File size: 26,657 Bytes
7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 51d2470 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 51d2470 7d20d33 51d2470 7d20d33 51d2470 7d20d33 7e7f067 7d20d33 4ba026e 7e7f067 4ba026e 7e7f067 7d20d33 4ba026e 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 4ba026e 7d20d33 4ba026e 7d20d33 4ba026e 7d20d33 4ba026e 7e7f067 4ba026e 7e7f067 4ba026e 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 7e7f067 7d20d33 4ba026e 7d20d33 4ba026e 7d20d33 7e7f067 4ba026e 7e7f067 4ba026e 7e7f067 7d20d33 7e7f067 7d20d33 4ba026e 7d20d33 7e7f067 7d20d33 4ba026e 7d20d33 7e7f067 4ba026e 7e7f067 4ba026e 7e7f067 4ba026e 7e7f067 4ba026e 7e7f067 4ba026e 7e7f067 7d20d33 | 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 | """
Comprehensive test suite for ViL Tracker.
16 tests covering all components:
1. mLSTM Cell (LinearHeadwiseExpand correctness + param count)
2. mLSTM Block (full block without MLP)
3. TMoE MLP
4. Backbone (standard, small depth)
5. Backbone (with TMoE + integrated FiLM, medium depth)
6. Prediction Heads
7. FiLM Temporal Modulation
8. Full Tracker (small depth for speed)
9. Loss Functions (all 6)
10. Kalman Filter (8-state, adaptive)
11. Dataset (synthetic)
12. Training Step (mini forward + backward with temporal)
13. Model Summary (FULL depth=24, constraint check)
14. Online Tracker (full inference pipeline)
15. Augmentation pipeline
16. ACL curriculum integration
"""
import sys
import time
import torch
import numpy as np
torch.manual_seed(42)
np.random.seed(42)
PASS = 0
FAIL = 0
def test(name, fn):
global PASS, FAIL
print(f"\nTest {PASS + FAIL + 1}: {name}...", flush=True)
try:
fn()
PASS += 1
print(f" ✅ PASSED")
except Exception as e:
FAIL += 1
print(f" ❌ FAILED: {e}")
import traceback
traceback.print_exc()
def count_params(model):
return sum(p.numel() for p in model.parameters())
# ============================================================
# Test 1: mLSTM Cell
# ============================================================
def test_mlstm_cell():
from vil_tracker.models.mlstm import mLSTMCell, LinearHeadwiseExpand
# Test LinearHeadwiseExpand
lhe = LinearHeadwiseExpand(768, num_heads=192, bias=False)
lhe_params = count_params(lhe)
assert lhe_params == 192 * 4 * 4, f"LHE params: {lhe_params} != {192*4*4}"
x = torch.randn(2, 10, 768)
y = lhe(x)
assert y.shape == (2, 10, 768), f"LHE output shape: {y.shape}"
# Test full mLSTM cell
cell = mLSTMCell(dim=384, proj_factor=2.0, qkv_proj_blocksize=4, num_heads=4)
cell_params = count_params(cell)
print(f" mLSTMCell params: {cell_params:,} ({cell_params/1e6:.3f}M)")
# Should be ~920K, not 2.66M
assert cell_params < 1_000_000, f"Cell has {cell_params:,} params (should be <1M)"
assert cell_params > 800_000, f"Cell has {cell_params:,} params (should be >800K)"
# Verify GroupNorm uses 192 groups (num_proj_heads), not 4 (num_heads)
assert cell.outnorm.num_groups == 192, f"GroupNorm should have 192 groups, got {cell.outnorm.num_groups}"
print(f" GroupNorm groups: {cell.outnorm.num_groups} (correct: per-projection-head)")
x = torch.randn(2, 20, 384)
y = cell(x)
assert y.shape == (2, 20, 384), f"Cell output shape: {y.shape}"
# Test reverse mode
y_rev = cell(x, reverse=True)
assert y_rev.shape == (2, 20, 384), f"Reverse output shape: {y_rev.shape}"
# Forward and reverse should produce different results
assert not torch.allclose(y, y_rev, atol=1e-3), "Forward and reverse should differ"
test("mLSTM Cell (LinearHeadwiseExpand)", test_mlstm_cell)
