File size: 27,069 Bytes
3cf4fff |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 |
"""
CBM models and utilities consolidated from the Video_cbm.ipynb notebook.
"""
from __future__ import annotations
import os
import random
import numpy as np
import torch
from typing import List, Optional, Dict, Tuple
import cv2
from PIL import Image
import torch.nn as nn
import torch.nn.functional as F
import math
from sklearn.preprocessing import LabelEncoder
import re
import pandas as pd
import glob
import matplotlib.pyplot as plt
import matplotlib as mpl
@torch.no_grad()
def explain_instance(
model: nn.Module,
window_embeddings: torch.Tensor,
key_padding_mask: Optional[torch.Tensor] = None,
channel_ids: Optional[Union[List[int], torch.Tensor]] = None,
window_ids: Optional[Union[List[int], torch.Tensor]] = None,
target_class: Optional[int] = None,
window_spans: Optional[List[Tuple[int, int]]] = None,
fps: Optional[float] = None,
):
# device + shape
device = next(model.parameters(), torch.empty(0)).device
x = window_embeddings.to(device)
if x.dim() == 2:
x = x.unsqueeze(0) # [1,T,C]
if key_padding_mask is not None and key_padding_mask.dim() == 1:
key_padding_mask = key_padding_mask.unsqueeze(0)
# single forward, reuse its tau/masking behavior
logits, concepts, concepts_t, sharpness = model(
x,
key_padding_mask=key_padding_mask,
channel_ids=channel_ids,
window_ids=window_ids,
) # logits:[B,K], concepts_t:[B,T,C]
# per-time logits (not returned by forward)
logits_t = model.classifier(concepts_t) # [B,T,K]
# choose class
if target_class is None:
target_class = int(logits[0].argmax().item())
# pull first item (assumes B=1 for explanation)
concepts_t_1 = concepts_t[0] # [T,C]
logits_t_1 = logits_t[0] # [T,K]
# class params
w = model.classifier.weight[target_class] # [C]
b = (
0.0
if model.classifier.bias is None
else float(model.classifier.bias[target_class].item())
)
# per-time contributions and scores
contrib_t = concepts_t_1 * w.unsqueeze(0) # [T,C]
score_per_time = contrib_t.sum(dim=1) + b # [T]
# time importance consistent with forward (LSE/softmax with tau and mask)
tau = float(model.lse_tau)
time_scores = logits_t_1[:, target_class] # [T]
if key_padding_mask is not None:
time_scores = time_scores.masked_fill(key_padding_mask[0], float("-inf"))
time_importance = torch.softmax(time_scores / tau, dim=0) # [T]
# time-weighted global concept contributions
contrib_global = (time_importance.unsqueeze(1) * contrib_t).sum(dim=0) # [C]
# package
res = {
"target_class": torch.tensor(target_class),
"logits": logits[0].detach().cpu(),
"logits_per_time": logits_t_1.detach().cpu(),
"concepts": concepts[0].detach().cpu(),
"concepts_per_time": concepts_t_1.detach().cpu(),
"time_importance": time_importance.detach().cpu(),
"score_per_time": score_per_time.detach().cpu(),
"concept_contributions_per_time": contrib_t.detach().cpu(),
"concept_contributions_global": contrib_global.detach().cpu(),
"sharpness": {
k: {m: v.detach().cpu() for m, v in d.items()} for k, d in sharpness.items()
},
}
# optional spans
if window_spans is not None and len(window_spans) == concepts_t_1.shape[0]:
res["frame_spans"] = torch.tensor(window_spans, dtype=torch.long)
if fps is not None and fps > 0:
res["second_spans"] = torch.tensor(
[(s / fps, e / fps) for (s, e) in window_spans], dtype=torch.float32
)
# optional per-layer attention if present
attn = [getattr(layer, "attn_weights", None) for layer in model.layers]
if any(a is not None for a in attn):
res["attn_per_layer"] = [
a[0].detach().cpu() if a is not None else None for a in attn
]
return res
def _bar(x, width=20):
# x in [0,1]
n = int(round(x * width))
return "█" * n + "·" * (width - n)
def print_explanation(
res: dict,
fps_frame: dict,
concepts_list: Optional[List[str]] = None,
top_k_times: int = 3,
top_k_concepts: int = 8,
by_abs: bool = True,
