File size: 10,635 Bytes
2d1810a | 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 | """Plot the learned Monge maps for the cached grid pair and FCOT-Separable models."""
from __future__ import annotations
import argparse
import sys
from pathlib import Path
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.lines import Line2D
import numpy as np
import torch
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
from optimal_transport.ot_fc_sep_map import FCOTSeparable
from tools.utils import (
L22_1d,
nL22_1d,
inverse_L22x,
inverse_nL22x,
)
DIM = 2
RADIUS = 4.0
X_ACCURACY = 2e-3
Y_ACCURACY = 2e-3
NY = int((2 * RADIUS) / Y_ACCURACY) + 1
N_PARAMS = NY * DIM
TITLE_FONT_SIZE = 20
AXIS_LABEL_FONT_SIZE = 14
FIG_SIZE = (18, 6.5)
LAYOUT_PAD = 1.0
OUTER_LR = 1e-2
TEMP_MIN = 1.0
TEMP_MAX = 60.0
TEMP_WARMUP_ITERS = 2_500
REACTIVATE_EVERY = 50
REACTIVATE_EPS = 1e-3
FULL_REFRESH_EVERY = 300
COARSE_X_FACTOR = 100
COARSE_TOP_K = 4
COARSE_WINDOW = 1
KERNEL_CONFIGS = {
"L22_1d": {
"kernel": L22_1d,
"inverse": inverse_L22x,
"label": "L22",
},
"nL22_1d": {
"kernel": nL22_1d,
"inverse": inverse_nL22x,
"label": "nL22",
},
}
def _latest_file(directory: Path, pattern: str) -> Path:
candidates = sorted(directory.glob(pattern))
if not candidates:
raise FileNotFoundError(f"No files matching {pattern} under {directory}")
return max(candidates, key=lambda p: p.stat().st_mtime)
def load_grid_XY(tmp_dir: Path) -> tuple[torch.Tensor, torch.Tensor]:
"""Load the cached X, Y pair used for the 2D grid experiments."""
grid_x_dir = tmp_dir / "grid_X"
grid_y_dir = tmp_dir / "grid_Y"
if not grid_x_dir.is_dir() or not grid_y_dir.is_dir():
raise FileNotFoundError("Grid directories tmp/grid_X or tmp/grid_Y are missing.")
latest_x = _latest_file(grid_x_dir, "*.npz")
latest_y = _latest_file(grid_y_dir, "*.npz")
X = torch.load(latest_x, map_location="cpu")["x"].float().clamp(-RADIUS, RADIUS)
Y_vert = torch.load(latest_y, map_location="cpu")["x"].float().clamp(-RADIUS, RADIUS)
Y = Y_vert[:, [1, 0]]
return X, Y
def build_solver(kernel_tag: str) -> FCOTSeparable:
"""Construct the FCOT-Separable solver architecture for the given kernel."""
config = KERNEL_CONFIGS[kernel_tag]
solver = FCOTSeparable.initialize_right_architecture(
dim=DIM,
radius=RADIUS,
n_params=N_PARAMS,
x_accuracy=X_ACCURACY,
kernel_1d=config["kernel"],
inverse_kx=config["inverse"],
outer_lr=OUTER_LR,
temp_min=TEMP_MIN,
temp_max=TEMP_MAX,
temp_warmup_iters=TEMP_WARMUP_ITERS,
reactivate_every=REACTIVATE_EVERY,
reactivate_eps=REACTIVATE_EPS,
full_refresh_every=FULL_REFRESH_EVERY,
cache_gradients=True,
coarse_x_factor=COARSE_X_FACTOR,
coarse_top_k=COARSE_TOP_K,
coarse_window=COARSE_WINDOW,
)
return solver
def load_solver_from_checkpoint(tmp_dir: Path, kernel_tag: str) -> tuple[FCOTSeparable, Path]:
"""Load the trained model checkpoint for the provided kernel tag."""
pattern = f"FCOTSeparable_dim{DIM}_kernel-{kernel_tag}_*.pt"
checkpoint = _latest_file(tmp_dir, pattern)
solver = build_solver(kernel_tag)
solver.load(checkpoint)
return solver, checkpoint
def compute_transport(solver: FCOTSeparable, X: torch.Tensor) -> torch.Tensor:
"""Apply the fitted Monge map to the input grid."""