# ============================================================
# Test 2: mLSTM Block
# ============================================================
def test_mlstm_block():
from vil_tracker.models.mlstm import mLSTMBlock
block = mLSTMBlock(dim=384, proj_factor=2.0, qkv_proj_blocksize=4,
num_heads=4, mlp_ratio=4.0)
params = count_params(block)
print(f" mLSTMBlock params: {params:,} ({params/1e6:.3f}M)")
# No separate MLP — should be ~920K, same as cell + LayerNorm
assert params < 1_050_000, f"Block has {params:,} params (should be <1.05M without MLP)"
x = torch.randn(2, 20, 384)
y = block(x)
assert y.shape == (2, 20, 384), f"Block output shape: {y.shape}"
# Residual connection: output should be close-ish to input at init
diff = (y - x).abs().mean().item()
print(f" Residual diff from input: {diff:.4f}")
test("mLSTM Block (no separate MLP)", test_mlstm_block)
# ============================================================
# Test 3: TMoE MLP
# ============================================================
def test_tmoe():
from vil_tracker.models.backbone import TMoEMLP
tmoe = TMoEMLP(dim=384, mlp_ratio=4.0, num_experts=4)
params = count_params(tmoe)
print(f" TMoEMLP params: {params:,} ({params/1e6:.3f}M)")
x = torch.randn(2, 20, 384)
y = tmoe(x)
assert y.shape == (2, 20, 384), f"TMoE output shape: {y.shape}"
# Test freezing shared expert
tmoe.freeze_shared_expert()
frozen = sum(1 for p in tmoe.shared_expert.parameters() if not p.requires_grad)
total_shared = sum(1 for p in tmoe.shared_expert.parameters())
assert frozen == total_shared, "Shared expert should be fully frozen"
test("TMoE MLP", test_tmoe)
# ============================================================
# Test 4: Backbone (standard, small depth)
# ============================================================
def test_backbone_small():
from vil_tracker.models.backbone import ViLBackbone
backbone = ViLBackbone(dim=384, depth=4, patch_size=16, tmoe_blocks=0)
params = count_params(backbone)
print(f" Backbone (depth=4, no TMoE) params: {params:,} ({params/1e6:.3f}M)")
template = torch.randn(2, 3, 128, 128)
search = torch.randn(2, 3, 256, 256)
t_feat, s_feat = backbone(template, search)
assert t_feat.shape == (2, 64, 384), f"Template feat shape: {t_feat.shape}"
assert s_feat.shape == (2, 256, 384), f"Search feat shape: {s_feat.shape}"
test("Backbone (standard, depth=4)", test_backbone_small)
# ============================================================
# Test 5: Backbone with TMoE + integrated FiLM
# ============================================================
def test_backbone_tmoe_film():
from vil_tracker.models.backbone import ViLBackbone
from vil_tracker.models.film_temporal import TemporalModulationManager
backbone = ViLBackbone(dim=384, depth=6, patch_size=16, tmoe_blocks=2,
num_experts=4, film_interval=3)
params = count_params(backbone)
print(f" Backbone (depth=6, TMoE=2) params: {params:,} ({params/1e6:.3f}M)")
# Create temporal modulation manager
temporal_mod = TemporalModulationManager(dim=384, num_blocks=6, modulation_interval=3)
template = torch.randn(1, 3, 128, 128)
search = torch.randn(1, 3, 256, 256)
# First pass: no temporal context yet
t_feat, s_feat = backbone(template, search, temporal_mod_manager=temporal_mod)
assert t_feat.shape == (1, 64, 384), f"Template feat shape: {t_feat.shape}"
assert s_feat.shape == (1, 256, 384), f"Search feat shape: {s_feat.shape}"
# Second pass: temporal context should be active now
t_feat2, s_feat2 = backbone(template, search, temporal_mod_manager=temporal_mod)
# Output should differ when temporal modulation is active
assert t_feat2.shape == (1, 64, 384)