positive_only: bool = True,
):
# pull & detach to CPU safely
def td(x):
return x.detach().cpu() if isinstance(x, torch.Tensor) else x
ti = td(res["time_importance"]).flatten() # [T]
spt = td(res["score_per_time"]).flatten() # [T]
cpt = td(res["concept_contributions_per_time"]) # [T, C]
cglob = td(res["concept_contributions_global"]).flatten() # [C]
tgt = res["target_class"]
target_class = int(tgt.item()) if hasattr(tgt, "item") else int(tgt)
T, C = ti.shape[0], cglob.shape[0]
if concepts_list is None:
concepts_list = [f"c{j}" for j in range(C)]
# optional spans
frame_spans = res.get("frame_spans", None)
second_spans = res.get("second_spans", None)
# normalizations
ti_norm = (ti - ti.min()) / (ti.max() - ti.min() + 1e-8)
spt_norm = (spt - spt.min()) / (spt.max() - spt.min() + 1e-8)
# global concepts: choose ranking
rank_vals = cglob.abs() if by_abs else cglob
if positive_only:
# Enforce positive-only based on original sign, even if by_abs=True
rank_vals = torch.where(cglob > 0, rank_vals, torch.zeros_like(rank_vals))
top_k_concepts = min(top_k_concepts, int((rank_vals > 0).sum().item()))
topc_vals, topc_idx = torch.topk(rank_vals, k=min(top_k_concepts, C))
print(f"Target class: {target_class}\n")
print("Top concepts (global):")
for _, j in zip(topc_vals, topc_idx):
j = int(j)
name = concepts_list[j] if j < len(concepts_list) else f"c{j}"
val = float(cglob[j]) # signed value
# bar by magnitude, show sign in number
mag = abs(val)
mag_norm = mag / (float(cglob.abs().max()) + 1e-8)
print(f" {name:30s} {val:+.3f} {_bar(mag_norm)}")
# top time steps
_, topt_idx = torch.topk(ti, k=min(top_k_times, T))
topt_idx = sorted(topt_idx.tolist(), key=lambda t: float(ti[t]), reverse=True)
print("\nImportant time steps:")
for t in topt_idx:
t_imp = float(ti[t])
extras = []
if frame_spans is not None:
fs = frame_spans[t]
extras.append(f"frames=[{int(fs[0])},{int(fs[1])}]")
if second_spans is not None:
ss = second_spans[t]
extras.append(f"sec=[{float(ss[0]):.2f},{float(ss[1]):.2f}]")
extra_str = (" " + " ".join(extras)) if extras else ""
start, end = fps_frame[t]
print(
f" t=[{int(start//60):02d}:{start%60:05.2f} - {int(end//60):02d}:{end%60:05.2f}] time_importance={t_imp:.3f} TI[{_bar(float(ti_norm[t]))}] Score[{_bar(float(spt_norm[t]))}]"
+ extra_str
)
ct = cpt[t] # [C]
# per-time top concepts (by abs or signed)
rank_vals_t = ct.abs() if by_abs else ct
if positive_only:
# Enforce positive-only based on original sign, even if by_abs=True
rank_vals_t = torch.where(ct > 0, rank_vals_t, torch.zeros_like(rank_vals_t))
k = min(top_k_concepts, int((rank_vals_t > 0).sum().item()), C)
else:
k = min(top_k_concepts, C)
vals, idxs = torch.topk(rank_vals_t, k=k)
# normalize bars by magnitude within this timestep for readability
denom = float(ct.abs().max()) + 1e-8
for j_rank in idxs:
j = int(j_rank)
name = concepts_list[j] if j < len(concepts_list) else f"c{j}"
val = float(ct[j]) # signed
print(f" - {name:30s} {val:+.3f} {_bar(abs(val)/denom)}")
def print_explanation_with_labels(
res: dict,
fps_frame: dict,
label_decoder: LabelEncoder,
true_label_idx: int,
positive_only: bool = True,
**kwargs
):
pred = res["target_class"]
pred_idx = int(pred.item()) if hasattr(pred, "item") else int(pred)
true_idx = int(true_label_idx)
pred_name = label_decoder.inverse_transform([pred_idx])[0]
true_name = label_decoder.inverse_transform([true_idx])[0]
print(f"Predicted: {pred_idx} ({pred_name}) | True: {true_idx} ({true_name})")
print_explanation(res, fps_frame, positive_only=positive_only, **kwargs)
# ---------------
# Batching utility
# ---------------
def pad_batch_sequences(
seqs: List[torch.Tensor], device: torch.device
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
"""
Pad a list of [T_i, C] tensors into a batch [B, T_max, C] and return
a key_padding_mask [B, T_max] with True for padded positions.
