Image-Text-to-Text
Transformers
Safetensors
mage_vl
multimodal
vision-language-model
mage-vl
video-understanding
streaming
conversational
custom_code
Instructions to use microsoft/Mage-VL with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use microsoft/Mage-VL with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="microsoft/Mage-VL", trust_remote_code=True) messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# Load model directly from transformers import AutoModelForImageTextToText model = AutoModelForImageTextToText.from_pretrained("microsoft/Mage-VL", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use microsoft/Mage-VL with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "microsoft/Mage-VL" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "microsoft/Mage-VL", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/microsoft/Mage-VL
- SGLang
How to use microsoft/Mage-VL with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "microsoft/Mage-VL" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "microsoft/Mage-VL", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "microsoft/Mage-VL" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "microsoft/Mage-VL", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use microsoft/Mage-VL with Docker Model Runner:
docker model run hf.co/microsoft/Mage-VL
File size: 21,471 Bytes
12acbba | 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 | #!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""Energy-based sampling and window selection algorithms."""
import math
from typing import Tuple, List, Dict, Any
import numpy as np
from .video_probe import ffprobe_packets_pb_energy_bins, ffprobe_sum_pkt_size
def pick_peak_frame_ids_from_pkt_size(
video_path: str,
fps: float,
total_frames: int,
bin_sec: float = 0.5,
peaks_cap: int = 8,
peaks_per_sec: float = 0.5,
neighbor: int = 1,
smooth_bins: int = 1,
) -> Tuple[List[int], Dict[str, Any]]:
"""Select peak frame ids using PB packet-size energy bins (exclude keyframes)."""
fps_use = float(fps) if (fps and fps > 0) else 30.0
total_frames = int(max(0, total_frames))
centers, energy, dbg_bins = ffprobe_packets_pb_energy_bins(
video_path=str(video_path),
bin_sec=float(bin_sec),
smooth_bins=int(smooth_bins),
)
dbg: Dict[str, Any] = {
"bin_sec": float(bin_sec),
"peaks_cap": int(peaks_cap),
"peaks_per_sec": float(peaks_per_sec),
"neighbor": int(neighbor),
"smooth_bins": int(smooth_bins),
"bins": dbg_bins,
"peak_bins": [],
"peak_times_sec": [],
"peak_frame_ids": [],
}
if energy.size == 0 or centers.size == 0:
dbg["error"] = "no_energy"
return [], dbg
duration_sec = float(total_frames) / float(fps_use) if total_frames > 0 else float(dbg_bins.get("duration_est_sec", 0.0))
duration_sec = max(duration_sec, 0.0)
k = int(max(1, round(duration_sec * float(peaks_per_sec)))) if duration_sec > 0 else 1
k = int(min(int(peaks_cap), k, int(energy.size)))
idx = np.argsort(-energy)[:k]
idx = [int(i) for i in idx.tolist()]
dbg["peak_bins"] = idx
peak_times = [float(centers[i]) for i in idx]
dbg["peak_times_sec"] = peak_times
out_set: set = set()
nb = int(max(0, int(neighbor)))
for t in peak_times:
fid0 = int(round(float(t) * float(fps_use)))
if total_frames > 0:
fid0 = max(0, min(total_frames - 1, fid0))
for d in range(-nb, nb + 1):
fid = int(fid0 + d)
if total_frames > 0:
fid = max(0, min(total_frames - 1, fid))
out_set.add(int(fid))
out = sorted(out_set)
dbg["peak_frame_ids"] = out
return out, dbg
def build_variable_length_gops_by_energy(
video_path: str,
total_frames: int,
fps: float,
target_num_gops: int,
bin_sec: float = 0.5,
smooth_bins: int = 1,
min_span_sec: float = 1.5,
max_span_sec: float = 6.0,
) -> Tuple[List[Dict[str, Any]], Dict[str, Any]]:
"""Build variable-length GOPs by cumulative PB packet-size energy.
