File size: 10,752 Bytes
a6b4c35 | 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 | #!/usr/bin/env python
"""
mamba_model.py
Mamba-DLoc: Replaces the ResNet bottleneck in DLoc with Mamba selective
state-space (S6) blocks for efficient long-range sequence modelling.
Architecture overview
---------------------
CNN Stem (4 -> 128 channels, 2x spatial downsample)
-> Tokenize (flatten spatial -> sequence)
-> 4x MambaResidualBlock (LayerNorm -> Mamba -> residual)
-> Reshape back to spatial
-> CNN Decoder (upsample back to original resolution)
Input: [B, 4, 161, 361] AoA heatmaps
Output: [B, 1, 161, 361] location heatmap
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
# ---------------------------------------------------------------------------
# Try to import the CUDA-optimised Mamba implementation (Linux + CUDA only).
# Fall back to a pure-PyTorch implementation on Windows / CPU.
# ---------------------------------------------------------------------------
HAS_MAMBA_SSM = False
try:
from mamba_ssm import Mamba as CudaMamba # type: ignore
HAS_MAMBA_SSM = True
except ImportError:
CudaMamba = None
# ===================================================================
# Pure-PyTorch Mamba (S6 selective scan) - works on any backend
# ===================================================================
class PurePytorchMamba(nn.Module):
"""
Minimal re-implementation of the Mamba selective-scan block (S6).
Interface: [B, L, D] -> [B, L, D]
Parameters
----------
d_model : int - input / output dimension (D)
d_state : int - SSM hidden state dimension (N)
d_conv : int - local convolution kernel width
expand : int - expansion factor for inner dimension (E = expand * D)
"""
def __init__(self, d_model: int = 128, d_state: int = 16,
d_conv: int = 4, expand: int = 2):
super().__init__()
self.d_model = d_model
self.d_state = d_state
self.d_conv = d_conv
self.expand = expand
self.d_inner = d_model * expand # E
self.dt_rank = math.ceil(d_model / 16) # rank for delta-t projection
# --- input projection: D -> 2*E (main path + gate) ---
self.in_proj = nn.Linear(d_model, 2 * self.d_inner, bias=False)
# --- causal depthwise conv on main path ---
self.conv1d = nn.Conv1d(
self.d_inner, self.d_inner,
kernel_size=d_conv, padding=d_conv - 1, # causal: trim later
groups=self.d_inner, bias=True,
)
# --- SSM parameter projections ---
# from the convolved main path (E) -> dt_rank + 2*N
self.x_proj = nn.Linear(self.d_inner, self.dt_rank + 2 * d_state, bias=False)
# dt_rank -> E (broadcast delta across inner dim)
self.dt_proj = nn.Linear(self.dt_rank, self.d_inner, bias=True)
# --- learnable SSM parameters ---
# A_log: [E, N] initialised as log(1..N) broadcast to each inner dim
A = torch.arange(1, d_state + 1, dtype=torch.float32).unsqueeze(0).expand(self.d_inner, -1)
self.A_log = nn.Parameter(torch.log(A)) # learned in log-space
# D (skip connection scalar per inner dim)
self.D = nn.Parameter(torch.ones(self.d_inner))
# --- output projection: E -> D ---
self.out_proj = nn.Linear(self.d_inner, d_model, bias=False)
# ---------------------------------------------------------------
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
x : [B, L, D]
returns : [B, L, D]
"""
B, L, _ = x.shape
# --- input projection & split ---
xz = self.in_proj(x) # [B, L, 2E]
x_main, z = xz.chunk(2, dim=-1) # each [B, L, E]
# --- causal conv1d on main path ---
# conv1d expects [B, E, L]
x_main = x_main.transpose(1, 2) # [B, E, L]
x_main = self.conv1d(x_main)[:, :, :L] # causal trim
x_main = x_main.transpose(1, 2) # [B, L, E]
x_main = F.silu(x_main)
# --- SSM parameters ---
x_ssm = self.x_proj(x_main) # [B, L, dt_rank + 2N]
dt, B_param, C_param = x_ssm.split(
[self.dt_rank, self.d_state, self.d_state], dim=-1
)
# dt: [B, L, dt_rank] -> [B, L, E]
dt = F.softplus(self.dt_proj(dt)) # [B, L, E]
# A: [E, N] (negative for stability)
A = -torch.exp(self.A_log) # [E, N]
# --- discretize ---
# dA: [B, L, E, N]
dA = torch.exp(dt.unsqueeze(-1) * A.unsqueeze(0).unsqueeze(0))
# dB: [B, L, E, N]
dB = dt.unsqueeze(-1) * B_param.unsqueeze(2) # broadcast E
# --- selective scan (sequential, kept simple for correctness) ---
# x_main: [B, L, E] -> weight input
# h: [B, E, N] SSM hidden state
h = torch.zeros(B, self.d_inner, self.d_state,
device=x.device, dtype=x.dtype)
ys = []
for t in range(L):
h = dA[:, t, :, :] * h + dB[:, t, :, :] * x_main[:, t, :].unsqueeze(-1)
y_t = (h * C_param[:, t, :].unsqueeze(1)).sum(dim=-1) # [B, E]
ys.append(y_t)
y = torch.stack(ys, dim=1) # [B, L, E]
# --- skip connection ---
y = y + x_main * self.D.unsqueeze(0).unsqueeze(0)
# --- gate ---
y = y * F.silu(z)
# --- output projection ---
return self.out_proj(y) # [B, L, D]
# ===================================================================
# Mamba residual block: LayerNorm -> Mamba -> residual add
# ===================================================================
class MambaResidualBlock(nn.Module):
"""
Pre-norm residual block wrapping a Mamba layer.