"""
if len(seqs) == 0:
raise ValueError("pad_batch_sequences received empty sequence list")
lengths = [int(s.shape[0]) for s in seqs]
C = int(seqs[0].shape[1])
T_max = int(max(lengths))
B = len(seqs)
batch = torch.zeros((B, T_max, C), dtype=torch.float32, device=device)
mask = torch.ones((B, T_max), dtype=torch.bool, device=device) # True=padded
for i, s in enumerate(seqs):
t = lengths[i]
batch[i, :t, :] = s.to(device)
mask[i, :t] = False
return batch, mask
def _cv_bar_img(frac: float, width: int = 160, height: int = 8) -> np.ndarray:
frac = float(max(0.0, min(1.0, frac)))
w = max(1, int(round(frac * width)))
bar = np.zeros((height, width, 3), dtype=np.uint8)
bar[:, :w, :] = 255
return bar
def _put_text_multiline(
img,
lines,
org,
line_h,
font=cv2.FONT_HERSHEY_SIMPLEX,
font_scale=0.40,
thickness=1,
color=(255, 255, 255),
):
x, y = org
for i, line in enumerate(lines):
cv2.putText(
img,
line,
(x, y + i * line_h),
font,
font_scale,
color,
thickness,
cv2.LINE_AA,
)
def _safe_paste_bar(frame: np.ndarray, x: int, y: int, bar: np.ndarray) -> None:
H, W = frame.shape[:2]
bh, bw = bar.shape[:2]
x1 = max(0, x)
y1 = max(0, y)
x2 = min(W, x + bw)
y2 = min(H, y + bh)
if x1 >= x2 or y1 >= y2:
return
bx1 = x1 - x
by1 = y1 - y
bx2 = bx1 + (x2 - x1)
by2 = by1 + (y2 - y1)
roi = frame[y1:y2, x1:x2]
bar_crop = bar[by1:by2, bx1:bx2]
np.maximum(roi, bar_crop, out=roi)
@torch.no_grad()
def render_explained_video_small_tl(
vid_path: str,
out_path: str,
res: dict, # from explain_instance(...)
fps_frame_seconds: List[Tuple[float, float]], # spans in SECONDS
label_decoder, # fitted LabelEncoder
true_label_idx: int,
concepts_list: Optional[List[str]] = None,
top_k_times: int = 3,
top_k_concepts: int = 4,
by_abs: bool = True,
up_scale: float = 2.0, # upscale factor
margin: int = 10,
panel_w_px: int = 300, # small box width
panel_alpha: float = 0.70,
font_scale: float = 0.40,
thickness: int = 1,
codec: str = "mp4v",
) -> str:
cap = cv2.VideoCapture(vid_path)
if not cap.isOpened():
raise RuntimeError(f"Could not open video: {vid_path}")
fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
W = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
H = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
F = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
outW = int(round(W * up_scale))
outH = int(round(H * up_scale))
writer = cv2.VideoWriter(
out_path, cv2.VideoWriter_fourcc(*codec), fps, (outW, outH)
)
if not writer.isOpened():
cap.release()
raise RuntimeError(f"Could not open writer for: {out_path}")
# tensors -> CPU
ti = res["time_importance"].detach().cpu().float() # [T]
cpt = res["concept_contributions_per_time"].detach().cpu() # [T,C]
tgt = res["target_class"]
pred_idx = int(tgt.item()) if hasattr(tgt, "item") else int(tgt)
T = ti.shape[0]
C = cpt.shape[1]
if concepts_list is None:
concepts_list = [f"c{j}" for j in range(C)]
try:
pred_name = label_decoder.inverse_transform([pred_idx])[0]
true_name = label_decoder.inverse_transform([int(true_label_idx)])[0]
except Exception:
pred_name = str(pred_idx)
true_name = str(true_label_idx)
# top-k windows
if top_k_times == 0:
top_k_times = T
kT = min(top_k_times, T)
_, topt_idx = torch.topk(ti, k=kT, largest=True, sorted=True)
important_t = set(int(i) for i in topt_idx.tolist())
# per-window top concepts (precompute)
per_t_top = []
for t in range(T):
ct = cpt[t]
rank_vals = ct.abs() if by_abs else ct
kk = min(top_k_concepts, C)
_, idxs = torch.topk(rank_vals, k=kk, largest=True, sorted=True)
denom = float(ct.abs().max().item()) + 1e-8
entries = []
for j in idxs.tolist():
name = concepts_list[j] if j < len(concepts_list) else f"c{j}"
sval = float(ct[j].item())
frac = min(1.0, abs(sval) / denom) if denom > 0 else 0.0
entries.append((name, sval, frac))
per_t_top.append(entries)
# map sec->frames on original fps
frame_to_t = [None] * F
for t, (ss, es) in enumerate(fps_frame_seconds):
fs = max(0, int(round(ss * fps)))
fe = min(F - 1, int(round(es * fps)))
for f in range(fs, fe + 1):