V2.2 implementation:
- compute PB-only packet-size energy bins
- accumulate smoothed energy over time
- cut a GOP when cumulative energy reaches a target threshold, subject to
min/max span constraints
- prefer a local low-energy valley near the threshold crossing so cut points
land on more stable boundaries
- each GOP records a start / end / anchor time
"""
total_frames = int(max(0, total_frames))
fps_use = float(fps) if (fps and fps > 0) else 30.0
target_num_gops = int(max(1, int(target_num_gops)))
bin_sec = float(bin_sec) if (bin_sec and bin_sec > 1e-6) else 0.5
smooth_bins = int(max(0, int(smooth_bins)))
min_span_sec = float(max(0.25, float(min_span_sec)))
max_span_sec = float(max(min_span_sec, float(max_span_sec)))
dbg: Dict[str, Any] = {
"mode": "variable_gop_cumulative_energy",
"target_num_gops": int(target_num_gops),
"bin_sec": float(bin_sec),
"smooth_bins": int(smooth_bins),
"min_span_sec": float(min_span_sec),
"max_span_sec": float(max_span_sec),
"segments": [],
"bins": {},
}
if total_frames <= 0:
seg = {
"gop_idx": 0,
"start_bin": 0,
"end_bin": 0,
"start_sec": 0.0,
"end_sec": 0.0,
"anchor_sec": 0.0,
"energy_sum": 0.0,
}
dbg["segments"] = [seg]
dbg["segment_count"] = 1
dbg["cut_energy_threshold"] = 0.0
return [seg], dbg
centers, energy, bins_dbg = ffprobe_packets_pb_energy_bins(
video_path=str(video_path),
bin_sec=float(bin_sec),
smooth_bins=int(smooth_bins),
)
dbg["bins"] = bins_dbg
duration_sec = float(total_frames) / float(fps_use) if total_frames > 0 else 0.0
duration_sec = max(duration_sec, 0.0)
if energy.size == 0 or centers.size == 0:
segs: List[Dict[str, Any]] = []
n = int(max(1, target_num_gops))
sec_edges = np.linspace(0.0, max(duration_sec, bin_sec), n + 1, dtype=np.float64).tolist()
for gi in range(n):
st_sec = float(sec_edges[gi])
ed_sec = float(sec_edges[gi + 1])
anchor_sec = 0.5 * (st_sec + ed_sec)
segs.append({
"gop_idx": int(gi),
"start_bin": int(gi),
"end_bin": int(gi),
"start_sec": float(st_sec),
"end_sec": float(ed_sec),
"anchor_sec": float(anchor_sec),
"energy_sum": 0.0,
})
dbg["segment_count"] = int(len(segs))
dbg["cut_energy_threshold"] = 0.0
dbg["fallback"] = "uniform_time"
dbg["segments"] = segs
return segs, dbg
energy = np.asarray(energy, dtype=np.float64)
energy = np.maximum(energy, 0.0)
nb = int(len(energy))
total_energy = float(energy.sum())
if (not np.isfinite(total_energy)) or total_energy <= 1e-12:
total_energy = float(nb)
energy = np.ones((nb,), dtype=np.float64)
cut_energy_threshold = float(total_energy) / float(max(1, target_num_gops))
min_bins = int(max(1, math.ceil(float(min_span_sec) / float(bin_sec))))
max_bins = int(max(min_bins, math.ceil(float(max_span_sec) / float(bin_sec))))
dbg["total_energy"] = float(total_energy)
dbg["cut_energy_threshold"] = float(cut_energy_threshold)
dbg["min_bins"] = int(min_bins)
dbg["max_bins"] = int(max_bins)
dbg["cut_points"] = []
segs: List[Dict[str, Any]] = []
start_bin = 0
acc = 0.0
gi = 0
valley_search_radius = int(max(1, min(3, max_bins // 3)))
def _best_cut_bin(start_idx: int, current_idx: int) -> int:
lo = int(max(int(start_idx) + int(min_bins) - 1, int(current_idx) - int(valley_search_radius)))
hi = int(min(int(start_idx) + int(max_bins) - 1, int(nb) - 1, int(current_idx) + int(valley_search_radius)))
if hi < lo:
return int(current_idx)
best = int(current_idx)
best_key = (float(energy[best]), abs(int(best) - int(current_idx)))
for bi in range(int(lo), int(hi) + 1):
cand_key = (float(energy[bi]), abs(int(bi) - int(current_idx)))
if cand_key < best_key:
best = int(bi)
best_key = cand_key
return int(best)
i = 0
while i < nb:
acc += float(energy[i])
span_bins = int(i - start_bin + 1)
reach_min = bool(span_bins >= min_bins)
reach_max = bool(span_bins >= max_bins)
reach_energy = bool(acc >= cut_energy_threshold)
should_cut = False
if reach_max:
should_cut = True
elif reach_min and reach_energy:
should_cut = True
if should_cut:
cut_bin = _best_cut_bin(int(start_bin), int(i))
st_sec = float(start_bin) * float(bin_sec)
ed_sec = min(duration_sec, float(cut_bin + 1) * float(bin_sec))
