File size: 24,218 Bytes
a95e79a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9af2926
 
 
 
a95e79a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9af2926
a95e79a
 
 
 
9af2926
a95e79a
 
 
 
 
9af2926
a95e79a
 
 
 
9af2926
a95e79a
 
 
 
9af2926
a95e79a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9af2926
 
 
 
 
 
 
 
a95e79a
 
 
 
 
 
 
 
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
"""EC-SimToken training script.

Adds existence head + synthetic null augmentation to SimToken.

Key differences from train.py:
  - Uses ECSimtoken_ForCausalLM (adds existence_head: Linear(256,1))
  - Audio-swap null augmentation: p_null fraction of batch items have
    their audio replaced with another sample's audio β†’ synthetic null
  - is_null tensor passed to model_forward to gate mask loss
  - test_n evaluation uses existence head (p_exist threshold) for Null S

Usage (training):
    python train_ec_simtoken.py \
        --data_dir data \
        --mllm Chat-UniVi/Chat-UniVi-7B-v1.5 \
        --vision_pretrained path/to/sam_vit_h_4b8939.pth \
        --name ec_simtoken_v1 \
        --epochs 10 \
        --batch_size 12 \
        --null_aug_prob 0.25 \
        --exist_loss_weight 1.0

Usage (eval only):
    python train_ec_simtoken.py --run eval \
        --saved_model checkpoints/ec_simtoken_v1.pth \
        --eval_splits test_s,test_u,test_n
"""

import argparse
import os
import random
import warnings
from functools import partial

import numpy as np
import torch
import torch.multiprocessing as mp
import transformers
from peft import LoraConfig, get_peft_model
from torch.optim import AdamW
from torch.utils.data import DataLoader
from tqdm import tqdm
from transformers import AutoConfig, get_cosine_schedule_with_warmup, logging

warnings.filterwarnings("ignore")
logging.set_verbosity_error()

import re

# ── Token constants ───────────────────────────────────────────────────────────

IGNORE_INDEX = -100
IMAGE_TOKEN_INDEX = -200
AUDIO_TOKEN_INDEX = -300


# ── Args: base (from configs) + EC-SimToken extensions ───────────────────────

from configs import args as base_args  # parses base SimToken args

_ec_parser = argparse.ArgumentParser(add_help=False)
_ec_parser.add_argument("--null_aug_prob", type=float, default=0.25,
                         help="Fraction of batch items with swapped audio (null aug)")
_ec_parser.add_argument("--exist_loss_weight", type=float, default=1.0,
                         help="Weight for BCE existence loss")
_ec_parser.add_argument("--exist_threshold", type=float, default=0.5,
                         help="p_exist sigmoid threshold for null classification")

ec_args, _ = _ec_parser.parse_known_args()
# Merge EC-SimToken args into the base args namespace
args = base_args
args.null_aug_prob = ec_args.null_aug_prob
args.exist_loss_weight = ec_args.exist_loss_weight
args.exist_threshold = ec_args.exist_threshold

os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id


from datasets import REFAVS
from models.ec_simtoken_model import ECSimtoken_ForCausalLM
from utils import utility


# ── Utilities ─────────────────────────────────────────────────────────────────

def set_seed(seed: int = 42):
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    os.environ["PYTHONHASHSEED"] = str(seed)
    # benchmark=True lets cuDNN find fastest conv algorithms (important for SAM)
    torch.backends.cudnn.benchmark = True


def seed_worker(worker_id):
    seed = torch.initial_seed() % 2 ** 32
    np.random.seed(seed)
    random.seed(seed)


def dict_to_cuda(d: dict) -> dict:
    for k, v in d.items():
        if isinstance(v, torch.Tensor):
            d[k] = v.cuda(non_blocking=True)
        elif isinstance(v, list) and v and isinstance(v[0], torch.Tensor):
            d[k] = [x.cuda(non_blocking=True) for x in v]
    return d


# ── Null augmentation ─────────────────────────────────────────────────────────

def apply_null_augmentation(
    audio_feats: list, p_null: float = 0.25
) -> tuple[list, torch.BoolTensor]:
    """Randomly replace some audio features with mismatched ones.

