File size: 14,612 Bytes
4edc9aa | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 | # %%
# =============================================================================
# CFM 2-D Toy Experiment β self-contained, no src path hacks
# =============================================================================
# Architecture contract (from your code):
# Decoder.forward(x, mu, t)
# x : (B, feat_dim, L)
# mu : (B, feat_dim, L)
# t : (B,) <-- scalar per sample, NOT (B,1,1)
# => out : (B, feat_dim, L)
#
# CFM.compute_loss(x1, mu)
# x1 : (B, feat_dim, L)
# mu : (B, feat_dim, L)
# Inside compute_loss, t is sampled as (B, 1, 1) and passed directly to
# estimator β BUT Decoder.time_emb expects (B,).
# FIX: squeeze t inside Decoder.forward, or patch compute_loss to pass t.squeeze().
# We patch the Decoder forward to handle both (B,) and (B,1,1).
# =============================================================================
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import matplotlib.pyplot as plt
from types import SimpleNamespace
from typing import List, Optional
from abc import ABC, abstractmethod
# ββ helpers ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def sinusoidal_pos_emb(t: torch.Tensor, dim: int) -> torch.Tensor:
"""t: (B,) -> (B, dim)"""
device = t.device
half = dim // 2
freqs = torch.exp(-torch.arange(half, device=device) * (np.log(10000) / (half - 1)))
args = t[:, None] * freqs[None]
return torch.cat([args.sin(), args.cos()], dim=-1)
class SinusoidalPosEmb(nn.Module):
def __init__(self, dim: int):
super().__init__()
self.dim = dim
def forward(self, t: torch.Tensor) -> torch.Tensor:
# accept (B,), (B,1), or (B,1,1) β always return (B, dim)
t = t.view(t.shape[0])
return sinusoidal_pos_emb(t, self.dim)
# ββ MLP block βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# %%
class MLP(nn.Module):
def __init__(self, in_c, hidden_c, out_c, time_emb_dim):
super().__init__()
self.time_net = nn.Sequential(nn.Linear(time_emb_dim, hidden_c), nn.Mish())
self.net1 = nn.Sequential(nn.Conv1d(in_c, hidden_c, 1), nn.ReLU())
self.net2 = nn.Sequential(nn.Conv1d(hidden_c, hidden_c, 1), nn.ReLU())
self.net3 = nn.Sequential(nn.Conv1d(hidden_c, hidden_c, 1), nn.ReLU())
self.out = nn.Conv1d(hidden_c, out_c, 1)
def forward(self, x, time_emb):
h = self.net1(x)
h = h + self.time_net(time_emb).unsqueeze(-1)
h = self.net2(h)
h = self.net3(h)
return self.out(h)
# class MLP(nn.Module):
# def __init__(self, in_c: int, hidden_c: int, out_c: int, time_emb_dim: int):
# super().__init__()
# self.time_net = nn.Sequential(nn.Linear(time_emb_dim, hidden_c), nn.Mish())
# self.net1 = nn.Sequential(nn.Linear(in_c, hidden_c), nn.ReLU())
# self.net2 = nn.Linear(hidden_c, out_c)
# def forward(self, x: torch.Tensor, time_emb: torch.Tensor) -> torch.Tensor:
# # x : (B, in_c, L)
# # time_emb : (B, time_emb_dim)
# x_t = x.transpose(1, 2) # (B, L, in_c) for Linear
# out = self.net1(x_t) # (B, L, hidden_c)
# out = out + self.time_net(time_emb).unsqueeze(1) # broadcast over L
# out = self.net2(out) # (B, L, out_c)
# return out.transpose(1, 2) # (B, out_c, L)
# %%
# ββ Decoder βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class Decoder(nn.Module):
"""
Lightweight MLP velocity estimator for toy 2-D flow-matching.
