File size: 15,753 Bytes
8065faa | 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 | #!/usr/bin/env python
"""Montage of qualitative examples across the pretraining and finetuned models.
Regenerates docs/assets/affmae_512_examples.png:
python scripts/render_examples.py \
--config configs/aff_base_finetune_512_fpw.yaml \
--checkpoint weights/segmentation/fpw_aff_base_ft_512_slits_pgbmi.pth \
--pretrain-config configs/aff_base_pretrain_0.4ds_0.5mask_last_local.yaml \
--pretrain-checkpoint weights/pretrain/ckpt_epoch_399_affmae_fpw.pth \
--samples docs/assets/sample{1,2,3,4}.png \
--output docs/assets/affmae_512_examples.png
One row per sample, one column per panel in ``--columns``. Seeded per sample, so
the command above reproduces the committed asset byte-for-byte.
Note:
Reconstruction panels use Perlin masking, matching pretraining
(``forward`` -> ``_forward_internal`` -> ``mask_and_embed``). The token
positions come from the reconstruction result, not a second
``token_layout`` call: that would draw a fresh mask, and only ~50% of the
tokens would land on a patch the masked image shows as visible.
"""
import argparse
import os
import sys
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt # noqa: E402
import numpy as np # noqa: E402
import torch # noqa: E402
from affmae.config import load_config # noqa: E402
from affmae.inference import AFFMAE # noqa: E402
from affmae.utils.paths import default_plot_path # noqa: E402
from affmae.viz import ( # noqa: E402
PAPER,
class_overlay,
draw_token_positions,
to_display_image,
)
#: Stage banners. The two training stages the figure contrasts.
STAGE_1 = "Stage 1: MAE pretraining"
STAGE_2 = "Stage 2: Supervised finetune"
#: Panel name -> (column title, needs the pretraining model, stage banner).
#: The banner groups columns under a training stage; "" means no banner, and
#: consecutive columns sharing one get a single banner spanning them with dashed
#: rules at the boundaries.
PANELS = {
"input": ("Input", False, ""),
"masked": ("Perlin-masked input", True, STAGE_1),
"tokens_sparse": ("Sparse tokens, encoder stage {sparse_stage}", True, STAGE_1),
"reconstruction": ("Reconstruction", True, STAGE_1),
"tokens": ("Final stage\ntoken locations", False, STAGE_2),
"truth": ("Ground truth", False, STAGE_2),
"prediction": ("Prediction", False, STAGE_2),
}
#: The sparse-token panel is available but not shown by default: the masked input
#: and the reconstruction already carry the pretraining story, and a token panel
#: on both sides of the figure invites a comparison between a sparse stage-2
#: layout and a dense final-stage one that is not like-for-like.
DEFAULT_COLUMNS = ("input", "masked", "reconstruction", "tokens", "prediction")
def column_groups(columns):
"""Contiguous runs of columns sharing a banner.
Returns:
list of ``(banner, first_index, last_index)``, in column order. A run is
broken by any change of banner, so reordering ``--columns`` cannot
produce a banner that spans a column belonging to another stage.
"""
groups = []
for index, name in enumerate(columns):
banner = PANELS[name][2]
if groups and groups[-1][0] == banner:
groups[-1][2] = index
else:
groups.append([banner, index, index])
return [tuple(g) for g in groups]
def draw_stage_banners(fig, axes, columns, groups, viz, rows):
"""Write the stage banners and the dashed rules between stages.
Called after ``tight_layout``, because it reads final axes positions. Works
in figure coordinates, so ``bbox_inches="tight"`` crops around the banners
instead of clipping them.
"""
from matplotlib.lines import Line2D
top = max(axes[0][c].get_position().y1 for c in range(len(columns)))
bottom = min(axes[rows - 1][c].get_position().y0 for c in range(len(columns)))
# Place the banner above the *rendered* column titles rather than above the
# axes plus a guess. Estimating from font points broke as soon as a title
# wrapped -- "Final stage\ntoken locations" is two lines, and a pad sized for
# one put the banner on top of it. Ask matplotlib where the text actually
# ended up instead.
fig.canvas.draw()
inverse = fig.transFigure.inverted()
title_top = top
for column in range(len(columns)):
title = axes[0][column].title
if not title.get_text():
continue
box = inverse.transform(title.get_window_extent(fig.canvas.get_renderer()))
title_top = max(title_top, box[1][1])
# One line of the banner's own font, as breathing room above the titles.
pad = (title_top - top) + (viz.font_size + 6) * 1.1 / (72.0 * fig.get_size_inches()[1])
for banner, first, last in groups:
if not banner:
continue
left = axes[0][first].get_position().x0
right = axes[0][last].get_position().x1
fig.text((left + right) / 2.0, top + pad, banner,
ha="center", va="bottom",
fontsize=viz.font_size + 6, fontweight="bold")
for previous, following in zip(groups, groups[1:]):
gap_left = axes[0][previous[2]].get_position().x1
gap_right = axes[0][following[1]].get_position().x0
x = (gap_left + gap_right) / 2.0
fig.add_artist(Line2D([x, x], [bottom, top + pad * 1.9],
transform=fig.transFigure,
color="0.35", linestyle="--", linewidth=2.0))
def load_truth(mask_dir, stem, num_classes):
"""Load a multi-channel mask and flatten it the way training does.
