File size: 20,331 Bytes
b4b2877
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Published baselines for T1 Scene Recognition, reproduced on DailyAct-5M.

Each method accepts a concatenated feature tensor (B, T, F_total) where F_total
is the sum of the active modality dims; the per-modality slices are recorded in
the `modality_dims` dict. Each method then uses the subset of modalities its
original paper intended.

All methods output an (B, num_classes) logit tensor.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F


def _slice(x, mod_dims, wanted):
    """Slice the concatenated feature tensor to keep only `wanted` modalities,
    in the order given. mod_dims is an ordered dict. Returns
    {name: tensor(B,T,d_name)} plus the concat."""
    parts = {}
    offset = 0
    for name, d in mod_dims.items():
        if name in wanted:
            parts[name] = x[..., offset:offset + d]
        offset += d
    assert len(parts) > 0, f"None of {wanted} in {list(mod_dims.keys())}"
    return parts


# ---------------------------------------------------------------------------
# 1) ST-GCN  (Yan et al., AAAI 2018)
#    Spatio-temporal graph CNN for skeleton action recognition.
#    We treat the 56-joint MoCap skeleton as the graph.
# ---------------------------------------------------------------------------

class STGCNBlock(nn.Module):
    def __init__(self, in_ch, out_ch, n_joints, stride=1, dropout=0.2):
        super().__init__()
        # Spatial graph conv: learnable adjacency (fully learned, no handcrafted A)
        self.A = nn.Parameter(torch.eye(n_joints) + 0.1 * torch.randn(n_joints, n_joints))
        self.spatial = nn.Conv2d(in_ch, out_ch, kernel_size=(1, 1), bias=False)
        self.spatial_bn = nn.BatchNorm2d(out_ch)
        self.temporal = nn.Conv2d(out_ch, out_ch, kernel_size=(9, 1),
                                  padding=(4, 0), stride=(stride, 1))
        self.temporal_bn = nn.BatchNorm2d(out_ch)
        self.dropout = nn.Dropout(dropout)
        if in_ch != out_ch or stride != 1:
            self.res = nn.Conv2d(in_ch, out_ch, kernel_size=1,
                                 stride=(stride, 1))
        else:
            self.res = nn.Identity()

    def forward(self, x):
        # x: (B, C, T, V)
        res = self.res(x)
        # spatial: aggregate along joints via A
        h = self.spatial(x)
        h = torch.einsum('bctv,vw->bctw', h, F.softmax(self.A, dim=-1))
        h = self.spatial_bn(h)
        h = F.relu(h)
        # temporal
        h = self.temporal(h)
        h = self.temporal_bn(h)
        h = self.dropout(h)
        return F.relu(h + res)


class STGCN(nn.Module):
    """ST-GCN on MoCap skeleton. We assume the MoCap modality is 620-dim
    (hip-relative + velocity) and reshape to ~56 joints."""
    def __init__(self, feat_dim_mocap, num_classes, hidden=64, n_joints=52):
        super().__init__()
        self.n_joints = n_joints
        # MoCap feat is (T, 620). 52 joints × 4 (xyz+quat_type), or we take per-joint xyz-only = 156.
        # In this repo, 620 = 52 markers * 4 cols + velocity features. We'll
        # reshape by slicing to 3*52=156 "primary" coords, padded if needed.
        self.coord_dim = 3  # we'll treat each joint as having 3 coords (XYZ)
        self.proj_in = nn.Linear(feat_dim_mocap, n_joints * self.coord_dim)

        self.blocks = nn.ModuleList([
            STGCNBlock(self.coord_dim, hidden, n_joints),
            STGCNBlock(hidden, hidden, n_joints),
            STGCNBlock(hidden, hidden * 2, n_joints, stride=2),
            STGCNBlock(hidden * 2, hidden * 2, n_joints),
            STGCNBlock(hidden * 2, hidden * 4, n_joints, stride=2),
            STGCNBlock(hidden * 4, hidden * 4, n_joints),
        ])
        self.head = nn.Sequential(
            nn.Dropout(0.3),
            nn.Linear(hidden * 4, num_classes),
        )