return solver.transport_X_to_Y(X, selection_mode="soft", snap_to_grid=False).cpu()
def _tensor_to_numpy(tensor: torch.Tensor) -> np.ndarray:
"""Detach tensor, move to CPU, and return NumPy view."""
return tensor.detach().cpu().numpy()
def _choose_indices(total: int, count: int) -> torch.Tensor:
"""Return up to `count` unique indices sampled from `[0, total)`."""
if count <= 0 or total == 0:
return torch.empty(0, dtype=torch.long)
if count >= total:
return torch.arange(total, dtype=torch.long)
return torch.randperm(total)[:count]
def plot_monge_map(
X: torch.Tensor,
Y: torch.Tensor,
T_L22: torch.Tensor,
T_nL22: torch.Tensor,
output_path: Path,
subsample_l22: int,
subsample_nl22: int,
) -> None:
"""Scatter the input/target data and overlay the learned transports."""
X_np = _tensor_to_numpy(X)
Y_np = _tensor_to_numpy(Y)
T_L22_np = _tensor_to_numpy(T_L22)
T_nL22_np = _tensor_to_numpy(T_nL22)
size = 2
markersize = 4.
markerscale = 2.
colors = {
"X": "blue",
"Y": "red",
"T": "yellow",
}
fig, axes = plt.subplots(1, 3, figsize=FIG_SIZE)
axes = axes.tolist()
axes[0].scatter(
X_np[:, 0], X_np[:, 1], s=size, alpha=0.6,
color=colors["X"],
)
axes[0].scatter(
Y_np[:, 0], Y_np[:, 1], s=size, alpha=0.6,
color=colors["Y"],
)
axes[0].set_title("X and Y marginals", fontsize=TITLE_FONT_SIZE)
axes[0].set_aspect("equal", "box")
legend_handles_0 = [
Line2D(
[0], [0], marker="o", color="black", markerfacecolor=colors["X"],
markeredgecolor="black", linewidth=0, markersize=markersize, label="X"
),
Line2D(
[0], [0], marker="o", color="black", markerfacecolor=colors["Y"],
markeredgecolor="black", linewidth=0, markersize=markersize, label="Y"
),
]
axes[0].legend(handles=legend_handles_0, fontsize=12, loc="lower left", markerscale=markerscale)
axes[0].set_xlabel("Dimension 1", fontsize=AXIS_LABEL_FONT_SIZE)
axes[0].set_ylabel("Dimension 2", fontsize=AXIS_LABEL_FONT_SIZE)
axes[1].scatter(
X_np[:, 0], X_np[:, 1], s=size, alpha=0.6,
color=colors["X"],
)
axes[1].scatter(
Y_np[:, 0], Y_np[:, 1], s=size, alpha=0.6,
color=colors["Y"],
)
axes[1].scatter(
T_L22_np[:, 0], T_L22_np[:, 1], s=size, alpha=0.6,
color=colors["T"],
)
axes[1].set_title("Monge map for $c(x,y)=||x-y||_2^2$", fontsize=TITLE_FONT_SIZE)
axes[1].set_aspect("equal", "box")
axes[1].set_xlabel("Dimension 1", fontsize=AXIS_LABEL_FONT_SIZE)
axes[1].set_ylabel("Dimension 2", fontsize=AXIS_LABEL_FONT_SIZE)
indices_l22 = _choose_indices(X.shape[0], subsample_l22)
for idx in indices_l22.tolist():
axes[1].plot(
[X_np[idx, 0], T_L22_np[idx, 0]],
[X_np[idx, 1], T_L22_np[idx, 1]],
color="black",
linewidth=0.5,
alpha=0.9,
)
legend_handles_1 = [
Line2D(
[0], [0], marker="o", color="black", markerfacecolor=colors["X"],
markeredgecolor="black", linewidth=0, markersize=markersize, label="X"
),
Line2D(
[0], [0], marker="o", color="black", markerfacecolor=colors["Y"],
markeredgecolor="black", linewidth=0, markersize=markersize, label="Y"
),
Line2D(
[0], [0], marker="o", color="black", markerfacecolor=colors["T"],
