File size: 11,159 Bytes
40a4412 | 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 | """
LiquidFlow Block — Hybrid CfC + Mamba-2 SSD architecture.
The core innovation: combine Liquid Neural Network dynamics (CfC)
with Mamba-2's efficient linear-time state space model.
Architecture per block:
Input → [CfC Gate → Mamba2 SSD → CfC Gate] → Output
↑ ↑
Adaptive gating Gated output
The CfC provides:
- Time-continuous adaptive gating (what to process/ignore)
- State initialization for the SSM (the "liquid" memory)
The Mamba-2 SSD provides:
- Efficient O(N) sequence processing
- Content-aware selection mechanism
- Parallelizable computation (no sequential bottleneck)
Together they create a "Liquid State Space Model" (LSSM):
h_t = σ(-f(x_t;θ_f)·t) ⊙ SSM(x_t, h_{t-1}) + (1-σ(...)) ⊙ h(x_t;θ_h)
Where SSM is the Mamba-2 selective state space model and the
CfC time-gates control how much the SSM output influences state.
This is inspired by:
- LNNs: adaptive time constants for state evolution
- Mamba-2: efficient selective state space models
- DiMSUM: multi-scan architecture for 2D images
- Gated SSM: gating mechanism from CfC applied to SSM
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from .cfc_cell import CfCCell
from .mamba2_ssd import Mamba2SSD
class LiquidMambaBlock(nn.Module):
"""
LiquidMamba: CfC-gated Mamba-2 SSD block.
The CfC cell acts as a learned gate on the Mamba-2 output,
creating a liquid time-constant mechanism for the SSM:
1. Input goes through Mamba-2 SSD (multi-directional scan)
2. CfC cell receives the SSM output + original input
3. CfC produces a time-gated output: σ(f)·SSM_out + (1-σ(f))·input
4. The CfC's liquid dynamics adaptively mix SSM features with raw input
This creates a "content-aware gating" that the CfC learns to
control based on both the input and the SSM's processed features.
"""
def __init__(self, dim, d_state=16, d_conv=4, expand=2, dropout=0.0):
super().__init__()
self.dim = dim
# LayerNorms
self.norm_in = nn.LayerNorm(dim)
self.norm_mamba = nn.LayerNorm(dim)
self.norm_out = nn.LayerNorm(dim)
# Mamba-2 SSD for efficient sequence processing
self.mamba = Mamba2SSD(dim=dim, d_state=d_state, d_conv=d_conv, expand=expand)
# CfC gate: controls the flow between Mamba output and residual
self.cfc_gate = CfCCell(dim=dim, backbone_dropout=dropout, use_conv=True)
# Feed-forward
ff_dim = dim * expand
self.ff = nn.Sequential(
nn.Linear(dim, ff_dim),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(ff_dim, dim),
nn.Dropout(dropout),
)
# Learnable mixing ratio init
self.gate_scale = nn.Parameter(torch.ones(1) * 0.5)
def forward(self, x):
"""
Args:
x: [B, C, H, W] (2D) or [B, L, C] (1D seq)
Returns:
Same shape as input
"""
is_2d = x.dim() == 4
if is_2d:
B, C, H, W = x.shape
L = H * W
x_flat = x.flatten(2).transpose(1, 2) # [B, HW, C]
else:
B, L, C = x.shape
x_flat = x
residual = x_flat
x_norm = self.norm_in(x_flat)
# Mamba-2 SSD processing with multi-directional scan
if is_2d:
# Reshape for 2D scanning
x_2d = x_norm.transpose(1, 2).reshape(B, C, H, W)
mamba_out = self._mamba_2d_scan(x_2d)
mamba_out = mamba_out.flatten(2).transpose(1, 2) # [B, HW, C]
else:
mamba_out = self.mamba(x_norm)
# CfC gating: liquid dynamics control the mix
mamba_norm = self.norm_mamba(mamba_out)
# CfC receives both the Mamba output and the residual
# This lets it learn when to trust the SSM vs the original signal
cfc_input = mamba_norm + residual
cfc_out = self.cfc_gate(cfc_input)
# Gated mix: CfC controls the blend
gate = torch.sigmoid(self.gate_scale * (cfc_out - mamba_out))
mixed = gate * mamba_out + (1 - gate) * residual + cfc_out
# Feed-forward + residual
out_norm = self.norm_out(mixed)
out = mixed + self.ff(out_norm)
if is_2d:
out = out.transpose(1, 2).reshape(B, C, H, W)
return out
def _mamba_2d_scan(self, x):
"""
Multi-directional Mamba-2 scan for 2D images.
Scans in forward and backward raster directions, then merges.
This preserves 2D spatial structure better than single-direction scan.
"""
B, C, H, W = x.shape
device = x.device
# Forward raster: left→right, top→bottom
fwd = x.flatten(2) # [B, C, HW]
fwd_seq = fwd.transpose(1, 2) # [B, HW, C]
fwd_out = self.mamba(fwd_seq)
# Backward raster: right→left, bottom→top
bwd = torch.flip(x.flatten(2), dims=[-1]) # [B, C, HW]
bwd_seq = bwd.transpose(1, 2)
bwd_out = self.mamba(bwd_seq)
bwd_out = torch.flip(bwd_out, dims=[1]) # Flip back
# Merge both directions
merged = (fwd_out + bwd_out) / 2
merged = merged.transpose(1, 2).reshape(B, C, H, W)
return merged
class LiquidFlowStage(nn.Module):
"""
A stage in LiquidFlow: multiple LiquidMamba blocks at the same resolution.
