File size: 20,861 Bytes
355f250 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 | """Paper-aligned SatMAE model components.
This is an original implementation of the architecture described in SatMAE.
The upstream repository was used only as a behavioral reference; no upstream
source text is incorporated here.
"""
import math
from functools import partial
import torch
from torch import nn
def _sincos_1d(values, dim):
"""Return a fixed sine-cosine embedding for arbitrary scalar positions."""
if dim <= 0:
return values.new_zeros((*values.shape, 0))
pairs = (dim + 1) // 2
omega = torch.arange(pairs, device=values.device, dtype=torch.float32)
omega = torch.exp(-math.log(10000.0) * omega / max(pairs - 1, 1))
phase = values.to(torch.float32).unsqueeze(-1) * omega
return torch.cat((phase.sin(), phase.cos()), dim=-1)[..., :dim]
def _sincos_2d(grid_size, dim):
"""Return a fixed row-major 2D sine-cosine position embedding."""
rows, cols = torch.meshgrid(
torch.arange(grid_size, dtype=torch.float32),
torch.arange(grid_size, dtype=torch.float32),
indexing="ij",
)
row_dim = dim // 2
return torch.cat(
(_sincos_1d(rows.reshape(-1), row_dim),
_sincos_1d(cols.reshape(-1), dim - row_dim)),
dim=-1,
)
def _timestamp_embedding(timestamps, dim):
"""Encode either scalar times or fMoW ``[year, month, hour]`` tuples."""
if timestamps.ndim == 2:
return _sincos_1d(timestamps, dim)
if timestamps.ndim != 3 or timestamps.shape[-1] != 3:
raise ValueError("timestamps must have shape [B, T] or [B, T, 3]")
field_dims = [dim // 3] * 3
for index in range(dim % 3):
field_dims[index] += 1
return torch.cat(
[_sincos_1d(timestamps[..., index], field_dim)
for index, field_dim in enumerate(field_dims)],
dim=-1,
)
class PatchEmbed(nn.Module):
def __init__(self, image_size, patch_size, in_channels, embed_dim):
super().__init__()
self.image_size = image_size
self.patch_size = patch_size
self.num_patches = (image_size // patch_size) ** 2
self.proj = nn.Conv2d(
in_channels, embed_dim, kernel_size=patch_size, stride=patch_size
)
def forward(self, images):
if images.shape[-2:] != (self.image_size, self.image_size):
raise ValueError(
f"expected {self.image_size}x{self.image_size} images, "
f"got {tuple(images.shape[-2:])}"
)
return self.proj(images).flatten(2).transpose(1, 2)
class TransformerBlock(nn.Module):
def __init__(self, dim, num_heads, mlp_ratio=4.0, norm_layer=nn.LayerNorm):
super().__init__()
self.norm1 = norm_layer(dim)
self.attention = nn.MultiheadAttention(
dim, num_heads, dropout=0.0, bias=True, batch_first=True
)
self.norm2 = norm_layer(dim)
hidden_dim = int(dim * mlp_ratio)
self.mlp = nn.Sequential(
nn.Linear(dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, dim)
)
def forward(self, tokens):
normalized = self.norm1(tokens)
tokens = tokens + self.attention(
normalized, normalized, normalized, need_weights=False
)[0]
return tokens + self.mlp(self.norm2(tokens))
class SatMAE(nn.Module):
"""Masked autoencoder for temporal or grouped multispectral imagery.
Temporal inputs use shape ``[B, T, C, H, W]`` and optional timestamps
``[B, T]``. Multispectral inputs use shape ``[B, C, H, W]``.