frame_to_t[f] = t
# small top-left panel geometry (after upscaling!)
# keep it compact: header(2 lines) + k concepts
line_h = 16
rows = 2 + top_k_concepts
panel_h_px = 18 + rows * line_h + 12
x0, y0 = margin, margin
panel_rect = (x0, y0, panel_w_px, panel_h_px)
fidx = 0
try:
while True:
ok, frame = cap.read()
if not ok:
break
# upscale first, so overlay stays small proportionally
frame = cv2.resize(frame, (outW, outH), interpolation=cv2.INTER_CUBIC)
t = frame_to_t[fidx] if fidx < len(frame_to_t) else None
if (t is not None) and (t in important_t):
# translucent panel
overlay = frame.copy()
x, y, pw, ph = panel_rect
cv2.rectangle(overlay, (x, y), (x + pw, y + ph), (0, 0, 0), -1)
cv2.addWeighted(overlay, panel_alpha, frame, 1 - panel_alpha, 0, frame)
# header (compressed)
sec = fidx / fps
ss, es = fps_frame_seconds[t]
header = [
f"Pred:{pred_name} | True:{true_name}",
f"t={t} TI={float(ti[t]):.3f} [{ss:.2f}-{es:.2f}]s",
]
_put_text_multiline(
frame,
header,
(x + 8, y + 18),
line_h,
font_scale=font_scale,
thickness=thickness,
)
# concepts (fewer, tight spacing)
y_cursor = y + 18 + line_h * len(header) + 2
for name, sval, frac in per_t_top[t]:
bar = _cv_bar_img(frac, width=120, height=8)
bx, by = x + 8, int(y_cursor - 8)
_safe_paste_bar(frame, bx, by, bar)
cv2.putText(
frame,
f"{name[:16]:16s} {sval:+.2f}",
(bx + bar.shape[1] + 8, int(y_cursor + 4)),
cv2.FONT_HERSHEY_SIMPLEX,
font_scale,
(255, 255, 255),
thickness,
cv2.LINE_AA,
)
y_cursor += line_h
writer.write(frame)
fidx += 1
finally:
cap.release()
writer.release()
return out_path
@torch.no_grad()
def print_temporal_dependencies(
res: dict,
top_k_times: int = 5, # how many query timesteps to print
top_k_links: int = 5, # how many strongest dependencies per timestep
concept_idx: Optional[
int
] = None, # pick a concept for per-channel attention; None -> mean over concepts
layer_agg: str = "mean", # "mean" or "max" across layers
head_or_concept_agg: str = "mean", # how to aggregate heads (full-attn) or concepts (per-channel): "mean" or "max"
focus_times: Optional[
List[int]
] = None, # if given, only print these query timesteps
by_abs: bool = False, # rank links by absolute weight (usually False)
):
"""
Print temporal dependencies (attention) between timesteps.
Handles both per-channel attention [C,T,T] and full-attention [H,T,T].
Strategy:
1) Load attention maps per layer.
2) If shape is [C,T,T] (per-channel), either select 'concept_idx' or aggregate across concepts.
If shape is [H,T,T] (full), aggregate across heads.
3) Aggregate across layers via mean/max.
4) Choose which timesteps to display:
- 'focus_times' if given,
- else top 'top_k_times' by res["time_importance"] (if available),
- else first 'top_k_times'.
5) For each chosen timestep t, print top 'top_k_links' target timesteps u with largest attention weight.
"""
attn_layers = res.get("attn_per_layer", None)
if not attn_layers or all(a is None for a in attn_layers):
print(
"[temporal] No attention maps available in 'res'. Ensure your model layers store 'attn_weights'."
