k = v = enc_r[:, lead, :, :] # (B, D, C)
key_padding_mask = mask[:, lead, :].bool() # (B, D)
_attn_self, _ = self.self_attn(q, k, v, key_padding_mask=key_padding_mask)
attn_self[:, lead, :, :] = _attn_self
# lead II -> other lead cross-attention
_attn_lead2, _ = self.cross_attn(
q, k2, v2, key_padding_mask=key_padding_mask2
)
attn_lead2[:, lead, :, :] = _attn_lead2
# print(f"in refineblock 2: {attn_out.shape}")
vc = enc_r + attn_self + attn_lead2 # residual: (B, 12, D, C)
vc_r = vc.permute(0, 3, 1, 2) # (B, C, 12, D)
# print(f"in refineblock 3: {refined.shape}")
# visible_kv_mean = visible_kv_raw.mean(dim=1) # (B,12,vis_len)
# return refined, visible_kv_mean
return vc_r
class ECGRecoverRandomMaskWithRS5(nn.Module):
def __init__(self, config, verbose=False):
super().__init__()
self.verbose = verbose
self.activation = get_activation(config.activation)
inplanes = int(config.inplanes)
kernel_size = tuple(config.kernel_size)
assert len(kernel_size) == 2, "len(kernel_size) must be 2"
assert kernel_size[0] % 2 == 1, "kernel_size[0] must be odd"
padding_1d = (kernel_size[1] - 1) // 2
padding_2d = [(k - 1) // 2 for k in kernel_size]
num_heads = int(config.num_heads)
num_depths_cfg = getattr(config, "num_depths_attn_start", 5)
if isinstance(num_depths_cfg, (tuple, list)):
self.num_depths, self.attn_start = num_depths_cfg
else:
self.num_depths = int(num_depths_cfg)
self.attn_start = self.num_depths # attention 없음
leaky_relu = float(config.leaky_relu)
dropout = float(config.dropout)
# self.output_size: int = config.output_size
self.convs_1d = nn.ModuleList()
self.convs_2d = nn.ModuleList()
self.vc_blocks = nn.ModuleDict() # mask + cross-attn + residual
for d in range(self.num_depths):
in_channels = 1 if d == 0 else inplanes * (2 ** (d - 1))
out_channels = inplanes * (2**d)
self.convs_1d.append(
Convolution1D_layer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size[1],
padding=padding_1d,
leaky_relu=leaky_relu,
dropout=dropout,
)
)
self.convs_2d.append(
Convolution2D_layer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
padding=padding_2d,
leaky_relu=leaky_relu,
dropout=dropout,
)
)
enc_channels = out_channels * 2 # concat(conv1d, conv2d)
if d >= self.attn_start:
self.vc_blocks[str(d)] = VCBlock(
channels=enc_channels, num_heads=num_heads
)
trans_channels = inplanes * (2**self.num_depths)
self.trans_block = nn.Sequential(
nn.ConvTranspose2d(
in_channels=trans_channels,
out_channels=trans_channels,
kernel_size=kernel_size,
stride=(1, 1),
padding=padding_2d,
),
nn.BatchNorm2d(trans_channels),
nn.LeakyReLU(leaky_relu),
)
self.deconvs = nn.ModuleList()
for d in reversed(range(self.num_depths)):
in_channels = (
trans_channels
if d == self.num_depths - 1
else inplanes * 2 * (2 ** (d + 1))
)
out_channels = 1 if d == 0 else inplanes * (2**d)
# print(f"Deconvolution2D_layer.__init__[{d}]: {in_channels} {out_channels}")
self.deconvs.append(
Deconvolution2D_layer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
padding=padding_2d,
leaky_relu=leaky_relu,
dropout=dropout,
)
)
# print(f"creating deconv: {in_channels} / {out_channels}")
def _downsample_mask(self, mask: torch.Tensor, target_D: int) -> torch.Tensor:
"""
mask: (B, 12, D) with 1=invisible, 0=visible
return: (B, 12, target_D) with 1/0 유지
"""
mask_down = mask.float()
mask_down = F.max_pool1d(
mask_down, kernel_size=2, stride=2
) # (B,12,floor(D/2))
if mask_down.shape[-1] != target_D:
mask_down = F.interpolate(mask_down, size=target_D, mode="nearest")
return (mask_down >= 0.5).to(mask.dtype)
def _log(self, name, x):
if self.verbose:
print(f"{name:<28}: {tuple(x.shape)}")