    Returns the (possibly mutated) audio_feats list and a bool tensor
    `is_null` where True means the sample's audio was swapped.
    """
    B = len(audio_feats)
    is_null = torch.zeros(B, dtype=torch.bool)
    if B < 2 or p_null <= 0.0:
        return audio_feats, is_null

    for i in range(B):
        if random.random() < p_null:
            candidates = [j for j in range(B) if j != i]
            j = random.choice(candidates)
            audio_feats[i] = audio_feats[j].clone()
            is_null[i] = True

    return audio_feats, is_null


# ── Collate (identical to train.py) ──────────────────────────────────────────

def tokenizer_image_audio_token(
    prompt, tokenizer,
    image_token_index=IMAGE_TOKEN_INDEX,
    audio_token_index=AUDIO_TOKEN_INDEX,
    num_frames=10,
    return_tensors=None,
):
    prompt_chunks = re.split(r'(<image>|<audio>|<video>)', prompt)
    prompt_chunks = [c for c in prompt_chunks if c]

    text_chunks, token_types = [], []
    for chunk in prompt_chunks:
        if chunk == "<image>":
            token_types.append("image")
        elif chunk == "<audio>":
            token_types.append("audio")
        elif chunk == "<video>":
            token_types.append("video")
        else:
            text_chunks.append(chunk)

    tokenized_chunks = [tokenizer(c).input_ids for c in text_chunks]

    input_ids = []
    offset = 0
    if tokenized_chunks and tokenized_chunks[0] and tokenized_chunks[0][0] == tokenizer.bos_token_id:
        offset = 1
        input_ids.append(tokenized_chunks[0][0])

    min_len = min(len(text_chunks), len(token_types))
    for i in range(min_len):
        input_ids.extend(tokenized_chunks[i][offset:])
        if token_types[i] == "image":
            input_ids.append(image_token_index)
        elif token_types[i] == "audio":
            input_ids.append(audio_token_index)
        elif token_types[i] == "video":
            input_ids.extend([image_token_index] * num_frames)
    if len(text_chunks) > min_len:
        input_ids.extend(tokenized_chunks[min_len][offset:])

    if return_tensors == "pt":
        return torch.tensor(input_ids, dtype=torch.long)
    return input_ids


def collate_fn(batch, tokenizer=None):
    vids, images, image_clips, masks, conversations = [], [], [], [], []
    audio_feats, image_feats, resizes, orgsizes = [], [], [], []
    refs, refs_num, fids = [], [], []

    for data in batch:
        vids.append(data["vid"])
        images.append(data["image"])
        image_clips.append(data["img_clip"])
        masks.append(data["mask"])
        conversations.append(data["conversation"])
        audio_feats.append(data["feat_aud"])
        resizes.append(data["resize"])
        orgsizes.append(data["orgsize"])
        image_feats.append(data["feat_sam"])
        refs_num.append(len(data["ref"]))
        fids.append(data["fids"])
        refs.append(data["ref"][0])

    input_ids = [
        tokenizer_image_audio_token(c, tokenizer, return_tensors="pt")
        for c in conversations
    ]
    input_ids = torch.nn.utils.rnn.pad_sequence(
        input_ids, batch_first=True, padding_value=tokenizer.pad_token_id
    )
    attention_masks = input_ids.ne(tokenizer.pad_token_id)

    ref_ids = [
        tokenizer_image_audio_token(r, tokenizer, return_tensors="pt") for r in refs
    ]

