File size: 20,176 Bytes
aea2f58
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""shout-out to https://github.com/lucidrains/x-transformers/tree/main/x_transformers"""
import torch
from torch import nn, einsum
import torch.nn.functional as F
from functools import partial
from inspect import isfunction
from collections import namedtuple
from einops import rearrange, repeat, reduce

# constants

DEFAULT_DIM_HEAD = 64

Intermediates = namedtuple('Intermediates', [
    'pre_softmax_attn',
    'post_softmax_attn'
])

LayerIntermediates = namedtuple('Intermediates', [
    'hiddens',
    'attn_intermediates'
])


class AbsolutePositionalEmbedding(nn.Module):
    def __init__(self, dim, max_seq_len):
        super().__init__()
        self.emb = nn.Embedding(max_seq_len, dim)
        self.init_()

    def init_(self):
        nn.init.normal_(self.emb.weight, std=0.02)

    def forward(self, x):
        n = torch.arange(x.shape[1], device=x.device)
        return self.emb(n)[None, :, :]


class FixedPositionalEmbedding(nn.Module):
    def __init__(self, dim):
        super().__init__()
        inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer('inv_freq', inv_freq)

    def forward(self, x, seq_dim=1, offset=0):
        t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset
        sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq)
        emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1)
        return emb[None, :, :]


# helpers

def exists(val):
    return val is not None


def default(val, d):
    if exists(val):
        return val
    return d() if isfunction(d) else d


def always(val):
    def inner(*args, **kwargs):
        return val
    return inner


def not_equals(val):
    def inner(x):
        return x != val
    return inner


def equals(val):
    def inner(x):
        return x == val
    return inner


def max_neg_value(tensor):
    return -torch.finfo(tensor.dtype).max


# keyword argument helpers

def pick_and_pop(keys, d):
    values = list(map(lambda key: d.pop(key), keys))
    return dict(zip(keys, values))


def group_dict_by_key(cond, d):
    return_val = [dict(), dict()]
    for key in d.keys():
        match = bool(cond(key))
        ind = int(not match)
        return_val[ind][key] = d[key]
    return (*return_val,)


def string_begins_with(prefix, str):
    return str.startswith(prefix)


def group_by_key_prefix(prefix, d):
    return group_dict_by_key(partial(string_begins_with, prefix), d)


def groupby_prefix_and_trim(prefix, d):
    kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d)
    kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
    return kwargs_without_prefix, kwargs


# classes
class Scale(nn.Module):
    def __init__(self, value, fn):
        super().__init__()
        self.value = value
        self.fn = fn

    def forward(self, x, **kwargs):
        x, *rest = self.fn(x, **kwargs)
        return (x * self.value, *rest)


class Rezero(nn.Module):
    def __init__(self, fn):
        super().__init__()
        self.fn = fn
        self.g = nn.Parameter(torch.zeros(1))

    def forward(self, x, **kwargs):
        x, *rest = self.fn(x, **kwargs)
        return (x * self.g, *rest)


class ScaleNorm(nn.Module):
    def __init__(self, dim, eps=1e-5):
        super().__init__()
        self.scale = dim ** -0.5
        self.eps = eps
        self.g = nn.Parameter(torch.ones(1))

    def forward(self, x):
        norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
        return x / norm.clamp(min=self.eps) * self.g


class RMSNorm(nn.Module):
    def __init__(self, dim, eps=1e-8):
        super().__init__()
        self.scale = dim ** -0.5
        self.eps = eps
        self.g = nn.Parameter(torch.ones(dim))

    def forward(self, x):
        norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
        return x / norm.clamp(min=self.eps) * self.g


class Residual(nn.Module):
    def forward(self, x, residual):
        return x + residual


class GRUGating(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.gru = nn.GRUCell(dim, dim)

    def forward(self, x, residual):
        gated_output = self.gru(
            rearrange(x, 'b n d -> (b n) d'),
            rearrange(residual, 'b n d -> (b n) d')
        )

        return gated_output.reshape_as(x)


# feedforward

class GEGLU(nn.Module):
    def __init__(self, dim_in, dim_out):
        super().__init__()
        self.proj = nn.Linear(dim_in, dim_out * 2)

    def forward(self, x):
        x, gate = self.proj(x).chunk(2, dim=-1)
        return x * F.gelu(gate)


class FeedForward(nn.Module):
    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
        super().__init__()
        inner_dim = int(dim * mult)
        dim_out = default(dim_out, dim)
        project_in = nn.Sequential(
            nn.Linear(dim, inner_dim),
            nn.GELU()
        ) if not glu else GEGLU(dim, inner_dim)

        self.net = nn.Sequential(
            project_in,
            nn.Dropout(dropout),
            nn.Linear(inner_dim, dim_out)
        )

    def forward(self, x):
        return self.net(x)


