File size: 4,397 Bytes
acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 0d8aa44 acb2ce3 | 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 | import torch
import torch.nn as nn
from typing import Optional
class CrossTokenShuffle(nn.Module):
"""Replace bag tokens with real-image donor tokens using vectorized tensor operations.
Real bags draw donors only from other real images. Generated bags may draw
from the full real-token pool. min_keep_original limits replacement, and
inserted tokens can be detached. The batch shape is preserved.
"""
def __init__(self):
super().__init__()
@torch.no_grad()
def forward(
self,
x: torch.Tensor, # (B, N, D)
y: torch.Tensor, # (B,) 0/1
replace_ratio: float = 0.5,
with_replacement: bool = True,
detach_inserted: bool = True,
generator: Optional[torch.Generator] = None,
min_keep_original: int = 0,
) -> torch.Tensor:
"""Return augmented tokens with the same shape as x, using the configured donor and replacement rules."""
assert x.dim() == 3, "x must be (B, N, D)"
B, N, D = x.shape
if B == 0 or N == 0:
return x
device = x.device
dtype = x.dtype
K = int(N * replace_ratio)
mk = int(min_keep_original)
if mk > 0:
K = min(K, max(0, N - mk))
if K <= 0:
return x
x_flat = x.reshape(B * N, D)
y = y.to(device).view(B) # (B,)
token_labels = y.repeat_interleave(N) # (B*N,)
real_flat = torch.nonzero(token_labels == 0, as_tuple=False).view(-1)
M = int(real_flat.numel())
if M == 0:
return x.clone()
owner = real_flat // N
b_ar = torch.arange(B, device=device, dtype=torch.long).unsqueeze(1) # (B, 1)
rows_fake = (y == 1).unsqueeze(1) # (B, 1)
allowed_m = rows_fake | (owner.unsqueeze(0) != b_ar) # (B, M)
has_cand = allowed_m.any(dim=1) # (B,)
r_pos = torch.rand(B, N, device=device, generator=generator)
replace_pos_idx = r_pos.argsort(dim=1)[:, :K].long()
cand_flat_idx = torch.zeros(B, K, device=device, dtype=torch.long)
if with_replacement:
rows_with = has_cand
if rows_with.any():
weights = allowed_m[rows_with].float() # (Rw, M)
idx_m = torch.multinomial(
weights, num_samples=K, replacement=True, generator=generator
) # (Rw, K) in [0..M-1]
cand_flat_idx[rows_with] = real_flat[idx_m]
else:
scores = torch.where(
allowed_m,
torch.zeros(B, M, device=device, dtype=torch.float32),
torch.full((B, M), float("-inf"), device=device, dtype=torch.float32),
)
U = torch.rand(B, M, device=device, generator=generator).clamp_(1e-6, 1 - 1e-6)
g = -torch.log(-torch.log(U))
cand_m = (scores + g).topk(k=K, dim=1).indices # (B, K) in [0..M-1]
cand_flat_idx = real_flat[cand_m]
inserted_tokens = x_flat.index_select(0, cand_flat_idx.reshape(-1)).reshape(B, K, D)
if detach_inserted:
inserted_tokens = inserted_tokens.detach()
x_out = x.clone()
if has_cand.any():
rows = torch.where(has_cand)[0]
pos_idx = replace_pos_idx.index_select(0, rows) # (Rw, K)
batch_idx = rows.unsqueeze(1).expand(-1, pos_idx.size(1))
ins = inserted_tokens.index_select(0, rows) # (Rw, K, D)
x_out[batch_idx, pos_idx, :] = ins
return x_out.to(dtype=dtype, device=device)
# -------------------------
# Quick self-check (toy)
# -------------------------
if __name__ == "__main__":
torch.manual_seed(0)
B, N, D = 256, 256, 1024
x = torch.arange(B*N*D, dtype=torch.float32).reshape(B, N, D)
y = torch.tensor([0, 1, 0, 1]*64) # real, fake, real, fake
aug = CrossTokenShuffle()
out = aug(x, y, replace_ratio=0.5, with_replacement=False, detach_inserted=True)
assert out.shape == x.shape
print("OK, shape:", out.shape) |