markeredgecolor="black", linewidth=0, markersize=markersize, label="T(X)"
),
Line2D([0], [0], color="black", linewidth=0.9, label="X--T(X)"),
]
axes[1].legend(handles=legend_handles_1, fontsize=12, loc="lower left", markerscale=markerscale)
axes[2].scatter(
X_np[:, 0], X_np[:, 1], s=size, alpha=0.6,
color=colors["X"],
)
axes[2].scatter(
Y_np[:, 0], Y_np[:, 1], s=size, alpha=0.6,
color=colors["Y"],
)
axes[2].scatter(
T_nL22_np[:, 0], T_nL22_np[:, 1], s=size, alpha=0.6,
color=colors["T"],
)
axes[2].set_title("Monge map for $c(x,y) = -||x-y||_2^2$", fontsize=TITLE_FONT_SIZE)
axes[2].set_aspect("equal", "box")
axes[2].set_xlabel("Dimension 1", fontsize=AXIS_LABEL_FONT_SIZE)
axes[2].set_ylabel("Dimension 2", fontsize=AXIS_LABEL_FONT_SIZE)
indices_nl22 = _choose_indices(X.shape[0], subsample_nl22)
for idx in indices_nl22.tolist():
axes[2].plot(
[X_np[idx, 0], T_nL22_np[idx, 0]],
[X_np[idx, 1], T_nL22_np[idx, 1]],
color="black",
linewidth=1,
alpha=0.6,
)
legend_handles_2 = [
Line2D(
[0], [0], marker="o", color="black", markerfacecolor=colors["X"],
markeredgecolor="black", linewidth=0, markersize=markersize, label="X"
),
Line2D(
[0], [0], marker="o", color="black", markerfacecolor=colors["Y"],
markeredgecolor="black", linewidth=0, markersize=markersize, label="Y"
),
Line2D(
[0], [0], marker="o", color="black", markerfacecolor=colors["T"],
markeredgecolor="black", linewidth=0, markersize=markersize, label="T(X)"
),
Line2D([0], [0], color="black", linewidth=0.9, label="X--T(X)"),
]
axes[2].legend(handles=legend_handles_2, fontsize=12, loc="lower left", markerscale=markerscale)
fig.tight_layout(pad=LAYOUT_PAD)
output_path.parent.mkdir(parents=True, exist_ok=True)
fig.savefig(output_path, dpi=150)
print(f"Saved Monge map figure to {output_path}")
def main() -> None:
parser = argparse.ArgumentParser(description="Plot Monge map for cached grid + FCOT-Separable models.")
parser.add_argument("--tmp-dir", type=Path, default=Path("tmp"), help="Directory containing cached data/models.")
parser.add_argument("--output", type=Path, default=None, help="Optional path for the figure.")
parser.add_argument("--subsample-l22", type=int, default=2000, help="Number of transports in the L22 panel to connect with black lines.")
parser.add_argument("--subsample-nl22", type=int, default=100, help="Number of transports in the nL22 panel to connect with black lines.")
args = parser.parse_args()
tmp_dir = args.tmp_dir
if not tmp_dir.is_dir():
raise FileNotFoundError(f"{tmp_dir} does not exist or is not a directory.")
X, Y = load_grid_XY(tmp_dir)
solver_L22, ckpt_L22 = load_solver_from_checkpoint(tmp_dir, "L22_1d")
solver_nL22, ckpt_nL22 = load_solver_from_checkpoint(tmp_dir, "nL22_1d")
print(f"Loaded checkpoints:\n L22: {ckpt_L22}\n nL22: {ckpt_nL22}")
T_L22 = compute_transport(solver_L22, X)
T_nL22 = compute_transport(solver_nL22, X)
output_path = args.output or tmp_dir / "monge_plots" / "monge_map.png"
plot_monge_map(
X,
Y,
T_L22,
T_nL22,
output_path,
args.subsample_l22,
args.subsample_nl22,
)
if __name__ == "__main__":
main()
|