Tensor contract
---------------
forward(x, mu, t) -> vel
x : (B, feat_dim, L)
mu : (B, feat_dim, L)
t : (B,) | (B,1) | (B,1,1) # all accepted
vel : (B, feat_dim, L)
"""
def __init__(
self,
in_c: int = 2,
hidden_dim: int = 128,
out_c: int = 2,
time_emb_dim: int = 64,
cond_dim: int = 0,
):
super().__init__()
self.time_emb = SinusoidalPosEmb(time_emb_dim)
self.time_mlp = nn.Sequential(
nn.Linear(time_emb_dim, time_emb_dim),
)
# concat(x, mu) along channel dim -> 2*feat_dim channels
self.net = MLP(
in_c=in_c * 2, hidden_c=hidden_dim, out_c=out_c, time_emb_dim=time_emb_dim
)
self._init_weights()
def _init_weights(self):
for m in self.modules():
if isinstance(m, nn.Linear):
nn.init.normal_(m.weight, 0.0, 0.02)
if m.bias is not None:
nn.init.zeros_(m.bias)
def forward(
self,
x: torch.Tensor,
mu: torch.Tensor,
t: torch.Tensor,
cond=None,
) -> torch.Tensor:
# normalise t to (B,) regardless of input shape
t_flat = t.reshape(x.shape[0]) # (B,)
t_emb = self.time_mlp(self.time_emb(t_flat)) # (B, time_emb_dim)
# concat along channel axis (B, 2*feat_dim, L)
xmu = torch.cat([x, mu], dim=1)
return self.net(xmu, t_emb) # (B, feat_dim, L)
# -- SourceGenerator
class SourceGenerator(nn.Module):
def __init__(self, feat_dim: int, hidden_dim: int = 64):
super().__init__()
# Outputs 2 * feat_dim to hold both mean and log_var
self.net = nn.Sequential(
nn.Conv1d(feat_dim, hidden_dim, 1),
nn.Mish(),
nn.Conv1d(hidden_dim, feat_dim * 2, 1),
)
def forward(self, mu: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
# mu: (B, feat_dim, L)
out = self.net(mu) # (B, 2*feat_dim, L)
mean_c, logvar_c = out.chunk(2, dim=1) # each (B, feat_dim, L)
return mean_c, logvar_c
# ββ BASECFM βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class BASECFM(nn.Module, ABC):
def __init__(self, feat_dim: int, cfm_params):
super().__init__()
self.feat_dim = feat_dim
self.sigma_min = cfm_params.sigma_min
self.estimator: Optional[nn.Module] = None
self.src_gen: Optional[nn.Module] = None
# ---- inference -----------------------------------------------------------
@torch.inference_mode()
def forward(
self,
mu: torch.Tensor, # (B, feat_dim, L)
n_timesteps: int,
temperature: float = 1.0,
) -> torch.Tensor:
z = self.src_gen(mu) * temperature
t_span = torch.linspace(0, 1, n_timesteps + 1, device=mu.device)
return self.solve_euler(z, t_span, mu)
def solve_euler(
self,
x: torch.Tensor, # (B, feat_dim, L)
t_span: torch.Tensor, # (n_steps+1,)
mu: torch.Tensor, # (B, feat_dim, L)
) -> torch.Tensor:
t = t_span[0]
dt = t_span[1] - t_span[0]
B = x.shape[0]
for step in range(1, len(t_span)):
t_batch = t.expand(B, device=device) # (B,)
dphi_dt = self.estimator(x, mu, t_batch)
x = x + dt * dphi_dt
t = t + dt
if step < len(t_span) - 1:
dt = t_span[step + 1] - t
return x
# ---- training ------------------------------------------------------------
def compute_loss(
self,
x1: torch.Tensor, # (B, feat_dim, L)
mu: torch.Tensor, # (B, feat_dim, L)
lambda_var: float = 1, # Hyperparameters from the paper
lambda_align: float = 0,
) -> tuple:
B = x1.shape[0]
# t sampled per sample, broadcast-ready for interpolation
t = torch.rand(B, 1, 1, device=mu.device, dtype=mu.dtype) # (B,1,1)
# z = torch.randn_like(mu) # (B, C, L)
mean_c, logvar_c = self.src_gen(mu) # (B, C, L)
eps = torch.randn_like(mean_c)
z = mean_c + torch.exp(0.5 * logvar_c) * eps
y = (1 - (1 - self.sigma_min) * t) * z + t * x1 # interpolant
u = x1 - (1 - self.sigma_min) * z # target velocity
# estimator expects t as (B,)
t_batch = t.reshape(B)
pred = self.estimator(y, mu, t_batch)
# 4. Standard Flow Matching Loss
loss_fm = F.mse_loss(pred, u)
# 5. Variance Regularization Loss [Eq. 9 in paper]
# D_KL( N(mu_c, sigma_c^2) || N(mu_c, I) ) = 0.5 * (sigma^2 - log(sigma^2) - 1)
loss_var = 0.5 * (torch.exp(logvar_c) - logvar_c - 1).mean()
# 6. Cosine Alignment Loss [Eq. 10 in paper]
sim = F.cosine_similarity(z.flatten(1), x1.flatten(1), dim=1)
loss_align = (1.0 - sim).mean()
# 7. Total Loss [Eq. 11 in paper]
loss_total = loss_fm + lambda_var * loss_var + lambda_align * loss_align
# Return total loss, and a dictionary for logging
loss_dict = {
"fm": loss_fm.item(),
"var": loss_var.item(),
"align": loss_align.item(),
}
return loss_total, loss_dict
class CFM(BASECFM):
def __init__(
self, feat_dim: int, cfm_params, decoder_params: dict, num_classes: int = 8
):
super().__init__(feat_dim=feat_dim, cfm_params=cfm_params)
self.estimator = Decoder(in_c=feat_dim, out_c=feat_dim, **decoder_params)
