Create robust_velocity_adapter.py
Browse files- robust_velocity_adapter.py +149 -0
robust_velocity_adapter.py
ADDED
@@ -0,0 +1,149 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
|
5 |
+
import math
|
6 |
+
import torch
|
7 |
+
import torch.nn as nn
|
8 |
+
import torch.nn.functional as F
|
9 |
+
|
10 |
+
class RobustVelocityAdapter(nn.Module):
|
11 |
+
"""
|
12 |
+
Fixed version: manual multi-head cross-attention emits [B, heads, Q, K] scores
|
13 |
+
so that _add_rel_pos_bias can unpack them correctly.
|
14 |
+
"""
|
15 |
+
def __init__(
|
16 |
+
self,
|
17 |
+
t5_dim: int = 512,
|
18 |
+
clip_dim: int = 768,
|
19 |
+
hidden_dim: int = 1024,
|
20 |
+
out_tokens: int = 64, # now aligned with your T5 finetune
|
21 |
+
self_attn_layers: int = 2,
|
22 |
+
cross_heads: int = 8,
|
23 |
+
max_rel_pos: int = 128,
|
24 |
+
):
|
25 |
+
super().__init__()
|
26 |
+
self.out_tokens = out_tokens
|
27 |
+
self.cross_heads = cross_heads
|
28 |
+
self.head_dim = t5_dim // cross_heads
|
29 |
+
self.max_rel_pos = max_rel_pos
|
30 |
+
|
31 |
+
# 1) Self-attention stack
|
32 |
+
self.self_attn = nn.ModuleList()
|
33 |
+
self.self_norm = nn.ModuleList()
|
34 |
+
for _ in range(self_attn_layers):
|
35 |
+
self.self_attn.append(nn.MultiheadAttention(t5_dim, cross_heads, batch_first=True))
|
36 |
+
self.self_norm.append(nn.LayerNorm(t5_dim))
|
37 |
+
|
38 |
+
# 2) Residual blocks
|
39 |
+
def resblock():
|
40 |
+
return nn.Sequential(
|
41 |
+
nn.LayerNorm(t5_dim),
|
42 |
+
nn.Linear(t5_dim, t5_dim),
|
43 |
+
nn.GELU(),
|
44 |
+
nn.Linear(t5_dim, t5_dim),
|
45 |
+
)
|
46 |
+
self.res1 = resblock()
|
47 |
+
self.res2 = resblock()
|
48 |
+
|
49 |
+
# 3) Learned queries for cross-attn
|
50 |
+
self.query_pos = nn.Parameter(torch.randn(out_tokens, t5_dim))
|
51 |
+
|
52 |
+
# 4) Projection heads
|
53 |
+
self.anchor_proj = nn.Sequential(
|
54 |
+
nn.Linear(t5_dim, hidden_dim), nn.ReLU(), nn.Linear(hidden_dim, clip_dim)
|
55 |
+
)
|
56 |
+
self.delta_proj = nn.Sequential(
|
57 |
+
nn.Linear(t5_dim, hidden_dim), nn.ReLU(), nn.Linear(hidden_dim, clip_dim)
|
58 |
+
)
|
59 |
+
self.var_proj = nn.Sequential(
|
60 |
+
nn.Linear(t5_dim, hidden_dim), nn.ReLU(), nn.Linear(hidden_dim, clip_dim)
|
61 |
+
)
|
62 |
+
self.gate_proj = nn.Sequential(
|
63 |
+
nn.Linear(t5_dim, hidden_dim), nn.ReLU(), nn.Linear(hidden_dim, clip_dim), nn.Sigmoid()
|
64 |
+
)
|
65 |
+
|
66 |
+
# 5) Relative-position bias table
|
67 |
+
self.rel_bias = nn.Parameter(torch.zeros(2*max_rel_pos-1, cross_heads))
|
68 |
+
|
69 |
+
# 6) Norm after cross-attn
|
70 |
+
self.cross_norm = nn.LayerNorm(t5_dim)
|
71 |
+
|
72 |
+
def _add_rel_pos_bias(self, attn_scores: torch.Tensor) -> torch.Tensor:
|
73 |
+
"""
|
74 |
+
attn_scores: [B, heads, Q, K]
|
75 |
+
returns: attn_scores + bias where bias is [B, heads, Q, K]