"""
def __init__(
self,
image_size=224,
patch_size=16,
in_channels=3,
frames=3,
embed_dim=1024,
encoder_depth=24,
encoder_heads=16,
decoder_dim=512,
decoder_depth=8,
decoder_heads=16,
mlp_ratio=4.0,
mode="temporal",
spectral_groups=None,
mask_ratio=0.75,
norm_pix_loss=False,
same_mask=False,
spatial_mask=False,
temporal_embed_dim=None,
decoder_temporal_embed_dim=None,
channel_embed_dim=None,
decoder_channel_embed_dim=None,
norm_layer=None,
):
super().__init__()
if image_size % patch_size:
raise ValueError("image_size must be divisible by patch_size")
if not 0.0 <= mask_ratio < 1.0:
raise ValueError("mask_ratio must be in [0, 1)")
if mode not in {"temporal", "multispectral"}:
raise ValueError("mode must be temporal or multispectral")
if embed_dim % encoder_heads or decoder_dim % decoder_heads:
raise ValueError("embedding dimensions must be divisible by head counts")
norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
self.image_size = image_size
self.patch_size = patch_size
self.in_channels = in_channels
self.frames = frames
self.embed_dim = embed_dim
self.decoder_dim = decoder_dim
self.mode = mode
self.mask_ratio = mask_ratio
self.norm_pix_loss = norm_pix_loss
self.same_mask = same_mask
self.spatial_mask = spatial_mask
self.grid_size = image_size // patch_size
self.num_patches = self.grid_size ** 2
if mode == "temporal":
self.spectral_groups = None
self.patch_embed = PatchEmbed(
image_size, patch_size, in_channels, embed_dim
)
self.token_groups = frames
semantic_dim = temporal_embed_dim
if semantic_dim is None:
semantic_dim = min(128, max(2, embed_dim // 4))
decoder_semantic_dim = decoder_temporal_embed_dim
if decoder_semantic_dim is None:
decoder_semantic_dim = min(64, max(2, decoder_dim // 4))
prediction_dims = [patch_size ** 2 * in_channels]
else:
groups = spectral_groups or [list(range(in_channels))]
flattened = [channel for group in groups for channel in group]
if sorted(flattened) != list(range(in_channels)):
raise ValueError("spectral_groups must partition all input channels")
self.spectral_groups = tuple(tuple(group) for group in groups)
self.patch_embed = nn.ModuleList(
PatchEmbed(image_size, patch_size, len(group), embed_dim)
for group in self.spectral_groups
)
self.token_groups = len(self.spectral_groups)
semantic_dim = channel_embed_dim
if semantic_dim is None:
semantic_dim = min(256, max(2, embed_dim // 4))
decoder_semantic_dim = decoder_channel_embed_dim
if decoder_semantic_dim is None:
decoder_semantic_dim = min(128, max(2, decoder_dim // 4))
prediction_dims = [patch_size ** 2 * len(g) for g in self.spectral_groups]
if not 0 < semantic_dim < embed_dim:
raise ValueError("encoder semantic embedding dimension is invalid")
if not 0 < decoder_semantic_dim < decoder_dim:
raise ValueError("decoder semantic embedding dimension is invalid")
self.semantic_dim = semantic_dim
self.decoder_semantic_dim = decoder_semantic_dim
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
self.mask_token = nn.Parameter(torch.zeros(1, 1, decoder_dim))
self.register_buffer(
"spatial_pos_embed",
_sincos_2d(self.grid_size, embed_dim - semantic_dim),
persistent=True,
)
self.register_buffer(
"decoder_spatial_pos_embed",
_sincos_2d(self.grid_size, decoder_dim - decoder_semantic_dim),
persistent=True,
)
if mode == "multispectral":
group_ids = torch.arange(self.token_groups, dtype=torch.float32)
self.register_buffer(
"group_embed", _sincos_1d(group_ids, semantic_dim), persistent=True