anchor_bin = int((start_bin + cut_bin) // 2)
anchor_sec = float(centers[max(0, min(nb - 1, anchor_bin))])
seg_energy = float(energy[start_bin:cut_bin + 1].sum())
segs.append({
"gop_idx": int(gi),
"start_bin": int(start_bin),
"end_bin": int(cut_bin),
"start_sec": float(st_sec),
"end_sec": float(ed_sec),
"anchor_sec": float(anchor_sec),
"energy_sum": float(seg_energy),
})
dbg["cut_points"].append({
"gop_idx": int(gi),
"threshold_bin": int(i),
"cut_bin": int(cut_bin),
"cut_sec": float(ed_sec),
})
gi += 1
start_bin = int(cut_bin + 1)
acc = 0.0
i = int(cut_bin + 1)
continue
i += 1
if start_bin < nb:
st_sec = float(start_bin) * float(bin_sec)
ed_sec = min(duration_sec, float(nb) * float(bin_sec))
anchor_bin = int((start_bin + (nb - 1)) // 2)
anchor_sec = float(centers[max(0, min(nb - 1, anchor_bin))])
tail_energy = float(energy[start_bin:nb].sum())
segs.append({
"gop_idx": int(gi),
"start_bin": int(start_bin),
"end_bin": int(nb - 1),
"start_sec": float(st_sec),
"end_sec": float(ed_sec),
"anchor_sec": float(anchor_sec),
"energy_sum": float(tail_energy),
})
# Merge last segment if too short
if len(segs) >= 2:
last = segs[-1]
span_last = float(last["end_sec"]) - float(last["start_sec"])
if span_last < float(min_span_sec) * 0.5:
prev = segs[-2]
prev["end_bin"] = int(last["end_bin"])
prev["end_sec"] = float(last["end_sec"])
prev["energy_sum"] = float(prev["energy_sum"]) + float(last["energy_sum"])
prev_anchor_bin = int((int(prev["start_bin"]) + int(prev["end_bin"])) // 2)
prev["anchor_sec"] = float(centers[max(0, min(nb - 1, prev_anchor_bin))])
segs = segs[:-1]
for gi, seg in enumerate(segs):
seg["gop_idx"] = int(gi)
dbg["segment_count"] = int(len(segs))
dbg["segments"] = [
{
"gop_idx": int(seg["gop_idx"]),
"start_bin": int(seg["start_bin"]),
"end_bin": int(seg["end_bin"]),
"start_sec": float(seg["start_sec"]),
"end_sec": float(seg["end_sec"]),
"anchor_sec": float(seg["anchor_sec"]),
"energy_sum": float(seg["energy_sum"]),
}
for seg in segs
]
return segs, dbg
def sample_frame_ids_by_energy_cdf(
video_path: str,
total_frames: int,
fps: float,
target_count: int,
bin_sec: float = 0.5,
smooth_bins: int = 1,
uniform_mix: float = 0.15,
max_per_bin: int = 16,
) -> Tuple[List[int], Dict[str, Any]]:
"""Sample frame ids with density proportional to PB packet-size energy.
Intuition:
- high-energy time bins get more sampled frames
- low-energy time bins get fewer sampled frames
- a small uniform prior prevents static-but-important segments from being starved
This effectively builds a new non-uniform timeline where "information-dense"
regions are sampled more densely.
"""
total_frames = int(max(0, total_frames))
target_count = int(max(0, target_count))
fps_use = float(fps) if (fps and fps > 0) else 30.0
bin_sec = float(bin_sec) if (bin_sec and bin_sec > 1e-6) else 0.5
smooth_bins = int(max(0, int(smooth_bins)))
uniform_mix = float(max(0.0, min(1.0, float(uniform_mix))))
max_per_bin = int(max(1, int(max_per_bin)))
dbg: Dict[str, Any] = {
"mode": "pkt_energy_cdf",
"target_count": int(target_count),
"bin_sec": float(bin_sec),
"smooth_bins": int(smooth_bins),
"uniform_mix": float(uniform_mix),
"max_per_bin": int(max_per_bin),
"bins": {},
"bin_count": 0,
"selected_before_pad": 0,
"selected_after_pad": 0,
"per_bin_counts": [],
}
if total_frames <= 0 or target_count <= 0:
dbg["error"] = "empty_input"
return [], dbg
centers, energy, bins_dbg = ffprobe_packets_pb_energy_bins(
video_path=str(video_path),
bin_sec=float(bin_sec),
smooth_bins=int(smooth_bins),
)
dbg["bins"] = bins_dbg
if energy.size == 0 or centers.size == 0:
dbg["error"] = "no_energy"
if total_frames == 1:
return [0] * target_count, dbg
out = np.linspace(0, total_frames - 1, target_count, dtype=np.int32).tolist()
dbg["selected_before_pad"] = int(len(out))
dbg["selected_after_pad"] = int(len(out))
return [int(x) for x in out], dbg
nb = int(len(energy))
dbg["bin_count"] = int(nb)
energy = np.asarray(energy, dtype=np.float64)
energy = np.maximum(energy, 0.0)
mean_e = float(energy.mean()) if energy.size > 0 else 0.0
if (not np.isfinite(mean_e)) or mean_e <= 0.0:
mean_e = 1.0
# Mix in a uniform prior so static-but-important regions still get sampled.
weights = (1.0 - uniform_mix) * energy + uniform_mix * mean_e
weights = np.maximum(weights, 1e-12)
weights = weights / float(weights.sum())
ideal = weights * float(target_count)
counts = np.floor(ideal).astype(np.int32)
frac = ideal - counts.astype(np.float64)
# Respect a per-bin cap so a single bursty region does not absorb everything.
counts = np.minimum(counts, int(max_per_bin)).astype(np.int32)
cur = int(counts.sum())
if cur < target_count:
order = np.argsort(-frac)
for idx in order.tolist():
if cur >= target_count:
break
if int(counts[idx]) >= int(max_per_bin):
continue
counts[idx] += 1
cur += 1
if cur < target_count:
order = np.argsort(-weights)
ptr = 0
while cur < target_count and len(order) > 0:
idx = int(order[ptr % len(order)])
if int(counts[idx]) < int(max_per_bin):
counts[idx] += 1
cur += 1
ptr += 1
if ptr > int(target_count * 8 + nb * 8):
break
frame_ids: List[int] = []
per_bin_counts: List[Tuple[int, int]] = []
for bi in range(nb):
c = int(counts[bi])
if c <= 0:
continue
per_bin_counts.append((int(bi), int(c)))
t0 = float(bi) * float(bin_sec)
t1 = float(bi + 1) * float(bin_sec)
f0 = int(math.floor(t0 * fps_use))
f1 = int(math.ceil(t1 * fps_use)) - 1
if total_frames > 0:
f0 = max(0, min(total_frames - 1, f0))
f1 = max(0, min(total_frames - 1, f1))
if f1 < f0:
f1 = f0
if c == 1 or f1 == f0:
picked = [int(round((f0 + f1) * 0.5))]
else:
picked = np.linspace(f0, f1, c, dtype=np.int32).tolist()
frame_ids.extend(int(x) for x in picked)
frame_ids = sorted(int(x) for x in frame_ids)
dbg["per_bin_counts"] = [{"bin": int(b), "count": int(c)} for (b, c) in per_bin_counts]
dbg["selected_before_pad"] = int(len(frame_ids))
if len(frame_ids) == 0:
if total_frames == 1:
frame_ids = [0] * target_count
else:
frame_ids = np.linspace(0, total_frames - 1, target_count, dtype=np.int32).tolist()
if len(frame_ids) < target_count:
if total_frames == 1:
extra = [0] * (target_count - len(frame_ids))
else:
extra = np.linspace(0, total_frames - 1, target_count - len(frame_ids), dtype=np.int32).tolist()
frame_ids = sorted([int(x) for x in frame_ids] + [int(x) for x in extra])
if len(frame_ids) > target_count:
idxs = np.linspace(0, len(frame_ids) - 1, target_count, dtype=np.int32).tolist()
frame_ids = [int(frame_ids[int(i)]) for i in idxs]
dbg["selected_after_pad"] = int(len(frame_ids))
return [int(x) for x in frame_ids], dbg
def pick_windows_by_energy(
video_path: str,
total_frames: int,
fps: float,
window_len_frames: int,
num_candidates: int,
top_k: int,
) -> Tuple[List[Tuple[int, int, int]], Dict[str, Any]]:
"""Return list of selected windows: [(start_f, end_f, score_pkt_sum), ...]
plus debug dict.