Mirrors ``EMDatasetMultiClass.__getitem__``: each channel is binarized at 10
and a pixel takes ``channel_position + 1``, so later channels win on overlap.
"""
import glob
import tifffile
matches = glob.glob(os.path.join(mask_dir, stem + ".*"))
if not matches:
raise FileNotFoundError(
f"no mask for {stem!r} in {mask_dir}. Ground truth needs a mask "
f"whose filename stem matches the image.")
mask = tifffile.imread(matches[0])
if mask.ndim == 4:
mask = mask[:, :, :, 0]
if mask.ndim == 2:
return torch.from_numpy(mask.astype(np.int64))
labels = np.zeros(mask.shape[1:], dtype=np.int64)
for position in range(min(mask.shape[0], num_classes - 1)):
labels[mask[position] > 10] = position + 1
return torch.from_numpy(labels)
def resize_labels(labels, shape):
"""Nearest-neighbour resize of a label map to ``shape``.
Masks are stored at the microscope's native resolution while predictions come
back at ``img_size``; overlaying the two unresized makes matplotlib size the
axes to whichever imshow came last, shrinking the background into a corner.
Nearest, not bilinear: interpolating class indices invents classes that were
never annotated.
"""
if tuple(labels.shape) == tuple(shape):
return labels
resized = torch.nn.functional.interpolate(
labels[None, None].float(), size=tuple(shape), mode="nearest")
return resized[0, 0].long()
def build_parser():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--config", required=True,
help="Finetuning config.")
parser.add_argument("--checkpoint", required=True,
help="Finetuned segmentation checkpoint.")
parser.add_argument("--pretrain-config", default=None,
help="Pretraining config; needed by the MAE panels.")
parser.add_argument("--pretrain-checkpoint", default=None,
help="Pretraining checkpoint; needed by the MAE panels.")
parser.add_argument("--samples", nargs="+", required=True,
help="Image paths, one row each.")
parser.add_argument("--columns", default=",".join(DEFAULT_COLUMNS),
help=f"comma-separated, from {sorted(PANELS)}")
parser.add_argument("--sparse-stage", type=int, default=2,
help="1-based encoder stage for the optional "
"tokens_sparse panel, which is off by default.")
parser.add_argument("--mask-ratio", type=float, default=0.5,
help="Fraction of patches hidden from the MAE.")
parser.add_argument("--masks", default=None,
help="Mask directory, required by the truth column.")
parser.add_argument("--seed", type=int, default=0,
help="Seeds the Perlin mask, so the figure is "
"reproducible. Pass -1 to leave the RNG alone.")
parser.add_argument("--output", default=None)
parser.add_argument("--device", default=None)
return parser
def main():
args = build_parser().parse_args()
columns = [name.strip() for name in args.columns.split(",") if name.strip()]
unknown = [name for name in columns if name not in PANELS]
if unknown:
raise SystemExit(f"unknown column(s) {unknown}; choose from {sorted(PANELS)}")
if "truth" in columns and not args.masks:
raise SystemExit("--columns includes 'truth', so --masks is required.")
needs_mae = any(PANELS[name][1] for name in columns)
if needs_mae and not (args.pretrain_config and args.pretrain_checkpoint):
wanted = [n for n in columns if PANELS[n][1]]
raise SystemExit(
f"column(s) {wanted} come from the MAE, so --pretrain-config and "
f"--pretrain-checkpoint are both required.")
if not 0.0 < args.mask_ratio < 1.0:
raise SystemExit(
f"--mask-ratio must be in (0, 1), got {args.mask_ratio}: nothing "
f"masked leaves the decoder no queries, everything masked leaves "
f"the encoder no input.")
config = load_config(args.config)
segmenter = AFFMAE.from_checkpoint(args.checkpoint, config=config,
device=args.device)
print(f"segmentation: {args.checkpoint} on {segmenter.device} "
f"({sorted(segmenter.capabilities)})")
mae = None
if needs_mae:
mae = AFFMAE.from_checkpoint(args.pretrain_checkpoint,
config=args.pretrain_config,
device=args.device)
print(f"pretraining: {args.pretrain_checkpoint} on {mae.device} "
f"({sorted(mae.capabilities)})")
if not mae.can_reconstruct:
raise SystemExit(
f"{args.pretrain_checkpoint} has no MAE head, so it cannot "
f"produce the masked/reconstruction panels.")