print(f" FiLM modulation active: features differ = {not torch.allclose(t_feat, t_feat2, atol=1e-5)}")
test("Backbone (TMoE + integrated FiLM)", test_backbone_tmoe_film)
# ============================================================
# Test 6: Prediction Heads
# ============================================================
def test_heads():
from vil_tracker.models.heads import CenterHead, UncertaintyHead, decode_predictions, create_hanning_window
center_head = CenterHead(dim=384, feat_size=16)
unc_head = UncertaintyHead(dim=384, feat_size=16)
print(f" CenterHead params: {count_params(center_head):,}")
print(f" UncertaintyHead params: {count_params(unc_head):,}")
search_feat = torch.randn(2, 256, 384)
preds = center_head(search_feat)
assert preds['heatmap'].shape == (2, 1, 16, 16), f"Heatmap shape: {preds['heatmap'].shape}"
assert preds['size'].shape == (2, 2, 16, 16), f"Size shape: {preds['size'].shape}"
assert preds['offset'].shape == (2, 2, 16, 16), f"Offset shape: {preds['offset'].shape}"
# Decode without Hanning
boxes, scores = decode_predictions(preds['heatmap'], preds['size'], preds['offset'])
assert boxes.shape == (2, 4), f"Boxes shape: {boxes.shape}"
assert scores.shape == (2,), f"Scores shape: {scores.shape}"
# Decode WITH Hanning window
hann = create_hanning_window(16)
assert hann.shape == (16, 16), f"Hanning shape: {hann.shape}"
assert abs(hann[8, 8].item() - 1.0) < 0.05, f"Hanning center should be ~1.0, got {hann[8, 8]}"
assert hann[0, 0].item() < 0.01, f"Hanning corner should be ~0, got {hann[0, 0]}"
boxes_h, scores_h = decode_predictions(preds['heatmap'], preds['size'], preds['offset'],
hanning_window=hann)
assert boxes_h.shape == (2, 4), f"Hanning boxes shape: {boxes_h.shape}"
print(f" Hanning window: center={hann[8,8]:.3f}, corner={hann[0,0]:.6f}")
print(f" Without Hanning: box={boxes[0].tolist()}, score={scores[0].item():.4f}")
print(f" With Hanning: box={boxes_h[0].tolist()}, score={scores_h[0].item():.4f}")
# Uncertainty
log_var = unc_head(search_feat)
assert log_var.shape == (2, 1, 16, 16), f"Log variance shape: {log_var.shape}"
test("Prediction Heads", test_heads)
# ============================================================
# Test 7: FiLM Temporal Modulation
# ============================================================
def test_film():
from vil_tracker.models.film_temporal import (
TemporalReliabilityCalibrator,
FiLMTemporalModulation,
TemporalModulationManager,
)
# Test individual components
calib = TemporalReliabilityCalibrator(384)
film = FiLMTemporalModulation(384)
x = torch.randn(2, 20, 384)
tc = torch.randn(2, 20, 384)
rel = calib(tc)
assert rel.shape == (2, 20, 1), f"Reliability shape: {rel.shape}"
assert (rel >= 0).all() and (rel <= 1).all(), "Reliability not in [0,1]"
modulated = film(x, tc, rel)
assert modulated.shape == (2, 20, 384), f"Modulated shape: {modulated.shape}"
# Test manager
manager = TemporalModulationManager(dim=384, num_blocks=24, modulation_interval=6)
print(f" TemporalModulationManager params: {count_params(manager):,}")
# First call: no temporal context yet, should return unchanged
y = manager.modulate(x, block_idx=5)
assert torch.allclose(y, x), "Should return unchanged without temporal context"
# Update context and try again
manager.update_temporal_context(x)
y = manager.modulate(x, block_idx=5) # block 5 → (5+1)%6==0, should modulate
assert y.shape == (2, 20, 384)
# Test reset
manager.reset()
y = manager.modulate(x, block_idx=5)
assert torch.allclose(y, x), "After reset, should return unchanged"