def make_default_group_center_mask_batch(
self, B: int, device=None, dtype=torch.int8
):
group_len = 1250 # 2.5s
vis_len = 625 # 1.25s
total_len = 5000
center_offset = (group_len - vis_len) // 2 # 312
# mask만 tensor로 생성
mask = torch.ones((12, total_len), device=device, dtype=dtype)
for g in range(4):
start = g * group_len + center_offset
end = start + vis_len
for lead in range(g * 3, g * 3 + 3):
mask[lead, start:end] = 0 # visible
return mask.unsqueeze(0).expand(B, -1, -1)
def forward(self, x: torch.Tensor) -> torch.Tensor:
input, mask = x
B, L, D = input.shape
if mask is None:
mask = self.make_default_group_center_mask_batch(
B, device=input.device, dtype=torch.float16
)
assert (
L == 12 and D == 5000
), "this network's input must be 12 lead 5000 points digitized signal"
input = input.unsqueeze(1) # make channel
out_1d = input
out_2d = input
encs = []
mask_down = mask
# encs_visible = []
# print(f"input: {input.shape}")
self._log("input", input)
for d in range(self.num_depths):
out_1d = self.convs_1d[d](out_1d)
out_2d = self.convs_2d[d](out_2d)
enc = torch.cat((out_1d, out_2d), dim=1) # (B, 2*C, 12, D)
self._log(f"enc[{d}]", enc)
mask_down = self._downsample_mask(mask_down, enc.shape[-1])
self._log(f"mask_down[{d}]", mask_down)
key = str(d)
if key in self.vc_blocks:
enc = self.vc_blocks[key](enc, mask_down)
self._log(f"enc_refined[{d}]", enc)
encs.append(enc)
trans = self.trans_block(encs[-1])
self._log("trans", trans)
out = self.deconvs[0](trans)
self._log("out initial", out)
# combine skip connection and visible context with encoding feature
for d in range(1, self.num_depths):
skip = encs[-(d + 1)] # 아래쪽 depth부터 사용
self._log(f"skip[{d}]", skip)
out = F.interpolate(out, skip.shape[-2:], mode="nearest")
self._log(f"out upsampled[{d}]", out)
out = torch.cat((out, skip), dim=1)
self._log(f"out concat[{d}]", out)
out = self.deconvs[d](out)
self._log(f"out deconv[{d}]", out)
out = F.interpolate(out, input.shape[-2:], mode="nearest")
self._log("out final upsampled", out)
out = out.squeeze(1)
self._log("out final", out)
# return out, encs_visible
return out
if __name__ == "__main__":
def get_model_size(model):
total_params = sum(p.numel() for p in model.parameters())
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
model_size_MB = total_params * 4 / (1024**2) # float32 기준 (4 bytes)
print(f"Total Parameters : {total_params:,}")
print(f"Trainable Parameters : {trainable_params:,}")
print(f"Estimated Model Size : {model_size_MB:.2f} MB")
return total_params, model_size_MB
class Config:
pass
class Activation:
pass
config = Config()
config.inplanes = 8
config.kernel_size = (7, 7)
config.num_depths_attn_start = (5, 2)
config.num_heads = 8
config.leaky_relu = 0.02
config.dropout = 0.2
config.activation = Activation()
config.activation.name = "Identity"
config.activation.params = None
input = torch.rand(size=(1, 12, 5000))
model = ECGRecoverRandomMaskWithRS5(config, True)
model.eval()
out = model([input, None])
print(out.shape)
from torchinfo import summary
# for i in range(len(encs_visible)):
# print(encs_visible[i].shape)
# summary(model, input_size=(1, 12, 5000), depth=4)
# get_model_size(model)
# from torchviz import make_dot
# # 그래프 생성
# dot = make_dot(
# out, params=dict(model.named_parameters()), show_attrs=False, show_saved=False
# )
# # 파일로 저장 (PNG)
# dot.render("ecgrecover_vc_filtermask", format="png")
# from torchview import draw_graph
# graph = draw_graph(
# model,
# input_size=(1, 12, 5000),
# expand_nested=False, # ← 내부 세부 구조 펼치지 않음 → 매우 간단
# graph_dir="TD", # top-down
# )
# graph.visual_graph.render("model_overview", format="png")
|