File size: 17,009 Bytes
8abd44b |
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 |
"""
Modular, block-based components for building autoencoders in PyTorch.
Core goals:
- Composable building blocks with consistent interfaces
- Support 2D (B, F) and 3D (B, T, F) tensors where applicable
- Simple configs to construct blocks and sequences
- Safe-by-default validation and helpful errors
This module is intentionally self-contained to allow gradual integration with
existing models. It does not mutate current behavior.
"""
from __future__ import annotations
from dataclasses import dataclass
from typing import Any, Dict, Iterable, List, Optional, Sequence, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
# Import config dataclasses that define block configurations
try:
from .configuration_autoencoder import (
BlockConfig,
LinearBlockConfig,
AttentionBlockConfig,
RecurrentBlockConfig,
ConvolutionalBlockConfig,
VariationalBlockConfig,
)
except Exception:
from configuration_autoencoder import (
BlockConfig,
LinearBlockConfig,
AttentionBlockConfig,
RecurrentBlockConfig,
ConvolutionalBlockConfig,
VariationalBlockConfig,
)
# Import shared utilities
try:
from .utils import _get_activation, _get_norm, _flatten_3d_to_2d, _maybe_restore_3d
except Exception:
from utils import _get_activation, _get_norm, _flatten_3d_to_2d, _maybe_restore_3d
# ---------------------------- Base Block ---------------------------- #
class BaseBlock(nn.Module):
"""Abstract base for all blocks.
All blocks should accept 2D (B, F) or 3D (B, T, F) tensors and return the
same rank, with last-dim equal to `output_dim`.
"""
def forward(self, x: torch.Tensor, **kwargs) -> torch.Tensor: # pragma: no cover - abstract
raise NotImplementedError
@property
def output_dim(self) -> int: # pragma: no cover - abstract
raise NotImplementedError
# ---------------------------- Residual Base ---------------------------- #
class ResidualBlock(BaseBlock):
"""Base class for blocks supporting residual connections.
Implements a safe residual add when input and output dims match; otherwise
falls back to a learned projection. Residuals can be scaled.
"""
def __init__(self, residual: bool = False, residual_scale: float = 1.0, proj_dim_in: Optional[int] = None, proj_dim_out: Optional[int] = None):
super().__init__()
self.use_residual = residual
self.residual_scale = residual_scale
self._proj: Optional[nn.Module] = None
if residual and proj_dim_in is not None and proj_dim_out is not None and proj_dim_in != proj_dim_out:
self._proj = nn.Linear(proj_dim_in, proj_dim_out)
def _apply_residual(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
if not self.use_residual:
return y
x2d, hint = _flatten_3d_to_2d(x)
y2d, _ = _flatten_3d_to_2d(y)
if x2d.shape[-1] != y2d.shape[-1]:
if self._proj is None:
self._proj = nn.Linear(x2d.shape[-1], y2d.shape[-1]).to(y2d.device)
x2d = self._proj(x2d)
out = x2d + self.residual_scale * y2d
return _maybe_restore_3d(out, hint)
# ---------------------------- LinearBlock ---------------------------- #
class LinearBlock(ResidualBlock):
"""Basic linear transformation with normalization and activation.
- Handles both 2D (B, F) and 3D (B, T, F) tensors
- Optional normalization: batch|layer|group|instance|none
- Configurable activation
- Optional dropout
- Optional residual connection (with auto projection)
"""
def __init__(self, cfg: LinearBlockConfig):
super().__init__(residual=cfg.use_residual, residual_scale=cfg.residual_scale, proj_dim_in=cfg.input_dim, proj_dim_out=cfg.output_dim)
self.cfg = cfg
self.linear = nn.Linear(cfg.input_dim, cfg.output_dim)
# Normalizations that expect N, C require 2D tensors; for 3D we flatten
# For LayerNorm, it supports last-dim directly
if cfg.normalization == "layer":
self.norm = nn.LayerNorm(cfg.output_dim)
else:
self.norm = _get_norm(cfg.normalization, cfg.output_dim)
self.act = _get_activation(cfg.activation)
self.drop = nn.Dropout(cfg.dropout_rate) if cfg.dropout_rate and cfg.dropout_rate > 0 else nn.Identity()
@property
def output_dim(self) -> int:
return self.cfg.output_dim
def forward(self, x: torch.Tensor) -> torch.Tensor:
x_in = x
x2d, hint = _flatten_3d_to_2d(x)
y = self.linear(x2d)
# Apply norm safely
if isinstance(self.norm, (nn.BatchNorm1d, nn.InstanceNorm1d, nn.GroupNorm)):
y = self.norm(y)
else:
# LayerNorm or Identity operates on last dim and supports both 2D/3D; we already have 2D
y = self.norm(y)
y = self.act(y)
y = self.drop(y)
y = _maybe_restore_3d(y, hint)
return self._apply_residual(x_in, y)
# ---------------------------- AttentionBlock ---------------------------- #
class AttentionBlock(BaseBlock):
"""Multi-head self-attention with optional FFN.