    labels = input_ids.clone()
    sep = "Sure, it is [SEG]"
    for conversation, target in zip(conversations, labels):
        parts = conversation.split(sep)
        cur_len = 1
        target[:cur_len] = IGNORE_INDEX
        sep_len = len(tokenizer_image_audio_token(sep, tokenizer)) - 1
        for i in range(len(parts) - 1):
            part_len = len(tokenizer_image_audio_token(parts[i], tokenizer)) - 2
            target[cur_len : cur_len + part_len] = IGNORE_INDEX
            cur_len += part_len + sep_len
        target[cur_len:] = IGNORE_INDEX

    return {
        "vids": vids,
        "images": images,
        "images_clip": image_clips,
        "masks": masks,
        "convs": conversations,
        "input_ids": input_ids,
        "attention_masks": attention_masks,
        "labels": labels,
        "audio_feats": audio_feats,
        "resizes": resizes,
        "orgsizes": orgsizes,
        "image_feats": image_feats,
        "ref_ids": ref_ids,
        "refs_num": refs_num,
        "fids": fids,
    }


# ── Build model ───────────────────────────────────────────────────────────────

def build_model(args, tokenizer, seg_token_idx) -> ECSimtoken_ForCausalLM:
    model_args = {
        "train_mask_decoder": True,
        "out_dim": 256,
        "ce_loss_weight": 1.0,
        "dice_loss_weight": 0.5,
        "bce_loss_weight": 2.0,
        "seg_token_idx": seg_token_idx,
        "vision_pretrained": args.vision_pretrained,
        "vision_tower": args.vision_tower,
        "use_im_start_end": False,
        "compress": args.compress,
        "start": args.start,
        "exist_loss_weight": args.exist_loss_weight,
    }
    model = ECSimtoken_ForCausalLM.from_pretrained(
        args.mllm, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, **model_args
    )
    model.config.eos_token_id = tokenizer.eos_token_id
    model.config.bos_token_id = tokenizer.bos_token_id
    model.config.pad_token_id = tokenizer.pad_token_id

    model.enable_input_require_grads()
    # Gradient checkpointing trades compute for memory (recomputes activations
    # during backward instead of storing them).  Measured memory at batch=16:
    # 78 GB / 82 GB β€” too close to OOM, so this must stay enabled.
    # model.gradient_checkpointing_enable()

    model.get_model().initialize_vision_modules(model.get_model().config)
    vision_tower = model.get_model().get_vision_tower()
    vision_tower.to(dtype=torch.bfloat16, device="cuda")

    cfg_pt = AutoConfig.from_pretrained(args.mllm)
    cfg_pt.use_cluster = True
    cfg_pt.freeze = False
    cfg_pt.mm_tune = True
    cfg_pt.spatial_cluster_rate0 = 64
    cfg_pt.spatial_cluster_rate1 = 32
    cfg_pt.spatial_cluster_rate2 = 16
    cfg_pt.temporal_cluster_rate = 0.0625
    cfg_pt.vision_tune = False
    model.get_model().initialize_cluster_modules(cfg_pt)
    model.get_model().initialize_lisa_modules(model.get_model().config)

    for p in vision_tower.parameters():
        p.requires_grad = False
    for p in model.get_model().mm_projector.parameters():
        p.requires_grad = False

    # LoRA
    lora_r = 8

    def find_linear_layers(m, targets):
        names = set()
        skip = {"visual_model", "vision_tower", "mm_projector",
                "text_hidden_fcs", "audio_feature_layer", "existence_head"}
        for name, mod in m.named_modules():
            if (
                isinstance(mod, torch.nn.Linear)
                and not any(s in name for s in skip)
                and any(t in name for t in targets)
            ):
                names.add(name)
        return sorted(names)

    lora_config = LoraConfig(
        r=lora_r,
        lora_alpha=16,
        target_modules=find_linear_layers(model, ["q_proj", "v_proj"]),
        lora_dropout=0.05,
        bias="none",
        task_type="CAUSAL_LM",
    )
    model = get_peft_model(model, lora_config)
    model.print_trainable_parameters()

    model = model.to("cuda")
    # Cast everything to bfloat16 β€” from_pretrained only converts checkpoint tensors;
    # modules added post-init (existence_head, audio_feature_layer) default to fp32.
    model = model.to(torch.bfloat16)
    model.resize_token_embeddings(len(tokenizer))