|
76 |
+
"""
|
77 |
+
B, H, Q, K = attn_scores.shape
|
78 |
+
device = attn_scores.device
|
79 |
+
|
80 |
+
# 1) Query & key position indices
|
81 |
+
idx_q = torch.arange(Q, device=device) # [Q]
|
82 |
+
idx_k = torch.arange(K, device=device) # [K]
|
83 |
+
|
84 |
+
# 2) Compute relative distances for every (q, k) pair
|
85 |
+
# rel[i,j] = idx_q[i] - idx_k[j]
|
86 |
+
rel = idx_q.unsqueeze(1) - idx_k.unsqueeze(0) # [Q, K]
|
87 |
+
|
88 |
+
# 3) Clamp & shift into bias table range [0, 2*max_rel-2]
|
89 |
+
max_rel = self.max_rel_pos
|
90 |
+
rel = rel.clamp(-max_rel+1, max_rel-1) + (max_rel - 1)
|
91 |
+
|
92 |
+
# 4) Lookup per-head biases
|
93 |
+
# self.rel_bias has shape [2*max_rel-1, H]
|
94 |
+
bias = self.rel_bias[rel] # [Q, K, H]
|
95 |
+
bias = bias.permute(2, 0, 1) # [H, Q, K]
|
96 |
+
|
97 |
+
# 5) Broadcast to [B, H, Q, K] and add
|
98 |
+
bias = bias.unsqueeze(0).expand(B, -1, -1, -1)
|
99 |
+
return attn_scores + bias
|
100 |
+
|
101 |
+
|
102 |
+
def forward(self, t5_seq: torch.Tensor):
|
103 |
+
"""
|
104 |
+
t5_seq: [B, L, t5_dim]
|
105 |
+
returns:
|
106 |
+
anchor: [B, out_tokens, clip_dim]
|
107 |
+
delta: [B, out_tokens, clip_dim]
|
108 |
+
sigma: [B, out_tokens, clip_dim]
|
109 |
+
"""
|
110 |
+
x = t5_seq
|
111 |
+
B, L, D = x.shape
|
112 |
+
|
113 |
+
# 1) Self-attention + residual
|
114 |
+
for attn, norm in zip(self.self_attn, self.self_norm):
|
115 |
+
res, _ = attn(x, x, x)
|
116 |
+
x = norm(x + res)
|
117 |
+
|
118 |
+
# 2) Residual blocks
|
119 |
+
x = x + self.res1(x)
|
120 |
+
x = x + self.res2(x)
|
121 |
+
|
122 |
+
# 3) Prepare queries & split heads
|
123 |
+
queries = self.query_pos.unsqueeze(0).expand(B, -1, -1) # [B, Q, D]
|
124 |
+
# reshape into heads
|
125 |
+
q = queries.view(B, self.out_tokens, self.cross_heads, self.head_dim).permute(0,2,1,3)
|
126 |
+
k = x.view(B, L, self.cross_heads, self.head_dim).permute(0,2,1,3)
|
127 |
+
v = k
|
128 |
+
|
129 |
+
# 4) Scaled dot-product to get [B, heads, Q, K]
|
130 |
+
scores = (q @ k.transpose(-2,-1)) / math.sqrt(self.head_dim)
|
131 |
+
scores = self._add_rel_pos_bias(scores)
|
132 |
+
probs = F.softmax(scores, dim=-1) # [B, H, Q, K]
|
133 |
+
|
134 |
+
# 5) Attend & merge heads → [B, Q, D]
|
135 |
+
ctx = probs @ v # [B, H, Q, head_dim]
|
136 |
+
ctx = ctx.permute(0,2,1,3).reshape(B, self.out_tokens, D)
|
137 |
+
ctx = self.cross_norm(ctx)
|
138 |
+
|
139 |
+
# 6) Project to anchor, delta_mean, delta_logvar, gate
|
140 |
+
anchor = self.anchor_proj(ctx)
|
141 |
+
delta_mean = self.delta_proj(ctx)
|
142 |
+
delta_logvar = self.var_proj(ctx)
|
143 |
+
gate = self.gate_proj(ctx)
|
144 |
+
|
145 |
+
# 7) Compute sigma & gated delta
|
146 |
+
sigma = torch.exp(0.5 * delta_logvar)
|
147 |
+
delta = delta_mean * gate
|
148 |
+
|
149 |
+
return anchor, delta, sigma
|