File size: 20,443 Bytes
960b1a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# coding: utf-8
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.init as init
import numpy as np
import math
from torch.nn.functional import silu
from torch.nn.functional import softplus
from einops import rearrange, einsum
from torch import Tensor
from torch_geometric.nn import GATConv, RGCNConv, TransformerConv

class PositionWiseFeedForward(nn.Module):
    def __init__(self, input_dim, hidden_dim, dropout=0.1):
        super().__init__()
        self.layer_1 = nn.Linear(input_dim, hidden_dim)
        self.layer_2 = nn.Linear(hidden_dim, input_dim)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        x = self.layer_1(x)
        x = F.gelu(x)  # Более плавная активация
        x = self.dropout(x)
        return self.layer_2(x)


class AddAndNorm(nn.Module):
    def __init__(self, input_dim, dropout=0.1):
        super().__init__()
        self.norm = nn.LayerNorm(input_dim)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x, residual):
        return self.norm(x + self.dropout(residual))


class PositionalEncoding(nn.Module):
    def __init__(self, d_model, dropout=0.1, max_len=5000):
        super().__init__()
        self.dropout = nn.Dropout(p=dropout)

        position = torch.arange(max_len).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model))
        pe = torch.zeros(max_len, d_model)
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)

        self.register_buffer("pe", pe)

    def forward(self, x):
        x = x + self.pe[: x.size(1)].detach()  # Отключаем градиенты
        return self.dropout(x)


class TransformerEncoderLayer(nn.Module):
    def __init__(self, input_dim, num_heads, dropout=0.1, positional_encoding=False):
        super().__init__()
        self.input_dim = input_dim
        self.self_attention = nn.MultiheadAttention(input_dim, num_heads, dropout=dropout, batch_first=True)
        # self.self_attention = MHA(
        #     embed_dim=input_dim,
        #     num_heads=num_heads,
        #     dropout=dropout,
        #     # bias=True,
        #     use_flash_attn=True
        # )
        self.feed_forward = PositionWiseFeedForward(input_dim, input_dim, dropout=dropout)
        self.add_norm_after_attention = AddAndNorm(input_dim, dropout=dropout)
        self.add_norm_after_ff = AddAndNorm(input_dim, dropout=dropout)
        self.positional_encoding = PositionalEncoding(input_dim) if positional_encoding else None

    def forward(self, key, value, query):
        if self.positional_encoding:
            key = self.positional_encoding(key)
            value = self.positional_encoding(value)
            query = self.positional_encoding(query)

        attn_output, _ = self.self_attention(query, key, value, need_weights=False)
        # attn_output = self.self_attention(query, key, value)

        x = self.add_norm_after_attention(attn_output, query)

        ff_output = self.feed_forward(x)
        x = self.add_norm_after_ff(ff_output, x)

        return x

class GAL(nn.Module):
    def __init__(self, input_dim_F1, input_dim_F2, gated_dim, dropout_rate):
        super(GAL, self).__init__()

        self.WF1 = nn.Parameter(torch.Tensor(input_dim_F1, gated_dim))
        self.WF2 = nn.Parameter(torch.Tensor(input_dim_F2, gated_dim))

        init.xavier_uniform_(self.WF1)
        init.xavier_uniform_(self.WF2)

        dim_size_f = input_dim_F1 + input_dim_F2

        self.WF = nn.Parameter(torch.Tensor(dim_size_f, gated_dim))

        init.xavier_uniform_(self.WF)

        self.dropout = nn.Dropout(dropout_rate)

    def forward(self, f1, f2):

        h_f1 = self.dropout(torch.tanh(torch.matmul(f1, self.WF1)))
        h_f2 = self.dropout(torch.tanh(torch.matmul(f2, self.WF2)))
        # print(h_f1.shape, h_f2.shape, self.WF.shape, torch.cat([f1, f2], dim=1).shape)
        z_f = torch.softmax(self.dropout(torch.matmul(torch.cat([f1, f2], dim=1), self.WF)), dim=1)
        h_f = z_f*h_f1 + (1 - z_f)*h_f2
        return h_f

class GraphFusionLayer(nn.Module):
    def __init__(self, hidden_dim, dropout=0.0, heads=2, out_mean=True):
        super().__init__()
        self.out_mean = out_mean
        # # Проекционные слои для признаков
        self.proj_audio = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim),
            nn.LayerNorm(hidden_dim),
            nn.Dropout(dropout)
        )
        self.proj_text = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim),
            nn.LayerNorm(hidden_dim),
            nn.Dropout(dropout)
        )