)
self.register_buffer(
"decoder_group_embed",
_sincos_1d(group_ids, decoder_semantic_dim),
persistent=True,
)
self.blocks = nn.ModuleList(
TransformerBlock(embed_dim, encoder_heads, mlp_ratio, norm_layer)
for _ in range(encoder_depth)
)
self.norm = norm_layer(embed_dim)
self.decoder_embed = nn.Linear(embed_dim, decoder_dim)
self.decoder_blocks = nn.ModuleList(
TransformerBlock(decoder_dim, decoder_heads, mlp_ratio, norm_layer)
for _ in range(decoder_depth)
)
self.decoder_norm = norm_layer(decoder_dim)
self.decoder_pred = nn.ModuleList(
nn.Linear(decoder_dim, output_dim) for output_dim in prediction_dims
)
self.initialize_weights()
def initialize_weights(self):
patch_embeds = (
[self.patch_embed]
if isinstance(self.patch_embed, PatchEmbed)
else self.patch_embed
)
for patch_embed in patch_embeds:
nn.init.xavier_uniform_(patch_embed.proj.weight.flatten(1))
if patch_embed.proj.bias is not None:
nn.init.zeros_(patch_embed.proj.bias)
nn.init.normal_(self.cls_token, std=0.02)
nn.init.normal_(self.mask_token, std=0.02)
for module in self.modules():
if isinstance(module, nn.Linear):
nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.LayerNorm):
nn.init.ones_(module.weight)
nn.init.zeros_(module.bias)
def patchify(self, images):
if images.ndim != 4:
raise ValueError("patchify expects [B, C, H, W]")
batch, channels, height, width = images.shape
patch = self.patch_size
if height != width or height != self.image_size:
raise ValueError(f"expected square images of size {self.image_size}")
patches = images.reshape(
batch, channels, height // patch, patch, width // patch, patch
)
patches = patches.permute(0, 2, 4, 1, 3, 5)
return patches.reshape(batch, self.num_patches, channels * patch ** 2)
def unpatchify(self, patches, channels=None):
channels = channels or self.in_channels
batch = patches.shape[0]
patch = self.patch_size
expected = channels * patch ** 2
if patches.shape[1:] != (self.num_patches, expected):
raise ValueError("patch tensor has incompatible shape")
images = patches.reshape(
batch, self.grid_size, self.grid_size, channels, patch, patch
)
images = images.permute(0, 3, 1, 4, 2, 5)
return images.reshape(batch, channels, self.image_size, self.image_size)
def _random_masking(self, tokens, mask_ratio, share_spatial_mask):
batch, length, dim = tokens.shape
if share_spatial_mask:
units = self.num_patches
len_keep_units = int(units * (1.0 - mask_ratio))
noise = torch.rand(batch, units, device=tokens.device)
spatial_order = noise.argsort(dim=1)
kept = [spatial_order[:, :len_keep_units] + g * units
for g in range(self.token_groups)]
removed = [spatial_order[:, len_keep_units:] + g * units
for g in range(self.token_groups)]
ids_shuffle = torch.cat(kept + removed, dim=1)
len_keep = len_keep_units * self.token_groups
else:
len_keep = int(length * (1.0 - mask_ratio))
ids_shuffle = torch.rand(batch, length, device=tokens.device).argsort(dim=1)
ids_restore = ids_shuffle.argsort(dim=1)
ids_keep = ids_shuffle[:, :len_keep]
visible = torch.gather(tokens, 1, ids_keep.unsqueeze(-1).expand(-1, -1, dim))
mask = torch.ones(batch, length, device=tokens.device)
mask[:, :len_keep] = 0
mask = torch.gather(mask, 1, ids_restore)
return visible, mask, ids_restore
def _temporal_tokens(self, images, timestamps):
if images.ndim != 5:
raise ValueError("temporal mode expects images shaped [B, T, C, H, W]")
batch, frames, channels, _, _ = images.shape