    def forward(self, x_mocap, mask=None):
        # x_mocap: (B, T, feat_dim_mocap)
        B, T, _ = x_mocap.shape
        h = self.proj_in(x_mocap)  # (B, T, n_joints * 3)
        h = h.reshape(B, T, self.n_joints, self.coord_dim).permute(0, 3, 1, 2)  # (B, C, T, V)
        for blk in self.blocks:
            h = blk(h)
        # Global mean pool over time & joints (with mask if provided)
        if mask is not None:
            # mask: (B, T), h: (B, C, T', V) where T' may be < T due to stride
            T_ = h.shape[2]
            m = mask[:, :T_].float().unsqueeze(1).unsqueeze(-1)  # (B, 1, T', 1)
            h = (h * m).sum(dim=(2, 3)) / (m.sum(dim=(2, 3)) * h.shape[3] + 1e-8)
        else:
            h = h.mean(dim=(2, 3))
        return self.head(h)


# ---------------------------------------------------------------------------
# 2) CTR-GCN  (Chen et al., ICCV 2021)
#    Channel-wise Topology Refinement GCN — learns a separate adjacency
#    matrix per channel group, known as SOTA for skeleton action recognition.
# ---------------------------------------------------------------------------

class CTRGC(nn.Module):
    """Simplified CTR-GC block: learnable per-channel topology refinement."""
    def __init__(self, in_ch, out_ch, n_joints, rel_reduction=4):
        super().__init__()
        self.n_joints = n_joints
        self.conv1 = nn.Conv2d(in_ch, out_ch // rel_reduction, 1)
        self.conv2 = nn.Conv2d(in_ch, out_ch // rel_reduction, 1)
        self.conv3 = nn.Conv2d(in_ch, out_ch, 1)
        self.alpha = nn.Parameter(torch.zeros(1))
        self.A = nn.Parameter(torch.eye(n_joints) + 0.1 * torch.randn(n_joints, n_joints))

    def forward(self, x):
        # x: (B, C, T, V)
        q = self.conv1(x).mean(dim=2)        # (B, C', V)
        k = self.conv2(x).mean(dim=2)        # (B, C', V)
        v = self.conv3(x)                    # (B, C_out, T, V)
        # Channel-specific topology refinement
        topology = F.softmax(torch.tanh(q.unsqueeze(-1) - k.unsqueeze(-2)), dim=-1)
        # topology: (B, C', V, V); we average across channels to get a shared (B, V, V)
        topology = topology.mean(dim=1)
        A = self.A.unsqueeze(0) + self.alpha * topology
        # apply A to v
        out = torch.einsum('bctv,bvw->bctw', v, A)
        return out


class CTRGCNBlock(nn.Module):
    def __init__(self, in_ch, out_ch, n_joints, stride=1):
        super().__init__()
        self.gc = CTRGC(in_ch, out_ch, n_joints)
        self.bn = nn.BatchNorm2d(out_ch)
        self.tcn = nn.Sequential(
            nn.Conv2d(out_ch, out_ch, (9, 1), padding=(4, 0), stride=(stride, 1)),
            nn.BatchNorm2d(out_ch),
        )
        if in_ch != out_ch or stride != 1:
            self.res = nn.Conv2d(in_ch, out_ch, 1, stride=(stride, 1))
        else:
            self.res = nn.Identity()

    def forward(self, x):
        res = self.res(x)
        h = self.gc(x)
        h = self.bn(h)
        h = F.relu(h)
        h = self.tcn(h)
        return F.relu(h + res)


class CTRGCN(nn.Module):
    def __init__(self, feat_dim_mocap, num_classes, hidden=64, n_joints=52):
        super().__init__()
        self.n_joints = n_joints
        self.coord_dim = 3
        self.proj_in = nn.Linear(feat_dim_mocap, n_joints * self.coord_dim)
        self.blocks = nn.ModuleList([
            CTRGCNBlock(self.coord_dim, hidden, n_joints),
            CTRGCNBlock(hidden, hidden, n_joints),
            CTRGCNBlock(hidden, hidden * 2, n_joints, stride=2),
            CTRGCNBlock(hidden * 2, hidden * 4, n_joints, stride=2),
        ])
        self.head = nn.Sequential(
            nn.Dropout(0.3),
            nn.Linear(hidden * 4, num_classes),
        )

    def forward(self, x_mocap, mask=None):
        B, T, _ = x_mocap.shape
        h = self.proj_in(x_mocap)
        h = h.reshape(B, T, self.n_joints, self.coord_dim).permute(0, 3, 1, 2)
        for blk in self.blocks:
            h = blk(h)
        h = h.mean(dim=(2, 3))
        return self.head(h)