Expects inputs as 3D (B, T, D) or 2D (B, D) which will be treated as (B, 1, D).
Supports optional attn mask and key padding mask via kwargs.
"""
def __init__(self, cfg: AttentionBlockConfig):
super().__init__()
self.cfg = cfg
d_model = cfg.input_dim
self.mha = nn.MultiheadAttention(d_model, num_heads=cfg.num_heads, dropout=cfg.dropout_rate, batch_first=True)
self.ln1 = nn.LayerNorm(d_model)
ffn_dim = cfg.ffn_dim or (4 * d_model)
self.ffn = nn.Sequential(
nn.Linear(d_model, ffn_dim),
_get_activation("gelu"),
nn.Dropout(cfg.dropout_rate),
nn.Linear(ffn_dim, d_model),
)
self.ln2 = nn.LayerNorm(d_model)
self.dropout = nn.Dropout(cfg.dropout_rate)
@property
def output_dim(self) -> int:
return self.cfg.input_dim
def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None, key_padding_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
if x.dim() == 2:
x = x.unsqueeze(1)
squeeze_back = True
else:
squeeze_back = False
# Self-attention
residual = x
attn_out, _ = self.mha(x, x, x, attn_mask=attn_mask, key_padding_mask=key_padding_mask, need_weights=False)
x = self.ln1(residual + self.dropout(attn_out))
# FFN
residual = x
x = self.ffn(x)
x = self.ln2(residual + self.dropout(x))
if squeeze_back:
x = x.squeeze(1)
return x
# ---------------------------- RecurrentBlock ---------------------------- #
class RecurrentBlock(BaseBlock):
"""RNN processing block supporting LSTM/GRU/RNN.
Input: 3D (B, T, F) preferred. If 2D, treated as (B, 1, F).
Output dim equals cfg.output_dim if set; otherwise hidden_size * directions.
"""
def __init__(self, cfg: RecurrentBlockConfig):
super().__init__()
self.cfg = cfg
rnn_type = cfg.rnn_type.lower()
rnn_cls = {"lstm": nn.LSTM, "gru": nn.GRU, "rnn": nn.RNN}.get(rnn_type)
if rnn_cls is None:
raise ValueError(f"Unknown rnn_type: {cfg.rnn_type}")
self.rnn = rnn_cls(
input_size=cfg.input_dim,
hidden_size=cfg.hidden_size,
num_layers=cfg.num_layers,
batch_first=True,
dropout=cfg.dropout_rate if cfg.num_layers > 1 else 0.0,
bidirectional=cfg.bidirectional,
)
out_dim = cfg.hidden_size * (2 if cfg.bidirectional else 1)
self._out_dim = cfg.output_dim or out_dim
self.proj = None if self._out_dim == out_dim else nn.Linear(out_dim, self._out_dim)
@property
def output_dim(self) -> int:
return self._out_dim
def forward(self, x: torch.Tensor, lengths: Optional[torch.Tensor] = None) -> torch.Tensor:
squeeze_back = False
if x.dim() == 2:
x = x.unsqueeze(1)
squeeze_back = True
if lengths is not None:
x = nn.utils.rnn.pack_padded_sequence(x, lengths, batch_first=True, enforce_sorted=False)
if isinstance(self.rnn, nn.LSTM):
out, (h, c) = self.rnn(x)
else:
out, h = self.rnn(x)
if lengths is not None:
out, _ = nn.utils.rnn.pad_packed_sequence(out, batch_first=True)
# Use last timestep
y = out[:, -1, :]
if self.proj is not None:
y = self.proj(y)
if squeeze_back:
# Keep 2D output
return y
# Return (B, 1, D) to keep 3D shape consistent with sequences
return y.unsqueeze(1)
# ---------------------------- ConvolutionalBlock ---------------------------- #
class ConvolutionalBlock(BaseBlock):
"""1D convolutional block for sequence-like data.
