File size: 5,831 Bytes
71d64bb | 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 | """
Temporal SRDA (T-SRDA) — PDAViT's Spatial-Reduction Dual-Attention
(Zhou et al., Neurocomputing 2026) adapted to the TEMPORAL axis.
Drop-in replacement for TemporalDaViTEncoder inside UTAE_ASPP.
Matches its calling convention exactly:
enc = TemporalSRDAEncoder(d_model=256, n_heads=8, R_schedule=(4, 4, 4))
out = enc(x, pos) # x: (N, T, d) pos: (N, T, d) date PE or None
Mechanism per block (1D mapping of CUSA + dual attention):
KV path : Linear d -> d/R ("point-wise conv")
unfold: group R consecutive timesteps -> (N, T/R, d)
First SA over T/R tokens (key-value "self-checking")
Q path : full-resolution T tokens
Second SA: cross-attention Q(T) x KV(T/R) -> GLOBAL temporal receptive
field in a single layer (Swin/DaViT window schedule needs the
whole [4,8,16]+shift stack to approximate this).
scMLP : FC -> GELU (+ parallel depth-wise Conv1d over T) -> FC
Design notes (do not change casually):
* NO internal gradient checkpointing — UTAE_ASPP already checkpoints at
the chunk level (_run_temporal_chunk). Layer-level checkpointing here
would recreate the double-checkpoint waste found in the DaViT encoder.
* Padding convention matches the existing pipeline: collate_fn zero-pads
to batch max_T and the windowed encoders attend over padded steps too.
We keep that convention for a clean apples-to-apples ablation. The
encoder internally pads T to a multiple of max(R_schedule) and crops
back, so any T works (33..61 all fine).
* Date PE MUST be added before the KV reduction (handled here: x = x+pos
at entry) — grouped KV tokens over irregular acquisitions are
meaningless without date information.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
class TemporalSRDABlock(nn.Module):
def __init__(self, d_model=256, n_heads=8, R=4, mlp_ratio=4.0,
attn_drop=0.0, proj_drop=0.0):
super().__init__()
assert d_model % R == 0, "d_model must be divisible by R"
assert d_model % n_heads == 0
self.d = d_model
self.R = R
self.h = n_heads
self.dh = d_model // n_heads
self.scale = self.dh ** -0.5
# ---- KV reduction path (CUSA, 1D) ---------------------------------
self.norm_kv = nn.LayerNorm(d_model)
self.pw = nn.Linear(d_model, d_model // R, bias=False) # point-wise conv
self.norm_first = nn.LayerNorm(d_model)
self.first_sa = nn.MultiheadAttention(
d_model, n_heads, dropout=attn_drop, batch_first=True)
# ---- Second (dual) attention ---------------------------------------
self.norm_q = nn.LayerNorm(d_model)
self.q_proj = nn.Linear(d_model, d_model, bias=False)
self.kv_proj = nn.Linear(d_model, 2 * d_model, bias=False)
self.attn_drop = nn.Dropout(attn_drop)
self.out_proj = nn.Linear(d_model, d_model)
self.proj_drop = nn.Dropout(proj_drop)
# ---- scMLP ----------------------------------------------------------
hidden = int(d_model * mlp_ratio)
self.norm_mlp = nn.LayerNorm(d_model)
self.fc1 = nn.Linear(d_model, hidden)
self.dw = nn.Conv1d(hidden, hidden, kernel_size=3, padding=1,
groups=hidden) # DW over T
self.fc2 = nn.Linear(hidden, d_model)
def forward(self, x):
"""x: (N, T, d) with T % R == 0 (guaranteed by the encoder wrapper)."""
N, T, d = x.shape
R = self.R
# ---- KV path: point-wise conv -> temporal unfold -> first SA -------
z = self.pw(self.norm_kv(x)) # (N, T, d/R)
z = z.reshape(N, T // R, R * (d // R)) # unfold -> (N, T/R, d)
zn = self.norm_first(z)
z_sa, _ = self.first_sa(zn, zn, zn, need_weights=False)
kv = z + z_sa # self-checked KV
# ---- second SA: full-T queries x reduced KV -------------------------
q = self.q_proj(self.norm_q(x))
k, v = self.kv_proj(kv).chunk(2, dim=-1)
q = q.reshape(N, T, self.h, self.dh).transpose(1, 2) # (N,h,T,dh)
k = k.reshape(N, -1, self.h, self.dh).transpose(1, 2) # (N,h,T/R,dh)
v = v.reshape(N, -1, self.h, self.dh).transpose(1, 2)
attn = (q @ k.transpose(-2, -1)) * self.scale # (N,h,T,T/R)
attn = self.attn_drop(attn.softmax(dim=-1))
out = (attn @ v).transpose(1, 2).reshape(N, T, d)
x = x + self.proj_drop(self.out_proj(out))
# ---- scMLP -----------------------------------------------------------
y = self.fc1(self.norm_mlp(x))
y = F.gelu(y) + self.dw(y.transpose(1, 2)).transpose(1, 2)
x = x + self.fc2(y)
return x
class TemporalSRDAEncoder(nn.Module):
"""
Signature-compatible replacement for TemporalDaViTEncoder.
forward(x, pos=None): x (N, T, d), pos (N, T, d) date positional encoding.
"""
def __init__(self, d_model=256, n_heads=8, R_schedule=(4, 4, 4), **kwargs):
super().__init__()
# **kwargs silently absorbs legacy windows=/shifts= if passed through
self.R_max = max(R_schedule)
self.blocks = nn.ModuleList(
TemporalSRDABlock(d_model, n_heads, R=r) for r in R_schedule)
def forward(self, x, pos=None):
if pos is not None:
x = x + pos # date PE BEFORE KV reduction
N, T, d = x.shape
pad = (-T) % self.R_max # pad T up to multiple of R_max
if pad:
x = F.pad(x, (0, 0, 0, pad)) # zero-pad — matches pipeline
for blk in self.blocks:
x = blk(x)
if pad:
x = x[:, :T]
return x
|