if frames != self.frames or channels != self.in_channels:
raise ValueError(
f"expected T={self.frames}, C={self.in_channels}; got T={frames}, C={channels}"
)
if timestamps is None:
timestamps = torch.arange(frames, device=images.device).expand(batch, -1)
if timestamps.shape[:2] != (batch, frames):
raise ValueError(
f"timestamps must start with shape {(batch, frames)}, "
f"got {tuple(timestamps.shape)}"
)
spatial = self.spatial_pos_embed.to(dtype=images.dtype)
time = _timestamp_embedding(timestamps, self.semantic_dim).to(dtype=images.dtype)
position = torch.cat(
(spatial.view(1, 1, self.num_patches, -1).expand(batch, frames, -1, -1),
time.unsqueeze(2).expand(-1, -1, self.num_patches, -1)),
dim=-1,
).reshape(batch, frames * self.num_patches, self.embed_dim)
tokens = torch.stack(
[self.patch_embed(images[:, frame]) for frame in range(frames)], dim=1
).reshape(batch, frames * self.num_patches, self.embed_dim)
return tokens + position, timestamps
def _multispectral_tokens(self, images):
if images.ndim != 4 or images.shape[1] != self.in_channels:
raise ValueError(
f"multispectral mode expects images shaped [B, {self.in_channels}, H, W]"
)
spatial = self.spatial_pos_embed.to(dtype=images.dtype)
group = self.group_embed.to(dtype=images.dtype)
positions = torch.cat(
(spatial.view(1, self.num_patches, -1).expand(self.token_groups, -1, -1),
group.view(self.token_groups, 1, -1).expand(-1, self.num_patches, -1)),
dim=-1,
).reshape(1, self.token_groups * self.num_patches, self.embed_dim)
tokens = torch.cat(
[embed(images[:, channels])
for embed, channels in zip(self.patch_embed, self.spectral_groups)],
dim=1,
)
return tokens + positions
def forward_encoder(self, images, timestamps=None, mask_ratio=None):
ratio = self.mask_ratio if mask_ratio is None else mask_ratio
if not 0.0 <= ratio < 1.0:
raise ValueError("mask_ratio must be in [0, 1)")
if self.mode == "temporal":
tokens, timestamps = self._temporal_tokens(images, timestamps)
shared = self.same_mask
else:
tokens = self._multispectral_tokens(images)
shared = self.spatial_mask
tokens, mask, ids_restore = self._random_masking(tokens, ratio, shared)
cls = self.cls_token.expand(tokens.shape[0], -1, -1)
tokens = torch.cat((cls, tokens), dim=1)
for block in self.blocks:
tokens = block(tokens)
return self.norm(tokens), mask, ids_restore, timestamps
def _decoder_positions(self, batch, timestamps, dtype, device):
spatial = self.decoder_spatial_pos_embed.to(device=device, dtype=dtype)
if self.mode == "temporal":
semantic = _timestamp_embedding(timestamps, self.decoder_semantic_dim).to(dtype=dtype)
else:
semantic = self.decoder_group_embed.to(device=device, dtype=dtype)
semantic = semantic.unsqueeze(0).expand(batch, -1, -1)
position = torch.cat(
(spatial.view(1, 1, self.num_patches, -1).expand(batch, self.token_groups, -1, -1),
semantic.unsqueeze(2).expand(-1, -1, self.num_patches, -1)),
dim=-1,
)
return position.reshape(batch, self.token_groups * self.num_patches, self.decoder_dim)
def forward_decoder(self, latent, ids_restore, timestamps=None):
tokens = self.decoder_embed(latent)
mask_tokens = self.mask_token.expand(
tokens.shape[0], ids_restore.shape[1] + 1 - tokens.shape[1], -1
)
restored = torch.cat((tokens[:, 1:], mask_tokens), dim=1)
restored = torch.gather(
restored, 1, ids_restore.unsqueeze(-1).expand(-1, -1, self.decoder_dim)
)
positions = self._decoder_positions(
tokens.shape[0], timestamps, tokens.dtype, tokens.device
)