Accepts 3D (B, T, F) or 2D (B, F) which is treated as (B, 1, F).
"""
def __init__(self, cfg: ConvolutionalBlockConfig):
super().__init__()
self.cfg = cfg
# Conv1d expects (B, C_in, L). We interpret features as channels and time as length.
# For inputs shaped (B, T, F): we transpose to (B, F, T), apply conv, transpose back.
padding = cfg.padding
if isinstance(padding, str) and padding == "same":
pad = cfg.kernel_size // 2
else:
pad = int(padding)
self.conv = nn.Conv1d(cfg.input_dim, cfg.output_dim, kernel_size=cfg.kernel_size, padding=pad)
# Norm: for Conv1d, use 1d norms over channels
if cfg.normalization == "layer":
self.norm = nn.GroupNorm(1, cfg.output_dim) # Layer-like over channels
else:
self.norm = _get_norm(cfg.normalization, cfg.output_dim)
self.act = _get_activation(cfg.activation)
self.drop = nn.Dropout(cfg.dropout_rate) if cfg.dropout_rate and cfg.dropout_rate > 0 else nn.Identity()
@property
def output_dim(self) -> int:
return self.cfg.output_dim
def forward(self, x: torch.Tensor) -> torch.Tensor:
squeeze_back = False
if x.dim() == 2:
x = x.unsqueeze(1)
squeeze_back = True
# x: (B, T, F) -> (B, F, T)
x = x.transpose(1, 2)
y = self.conv(x)
if isinstance(self.norm, (nn.BatchNorm1d, nn.InstanceNorm1d, nn.GroupNorm)):
y = self.norm(y)
y = self.act(y)
y = self.drop(y)
y = y.transpose(1, 2)
if squeeze_back:
y = y.squeeze(1)
return y
# ---------------------------- VariationalBlock ---------------------------- #
class VariationalBlock(BaseBlock):
"""Encapsulates mu/logvar projection and reparameterization.
Input can be 2D (B, F) or 3D (B, T, F); for 3D, operates per timestep and returns same rank.
Stores mu/logvar on the module for downstream loss usage.
"""
def __init__(self, cfg: VariationalBlockConfig):
super().__init__()
self.cfg = cfg
self.fc_mu = nn.Linear(cfg.input_dim, cfg.latent_dim)
self.fc_logvar = nn.Linear(cfg.input_dim, cfg.latent_dim)
self._mu: Optional[torch.Tensor] = None
self._logvar: Optional[torch.Tensor] = None
@property
def output_dim(self) -> int:
return self.cfg.latent_dim
def forward(self, x: torch.Tensor, training: Optional[bool] = None) -> torch.Tensor:
if training is None:
training = self.training
x2d, hint = _flatten_3d_to_2d(x)
mu = self.fc_mu(x2d)
logvar = self.fc_logvar(x2d)
if training:
std = torch.exp(0.5 * logvar)
eps = torch.randn_like(std)
z = mu + eps * std
else:
z = mu
self._mu = mu
self._logvar = logvar
z = _maybe_restore_3d(z, hint)
return z
# ---------------------------- BlockSequence ---------------------------- #
class BlockSequence(nn.Module):
"""Compose multiple blocks into a validated sequence.