Architecture:
[LiquidMamba Block] × num_blocks
[Optional Downsample/Upsample]
This mirrors the hierarchical design from DiT/DiMSUM but with
liquid neural network dynamics in every block.
"""
def __init__(self, dim, num_blocks=4, d_state=16, expand=2, dropout=0.0):
super().__init__()
self.dim = dim
self.blocks = nn.ModuleList([
LiquidMambaBlock(dim=dim, d_state=d_state, expand=expand, dropout=dropout)
for _ in range(num_blocks)
])
def forward(self, x):
for block in self.blocks:
x = block(x)
return x
class LiquidFlowBackbone(nn.Module):
"""
Complete LiquidFlow backbone for image generation.
Architecture:
Input (noisy latent) [B, C, H, W]
↓
[Patch Embed + Positional Encoding]
↓
[LiquidMamba Stages × N] (at uniform resolution)
↓
[Output Head] → predicted noise
This is designed as a DiT-style noise predictor for diffusion models.
Args:
in_channels: Input channels (latent dim from VAE)
hidden_dim: Hidden dimension
num_stages: Number of processing stages
blocks_per_stage: Number of blocks per stage
d_state: SSM state dimension
expand: Expansion factor
dropout: Dropout rate
"""
def __init__(
self,
in_channels=4,
hidden_dim=256,
num_stages=4,
blocks_per_stage=4,
d_state=16,
expand=2,
dropout=0.0,
):
super().__init__()
self.in_channels = in_channels
self.hidden_dim = hidden_dim
self.num_stages = num_stages
# Input embedding: patch embedding
self.patch_size = 2 # Fixed patch size
self.in_proj = nn.Conv2d(in_channels, hidden_dim, kernel_size=1)
# Time embedding (for diffusion timestep)
self.time_embed = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim * 4),
nn.SiLU(),
nn.Linear(hidden_dim * 4, hidden_dim),
)
# Learnable positional encoding
# For 128×128 with patch_size=2: 64×64 = 4096 positions
self.pos_embed = nn.Parameter(torch.randn(1, 4096, hidden_dim) * 0.02)
# LiquidFlow stages
self.stages = nn.ModuleList([
LiquidFlowStage(
dim=hidden_dim,
num_blocks=blocks_per_stage,
d_state=d_state,
expand=expand,
dropout=dropout,
)
for _ in range(num_stages)
])
# Output head
self.out_norm = nn.LayerNorm(hidden_dim)
self.out_proj = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.GELU(),
nn.Linear(hidden_dim, in_channels * self.patch_size * self.patch_size),
)
# Timestep conditioner (modulated conv trick)
self.t_conditioner = nn.Sequential(
nn.SiLU(),
nn.Linear(hidden_dim, hidden_dim * 2), # scale, shift
)
def _get_timestep_embedding(self, timesteps, dim, max_period=10000):
"""Sinusoidal timestep embedding (from DiT)."""
half = dim // 2
freqs = torch.exp(
-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
).to(timesteps.device)
args = timesteps.float().unsqueeze(-1) * freqs.unsqueeze(0)
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
if dim % 2:
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
return embedding
def forward(self, x, t):
"""
Args:
x: Noisy latent [B, C, H, W]
t: Diffusion timesteps [B]
Returns:
Predicted noise [B, C, H, W]
"""
B, C, H, W = x.shape
device = x.device
L = (H // self.patch_size) * (W // self.patch_size)
# Input projection
x = self.in_proj(x) # [B, hidden_dim, H, W]
# Flatten and add positional encoding
x_flat = x.flatten(2).transpose(1, 2) # [B, H*W, hidden_dim]
# Time embedding
t_emb = self._get_timestep_embedding(t, self.hidden_dim)
t_emb = self.time_embed(t_emb) # [B, hidden_dim]
# Add time conditioning as bias to input
t_cond = self.t_conditioner(t_emb) # [B, hidden_dim * 2]
t_scale, t_shift = t_cond.chunk(2, dim=-1)
x_flat = x_flat * (1 + t_scale.unsqueeze(1)) + t_shift.unsqueeze(1)
# Add positional encoding
x_flat = x_flat + self.pos_embed[:, :L, :]
# Reshape back to 2D for processing
x_2d = x_flat.transpose(1, 2).reshape(B, self.hidden_dim, H, W)
# Process through all stages
for stage in self.stages:
x_2d = stage(x_2d)
# Output head
x_out = x_2d.flatten(2).transpose(1, 2) # [B, H*W, hidden_dim]
x_out = self.out_norm(x_out)
x_out = self.out_proj(x_out) # [B, H*W, C * patch²]
# Reshape to image
x_out = x_out.reshape(B, H, W, C, self.patch_size, self.patch_size)
x_out = x_out.permute(0, 3, 1, 4, 2, 5).reshape(B, C, H * self.patch_size, W * self.patch_size)
return x_out
import math
|