tokens = torch.cat((tokens[:, :1], restored + positions), dim=1)
for block in self.decoder_blocks:
tokens = block(tokens)
decoded = self.decoder_norm(tokens)[:, 1:]
if self.mode == "temporal":
return [self.decoder_pred[0](decoded)]
decoded = decoded.reshape(
decoded.shape[0], self.token_groups, self.num_patches, self.decoder_dim
)
return [head(decoded[:, index]) for index, head in enumerate(self.decoder_pred)]
def _targets(self, images):
if self.mode == "temporal":
return [torch.cat(
[self.patchify(images[:, frame]) for frame in range(self.frames)], dim=1
)]
return [self.patchify(images[:, group]) for group in self.spectral_groups]
def forward_loss(self, targets, predictions, mask):
losses = []
if self.mode == "temporal":
pairs = [(targets[0], predictions[0], mask)]
else:
group_mask = mask.reshape(mask.shape[0], self.token_groups, self.num_patches)
pairs = [
(target, prediction, group_mask[:, index])
for index, (target, prediction) in enumerate(zip(targets, predictions))
]
removed = mask.new_zeros(())
total = mask.new_zeros(())
for target, prediction, patch_mask in pairs:
patch_loss = (prediction - target).square().mean(dim=-1)
total = total + (patch_loss * patch_mask).sum()
removed = removed + patch_mask.sum()
losses.append(patch_loss)
return total / removed.clamp_min(1), losses
def _normalize_targets(self, targets):
if not self.norm_pix_loss:
return targets
normalized = []
for target in targets:
mean = target.mean(dim=-1, keepdim=True)
variance = target.var(dim=-1, keepdim=True, unbiased=False)
normalized.append((target - mean) / torch.sqrt(variance + 1e-6))
return normalized
def _padded_outputs(self, tensors):
if self.mode == "temporal":
return tensors[0]
width = max(tensor.shape[-1] for tensor in tensors)
padded = []
for tensor in tensors:
if tensor.shape[-1] < width:
tensor = torch.nn.functional.pad(tensor, (0, width - tensor.shape[-1]))
padded.append(tensor)
return torch.cat(padded, dim=1)
def forward(self, images, timestamps=None, mask_ratio=None):
latent, mask, ids_restore, timestamps = self.forward_encoder(
images, timestamps, mask_ratio
)
predictions = self.forward_decoder(latent, ids_restore, timestamps)
targets = self._normalize_targets(self._targets(images))
loss, patch_losses = self.forward_loss(targets, predictions, mask)
return {
"loss": loss,
"prediction": self._padded_outputs(predictions),
"target": self._padded_outputs(targets),
"mask": mask.bool(),
"features": latent,
"ids_restore": ids_restore,
"group_predictions": predictions,
"group_targets": targets,
"patch_losses": patch_losses,
}
def satmae_vit_base_patch16(**kwargs):
return SatMAE(
patch_size=16, embed_dim=768, encoder_depth=12, encoder_heads=12,
decoder_dim=512, decoder_depth=8, decoder_heads=16,
temporal_embed_dim=128, decoder_temporal_embed_dim=64,
channel_embed_dim=256, decoder_channel_embed_dim=128, **kwargs
)
def satmae_vit_large_patch16(**kwargs):
return SatMAE(
patch_size=16, embed_dim=1024, encoder_depth=24, encoder_heads=16,
decoder_dim=512, decoder_depth=8, decoder_heads=16,
temporal_embed_dim=128, decoder_temporal_embed_dim=64,
channel_embed_dim=256, decoder_channel_embed_dim=128, **kwargs
)
def satmae_vit_huge_patch14(**kwargs):
return SatMAE(
patch_size=14, embed_dim=1280, encoder_depth=32, encoder_heads=16,
decoder_dim=512, decoder_depth=8, decoder_heads=16,
temporal_embed_dim=128, decoder_temporal_embed_dim=64,
channel_embed_dim=256, decoder_channel_embed_dim=128, **kwargs
)
|