self.label_emb = nn.Embedding(num_classes, feat_dim)
self.src_gen = SourceGenerator(feat_dim=feat_dim)
# %%
# =============================================================================
# Experiment: Gaussian -> 8-Gaussians
# =============================================================================
np.random.seed(42)
torch.manual_seed(42)
# ---- GPU setup ------
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")
n_samples = 4000
scale = 4.0
centers = np.array(
[
(np.cos(t) * scale, np.sin(t) * scale)
for t in np.linspace(0, 2 * np.pi, 8, endpoint=False)
]
)
assignments = np.random.randint(0, 8, size=n_samples)
gaussians_x = centers[assignments] + np.random.randn(n_samples, 2) * 0.4
target_tensor = torch.tensor(gaussians_x, dtype=torch.float32, device=device)
goal_dist = (target_tensor - target_tensor.mean(0)) / target_tensor.std(0)
# ---- build model ------------------------------------------------------------
cfm_params = SimpleNamespace(sigma_min=1e-4, solver="euler")
decoder_params = dict(hidden_dim=256, time_emb_dim=128, cond_dim=0)
model = CFM(feat_dim=2, cfm_params=cfm_params, decoder_params=decoder_params).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
# ---- training loop ----------------------------------------------------------
epochs, batch_size = 3000, 512
losses = []
model.train()
for epoch in range(epochs):
idx = torch.randint(0, n_samples, (batch_size,))
x1 = goal_dist[idx].unsqueeze(-1) # (B, 2, 1)
# Conditional -> cluster embedding conditioning
labels = torch.tensor(assignments[idx], dtype=torch.long, device=device)
mu = model.label_emb(labels).unsqueeze(-1)
loss, loss_dict = model.compute_loss(x1, mu)
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses.append(loss.item())
if (epoch + 1) % 1000 == 0:
print(
f"Epoch {epoch+1:5d} loss={loss.item():.5f} | "
f"FM={loss_dict['fm']:.5f} | "
f"Var={loss_dict['var']:.5f} | "
f"Align={loss_dict['align']:.5f}"
)
# %%
# ---- inference -------------------------------------------------------------
model.eval()
n_eval = 1000
eval_labels = torch.arange(8, device=device).repeat_interleave(n_eval // 8 + 1)[
:n_eval
] # TODO: investigate
mu_eval = model.label_emb(eval_labels).unsqueeze(-1).detach()
steps = 100
t_span = torch.linspace(0, 1, steps + 1, device=device)
trajectories = []
with torch.no_grad():
x = torch.randn(mu_eval.size(), device=device)
trajectories.append(x.squeeze(-1).cpu().numpy().copy())
t = t_span[0]
dt = t_span[1] - t_span[0]
snap_at = {0, 20, 40, 60, 80, 100}
for step in range(1, len(t_span)):
t_batch = t.expand(n_eval)
dphi_dt = model.estimator(x, mu_eval, t_batch)
x = x + dt * dphi_dt
t = t + dt
if step < len(t_span) - 1:
dt = t_span[step + 1] - t
if step in snap_at:
trajectories.append(x.squeeze(-1).cpu().numpy().copy())
print(x.max(), " -- ", x.min())
# ---- plot ------------------------------------------------------------------
fig, axes = plt.subplots(1, 7, figsize=(21, 3))
fig.suptitle(
"OT-CFM: Gaussian β 8 Gaussians (conditional on cluster label)",
fontsize=13,
y=1.04,
)
times = [0.0, 0.2, 0.4, 0.6, 0.8, 1.0, "target"]
colors = ["#636EFA", "#7A89FB", "#9BA4FC", "#BCBFFD", "#DDDAFE", "#EF553B", "#00CC96"]
for ax, traj, label, c in zip(axes, trajectories, times, colors):
ax.scatter(traj[:, 0], traj[:, 1], s=4, alpha=0.6, color=c, linewidths=0)
ax.set_xlim(-3.5, 3.5)
ax.set_ylim(-3.5, 3.5)
ax.set_xlabel("X", fontsize=9)
ax.set_ylabel("Y", fontsize=9)
ax.set_title(f"t = {label}" if isinstance(label, float) else label, fontsize=10)
ax.axis("off")
# last panel: overlay ground-truth
gt = goal_dist[:1000].cpu().numpy()
axes[-1].scatter(gt[:, 0], gt[:, 1], s=4, alpha=0.3, color="#00CC96", linewidths=0)
axes[-1].set_xlim(-3.5, 3.5)
axes[-1].set_ylim(-3.5, 3.5)
axes[-1].set_xlabel("X", fontsize=9)
axes[-1].set_ylabel("Y", fontsize=9)
axes[-1].set_title("target", fontsize=10)
axes[-1].axis("off")
# loss curve panel
fig2, ax2 = plt.subplots(figsize=(7, 3))
ax2.plot(
np.convolve(losses, np.ones(50) / 50, mode="valid"), linewidth=1.2, color="#636EFA"
)
ax2.set_xlabel("Epoch")
ax2.set_ylabel("MSE Loss")
ax2.set_title("CFM Training Loss (50-epoch moving avg)")
ax2.spines[["top", "right"]].set_visible(False)
plt.tight_layout()
fig.savefig("cfm_trajectories.png", dpi=130, bbox_inches="tight")
fig2.savefig("cfm_loss.png", dpi=130, bbox_inches="tight")
print("Saved cfm_trajectories.png and cfm_loss.png")
# %%
from torchinfo import summary
print(summary(model))
# %%
print(goal_dist.max(), goal_dist.min())
# %%
|