File size: 12,375 Bytes
112139c | 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 | #!/usr/bin/env python3
"""Causal global/high branch ablation for a RenderMe INS checkpoint."""
from __future__ import annotations
import argparse
import json
import sys
from contextlib import contextmanager
from pathlib import Path
from types import SimpleNamespace
from typing import Iterator, Optional
import torch
import torch.nn.functional as F
from torch import Tensor
from torchvision.utils import make_grid, save_image
REPOSITORY_ROOT = Path(__file__).resolve().parents[1]
if str(REPOSITORY_ROOT) not in sys.path:
sys.path.insert(0, str(REPOSITORY_ROOT))
from datasets.renderme_ins_dataset import RenderMeINSViewDataset
from scripts.evaluate_ins_stage1_renderme import (
build_model_from_checkpoint,
foreground_metrics,
read_checkpoint,
)
from scripts.train_ins_stage1 import (
DistributedContext,
build_auxiliary_losses,
resize_target,
)
CONDITIONS = ("baseline", "global_zero", "high_zero")
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description="Render baseline/global=0/high=0 from one INS checkpoint."
)
parser.add_argument("--checkpoint", type=Path, required=True)
parser.add_argument("--data-root", type=Path, default=None)
parser.add_argument("--frame-index", type=int, default=None)
parser.add_argument("--camera-id", default=None)
parser.add_argument("--source-camera-id", default=None)
parser.add_argument(
"--mode",
choices=("auto", "same-view", "fixed-source-reuse"),
default="auto",
)
parser.add_argument("--resolution", type=int, default=512)
parser.add_argument("--amp", choices=("auto", "none", "bf16"), default="auto")
parser.add_argument("--lpips", action=argparse.BooleanOptionalAction, default=True)
parser.add_argument(
"--lpips-net", choices=("alex", "squeeze", "vgg"), default="alex"
)
parser.add_argument("--device", default="cuda:0")
parser.add_argument(
"--output-dir",
type=Path,
default=Path("/workspace/cpfs-data/tmp/ins_branch_ablation"),
)
args = parser.parse_args()
if not args.checkpoint.is_file():
raise FileNotFoundError(args.checkpoint)
if args.data_root is not None and not args.data_root.is_dir():
raise FileNotFoundError(args.data_root)
if args.frame_index is not None and args.frame_index < 0:
raise ValueError("--frame-index cannot be negative.")
if args.resolution <= 0:
raise ValueError("--resolution must be positive.")
return args
def saved_args(checkpoint: dict[str, object]) -> dict[str, object]:
value = checkpoint.get("args", {})
return value if isinstance(value, dict) else {}
def resolve_config(
args: argparse.Namespace,
checkpoint: dict[str, object],
) -> dict[str, object]:
saved = saved_args(checkpoint)
root = args.data_root or Path(
saved.get(
"data_root",
"/workspace/cpfs-data/data/renderme_all/0026/0026_e0",
)
)
frame_index = (
int(saved.get("frame_index", 0))
if args.frame_index is None
else args.frame_index
)
mode = str(saved.get("mode", "same-view")) if args.mode == "auto" else args.mode
if mode not in {"same-view", "fixed-source-reuse"}:
raise ValueError(f"Unsupported checkpoint mode: {mode}")
source_camera_id = args.source_camera_id
if source_camera_id is None:
source_camera_id = saved.get("source_camera_id")
checkpoint_amp = str(saved.get("amp", "none"))
amp = checkpoint_amp if args.amp == "auto" else args.amp
if amp not in {"none", "bf16"}:
raise ValueError(
f"Ablation supports none/bf16, but checkpoint requested {amp!r}."
)
return {
"root": Path(root),
"frame_index": frame_index,
"mode": mode,
"source_camera_id": (
None if source_camera_id is None else str(source_camera_id)
),
"canonical_radius": float(saved.get("canonical_radius", 2.7)),
"canonical_focus": tuple(
float(value)
for value in saved.get("canonical_focus", (0.0, 0.0, 0.2))
),
"background": tuple(
float(value) for value in saved.get("background", (1.0, 1.0, 1.0))
),
"amp": amp,
}
@contextmanager
def ablate_branch(model, condition: str) -> Iterator[None]:
"""Temporarily replace one branch output with an all-zero tensor."""
if condition not in CONDITIONS:
raise ValueError(f"Unknown branch condition: {condition}")
if condition == "baseline":
yield
return
module = (
model.encoder.f_encoder
if condition == "global_zero"
else model.encoder.high_encoder
)
def zero_output(_module, _inputs, output):
if not torch.is_tensor(output):
raise TypeError("INS encoder branch output must be a tensor.")
return torch.zeros_like(output)
handle = module.register_forward_hook(zero_output)
try:
yield
finally:
handle.remove()
def masked_mean(value: Tensor, mask: Tensor) -> Tensor:
if mask.shape[1] == 1 and value.shape[1] != 1:
mask = mask.expand(-1, value.shape[1], -1, -1)
return (value * mask).sum() / mask.sum().clamp_min(1.0)
@torch.inference_mode()
def evaluate(args: argparse.Namespace) -> dict[str, object]:
device = torch.device(args.device)
if device.type != "cuda" or not torch.cuda.is_available():
raise RuntimeError("INS branch ablation requires CUDA.")