    # Ensure key modules are trainable
    for n, p in model.named_parameters():
        if any(x in n for x in ["lm_head", "embed_tokens", "mask_decoder",
                                  "text_hidden_fcs", "audio_feature_layer",
                                  "existence_head"]):
            p.requires_grad = True

    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    print(f"Trainable parameters: {trainable:,}")
    return model


# ── Evaluation ────────────────────────────────────────────────────────────────

@torch.no_grad()
def evaluate(model, dataloader, split_name: str, exist_threshold: float = 0.5):
    model.eval()
    total_iou = total_fscore = count = 0.0
    # For test_n: existence-gated null metric (lower Null_S is better)
    total_null_metric = null_count = 0.0
    null_tp = 0  # correctly predicted null (p_exist < threshold)
    null_fn = 0  # missed null detection (p_exist >= threshold)

    for batch in tqdm(dataloader, desc=f"Eval {split_name}", leave=False):
        batch = dict_to_cuda(batch)
        with torch.autocast("cuda", dtype=torch.bfloat16):
            out = model.forward(
                images=batch["images"],
                images_clip=batch["images_clip"],
                audio_features=batch["audio_feats"],
                image_features=batch["image_feats"],
                input_ids=batch["input_ids"],
                labels=batch["labels"],
                attention_masks=batch["attention_masks"],
                masks_list=batch["masks"],
                resize_list=batch["resizes"],
                orgsize_list=batch["orgsizes"],
                conversation_list=batch["convs"],
                refs_num=batch["refs_num"],
                fids=batch["fids"],
                vids=batch["vids"],
                ref_ids=batch["ref_ids"],
                inference=True,
            )
        pred_masks = out["pred_masks"]
        gt_masks = out["gt_masks"]
        # exist_logit shape [seg_num, 1]; refs_num==1 per sample so seg_num==B
        p_exist = torch.sigmoid(out["exist_logit"]).squeeze(-1).cpu()  # [B]

        for i in range(len(pred_masks)):
            pred_i = pred_masks[i]
            gt_i = gt_masks[i]
            pe = p_exist[i].item()

            if split_name == "test_n":
                # p_exist < threshold β†’ correctly detect null β†’ output empty mask
                if pe < exist_threshold:
                    null_score = utility.metric_s_for_null(torch.zeros_like(pred_i))
                    null_tp += 1
                else:
                    null_score = utility.metric_s_for_null(pred_i)
                    null_fn += 1
                total_null_metric += null_score * pred_i.shape[0] * pred_i.shape[1]
                null_count += pred_i.shape[0] * pred_i.shape[1]
            else:
                iou = utility.mask_iou(pred_i, gt_i)
                fscore = utility.Eval_Fmeasure(pred_i, gt_i, None)
                n = pred_i.shape[0] * pred_i.shape[1]
                total_iou += iou * n
                total_fscore += fscore * n
                count += n

    if split_name == "test_n":
        null_s = total_null_metric / (null_count + 1e-8)
        total_n = null_tp + null_fn
        print(f"\n  [{split_name}]  Null_S={null_s:.4f}  "
              f"null_tp={null_tp}/{total_n}  null_fn={null_fn}/{total_n}")
        return {"null_s": null_s, "null_tp": null_tp, "null_fn": null_fn}
    else:
        miou = total_iou / (count + 1e-8)
        fscore = total_fscore / (count + 1e-8)
        print(f"\n  [{split_name}]  mIoU={miou:.4f}  F={fscore:.4f}")
        return {"miou": miou, "fscore": fscore}