        # Графовые слои
        self.gat1 = GATConv(hidden_dim, hidden_dim, heads=heads)
        self.gat2 = GATConv(hidden_dim*heads, hidden_dim)

        # Финальная проекция
        self.fc = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim),
            nn.LayerNorm(hidden_dim),
            nn.Dropout(dropout)
        )

    def build_complete_graph(self, num_nodes):
        # Создаем полный граф (каждый узел соединен со всеми)
        edge_index = []
        for i in range(num_nodes):
            for j in range(num_nodes):
                if i != j:
                    edge_index.append([i, j])
        return torch.tensor(edge_index).t().contiguous()

    def forward(self, audio_stats, text_stats):
        """
        audio_stats: [batch_size, hidden_dim]
        text_stats: [batch_size, hidden_dim]
        """
        batch_size = audio_stats.size(0)

        # Проекция признаков
        x_audio = F.relu(self.proj_audio(audio_stats))  # [batch_size, hidden_dim]
        x_text = F.relu(self.proj_text(text_stats))    # [batch_size, hidden_dim]

        # Объединение узлов (аудио и текст попеременно)
        nodes = torch.stack([x_audio, x_text], dim=1)  # [batch_size, 2, hidden_dim]
        nodes = nodes.view(-1, nodes.size(-1))        # [batch_size*2, hidden_dim]

        # Построение графа (полный граф для каждого элемента батча)
        edge_index = self.build_complete_graph(2)  # Граф для одной пары аудио-текст
        edge_index = edge_index.to(audio_stats.device)

        # Применение GAT
        x = F.relu(self.gat1(nodes, edge_index))
        x = self.gat2(x, edge_index)

        # Разделяем обратно аудио и текст
        x = x.view(batch_size, 2, -1)  # [batch_size, 2, hidden_dim]

        if self.out_mean:
            # Усреднение по модальностям
            fused = torch.mean(x, dim=1)   # [batch_size, hidden_dim]

            return self.fc(fused)
        else:
            return x

class GraphFusionLayerAtt(nn.Module):
    def __init__(self, hidden_dim, heads=2):
        super().__init__()
        # Проекционные слои для признаков
        self.proj_audio = nn.Linear(hidden_dim, hidden_dim)
        self.proj_text = nn.Linear(hidden_dim, hidden_dim)

        # Графовые слои
        self.gat1 = GATConv(hidden_dim, hidden_dim, heads=heads)
        self.gat2 = GATConv(hidden_dim*heads, hidden_dim)

        self.attention_fusion = nn.Linear(hidden_dim, 1)

        # Финальная проекция
        self.fc = nn.Linear(hidden_dim, hidden_dim)

    def build_complete_graph(self, num_nodes):
        # Создаем полный граф (каждый узел соединен со всеми)
        edge_index = []
        for i in range(num_nodes):
            for j in range(num_nodes):
                if i != j:
                    edge_index.append([i, j])
        return torch.tensor(edge_index).t().contiguous()

    def forward(self, audio_stats, text_stats):
        """
        audio_stats: [batch_size, hidden_dim]
        text_stats: [batch_size, hidden_dim]
        """
        batch_size = audio_stats.size(0)

        # Проекция признаков
        x_audio = F.relu(self.proj_audio(audio_stats))  # [batch_size, hidden_dim]
        x_text = F.relu(self.proj_text(text_stats))    # [batch_size, hidden_dim]