# ---------------------------------------------------------------------------
# 3) LIMU-BERT  (Xu et al., SenSys 2021)
#    IMU self-supervised pretraining via masked reconstruction + fine-tune.
#    We implement a simpler variant: BERT-style encoder with optional
#    pretraining head.
# ---------------------------------------------------------------------------

class LIMUBertEncoder(nn.Module):
    def __init__(self, feat_dim_imu, hidden=128, n_layers=4, n_heads=4, dropout=0.1):
        super().__init__()
        self.in_proj = nn.Linear(feat_dim_imu, hidden)
        self.pos = nn.Parameter(torch.zeros(1, 4096, hidden))
        nn.init.trunc_normal_(self.pos, std=0.02)
        layer = nn.TransformerEncoderLayer(
            d_model=hidden, nhead=n_heads, dim_feedforward=4 * hidden,
            dropout=dropout, batch_first=True, activation='gelu',
        )
        self.encoder = nn.TransformerEncoder(layer, num_layers=n_layers)

    def forward(self, x, mask):
        T = x.size(1)
        h = self.in_proj(x) + self.pos[:, :T, :]
        h = self.encoder(h, src_key_padding_mask=~mask)
        return h


class LIMUBert(nn.Module):
    """Supervised-only variant: encoder + classifier head. Paper's
    pretraining is a masked-recon objective; for simplicity we report the
    supervised-only baseline here."""
    def __init__(self, feat_dim_imu, num_classes, hidden=128, n_layers=4,
                 n_heads=4, dropout=0.1):
        super().__init__()
        self.encoder = LIMUBertEncoder(feat_dim_imu, hidden, n_layers, n_heads, dropout)
        self.head = nn.Sequential(
            nn.LayerNorm(hidden),
            nn.Dropout(dropout),
            nn.Linear(hidden, num_classes),
        )

    def forward(self, x_imu, mask):
        h = self.encoder(x_imu, mask)
        m = mask.unsqueeze(-1).float()
        pooled = (h * m).sum(dim=1) / m.sum(dim=1).clamp(min=1.0)
        return self.head(pooled)


# ---------------------------------------------------------------------------
# 4) EMG-CNN  (standard 1D CNN baseline from sEMG classification literature)
#    E.g. Atzori et al. — multi-layer CNN with moving-window input.
# ---------------------------------------------------------------------------

class EMGCNN(nn.Module):
    def __init__(self, feat_dim_emg, num_classes, hidden=64):
        super().__init__()
        self.cnn = nn.Sequential(
            nn.Conv1d(feat_dim_emg, hidden, 7, padding=3),
            nn.BatchNorm1d(hidden), nn.ReLU(), nn.Dropout(0.3),
            nn.Conv1d(hidden, hidden * 2, 5, padding=2),
            nn.BatchNorm1d(hidden * 2), nn.ReLU(), nn.Dropout(0.3),
            nn.Conv1d(hidden * 2, hidden * 4, 3, padding=1),
            nn.BatchNorm1d(hidden * 4), nn.ReLU(),
        )
        self.head = nn.Linear(hidden * 4, num_classes)

    def forward(self, x_emg, mask):
        # (B, T, 8) -> (B, 8, T) for conv1d
        h = self.cnn(x_emg.transpose(1, 2))
        # Masked pool
        m = mask.unsqueeze(1).float()
        T_ = h.size(2)
        if m.size(2) != T_:
            m = F.adaptive_avg_pool1d(m, T_)
            m = (m > 0.5).float()
        pooled = (h * m).sum(dim=2) / m.sum(dim=2).clamp(min=1.0)
        return self.head(pooled)