test("FiLM Temporal Modulation", test_film)
# ============================================================
# Test 8: Full Tracker (small depth for speed)
# ============================================================
def test_full_tracker_small():
from vil_tracker.models.tracker import ViLTracker, get_default_config
config = get_default_config()
config['depth'] = 4
config['tmoe_blocks'] = 1
config['film_interval'] = 2
tracker = ViLTracker(config)
params = count_params(tracker)
print(f" Tracker (depth=4) params: {params:,} ({params/1e6:.3f}M)")
B, K = 2, 3
template = torch.randn(B, 3, 128, 128)
# Test single-frame (backward compat)
search_single = torch.randn(B, 3, 256, 256)
output_s = tracker(template, search_single, use_temporal=False)
assert output_s['heatmap'].shape == (B, 1, 16, 16), f"Single heatmap: {output_s['heatmap'].shape}"
assert output_s['boxes'].shape == (B, 4), f"Single boxes: {output_s['boxes'].shape}"
assert output_s['scores'].shape == (B,), f"Single scores: {output_s['scores'].shape}"
print(f" Single-frame: boxes={output_s['boxes'][0].tolist()}")
# Test multi-frame sequence
searches = torch.randn(B, K, 3, 256, 256)
output_m = tracker(template, searches, use_temporal=True)
assert output_m['heatmap'].shape == (B, K, 1, 16, 16), f"Multi heatmap: {output_m['heatmap'].shape}"
assert output_m['boxes'].shape == (B, K, 4), f"Multi boxes: {output_m['boxes'].shape}"
assert output_m['scores'].shape == (B, K), f"Multi scores: {output_m['scores'].shape}"
assert output_m['search_feats'].shape == (B, K, 256, 384), f"Multi feats: {output_m['search_feats'].shape}"
print(f" Multi-frame (K={K}): frame 0 box={output_m['boxes'][0,0].tolist()}")
print(f" frame 2 box={output_m['boxes'][0,2].tolist()}")
tracker.reset_temporal()
test("Full Tracker (single + multi-frame)", test_full_tracker_small)
# ============================================================
# Test 9: Loss Functions (all 6)
# ============================================================
def test_losses():
from vil_tracker.training.losses import (
FocalLoss, GIoULoss, UncertaintyNLLLoss,
MemoryContrastiveLoss, AFKDDistillationLoss,
ADWLoss, CombinedTrackingLoss,
)
B = 4
# Focal loss
focal = FocalLoss()
pred_hm = torch.randn(B, 1, 16, 16)
gt_hm = torch.zeros(B, 1, 16, 16)
gt_hm[:, :, 8, 8] = 1.0
fl = focal(pred_hm, gt_hm)
print(f" Focal loss: {fl.item():.4f}")
assert fl.item() > 0, "Focal loss should be positive"
# GIoU loss
giou = GIoULoss()
pred_box = torch.tensor([[128.0, 128.0, 50.0, 50.0]] * B)
gt_box = torch.tensor([[130.0, 130.0, 48.0, 48.0]] * B)
gl = giou(pred_box, gt_box)
print(f" GIoU loss: {gl.item():.4f}")
assert 0 <= gl.item() <= 2, f"GIoU loss out of range: {gl.item()}"
# Uncertainty NLL loss
unc = UncertaintyNLLLoss()
pred_v = torch.randn(B, 4)
target_v = torch.randn(B, 4)
log_var = torch.zeros(B, 4) # unit variance
ul = unc(pred_v, target_v, log_var)
print(f" Uncertainty NLL loss: {ul.item():.4f}")
assert ul.item() > 0
# Contrastive loss
contrastive = MemoryContrastiveLoss()
feat_a = torch.randn(B, 384)
feat_b = feat_a + torch.randn(B, 384) * 0.1
cl = contrastive(feat_a, feat_b)
print(f" Contrastive loss: {cl.item():.4f}")
# AFKD distillation loss
afkd = AFKDDistillationLoss(student_dim=384, teacher_dim=768)
student_feat = torch.randn(B, 256, 384)
teacher_feat = torch.randn(B, 256, 768)
dl = afkd(student_feat, teacher_feat)
print(f" AFKD distillation loss: {dl.item():.4f}")
assert dl.item() > 0
# ADW loss
adw = ADWLoss(num_tasks=3)
losses = [torch.tensor(1.0), torch.tensor(0.5), torch.tensor(2.0)]
al = adw(losses)