viz = PAPER
stage_index = args.sparse_stage - 1
rows = []
for index, path in enumerate(args.samples):
# Re-seed per sample so a row's mask does not depend on how many rows
# precede it -- otherwise dropping one sample changes all the others.
if args.seed >= 0:
torch.manual_seed(args.seed + index)
row = {"stem": os.path.splitext(os.path.basename(path))[0]}
row["result"] = segmenter.segment(path)
_, dense = segmenter.token_layout(path)
row["dense_tokens"] = dense[-1]
if mae is not None:
# One pass. Taking the tokens from a second call to
# token_layout(mask_ratio=...) drew a *fresh* Perlin mask, so the
# figure showed one mask's tokens over another mask's image -- only
# ~50% of them landed on a visible patch, which is chance. The
# reconstruction result carries the tokens from its own pass, so
# they and the masked image agree by construction (measured: 100%).
row["recon"] = mae.reconstruct(path, mask_ratio=args.mask_ratio)
if "tokens_sparse" in columns:
sparse = row["recon"].locations
if not sparse:
raise SystemExit(
f"{args.pretrain_checkpoint} exposes no per-stage token "
f"positions, so the sparse-token panel cannot be drawn.")
if not -len(sparse) <= stage_index < len(sparse):
raise SystemExit(
f"--sparse-stage {args.sparse_stage} is out of range; "
f"the encoder has {len(sparse)} stages.")
row["sparse_tokens"] = sparse[stage_index]
rows.append(row)
counts = row["result"].class_pixel_counts
print(f" {row['stem'][:44]}: classes {sorted(counts)}, "
f"dense {row['dense_tokens'].shape[0]} tokens"
+ (f", sparse {row['sparse_tokens'].shape[0]}"
if "sparse_tokens" in row else ""))
n_rows, n_cols = len(rows), len(columns)
size = viz.figsize_per_cell
fig, axes = plt.subplots(n_rows, n_cols,
figsize=(size * n_cols, size * n_rows),
squeeze=False)
for r, row in enumerate(rows):
result = row["result"]
dense_bg = to_display_image(result.image, viz)
for c, column in enumerate(columns):
ax = axes[r][c]
if column == "input":
ax.imshow(dense_bg, cmap=viz.cmap, vmin=0, vmax=1)
elif column == "masked":
ax.imshow(to_display_image(row["recon"].masked, viz),
cmap=viz.cmap, vmin=0, vmax=1)
elif column == "reconstruction":
# reconstructions[-1] is the finest head, and AFFMAE.reconstruct
# already stitches the visible original patches back in, so only
# the masked region is model output.
ax.imshow(to_display_image(row["recon"].reconstructions[-1], viz),
cmap=viz.cmap, vmin=0, vmax=1)
elif column == "tokens_sparse":
ax.imshow(draw_token_positions(
row["recon"].masked, row["sparse_tokens"],
config.patch_size, viz))
elif column == "tokens":
ax.imshow(draw_token_positions(
result.image, row["dense_tokens"], config.patch_size, viz))
else:
ax.imshow(dense_bg, cmap=viz.cmap, vmin=0, vmax=1)
if column == "prediction":
ax.imshow(class_overlay(result.labels, result.num_classes, viz))
elif column == "truth":
truth = load_truth(args.masks, row["stem"], result.num_classes)
ax.imshow(class_overlay(
resize_labels(truth, result.labels.shape),
result.num_classes, viz))
ax.set_xticks([])
ax.set_yticks([])
if r == 0:
ax.set_title(
PANELS[column][0].format(sparse_stage=args.sparse_stage),
fontsize=viz.font_size + 2)
if c == 0:
label = row["stem"]
if len(label) > 24: # microscope filenames run past 60 chars
label = label[:11] + "..." + label[-10:]
ax.set_ylabel(label, fontsize=viz.font_size)
fig.tight_layout()
draw_stage_banners(fig, axes, columns, column_groups(columns), viz, n_rows)
output = args.output or default_plot_path("affmae_examples.png")
os.makedirs(os.path.dirname(os.path.abspath(output)), exist_ok=True)
fig.savefig(output, dpi=viz.dpi, bbox_inches="tight")
plt.close(fig)
print(f"Wrote {output} ({n_rows} rows x {n_cols} cols: {', '.join(columns)})")
if __name__ == "__main__":
main()
|