# ── Main ──────────────────────────────────────────────────────────────────────

if __name__ == "__main__":
    mp.set_start_method("spawn")
    set_seed(42)

    os.makedirs(args.log_root, exist_ok=True)
    os.makedirs(args.checkpoint_root, exist_ok=True)

    tokenizer = transformers.AutoTokenizer.from_pretrained(
        args.mllm,
        cache_dir=None,
        model_max_length=2048,
        padding_side="right",
        use_fast=False,
    )
    tokenizer.pad_token = tokenizer.unk_token
    tokenizer.add_tokens("[SEG]")
    seg_token_idx = tokenizer("[SEG]", add_special_tokens=False).input_ids[0]
    print(f"seg_token_idx: {seg_token_idx}")

    # ── Datasets ──────────────────────────────────────────────────────────────
    train_dataset = REFAVS("train", args, tokenizer, input_type="refer")
    val_dataset_s = REFAVS("test_s", args, tokenizer, input_type="refer")
    val_dataset_u = REFAVS("test_u", args, tokenizer, input_type="refer")
    val_dataset_n = REFAVS("test_n", args, tokenizer, input_type="refer")

    g = torch.Generator()
    g.manual_seed(42)

    train_loader = DataLoader(
        train_dataset,
        batch_size=args.batch_size,
        shuffle=True,
        num_workers=4,
        worker_init_fn=seed_worker,
        collate_fn=partial(collate_fn, tokenizer=tokenizer),
        generator=g,
        pin_memory=True,
        persistent_workers=False,   # True caused worker-restart deadlock after shm error
        prefetch_factor=2,
    )
    val_loader_s = DataLoader(
        val_dataset_s, batch_size=4, shuffle=False, num_workers=4,
        collate_fn=partial(collate_fn, tokenizer=tokenizer),
        pin_memory=True, persistent_workers=False,
    )
    val_loader_u = DataLoader(
        val_dataset_u, batch_size=4, shuffle=False, num_workers=4,
        collate_fn=partial(collate_fn, tokenizer=tokenizer),
        pin_memory=True, persistent_workers=False,
    )
    val_loader_n = DataLoader(
        val_dataset_n, batch_size=4, shuffle=False, num_workers=4,
        collate_fn=partial(collate_fn, tokenizer=tokenizer),
        pin_memory=True, persistent_workers=False,
    )

    # ── Model ─────────────────────────────────────────────────────────────────
    model = build_model(args, tokenizer, seg_token_idx)

    if args.saved_model and os.path.exists(args.saved_model):
        ckpt = torch.load(args.saved_model, map_location="cuda")
        # Support both raw state dict and {"model": ...} dicts
        state = ckpt.get("model", ckpt)
        missing, unexpected = model.load_state_dict(state, strict=False)
        print(f"Loaded {args.saved_model}  missing={len(missing)}  unexpected={len(unexpected)}")

    if args.run == "eval":
        for split, loader in [("test_s", val_loader_s),
                               ("test_u", val_loader_u),
                               ("test_n", val_loader_n)]:
            if split in args.eval_splits:
                evaluate(model, loader, split, args.exist_threshold)
        exit(0)

    # ── Training ──────────────────────────────────────────────────────────────
    model.train()
    optimizer = AdamW(model.parameters(), lr=args.lr, betas=(0.9, 0.95), weight_decay=0.01)

    gradient_accumulation_steps = max(1, int(16 // args.batch_size))
    steps_per_epoch = len(train_loader) // gradient_accumulation_steps
    total_steps = args.epochs * steps_per_epoch
    warmup_steps = max(1, int(total_steps * 0.1))

    scheduler = get_cosine_schedule_with_warmup(
        optimizer,
        num_warmup_steps=warmup_steps,
        num_training_steps=total_steps,
    )

    log_path = os.path.join(args.log_root, f"{args.name}.txt")

    for epoch in range(args.epochs):
        model.train()
        optimizer.zero_grad()
        running = {"loss": 0.0, "ce": 0.0, "mask": 0.0, "exist": 0.0}
        n_steps = 0