print(f" ADW loss: {al.item():.4f}")
# Combined loss
combined = CombinedTrackingLoss()
pred = {
'heatmap': pred_hm,
'size': torch.rand(B, 2, 16, 16),
'boxes': pred_box,
'log_variance': torch.randn(B, 1, 16, 16),
}
loss_dict = combined(pred, gt_hm, torch.tensor([[0.2, 0.2]] * B), gt_box)
print(f" Combined loss: {loss_dict['total'].item():.4f}")
assert loss_dict['total'].item() > 0
test("Loss Functions (all 6)", test_losses)
# ============================================================
# Test 10: Kalman Filter
# ============================================================
def test_kalman():
from vil_tracker.inference.kalman import KalmanFilter
kf = KalmanFilter()
assert not kf.initialized
# Initialize
init_box = np.array([100.0, 100.0, 50.0, 50.0])
kf.initialize(init_box)
assert kf.initialized
# Predict + update cycle with moving target
for i in range(10):
pred = kf.predict()
assert len(pred) == 4, f"Prediction length: {len(pred)}"
# Simulate noisy measurement of linearly moving target
noise = np.random.randn(4) * 2
meas = init_box + np.array([i * 2, i * 1, 0, 0]) + noise
kf.update(meas, uncertainty=1.0)
state = kf.get_state()
print(f" Final state: cx={state[0]:.1f}, cy={state[1]:.1f}, w={state[2]:.1f}, h={state[3]:.1f}")
assert state[2] > 0 and state[3] > 0, "Width/height should be positive"
# Test outlier rejection (chi-squared gating)
kf.update(np.array([500.0, 500.0, 50.0, 50.0]), uncertainty=1.0) # Far outlier
state_after = kf.get_state()
# State should NOT have jumped to 500,500
assert state_after[0] < 200, f"Outlier should be rejected, cx={state_after[0]}"
test("Kalman Filter (8-state, adaptive)", test_kalman)
# ============================================================
# Test 11: Dataset (synthetic)
# ============================================================
def test_dataset():
from vil_tracker.data.dataset import SyntheticTrackingDataset, TrackingDataset
ds = SyntheticTrackingDataset(length=100, clip_length=3)
assert len(ds) == 100
sample = ds[0]
assert sample['template'].shape == (3, 128, 128), f"Template shape: {sample['template'].shape}"
assert sample['searches'].shape == (3, 3, 256, 256), f"Searches shape: {sample['searches'].shape}"
assert sample['heatmaps'].shape == (3, 1, 16, 16), f"Heatmaps shape: {sample['heatmaps'].shape}"
assert sample['sizes'].shape == (3, 2), f"Sizes shape: {sample['sizes'].shape}"
assert sample['boxes'].shape == (3, 4), f"Boxes shape: {sample['boxes'].shape}"
# Verify target moves across frames (not static)
cx_f0 = sample['boxes'][0, 0].item()
cx_f2 = sample['boxes'][2, 0].item()
print(f" Frame 0 cx: {cx_f0:.1f}, Frame 2 cx: {cx_f2:.1f} (moving target)")
# Check ACL difficulty changes motion magnitude
ds.set_acl_difficulty(0.0)
easy_sample = ds[42]
ds.set_acl_difficulty(1.0)
hard_sample = ds[42]
print(f" Easy frame spread: {(easy_sample['boxes'][:, 0].max() - easy_sample['boxes'][:, 0].min()).item():.1f} px")
print(f" Hard frame spread: {(hard_sample['boxes'][:, 0].max() - hard_sample['boxes'][:, 0].min()).item():.1f} px")
# Test backward-compatible alias
ds2 = TrackingDataset(synthetic=True, synthetic_length=50, clip_length=3)
assert len(ds2) == 50
sample2 = ds2[0]
assert sample2['searches'].shape[0] == 3, "Clip length should be 3"
test("Dataset (synthetic + backward compat)", test_dataset)
# ============================================================
# Test 12: Training Step (with temporal modulation)
# ============================================================
def test_training_step():
from vil_tracker.models.tracker import ViLTracker, get_default_config