)
return
# Collect valid layers and ensure tensor type
mats = []
for a in attn_layers:
if a is None:
continue
# a can be [C,T,T] (per-channel) OR [H,T,T] (full multi-head)
if not torch.is_tensor(a):
a = torch.as_tensor(a)
mats.append(a.float())
if len(mats) == 0:
print("[temporal] No attention maps available after filtering.")
return
# Determine shape kind
# Each layer mat has shape [G, T, T], where G = C (per-channel) or H (heads)
G, T, T2 = mats[0].shape
assert T == T2, f"Expected square attention [G,T,T], got {mats[0].shape}"
# Aggregate across concepts/heads (dim 0)
def agg_g(x: torch.Tensor) -> torch.Tensor: # x: [G,T,T] -> [T,T]
if concept_idx is not None and x.shape[0] > concept_idx:
return x[concept_idx]
if head_or_concept_agg == "max":
return x.max(dim=0).values
return x.mean(dim=0)
mats_agg_g = [agg_g(a) for a in mats] # list of [T,T]
# Aggregate across layers -> [T,T]
stack = torch.stack(mats_agg_g, dim=0) # [L,T,T]
if layer_agg == "max":
A = stack.max(dim=0).values
else:
A = stack.mean(dim=0).values if hasattr(stack, "values") else stack.mean(dim=0)
if isinstance(A, torch.return_types.max):
A = A.values
# Sanity: normalize rows (optional; attention should already be row-softmaxed)
# A = A / (A.sum(dim=-1, keepdim=True) + 1e-9)
# Decide which timesteps to print
if focus_times is not None and len(focus_times) > 0:
query_times = [t for t in focus_times if 0 <= t < T]
else:
ti = res.get("time_importance", None)
if isinstance(ti, torch.Tensor) and ti.numel() == T:
vals, idx = torch.topk(ti, k=min(top_k_times, T))
query_times = idx.tolist()
# Sort by decreasing importance
query_times = sorted(query_times, key=lambda t: float(ti[t]), reverse=True)
else:
query_times = list(range(min(top_k_times, T)))
# Optional second spans
second_spans = res.get("second_spans", None) # [T,2] if present
def _fmt_time(ti_):
if (
second_spans is not None
and hasattr(second_spans, "__len__")
and len(second_spans) == T
):
ss, es = second_spans[ti_]
return f"t={ti_} [{float(ss):.2f}-{float(es):.2f}s]"
return f"t={ti_}"
# Print header context
tgt = res.get("target_class", None)
if tgt is not None:
tc = int(tgt.item()) if hasattr(tgt, "item") else int(tgt)
print(f"[temporal] Target class: {tc}")
if concept_idx is not None:
print(f"[temporal] Using per-channel attention for concept c={concept_idx}")
else:
print(
f"[temporal] Aggregation over {'concepts' if G==A.shape[0] else 'heads'}: {head_or_concept_agg}, layers: {layer_agg}"
)
# For each chosen query timestep, print its strongest links
for t in query_times:
row = A[t] # [T]
# row = row.clone(); row[t] = 0.0
rank_vals = row.abs() if by_abs else row
k = min(top_k_links, T)
vals, idxs = torch.topk(rank_vals, k=k, largest=True, sorted=True)
# Pretty print
print(f"\n{_fmt_time(t)} (row-softmaxed attention to other timesteps)")
# Normalize for bar length
denom = float(rank_vals[idxs[0]] + 1e-12)
for j, v in zip(idxs.tolist(), vals.tolist()):
w = float(row[j])
rel = max(0.0, min(1.0, float(abs(v) / denom)))
bar = (
_bar(rel)
if " _bar" in globals() or "_bar" in locals()
else f"{rel:.2f}"
)
if second_spans is not None and len(second_spans) == T:
ss, es = second_spans[j]
target_str = f"u={j} [{float(ss):.2f}-{float(es):.2f}s]"
else:
target_str = f"u={j}"
print(f" -> {target_str:18s} w={w:+.4f} {bar}")
def _fmt_sec(sec: float) -> str:
# 0:00.00 style for readability
m = int(sec // 60)
s = sec - 60 * m
return f"{m}:{s:05.2f}s" if m else f"{s:.2f}s"
@torch.no_grad()
def plot_attention_heatmaps(
res: dict,
concept_idx: Optional[
int
] = None, # per-channel if set; else aggregate across concepts/heads
concept_names: Optional[List[str]] = None, # usually concepts.text_concepts
layer_idxs: Optional[List[int]] = None, # which layers to plot; None -> all
layer_agg: Optional[str] = None, # None | "mean" | "max"
head_or_concept_agg: str = "mean", # "mean" | "max"
normalize_rows: bool = True,
show_seconds: bool = True,
cmap: str = "magma",
figsize: Tuple[int, int] = (5, 4),
savepath: Optional[str] = None,
title_prefix: str = "Attention",
):
rc = {
"font.family": "serif",
"font.serif": ["Times New Roman", "Times", "DejaVu Serif", "Liberation Serif"],
"mathtext.fontset": "stix",
}
attn_layers = res.get("attn_per_layer", None)
if not attn_layers or all(a is None for a in attn_layers):
print("[heatmap] No attention maps in 'res'.")