# ---------------------------------------------------------------------------
# 5) ActionSense baseline  (DelPreto et al., NeurIPS '22)
#    Simple 3-layer MLP per modality + shared LSTM + classifier.
# ---------------------------------------------------------------------------

class ActionSenseLSTM(nn.Module):
    def __init__(self, modality_dims: dict, num_classes, hidden=128):
        super().__init__()
        self.mod_names = list(modality_dims.keys())
        self.mod_dims = modality_dims
        self.per_mod = nn.ModuleDict({
            name: nn.Sequential(
                nn.Linear(d, hidden), nn.ReLU(), nn.Dropout(0.2),
                nn.Linear(hidden, hidden), nn.ReLU(),
            ) for name, d in modality_dims.items()
        })
        concat_dim = hidden * len(modality_dims)
        self.lstm = nn.LSTM(concat_dim, hidden, num_layers=2,
                            batch_first=True, bidirectional=True, dropout=0.2)
        self.head = nn.Linear(hidden * 2, num_classes)

    def forward(self, x, mask):
        # x: (B, T, F_total), slice by modality
        offset = 0
        feats = []
        for name in self.mod_names:
            d = self.mod_dims[name]
            x_m = x[..., offset:offset + d]
            offset += d
            feats.append(self.per_mod[name](x_m))
        h = torch.cat(feats, dim=-1)  # (B, T, hidden * M)
        h, _ = self.lstm(h)
        m = mask.unsqueeze(-1).float()
        pooled = (h * m).sum(dim=1) / m.sum(dim=1).clamp(min=1.0)
        return self.head(pooled)


# ---------------------------------------------------------------------------
# 6) MulT  (Multimodal Transformer, Tsai et al., ACL 2019)
#    Core idea: cross-modal attention between every pair of modalities.
#    For a 3-modality input (A, B, C), produce
#    {A->B, A->C, B->A, B->C, C->A, C->B} via directed cross-attention.
# ---------------------------------------------------------------------------

class CrossModalTransformer(nn.Module):
    def __init__(self, d_model, n_heads=4, n_layers=2, dropout=0.1):
        super().__init__()
        self.layers = nn.ModuleList([
            nn.TransformerDecoderLayer(
                d_model=d_model, nhead=n_heads, dim_feedforward=4 * d_model,
                dropout=dropout, batch_first=True, activation='gelu',
            ) for _ in range(n_layers)
        ])

    def forward(self, q, kv, q_mask, kv_mask):
        # q: (B, T_q, D), kv: (B, T_kv, D)
        h = q
        for layer in self.layers:
            h = layer(h, kv,
                      tgt_key_padding_mask=~q_mask,
                      memory_key_padding_mask=~kv_mask)
        return h


class MulT(nn.Module):
    """Multimodal Transformer. Uses MoCap + EMG + IMU as 3 modalities
    (EyeTrack/Pressure omitted to match original 3-mod paper design)."""
    def __init__(self, modality_dims: dict, num_classes, d_model=128,
                 n_layers=2, n_heads=4, dropout=0.1):
        super().__init__()
        self.mod_names = [m for m in ['mocap', 'emg', 'imu'] if m in modality_dims]
        if len(self.mod_names) < 2:
            self.mod_names = list(modality_dims.keys())[:3]
        self.mod_dims = {m: modality_dims[m] for m in self.mod_names}
        self.in_proj = nn.ModuleDict({
            m: nn.Linear(d, d_model) for m, d in self.mod_dims.items()
        })
        # Pairwise cross-attention
        self.cross = nn.ModuleDict({
            f"{a}_to_{b}": CrossModalTransformer(d_model, n_heads, n_layers, dropout)
            for a in self.mod_names for b in self.mod_names if a != b
        })
        # Self-attention after cross
        self.self_tx = nn.ModuleDict({
            m: nn.TransformerEncoder(
                nn.TransformerEncoderLayer(
                    d_model=d_model, nhead=n_heads,
                    dim_feedforward=4 * d_model, dropout=dropout,
                    batch_first=True, activation='gelu',
                ), num_layers=1,
            ) for m in self.mod_names
        })
        total_dim = d_model * len(self.mod_names) * len(self.mod_names)
        self.head = nn.Sequential(
            nn.LayerNorm(total_dim),
            nn.Dropout(dropout),
            nn.Linear(total_dim, num_classes),
        )

    def forward(self, x, mask):
        # Slice modalities from x
        offset = 0
        projs = {}
        # Walk through all known mod_dims to find offsets
        # We need the FULL modality_dims order, which we don't have here;
        # expect caller to already supply x with exactly mod_names in order.
        # Workaround: assume caller passes mod_names order matching projection.
        for m in self.mod_names:
            d = self.mod_dims[m]
            projs[m] = self.in_proj[m](x[..., offset:offset + d])
            offset += d