        loop = tqdm(train_loader, desc=f"Epoch {epoch+1}/{args.epochs}")
        for step, batch in enumerate(loop):
            # ── Null augmentation ──────────────────────────────────────
            batch["audio_feats"], is_null = apply_null_augmentation(
                batch["audio_feats"], p_null=args.null_aug_prob
            )
            batch = dict_to_cuda(batch)
            is_null = is_null.cuda()

            with torch.autocast("cuda", dtype=torch.bfloat16):
                out = model.forward(
                    images=batch["images"],
                    images_clip=batch["images_clip"],
                    audio_features=batch["audio_feats"],
                    image_features=batch["image_feats"],
                    input_ids=batch["input_ids"],
                    labels=batch["labels"],
                    attention_masks=batch["attention_masks"],
                    masks_list=batch["masks"],
                    resize_list=batch["resizes"],
                    orgsize_list=batch["orgsizes"],
                    conversation_list=batch["convs"],
                    refs_num=batch["refs_num"],
                    fids=batch["fids"],
                    vids=batch["vids"],
                    ref_ids=batch["ref_ids"],
                    epoch=epoch,
                    inference=False,
                    contrast=args.ct_weight,
                    is_null=is_null,
                )

            loss = out["loss"] / gradient_accumulation_steps
            loss.backward()

            for k, key in [("loss", "loss"), ("ce", "ce_loss"),
                            ("mask", "mask_loss"), ("exist", "exist_loss")]:
                v = out.get(key, torch.tensor(0.0))
                running[k] += v.item() if isinstance(v, torch.Tensor) else v

            if (step + 1) % gradient_accumulation_steps == 0:
                optimizer.step()
                scheduler.step()
                optimizer.zero_grad()
                n_steps += 1
                lr = scheduler.get_last_lr()[0]
                avg = {k: running[k] / n_steps for k in running}
                loop.set_postfix(
                    lr=f"{lr:.2e}",
                    loss=f"{avg['loss']:.4f}",
                    exist=f"{avg['exist']:.4f}",
                )

        # ── End of epoch eval ────────────────────────────────────────
        denom = max(n_steps, 1)
        epoch_loss = running["loss"] / denom
        print(
            f"Epoch {epoch+1}  loss={epoch_loss:.4f}  "
            f"ce={running['ce']/denom:.4f}  "
            f"mask={running['mask']/denom:.4f}  "
            f"exist={running['exist']/denom:.4f}  "
            f"lr={scheduler.get_last_lr()[0]:.2e}"
        )

        with open(log_path, "a") as f:
            f.write(
                f"epoch={epoch+1}  loss={epoch_loss:.4f}  "
                f"ce={running['ce']/denom:.4f}  "
                f"mask={running['mask']/denom:.4f}  "
                f"exist={running['exist']/denom:.4f}\n"
            )

        # Per-epoch checkpoint β€” keep last 2 to save disk space.
        ckpt_ep = os.path.join(args.checkpoint_root, f"{args.name}_ep{epoch+1}.pth")
        torch.save(model.state_dict(), ckpt_ep)
        print(f"Saved: {ckpt_ep}")
        prev_ckpt = os.path.join(args.checkpoint_root, f"{args.name}_ep{epoch-1}.pth")
        if epoch >= 2 and os.path.exists(prev_ckpt):
            os.remove(prev_ckpt)

        evaluate(model, val_loader_s, "test_s", args.exist_threshold)
        evaluate(model, val_loader_u, "test_u", args.exist_threshold)
        evaluate(model, val_loader_n, "test_n", args.exist_threshold)

    # ── Save ──────────────────────────────────────────────────────────────────
    ckpt_path = os.path.join(args.checkpoint_root, f"{args.name}.pth")
    torch.save(model.state_dict(), ckpt_path)
    print(f"Saved: {ckpt_path}")