"""
total_frames = int(total_frames)
window_len_frames = int(window_len_frames)
num_candidates = int(num_candidates)
top_k = int(top_k)
dbg: Dict[str, Any] = {
"total_frames": total_frames,
"fps": float(fps),
"window_len_frames": window_len_frames,
"num_candidates": num_candidates,
"top_k": top_k,
"candidates": [],
"selected": [],
}
if total_frames <= 0:
return [(0, 0, 0)], dbg
if window_len_frames <= 0 or window_len_frames >= total_frames:
# treat whole video as one window
return [(0, total_frames - 1, 0)], dbg
# candidate window starts evenly spaced in [0, total_frames - window_len]
max_start = max(0, total_frames - window_len_frames)
if num_candidates <= 1:
starts = [0]
else:
starts = np.linspace(0, max_start, num_candidates, dtype=np.int32).tolist()
fps_use = fps if fps and fps > 0 else 30.0
dur_sec = float(window_len_frames) / float(fps_use)
cand = []
for st in starts:
st = int(st)
ed = int(min(total_frames - 1, st + window_len_frames - 1))
start_sec = float(st) / float(fps_use)
score = ffprobe_sum_pkt_size(video_path, start_sec, dur_sec)
cand.append((st, ed, int(score)))
cand_sorted = sorted(cand, key=lambda x: x[2], reverse=True)
dbg["candidates"] = cand_sorted
chosen = cand_sorted[: max(1, top_k)]
dbg["selected"] = chosen
return chosen, dbg
def allocate_frames_across_windows(
windows: List[Tuple[int, int, int]],
seq_len: int,
) -> List[int]:
"""Allocate seq_len frames across windows proportional to scores (fallback equal),
sample uniformly within each window, then merge & pad/truncate to seq_len.
"""
seq_len = int(seq_len)
if seq_len <= 0:
return []
# if only one window
if len(windows) == 1:
st, ed, _ = windows[0]
return np.linspace(st, ed, seq_len, dtype=np.int32).tolist()
scores = np.array([max(0, w[2]) for w in windows], dtype=np.float64)
if scores.sum() <= 0:
# equal allocation
weights = np.ones_like(scores) / float(len(scores))
else:
weights = scores / scores.sum()
# initial per-window counts (at least 1)
counts = np.maximum(1, np.floor(weights * seq_len).astype(int))
# adjust to sum == seq_len
while counts.sum() > seq_len:
i = int(np.argmax(counts))
if counts[i] > 1:
counts[i] -= 1
else:
break
while counts.sum() < seq_len:
i = int(np.argmax(weights))
counts[i] += 1
frame_ids: List[int] = []
for (st, ed, _), c in zip(windows, counts.tolist()):
if c <= 0:
continue
frame_ids.extend(np.linspace(int(st), int(ed), int(c), dtype=np.int32).tolist())
frame_ids = sorted([int(x) for x in frame_ids])
# dedup but keep order (dedup can reduce length)
dedup = []
last = None
for x in frame_ids:
if last is None or x != last:
dedup.append(x)
last = x
frame_ids = dedup
# pad/truncate to seq_len
if len(frame_ids) == 0:
frame_ids = [0] * seq_len
if len(frame_ids) < seq_len:
frame_ids = (frame_ids + [frame_ids[-1]] * (seq_len - len(frame_ids)))[:seq_len]
else:
frame_ids = frame_ids[:seq_len]
return frame_ids
def enforce_time_coverage_frame_ids(
frame_ids: List[int],
total_frames: int,
fps: float,
seq_len: int,
stride_sec: float = 1.0,
) -> List[int]:
"""Best-effort ensure at least one sampled frame per time-bin.
We create mandatory frame ids at roughly the center of each `stride_sec` bin,
merge with existing `frame_ids`, then truncate/pad to `seq_len`.
Note: If duration_bins > seq_len, it is impossible to cover all bins. In that
case we cover as many earliest bins as possible.
"""
seq_len = int(seq_len)
if seq_len <= 0:
return []
fps_use = float(fps) if (fps and fps > 0) else 30.0
total_frames = int(max(0, total_frames))
if total_frames <= 0:
return frame_ids[:seq_len]
stride = float(stride_sec) if (stride_sec and stride_sec > 0) else 1.0
duration_sec = float(total_frames) / float(fps_use)
n_bins = int(math.floor(duration_sec / stride + 1e-9))
n_bins = max(1, n_bins)
# If bins exceed seq_len, we can only cover the first `seq_len` bins.
n_bins_cover = min(int(n_bins), int(seq_len))
mandatory: List[int] = []
for b in range(n_bins_cover):
t_center = (float(b) + 0.5) * stride
fid = int(round(t_center * fps_use))
if total_frames > 0:
fid = max(0, min(total_frames - 1, fid))
mandatory.append(fid)
merged = sorted(set(mandatory + list(frame_ids)))
if len(merged) < seq_len:
# pad with duplicates of the last frame
merged = merged + [merged[-1]] * (seq_len - len(merged))
elif len(merged) > seq_len:
# truncate, but try to keep the mandatory ones
kept = []
mset = set(mandatory)
for x in merged:
if len(kept) >= seq_len:
break
kept.append(x)
merged = kept
return [int(x) for x in merged[:seq_len]]
|