from vil_tracker.training.losses import CombinedTrackingLoss, MemoryContrastiveLoss
from vil_tracker.models.heads import generate_heatmap
config = get_default_config()
config['depth'] = 2
config['tmoe_blocks'] = 0
config['film_interval'] = 2
model = ViLTracker(config)
model.train()
loss_fn = CombinedTrackingLoss()
contrastive_loss = MemoryContrastiveLoss()
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
B, K = 2, 3
template = torch.randn(B, 3, 128, 128)
searches = torch.randn(B, K, 3, 256, 256)
# GT targets for K frames
gt_heatmaps = torch.zeros(B, K, 1, 16, 16)
gt_heatmaps[:, :, :, 8, 8] = 1.0 # center
gt_sizes = torch.tensor([[[0.2, 0.3]] * K] * B)
gt_boxes = torch.tensor([[[128.0, 128.0, 51.2, 76.8]] * K] * B)
# Forward WITH temporal modulation, multi-frame
pred = model(template, searches, use_temporal=True)
assert pred['heatmap'].shape == (B, K, 1, 16, 16), f"Heatmap shape: {pred['heatmap'].shape}"
assert pred['boxes'].shape == (B, K, 4), f"Boxes shape: {pred['boxes'].shape}"
assert pred['scores'].shape == (B, K), f"Scores shape: {pred['scores'].shape}"
assert pred['search_feats'].shape == (B, K, 256, 384), f"Search feats: {pred['search_feats'].shape}"
# Accumulate loss over K frames
total_loss = torch.tensor(0.0)
for k in range(K):
pred_k = {
'heatmap': pred['heatmap'][:, k],
'size': pred['size'][:, k],
'boxes': pred['boxes'][:, k],
}
if 'log_variance' in pred:
pred_k['log_variance'] = pred['log_variance'][:, k]
loss_dict = loss_fn(pred_k, gt_heatmaps[:, k], gt_sizes[:, k], gt_boxes[:, k])
total_loss = total_loss + loss_dict['total']
total_loss = total_loss / K
# Add contrastive loss
t_pooled = pred['template_feat'].mean(dim=1)
s_pooled = pred['search_feats'][:, -1].mean(dim=1)
c_loss = contrastive_loss(t_pooled, s_pooled)
total_loss = total_loss + 0.1 * c_loss
# Backward
total_loss.backward()
has_grads = sum(1 for p in model.parameters() if p.grad is not None)
total_params_count = sum(1 for p in model.parameters())
print(f" Total loss: {total_loss.item():.4f} (K={K} frames, contr={c_loss.item():.4f})")
print(f" Params with gradients: {has_grads}/{total_params_count}")
optimizer.step()
optimizer.zero_grad()
assert total_loss.item() > 0
assert has_grads > 0
test("Training Step (K=3 sequence + contrastive)", test_training_step)
# ============================================================
# Test 13: Model Summary (FULL depth=24, constraint check)
# ============================================================
def test_model_summary():
from vil_tracker.models.tracker import ViLTracker, get_default_config
from vil_tracker.utils.helpers import print_model_summary
config = get_default_config()
model = ViLTracker(config)
summary = print_model_summary(model, config)
total_m = summary['total_params'] / 1e6
# HARD CONSTRAINTS
assert summary['param_ok'], f"FAIL: {total_m:.2f}M params exceeds 50M limit"
assert summary['size_ok'], f"FAIL: {summary['size_fp16_mb']:.1f}MB exceeds 500MB limit"
# GFLOPs is approximate, warn but don't fail if close
if not summary['flop_ok']:
print(f" ⚠️ GFLOPs estimate ({summary['gflops']:.2f}) exceeds 20, but this is approximate")
test("Model Summary (full depth=24)", test_model_summary)
# ============================================================
# Test 14: Online Tracker (inference pipeline)
# ============================================================
def test_online_tracker():
from vil_tracker.models.tracker import ViLTracker, get_default_config
from vil_tracker.inference.online_tracker import OnlineTracker