torch.cuda.set_device(device)
checkpoint = read_checkpoint(args.checkpoint)
config = resolve_config(args, checkpoint)
model = build_model_from_checkpoint(checkpoint, args.resolution, device)
dataset = RenderMeINSViewDataset(
root=config["root"],
frame_index=config["frame_index"],
image_size=512,
camera_ids=None if args.camera_id is None else [args.camera_id],
source_camera_id=config["source_camera_id"],
canonical_radius=config["canonical_radius"],
canonical_focus=config["canonical_focus"],
background=config["background"],
)
target_camera_id = str(args.camera_id or dataset.source_camera_id)
target_sample = dataset.get_view(target_camera_id)
source_camera_id = str(dataset.source_camera_id)
source_sample = dataset.get_view(source_camera_id)
if config["mode"] == "same-view":
encoder_sample = target_sample
else:
encoder_sample = source_sample
encoder_image = encoder_sample["image"].unsqueeze(0).to(device)
encoder_camera = encoder_sample["camera"].unsqueeze(0).to(device)
render_camera = target_sample["camera"].unsqueeze(0).to(device)
target_source = target_sample["image"].unsqueeze(0).to(device)
target = resize_target(target_source, args.resolution)
foreground_mask = target_sample["foreground_mask"].unsqueeze(0).to(device)
background = torch.tensor(config["background"], device=device)
context = DistributedContext(
rank=0,
local_rank=device.index or 0,
world_size=1,
distributed=False,
device=device,
)
metric_args = SimpleNamespace(
lpips_weight=1.0 if args.lpips else 0.0,
id_weight=0.0,
lpips_net=args.lpips_net,
)
lpips_metric, _ = build_auxiliary_losses(metric_args, context)
predictions: dict[str, Tensor] = {}
masks: dict[str, Tensor] = {}
metrics: dict[str, dict[str, float]] = {}
for condition in CONDITIONS:
with ablate_branch(model, condition):
with torch.autocast(
device_type="cuda",
dtype=torch.bfloat16,
enabled=config["amp"] == "bf16",
):
output = model(
encoder_image,
render_camera,
sampling_camera=encoder_camera,
encoder_camera=encoder_camera,
render_resolution=args.resolution,
deterministic_sampling=True,
)
raw_prediction = output["image"].float().clamp(0.0, 1.0)
condition_metrics, composited, resized_mask = foreground_metrics(
raw_prediction,
target,
foreground_mask,
background,
lpips_metric,
identity_metric=None,
)
gaussians = output["gaussians"]
condition_metrics.update(
{
"opacity_mean": float(gaussians["opacities"].float().mean()),
"scale_mean": float(gaussians["scales"].float().mean()),
"scale_max": float(gaussians["scales"].float().max()),
}
)
predictions[condition] = composited
masks[condition] = resized_mask
metrics[condition] = condition_metrics
baseline = predictions["baseline"]
mask = masks["baseline"]
for condition in ("global_zero", "high_zero"):
metrics[condition]["foreground_render_change_mae"] = float(
masked_mean((predictions[condition] - baseline).abs(), mask)
)
metrics[condition]["delta_foreground_psnr"] = (
metrics[condition]["foreground_psnr"]
- metrics["baseline"]["foreground_psnr"]
)
if "lpips" in metrics[condition]:
metrics[condition]["delta_lpips"] = (
metrics[condition]["lpips"] - metrics["baseline"]["lpips"]
)
args.output_dir.mkdir(parents=True, exist_ok=True)
encoder_01 = resize_target(encoder_image, args.resolution).cpu()
target_cpu = target.cpu()
mask_rgb = mask.expand(-1, 3, -1, -1).cpu()
prediction_row = [predictions[name].cpu() for name in CONDITIONS]
error_row = [(prediction - target_cpu).abs() * mask_rgb for prediction in prediction_row]
overview = make_grid(
torch.cat(
[
encoder_01,
target_cpu,
prediction_row[0],
prediction_row[1],
mask_rgb,
error_row[0],
error_row[1],
error_row[2],
prediction_row[2],
],
dim=0,
),
nrow=4,
padding=3,
)
# The final high-zero prediction occupies a third row because the first two
# rows are kept as four aligned columns. Individual renders below are the
# unambiguous source for numerical/visual comparison.
overview_path = args.output_dir / "ablation_overview.png"
save_image(overview, overview_path)
for name, prediction in predictions.items():
save_image(prediction.cpu(), args.output_dir / f"{name}.png")
save_image(
(prediction.cpu() - target_cpu).abs() * mask_rgb,
args.output_dir / f"{name}_error.png",
)
result: dict[str, object] = {
"checkpoint": str(args.checkpoint.resolve()),
"checkpoint_step": int(checkpoint.get("global_step", 0)),
"protocol": config["mode"],
"amp": config["amp"],
"encoder_camera_id": str(encoder_sample["camera_id"]),
"target_camera_id": target_camera_id,
"resolution": args.resolution,
"metrics": metrics,
"interpretation": {
"foreground_render_change_mae": (
"Mean absolute foreground render change relative to baseline."
),
"delta_foreground_psnr": (
"A large negative value means the ablated branch was important."
),
"delta_lpips": (
"A large positive value means the ablated branch was important."
),
},
"outputs": {
"overview": str(overview_path),
**{
name: str(args.output_dir / f"{name}.png")
for name in CONDITIONS
},
},
}
metrics_path = args.output_dir / "metrics.json"
with metrics_path.open("w", encoding="utf-8") as file:
json.dump(result, file, indent=2)
file.write("\n")
print(json.dumps(result, indent=2), flush=True)
print(f"overview: {overview_path}", flush=True)
print(f"metrics: {metrics_path}", flush=True)
return result
def main() -> int:
evaluate(parse_args())
return 0
if __name__ == "__main__":
raise SystemExit(main())
|