        # Объединение узлов (аудио и текст попеременно)
        nodes = torch.stack([x_audio, x_text], dim=1)  # [batch_size, 2, hidden_dim]
        nodes = nodes.view(-1, nodes.size(-1))        # [batch_size*2, hidden_dim]

        # Построение графа (полный граф для каждого элемента батча)
        edge_index = self.build_complete_graph(2)  # Граф для одной пары аудио-текст
        edge_index = edge_index.to(audio_stats.device)

        # Применение GAT
        x = F.relu(self.gat1(nodes, edge_index))
        x = self.gat2(x, edge_index)

        # Разделяем обратно аудио и текст
        x = x.view(batch_size, 2, -1)  # [batch_size, 2, hidden_dim]

        # Усреднение по модальностям
        # fused = torch.mean(x, dim=1)   # [batch_size, hidden_dim]

        weights = F.softmax(self.attention_fusion(x), dim=1)
        fused = torch.sum(weights * x, dim=1)  # [batch_size, hidden_dim]

        return self.fc(fused)

# Full code see https://github.com/leson502/CORECT_EMNLP2023/tree/master/corect/model

class GNN(nn.Module):
    def __init__(self, g_dim, h1_dim, h2_dim, num_relations, num_modals, gcn_conv, use_graph_transformer, graph_transformer_nheads):
        super(GNN, self).__init__()
        self.gcn_conv = gcn_conv
        self.use_graph_transformer=use_graph_transformer

        self.num_modals = num_modals

        if self.gcn_conv == "rgcn":
            print("GNN --> Use RGCN")
            self.conv1 = RGCNConv(g_dim, h1_dim, num_relations)

        if self.use_graph_transformer:
            print("GNN --> Use Graph Transformer")

            in_dim = h1_dim

            self.conv2 = TransformerConv(in_dim, h2_dim, heads=graph_transformer_nheads, concat=True)
            self.bn = nn.BatchNorm1d(h2_dim * graph_transformer_nheads)


    def forward(self, node_features, node_type, edge_index, edge_type):
        print(node_features.shape, edge_index.shape, edge_type.shape)

        if self.gcn_conv == "rgcn":
            x = self.conv1(node_features, edge_index, edge_type)

        if self.use_graph_transformer:
            x = nn.functional.leaky_relu(self.bn(self.conv2(x, edge_index)))

        return x

class GraphModel(nn.Module):
    def __init__(self, g_dim, h1_dim, h2_dim, device, modalities, wp, wf, edge_type, gcn_conv, use_graph_transformer, graph_transformer_nheads):
        super(GraphModel, self).__init__()

        self.n_modals = len(modalities)
        self.wp = wp
        self.wf = wf
        self.device = device
        self.gcn_conv=gcn_conv
        self.use_graph_transformer=use_graph_transformer

        print(f"GraphModel --> Edge type: {edge_type}")
        print(f"GraphModel --> Window past: {wp}")
        print(f"GraphModel --> Window future: {wf}")
        edge_temp = "temp" in edge_type
        edge_multi = "multi" in edge_type

        edge_type_to_idx = {}

        if edge_temp:
            temporal = [-1, 1, 0]
            for j in temporal:
                for k in range(self.n_modals):
                    edge_type_to_idx[str(j) + str(k) + str(k)] = len(edge_type_to_idx)
        else:
            for j in range(self.n_modals):
                edge_type_to_idx['0' + str(j) + str(j)] = len(edge_type_to_idx)

        if edge_multi:
            for j in range(self.n_modals):
                for k in range(self.n_modals):
                    if (j != k):
                        edge_type_to_idx['0' + str(j) + str(k)] = len(edge_type_to_idx)

        self.edge_type_to_idx = edge_type_to_idx
        self.num_relations = len(edge_type_to_idx)
        self.edge_multi = edge_multi
        self.edge_temp = edge_temp

        self.gnn = GNN(g_dim, h1_dim, h2_dim, self.num_relations, self.n_modals, self.gcn_conv, self.use_graph_transformer, graph_transformer_nheads)