return
mats = []
for a in attn_layers:
if a is None:
continue
a = torch.as_tensor(a, dtype=torch.float32)
assert (
a.ndim == 3 and a.shape[-1] == a.shape[-2]
), f"Expected [G,T,T], got {tuple(a.shape)}"
mats.append(a)
if not mats:
print("[heatmap] No usable attention maps.")
return
if layer_idxs is not None:
mats = [mats[i] for i in layer_idxs if 0 <= i < len(mats)]
if not mats:
print("[heatmap] Selected layer_idxs produced empty set.")
return
G, T, _ = mats[0].shape
second_spans = res.get("second_spans", None)
def agg_g(x: torch.Tensor) -> torch.Tensor:
if concept_idx is not None:
if not (0 <= concept_idx < x.shape[0]):
raise IndexError(
f"concept_idx={concept_idx} out of range [0,{x.shape[0]-1}]."
)
return x[concept_idx]
return x.max(dim=0).values if head_or_concept_agg == "max" else x.mean(dim=0)
per_layer = [agg_g(L) for L in mats]
plots = []
if layer_agg in (None, ""):
for Li, A in enumerate(per_layer):
plots.append((Li, A))
elif layer_agg == "mean":
plots.append(("mean", torch.stack(per_layer, dim=0).mean(dim=0)))
elif layer_agg == "max":
plots.append(("max", torch.stack(per_layer, dim=0).max(dim=0).values))
else:
raise ValueError("layer_agg must be None, 'mean', or 'max'.")
def row_norm(A: torch.Tensor) -> torch.Tensor:
if not normalize_rows:
return A
denom = A.sum(dim=-1, keepdim=True).clamp_min(1e-12)
return A / denom
def make_ticks(T: int):
step = max(1, T // 8)
idxs = list(range(0, T, step))
if idxs[-1] != T - 1:
idxs.append(T - 1)
if (
show_seconds
and isinstance(second_spans, torch.Tensor)
and second_spans.shape[0] == T
):
lbls = []
for i in idxs:
ss, es = second_spans[i].tolist()
mid = 0.5 * (float(ss) + float(es))
lbls.append(f"u={i} · {_fmt_sec(mid)}")
else:
lbls = [f"u={i}" for i in idxs]
return idxs, lbls
figs = []
# apply Times New Roman only for the plotting block
with mpl.rc_context(rc):
for tag, A in plots:
A = row_norm(A.detach().cpu())
fig, ax = plt.subplots(figsize=figsize)
im = ax.imshow(
A,
origin="lower",
interpolation="nearest",
cmap=cmap,
vmin=0.0,
vmax=float(A.max().item()) or None,
)
ax.set_xlabel("Key time u (source/context)", fontsize=16)
ax.set_ylabel("Query time t (target/current)", fontsize=16)
xt, xl = make_ticks(T)
yt, yl = make_ticks(T)
yl = [lbl.replace("u=", "t=") for lbl in yl]
ax.set_xticks(xt)
ax.set_xticklabels(xl, rotation=45, ha="right", fontsize=13)
ax.set_yticks(yt)
ax.set_yticklabels(yl, fontsize=13)
cname = None
if (
concept_idx is not None
and concept_names
and 0 <= concept_idx < len(concept_names)
):
cname = concept_names[concept_idx]
tag_str = f"layer={tag}" if isinstance(tag, (int, str)) else str(tag)
if concept_idx is not None:
title = f" ({cname})" if cname else ""
else:
title = f" (agg over {'concepts' if G==A.shape[0] else 'heads'})"
ax.set_title(title, fontsize=18)
cbar = fig.colorbar(im, ax=ax, fraction=0.046, pad=0.04)
cbar.set_label("Attention weight", fontsize=16)
fig.tight_layout()
if savepath:
p = savepath
if len(plots) > 1:
stem, ext = (savepath.rsplit(".", 1) + ["png"])[:2]
p = f"{stem}_{tag_str}.{ext}"
fig.savefig(p, dpi=150, bbox_inches="tight")
figs.append(fig)
plt.show()
return figs
|