- Validates dimension flow between blocks
- Supports gradient checkpointing (per-block) via forward(checkpoint=True)
- Supports optional skip connections: pass `skips` as list of (src_idx, dst_idx)
"""
def __init__(self, blocks: Sequence[BaseBlock], validate_dims: bool = True, skips: Optional[List[Tuple[int, int]]] = None):
super().__init__()
self.blocks = nn.ModuleList(blocks)
self.skips = skips or []
if validate_dims and len(blocks) > 1:
for i in range(1, len(blocks)):
prev = blocks[i - 1]
cur = blocks[i]
if getattr(prev, "output_dim", None) is None or getattr(cur, "output_dim", None) is None:
continue
if prev.output_dim != cur.output_dim and not isinstance(cur, LinearBlock):
# Allow LinearBlock to change dims; others must preserve unless they project internally
pass # Only warn; users may know what they're doing
def forward(self, x: torch.Tensor, checkpoint: bool = False, **kwargs) -> torch.Tensor:
activations: Dict[int, torch.Tensor] = {}
for i, block in enumerate(self.blocks):
if checkpoint and x.requires_grad:
x = torch.utils.checkpoint.checkpoint(lambda inp: block(inp, **kwargs), x)
else:
x = block(x, **kwargs)
activations[i] = x
# Apply any pending skips to this idx
for src, dst in self.skips:
if dst == i and src in activations:
x = x + activations[src]
return x
# ---------------------------- Factory ---------------------------- #
class BlockFactory:
"""Factory to build blocks/sequences from configs.
This is intentionally minimal; extend as needed.
"""
@staticmethod
def build_block(cfg: Union[BlockConfig, Dict[str, Any]]) -> BaseBlock:
# Allow dict-like
if isinstance(cfg, dict):
type_name = cfg.get("type")
# copy and remove 'type' to satisfy dataclass init
params = dict(cfg)
params.pop("type", None)
if type_name == "linear":
return LinearBlock(LinearBlockConfig(**params))
if type_name == "attention":
return AttentionBlock(AttentionBlockConfig(**params))
if type_name == "recurrent":
return RecurrentBlock(RecurrentBlockConfig(**params))
if type_name == "conv1d":
return ConvolutionalBlock(ConvolutionalBlockConfig(**params))
raise ValueError(f"Unsupported block type in dict cfg: {type_name} cfg={cfg}")
# Dataclass path
if isinstance(cfg, LinearBlockConfig) or getattr(cfg, "type", None) == "linear":
if not isinstance(cfg, LinearBlockConfig):
cfg = LinearBlockConfig(**cfg.__dict__) # type: ignore[arg-type]
return LinearBlock(cfg)
if isinstance(cfg, AttentionBlockConfig) or getattr(cfg, "type", None) == "attention":
if not isinstance(cfg, AttentionBlockConfig):
cfg = AttentionBlockConfig(**cfg.__dict__) # type: ignore[arg-type]
return AttentionBlock(cfg)
if isinstance(cfg, RecurrentBlockConfig) or getattr(cfg, "type", None) == "recurrent":
if not isinstance(cfg, RecurrentBlockConfig):
cfg = RecurrentBlockConfig(**cfg.__dict__) # type: ignore[arg-type]
return RecurrentBlock(cfg)
if isinstance(cfg, ConvolutionalBlockConfig) or getattr(cfg, "type", None) == "conv1d":
if not isinstance(cfg, ConvolutionalBlockConfig):
cfg = ConvolutionalBlockConfig(**cfg.__dict__) # type: ignore[arg-type]
return ConvolutionalBlock(cfg)
if isinstance(cfg, VariationalBlockConfig) or getattr(cfg, "type", None) == "variational":
if not isinstance(cfg, VariationalBlockConfig):
cfg = VariationalBlockConfig(**cfg.__dict__) # type: ignore[arg-type]
return VariationalBlock(cfg)
raise ValueError(f"Unsupported block type: {cfg}")
@staticmethod
def build_sequence(configs: Sequence[Union[BlockConfig, Dict[str, Any]]]) -> BlockSequence:
blocks: List[BaseBlock] = [BlockFactory.build_block(c) for c in configs]
return BlockSequence(blocks)
__all__ = [
"BlockConfig",
"LinearBlockConfig",
"AttentionBlockConfig",
"RecurrentBlockConfig",
"ConvolutionalBlockConfig",
"VariationalBlockConfig",
"BaseBlock",
"ResidualBlock",
"LinearBlock",
"AttentionBlock",
"RecurrentBlock",
"ConvolutionalBlock",
"VariationalBlock",
"BlockSequence",
"BlockFactory",
]
|