    def forward(self, x, lengths):
        # print(f"x shape: {x.shape}, lengths: {lengths}, lengths.shape: {lengths.shape}")

        node_features = feature_packing(x, lengths)

        node_type, edge_index, edge_type, edge_index_lengths = \
            self.batch_graphify(lengths)

        out_gnn = self.gnn(node_features, node_type, edge_index, edge_type)
        out_gnn = multi_concat(out_gnn, lengths, self.n_modals)

        return out_gnn

    def batch_graphify(self, lengths):

        node_type, edge_index, edge_type, edge_index_lengths = [], [], [], []
        edge_type_lengths = [0] * len(self.edge_type_to_idx)

        lengths = lengths.tolist()

        sum_length = 0
        total_length = sum(lengths)
        batch_size = len(lengths)

        for k in range(self.n_modals):
            for j in range(batch_size):
                cur_len = lengths[j]
                node_type.extend([k] * cur_len)

        for j in range(batch_size):
            cur_len = lengths[j]

            perms = self.edge_perms(cur_len, total_length)
            edge_index_lengths.append(len(perms))

            for item in perms:
                vertices = item[0]
                neighbor = item[1]
                edge_index.append(torch.tensor([vertices + sum_length, neighbor + sum_length]))

                if vertices % total_length > neighbor % total_length:
                    temporal_type = 1
                elif vertices % total_length < neighbor % total_length:
                    temporal_type = -1
                else:
                    temporal_type = 0
                edge_type.append(self.edge_type_to_idx[str(temporal_type)
                                                + str(node_type[vertices + sum_length])
                                                + str(node_type[neighbor + sum_length])])

            sum_length += cur_len

        node_type = torch.tensor(node_type).long().to(self.device)
        edge_index = torch.stack(edge_index).t().contiguous().to(self.device)  # [2, E]
        edge_type = torch.tensor(edge_type).long().to(self.device)  # [E]
        edge_index_lengths = torch.tensor(edge_index_lengths).long().to(self.device)  # [B]

        return node_type, edge_index, edge_type, edge_index_lengths

    def edge_perms(self, length, total_lengths):

        all_perms = set()
        array = np.arange(length)
        for j in range(length):
            if self.wp == -1 and self.wf == -1:
                eff_array = array
            elif self.wp == -1:  # use all past context
                eff_array = array[: min(length, j + self.wf)]
            elif self.wf == -1:  # use all future context
                eff_array = array[max(0, j - self.wp) :]
            else:
                eff_array = array[
                    max(0, j - self.wp) : min(length, j + self.wf)
                ]
            perms = set()


            for k in range(self.n_modals):
                node_index = j + k * total_lengths
                if self.edge_temp == True:
                    for item in eff_array:
                        perms.add((node_index, item + k * total_lengths))
                else:
                    perms.add((node_index, node_index))
                if self.edge_multi == True:
                    for l in range(self.n_modals):
                        if l != k:
                            perms.add((node_index, j + l * total_lengths))

            all_perms = all_perms.union(perms)

        return list(all_perms)

def feature_packing(multimodal_feature, lengths):
        batch_size = lengths.size(0)
        # print(multimodal_feature.shape, batch_size, lengths.shape)
        node_features = []

        for feature in multimodal_feature:
            for j in range(batch_size):
                cur_len = lengths[j].item()
                # print(f"feature.shape: {feature.shape}, j: {j}, cur_len: {cur_len}")
                node_features.append(feature[j,:cur_len])

        node_features = torch.cat(node_features, dim=0)

        return node_features

def multi_concat(nodes_feature, lengths, n_modals):
    sum_length = lengths.sum().item()
    feature = []
    for j in range(n_modals):
        feature.append(nodes_feature[j * sum_length : (j + 1) * sum_length])

    feature = torch.cat(feature, dim=-1)

    return feature

class RMSNorm(nn.Module):
    def __init__(self, d_model: int, eps: float = 1e-8) -> None:
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(d_model))