        # Cross-attention: each modality attends to each other
        fused = {m: [] for m in self.mod_names}
        for a in self.mod_names:
            for b in self.mod_names:
                if a == b:
                    fused[a].append(projs[a])
                else:
                    out = self.cross[f"{a}_to_{b}"](projs[a], projs[b], mask, mask)
                    fused[a].append(out)

        # Self-attention + pool per modality
        pooled = []
        for a in self.mod_names:
            # Concat all attended-to representations along feature dim
            cat = torch.cat(fused[a], dim=-1)  # (B, T, D * M)
            # Actually re-project back to D per stream, then self-attn on stacked
            # Simplified: self-attention over concatenated, pool, flatten
            # Here we just pool each separately
            for i, rep in enumerate(fused[a]):
                rep = self.self_tx[a](rep)
                m = mask.unsqueeze(-1).float()
                p = (rep * m).sum(dim=1) / m.sum(dim=1).clamp(min=1.0)
                pooled.append(p)

        h = torch.cat(pooled, dim=-1)
        return self.head(h)


# ---------------------------------------------------------------------------
# 7) Perceiver IO  (Jaegle et al., ICML 2021)
#    Cross-attention from a fixed-size latent query set to all input tokens,
#    repeated for a few iterations.
# ---------------------------------------------------------------------------

class PerceiverBlock(nn.Module):
    def __init__(self, latent_dim, n_heads, dropout):
        super().__init__()
        self.ca = nn.MultiheadAttention(
            latent_dim, n_heads, dropout=dropout, batch_first=True,
        )
        self.norm1 = nn.LayerNorm(latent_dim)
        self.sa = nn.TransformerEncoderLayer(
            d_model=latent_dim, nhead=n_heads,
            dim_feedforward=4 * latent_dim, dropout=dropout,
            batch_first=True, activation='gelu',
        )

    def forward(self, latents, inputs, input_kpm):
        # Cross-attn: latents attend to inputs
        h, _ = self.ca(latents, inputs, inputs, key_padding_mask=input_kpm)
        latents = self.norm1(latents + h)
        # Self-attn on latents
        latents = self.sa(latents)
        return latents


class PerceiverIO(nn.Module):
    """Perceiver with N learnable latent queries; supports any modality mix."""
    def __init__(self, modality_dims: dict, num_classes,
                 latent_dim=128, n_latents=32, n_layers=3, n_heads=4, dropout=0.1):
        super().__init__()
        self.mod_names = list(modality_dims.keys())
        self.mod_dims = modality_dims
        # Per-modality input projection to latent_dim, with modality-id embedding
        self.in_proj = nn.ModuleDict({
            m: nn.Linear(d, latent_dim) for m, d in modality_dims.items()
        })
        self.mod_emb = nn.Parameter(torch.randn(len(self.mod_names), latent_dim) * 0.02)
        # Positional encoding (shared)
        self.pos = nn.Parameter(torch.zeros(1, 4096, latent_dim))
        nn.init.trunc_normal_(self.pos, std=0.02)
        # Learnable latents
        self.latents = nn.Parameter(torch.randn(n_latents, latent_dim) * 0.02)
        self.blocks = nn.ModuleList([
            PerceiverBlock(latent_dim, n_heads, dropout) for _ in range(n_layers)
        ])
        self.head = nn.Sequential(
            nn.LayerNorm(latent_dim),
            nn.Linear(latent_dim, num_classes),
        )

    def forward(self, x, mask):
        B, T, _ = x.shape
        # Project each modality + add modality embedding
        offset = 0
        tokens = []
        for i, m in enumerate(self.mod_names):
            d = self.mod_dims[m]
            tok = self.in_proj[m](x[..., offset:offset + d])  # (B, T, D)
            tok = tok + self.mod_emb[i]
            offset += d
            tokens.append(tok)
        # Concatenate along TIME dim, add shared pos enc per-modality
        # Each modality gets its own time sequence concatenated
        # Simpler: sum across modalities (like early fusion in latent space) + pos
        h = torch.stack(tokens, dim=2).mean(dim=2)  # (B, T, D)
        h = h + self.pos[:, :T, :]
        input_kpm = ~mask  # (B, T), True = ignore
        # Iterative cross-attention
        latents = self.latents.unsqueeze(0).expand(B, -1, -1)  # (B, N, D)
        for blk in self.blocks:
            latents = blk(latents, h, input_kpm)
        # Mean-pool latents
        pooled = latents.mean(dim=1)
        return self.head(pooled)