config = get_default_config()
config['depth'] = 2
config['tmoe_blocks'] = 0
config['film_interval'] = 2
model = ViLTracker(config)
model.eval()
tracker = OnlineTracker(model, device='cpu', template_size=128, search_size=256)
# Simulate a sequence: 480x640 frames with a moving rectangle
H, W = 480, 640
init_bbox = [200, 200, 60, 80] # [x, y, w, h]
# First frame
frame0 = np.random.randint(0, 255, (H, W, 3), dtype=np.uint8)
# Draw target
x, y, w, h = init_bbox
frame0[y:y+h, x:x+w] = [255, 0, 0] # Red rectangle
tracker.initialize(frame0, init_bbox)
# Track for 5 frames
for i in range(1, 6):
frame = np.random.randint(0, 255, (H, W, 3), dtype=np.uint8)
# Move target
nx = x + i * 5
ny = y + i * 3
frame[ny:ny+h, nx:nx+w] = [255, 0, 0]
bbox = tracker.track(frame)
assert len(bbox) == 4, f"Bbox should have 4 elements, got {len(bbox)}"
assert all(isinstance(v, (int, float, np.floating)) for v in bbox), f"Bbox values: {bbox}"
print(f" Frame {i}: predicted [{bbox[0]:.1f}, {bbox[1]:.1f}, {bbox[2]:.1f}, {bbox[3]:.1f}]")
print(f" Online tracker completed 5-frame sequence")
test("Online Tracker (inference pipeline)", test_online_tracker)
# ============================================================
# Test 15: Augmentation pipeline
# ============================================================
def test_augmentation():
from vil_tracker.data.dataset import TrackingAugmentation
aug = TrackingAugmentation(
brightness=0.2,
contrast=0.2,
horizontal_flip_prob=1.0, # Force flip to test bbox update
grayscale_prob=0.0,
blur_prob=0.0,
)
template = torch.rand(3, 128, 128)
search = torch.rand(3, 256, 256)
bbox = torch.tensor([128.0, 128.0, 50.0, 50.0]) # [cx, cy, w, h]
t_aug, s_aug, b_aug = aug(template, search, bbox)
assert t_aug.shape == (3, 128, 128), f"Aug template shape: {t_aug.shape}"
assert s_aug.shape == (3, 256, 256), f"Aug search shape: {s_aug.shape}"
assert b_aug.shape == (4,), f"Aug bbox shape: {b_aug.shape}"
# With flip_prob=1.0, cx should be flipped: new_cx = W - old_cx = 256 - 128 = 128
print(f" Original bbox: {bbox.tolist()}")
print(f" Augmented bbox: {b_aug.tolist()}")
assert abs(b_aug[0].item() - (256 - 128)) < 1.0, f"Flipped cx should be ~128, got {b_aug[0]}"
test("Augmentation pipeline", test_augmentation)
# ============================================================
# Test 16: ACL curriculum integration
# ============================================================
def test_acl_curriculum():
from vil_tracker.data.dataset import SyntheticTrackingDataset
ds = SyntheticTrackingDataset(length=100, acl_difficulty=0.0, clip_length=3)
# Easy: targets barely move
easy_spreads = []
for i in range(20):
sample = ds[i]
spread = (sample['boxes'][:, 0].max() - sample['boxes'][:, 0].min()).item()
easy_spreads.append(spread)
ds.set_acl_difficulty(1.0)
hard_spreads = []
for i in range(20):
sample = ds[i]
spread = (sample['boxes'][:, 0].max() - sample['boxes'][:, 0].min()).item()
hard_spreads.append(spread)
avg_easy = np.mean(easy_spreads)
avg_hard = np.mean(hard_spreads)
print(f" Avg cx spread (easy, d=0.0): {avg_easy:.1f} px")
print(f" Avg cx spread (hard, d=1.0): {avg_hard:.1f} px")
print(f" Hard > Easy: {avg_hard > avg_easy}")
test("ACL curriculum integration", test_acl_curriculum)
# ============================================================
# Summary
# ============================================================
print("\n" + "=" * 60)
print(f"Results: {PASS}/{PASS + FAIL} tests passed")
if FAIL > 0:
print(f" ❌ {FAIL} test(s) FAILED")
sys.exit(1)
else:
print(" ✅ All tests passed!")
sys.exit(0)
|