    def forward(self, x: Tensor) -> Tensor:
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim = True) + self.eps) * self.weight

class Mamba(nn.Module):
    def __init__(self, num_layers, d_input, d_model, d_state=16, d_discr=None, ker_size=4, num_classes=7, pooling=None):
        super().__init__()
        mamba_par = {
            'd_input' : d_input,
            'd_model' : d_model,
            'd_state' : d_state,
            'd_discr' : d_discr,
            'ker_size': ker_size
        }
        self.layers = nn.ModuleList([nn.ModuleList([MambaBlock(**mamba_par), RMSNorm(d_input)]) for _ in range(num_layers)])
        self.fc_out = nn.Linear(d_input, num_classes)
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    def forward(self, seq, cache=None):
        seq = torch.tensor(self.embedding(seq)).to(self.device)
        for mamba, norm in self.layers:
            out, cache = mamba(norm(seq), cache)
            seq = out + seq
        return self.fc_out(seq.mean(dim = 1))

class MambaBlock(nn.Module):
    def __init__(self, d_input, d_model, d_state=16, d_discr=None, ker_size=4):
        super().__init__()
        d_discr = d_discr if d_discr is not None else d_model // 16
        self.in_proj  = nn.Linear(d_input, 2 * d_model, bias=False)
        self.out_proj = nn.Linear(d_model, d_input, bias=False)
        self.s_B = nn.Linear(d_model, d_state, bias=False)
        self.s_C = nn.Linear(d_model, d_state, bias=False)
        self.s_D = nn.Sequential(nn.Linear(d_model, d_discr, bias=False), nn.Linear(d_discr, d_model, bias=False),)
        self.conv = nn.Conv1d(
            in_channels=d_model,
            out_channels=d_model,
            kernel_size=ker_size,
            padding=ker_size - 1,
            groups=d_model,
            bias=True,
        )
        self.A = nn.Parameter(torch.arange(1, d_state + 1, dtype=torch.float).repeat(d_model, 1))
        self.D = nn.Parameter(torch.ones(d_model, dtype=torch.float))
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    def forward(self, seq, cache=None):
        b, l, d = seq.shape
        (prev_hid, prev_inp) = cache if cache is not None else (None, None)
        a, b = self.in_proj(seq).chunk(2, dim=-1)
        x = rearrange(a, 'b l d -> b d l')
        x = x if prev_inp is None else torch.cat((prev_inp, x), dim=-1)
        a = self.conv(x)[..., :l]
        a = rearrange(a, 'b d l -> b l d')
        a = silu(a)
        a, hid = self.ssm(a, prev_hid=prev_hid)
        b = silu(b)
        out = a * b
        out =  self.out_proj(out)
        if cache:
            cache = (hid.squeeze(), x[..., 1:])
        return out, cache

    def ssm(self, seq, prev_hid):
        A = -self.A
        D = +self.D
        B = self.s_B(seq)
        C = self.s_C(seq)
        s = softplus(D + self.s_D(seq))
        A_bar = einsum(torch.exp(A), s, 'd s,   b l d -> b l d s')
        B_bar = einsum(          B,  s, 'b l s, b l d -> b l d s')
        X_bar = einsum(B_bar, seq, 'b l d s, b l d -> b l d s')
        hid = self._hid_states(A_bar, X_bar, prev_hid=prev_hid)
        out = einsum(hid, C, 'b l d s, b l s -> b l d')
        out = out + D * seq
        return out, hid

    def _hid_states(self, A, X, prev_hid=None):
        b, l, d, s = A.shape
        A = rearrange(A, 'b l d s -> l b d s')
        X = rearrange(X, 'b l d s -> l b d s')
        if prev_hid is not None:
            return rearrange(A * prev_hid + X, 'l b d s -> b l d s')
        h = torch.zeros(b, d, s, device=self.device)
        return torch.stack([h := A_t * h + X_t for A_t, X_t in zip(A, X)], dim=1)