Interface: [B, L, D] -> [B, L, D]
Uses the CUDA Mamba kernel when available, otherwise falls back to
PurePytorchMamba.
"""
def __init__(self, d_model: int = 128, d_state: int = 16,
d_conv: int = 4, expand: int = 2):
super().__init__()
self.norm = nn.LayerNorm(d_model)
if HAS_MAMBA_SSM and CudaMamba is not None:
self.mamba = CudaMamba(
d_model=d_model, d_state=d_state,
d_conv=d_conv, expand=expand,
)
else:
self.mamba = PurePytorchMamba(
d_model=d_model, d_state=d_state,
d_conv=d_conv, expand=expand,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return x + self.mamba(self.norm(x))
# ===================================================================
# MambaDLocNet - full end-to-end model
# ===================================================================
class MambaDLocNet(nn.Module):
"""
Mamba-DLoc: CNN stem -> Mamba sequence blocks -> CNN decoder.
Input: [B, 4, 161, 361]
Output: [B, 1, 161, 361]
"""
# Spatial sizes at each resolution level (full, mid, low)
SPATIAL_SIZES = [(161, 361), (81, 181), (41, 91)]
def __init__(self, input_nc: int = 4, d_model: int = 128,
n_mamba_blocks: int = 4,
d_state: int = 16, d_conv: int = 4, expand: int = 2):
super().__init__()
self.d_model = d_model
# ------ CNN Stem (encoder) ------
self.stem = nn.Sequential(
# [B,input_nc,161,361] -> [B,32,161,361]
nn.Conv2d(input_nc, 32, kernel_size=3, stride=1, padding=1, bias=False),
nn.BatchNorm2d(32),
nn.ReLU(inplace=True),
# [B,32,161,361] -> [B,64,81,181]
nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
# [B,64,81,181] -> [B,128,41,91]
nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
)
# ------ Mamba sequence blocks ------
self.mamba_blocks = nn.Sequential(*[
MambaResidualBlock(d_model=d_model, d_state=d_state,
d_conv=d_conv, expand=expand)
for _ in range(n_mamba_blocks)
])
# ------ CNN Decoder ------
# Stage 1: [B,128,41,91] -> interpolate to (81,181) -> conv 128->64
self.dec_conv1 = nn.Sequential(
nn.Conv2d(128, 64, kernel_size=3, stride=1, padding=1, bias=False),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
)
# Stage 2: [B,64,81,181] -> interpolate to (161,361) -> conv 64->32
self.dec_conv2 = nn.Sequential(
nn.Conv2d(64, 32, kernel_size=3, stride=1, padding=1, bias=False),
nn.BatchNorm2d(32),
nn.ReLU(inplace=True),
)
# Stage 3: 1x1 head -> sigmoid
self.dec_head = nn.Sequential(
nn.Conv2d(32, 1, kernel_size=1, bias=True),
nn.Sigmoid(),
)
# ---------------------------------------------------------------
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
x : [B, 4, 161, 361]
returns : [B, 1, 161, 361]
"""
B = x.size(0)
# --- CNN Stem ---
feat = self.stem(x) # [B, 128, 41, 91]
# --- Tokenize: flatten spatial dims into sequence ---
_, C, H, W = feat.shape
tokens = feat.reshape(B, C, H * W).permute(0, 2, 1) # [B, 3731, 128]
# --- Mamba blocks ---
tokens = self.mamba_blocks(tokens) # [B, 3731, 128]
# --- Reshape back to spatial ---
feat = tokens.permute(0, 2, 1).reshape(B, C, H, W) # [B, 128, 41, 91]
# --- CNN Decoder ---
feat = F.interpolate(feat, size=self.SPATIAL_SIZES[1],
mode='bilinear', align_corners=False) # [B,128,81,181]
feat = self.dec_conv1(feat) # [B,64,81,181]
feat = F.interpolate(feat, size=self.SPATIAL_SIZES[0],
mode='bilinear', align_corners=False) # [B,64,161,361]
feat = self.dec_conv2(feat) # [B,32,161,361]
out = self.dec_head(feat) # [B,1,161,361]
return out
|