# attention.
class Attention(nn.Module):
    def __init__(
            self,
            dim,
            dim_head=DEFAULT_DIM_HEAD,
            heads=8,
            causal=False,
            mask=None,
            talking_heads=False,
            sparse_topk=None,
            use_entmax15=False,
            num_mem_kv=0,
            dropout=0.,
            on_attn=False
    ):
        super().__init__()
        if use_entmax15:
            raise NotImplementedError("Check out entmax activation instead of softmax activation!")
        self.scale = dim_head ** -0.5
        self.heads = heads
        self.causal = causal
        self.mask = mask

        inner_dim = dim_head * heads

        self.to_q = nn.Linear(dim, inner_dim, bias=False)
        self.to_k = nn.Linear(dim, inner_dim, bias=False)
        self.to_v = nn.Linear(dim, inner_dim, bias=False)
        self.dropout = nn.Dropout(dropout)

        # talking heads
        self.talking_heads = talking_heads
        if talking_heads:
            self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads))
            self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads))

        # explicit topk sparse attention
        self.sparse_topk = sparse_topk

        # entmax
        #self.attn_fn = entmax15 if use_entmax15 else F.softmax
        self.attn_fn = F.softmax

        # add memory key / values
        self.num_mem_kv = num_mem_kv
        if num_mem_kv > 0:
            self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
            self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))

        # attention on attention
        self.attn_on_attn = on_attn
        self.to_out = nn.Sequential(nn.Linear(inner_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(inner_dim, dim)

    def forward(
            self,
            x,
            context=None,
            mask=None,
            context_mask=None,
            rel_pos=None,
            sinusoidal_emb=None,
            prev_attn=None,
            mem=None
    ):
        b, n, _, h, talking_heads, device = *x.shape, self.heads, self.talking_heads, x.device
        kv_input = default(context, x)

        q_input = x
        k_input = kv_input
        v_input = kv_input

        if exists(mem):
            k_input = torch.cat((mem, k_input), dim=-2)
            v_input = torch.cat((mem, v_input), dim=-2)

        if exists(sinusoidal_emb):
            # in shortformer, the query would start at a position offset depending on the past cached memory
            offset = k_input.shape[-2] - q_input.shape[-2]
            q_input = q_input + sinusoidal_emb(q_input, offset=offset)
            k_input = k_input + sinusoidal_emb(k_input)

        q = self.to_q(q_input)
        k = self.to_k(k_input)
        v = self.to_v(v_input)

        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v))

        input_mask = None
        if any(map(exists, (mask, context_mask))):
            q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool())
            k_mask = q_mask if not exists(context) else context_mask
            k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool())
            q_mask = rearrange(q_mask, 'b i -> b () i ()')
            k_mask = rearrange(k_mask, 'b j -> b () () j')
            input_mask = q_mask * k_mask

        if self.num_mem_kv > 0:
            mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v))
            k = torch.cat((mem_k, k), dim=-2)
            v = torch.cat((mem_v, v), dim=-2)
            if exists(input_mask):
                input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True)

        dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
        mask_value = max_neg_value(dots)

        if exists(prev_attn):
            dots = dots + prev_attn

        pre_softmax_attn = dots

        if talking_heads:
            dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous()

        if exists(rel_pos):
            dots = rel_pos(dots)

        if exists(input_mask):
            dots.masked_fill_(~input_mask, mask_value)
            del input_mask

        if self.causal:
            i, j = dots.shape[-2:]
            r = torch.arange(i, device=device)
            mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j')
            mask = F.pad(mask, (j - i, 0), value=False)
            dots.masked_fill_(mask, mask_value)
            del mask

        if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]:
            top, _ = dots.topk(self.sparse_topk, dim=-1)
            vk = top[..., -1].unsqueeze(-1).expand_as(dots)
            mask = dots < vk
            dots.masked_fill_(mask, mask_value)
            del mask

        attn = self.attn_fn(dots, dim=-1)
        post_softmax_attn = attn

        attn = self.dropout(attn)

        if talking_heads:
            attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous()

        out = einsum('b h i j, b h j d -> b h i d', attn, v)
        out = rearrange(out, 'b h n d -> b n (h d)')

        intermediates = Intermediates(
            pre_softmax_attn=pre_softmax_attn,
            post_softmax_attn=post_softmax_attn
        )

        return self.to_out(out), intermediates


class AttentionLayers(nn.Module):
    def __init__(
            self,
            dim,
            depth,
            heads=8,
            causal=False,
            cross_attend=False,
            only_cross=False,
            use_scalenorm=False,
            use_rmsnorm=False,
            use_rezero=False,
            rel_pos_num_buckets=32,
            rel_pos_max_distance=128,
            position_infused_attn=False,
            custom_layers=None,
            sandwich_coef=None,
            par_ratio=None,
            residual_attn=False,
            cross_residual_attn=False,
            macaron=False,
            pre_norm=True,
            gate_residual=False,
            **kwargs
    ):
        super().__init__()
        ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs)
        attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs)

        dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD)

        self.dim = dim
        self.depth = depth
        self.layers = nn.ModuleList([])

        self.has_pos_emb = position_infused_attn
        self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None
        self.rotary_pos_emb = always(None)

        assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance'
        self.rel_pos = always(None)

        self.pre_norm = pre_norm

        self.residual_attn = residual_attn
        self.cross_residual_attn = cross_residual_attn

        norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm
        norm_class = RMSNorm if use_rmsnorm else norm_class
        norm_fn = partial(norm_class, dim)

        norm_fn = nn.Identity if use_rezero else norm_fn
        branch_fn = Rezero if use_rezero else None

        if cross_attend and not only_cross:
            default_block = ('a', 'c', 'f')
        elif cross_attend and only_cross:
            default_block = ('c', 'f')
        else:
            default_block = ('a', 'f')

        if macaron:
            default_block = ('f',) + default_block

        if exists(custom_layers):
            layer_types = custom_layers
        elif exists(par_ratio):
            par_depth = depth * len(default_block)
            assert 1 < par_ratio <= par_depth, 'par ratio out of range'
            default_block = tuple(filter(not_equals('f'), default_block))
            par_attn = par_depth // par_ratio
            depth_cut = par_depth * 2 // 3  # 2 / 3 attention layer cutoff suggested by PAR paper
            par_width = (depth_cut + depth_cut // par_attn) // par_attn
            assert len(default_block) <= par_width, 'default block is too large for par_ratio'
            par_block = default_block + ('f',) * (par_width - len(default_block))
            par_head = par_block * par_attn
            layer_types = par_head + ('f',) * (par_depth - len(par_head))
        elif exists(sandwich_coef):
            assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth'
            layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef
        else:
            layer_types = default_block * depth

        self.layer_types = layer_types
        self.num_attn_layers = len(list(filter(equals('a'), layer_types)))

        for layer_type in self.layer_types:
            if layer_type == 'a':
                layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs)
            elif layer_type == 'c':
                layer = Attention(dim, heads=heads, **attn_kwargs)
            elif layer_type == 'f':
                layer = FeedForward(dim, **ff_kwargs)
                layer = layer if not macaron else Scale(0.5, layer)
            else:
                raise Exception(f'invalid layer type {layer_type}')

            if isinstance(layer, Attention) and exists(branch_fn):
                layer = branch_fn(layer)

            if gate_residual:
                residual_fn = GRUGating(dim)
            else:
                residual_fn = Residual()

            self.layers.append(nn.ModuleList([
                norm_fn(),
                layer,
                residual_fn
            ]))

    def forward(
            self,
            x,
            context=None,
            mask=None,
            context_mask=None,
            mems=None,
            return_hiddens=False
    ):
        hiddens = []
        intermediates = []
        prev_attn = None
        prev_cross_attn = None

        mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers

        for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)):
            is_last = ind == (len(self.layers) - 1)

            if layer_type == 'a':
                hiddens.append(x)
                layer_mem = mems.pop(0)

            residual = x

            if self.pre_norm:
                x = norm(x)

            if layer_type == 'a':
                out, inter = block(x, mask=mask, sinusoidal_emb=self.pia_pos_emb, rel_pos=self.rel_pos,
                                   prev_attn=prev_attn, mem=layer_mem)
            elif layer_type == 'c':
                out, inter = block(x, context=context, mask=mask, context_mask=context_mask, prev_attn=prev_cross_attn)
            elif layer_type == 'f':
                out = block(x)

            x = residual_fn(out, residual)

            if layer_type in ('a', 'c'):
                intermediates.append(inter)

            if layer_type == 'a' and self.residual_attn:
                prev_attn = inter.pre_softmax_attn
            elif layer_type == 'c' and self.cross_residual_attn:
                prev_cross_attn = inter.pre_softmax_attn

            if not self.pre_norm and not is_last:
                x = norm(x)

        if return_hiddens:
            intermediates = LayerIntermediates(
                hiddens=hiddens,
                attn_intermediates=intermediates
            )

            return x, intermediates

        return x


class Encoder(AttentionLayers):
    def __init__(self, **kwargs):
        assert 'causal' not in kwargs, 'cannot set causality on encoder'
        super().__init__(causal=False, **kwargs)



class TransformerWrapper(nn.Module):
    def __init__(
            self,
            *,
            num_tokens,
            max_seq_len,
            attn_layers,
            emb_dim=None,
            max_mem_len=0.,
            emb_dropout=0.,
            num_memory_tokens=None,
            tie_embedding=False,
            use_pos_emb=True
    ):
        super().__init__()
        assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder'

        dim = attn_layers.dim
        emb_dim = default(emb_dim, dim)

        self.max_seq_len = max_seq_len
        self.max_mem_len = max_mem_len
        self.num_tokens = num_tokens

        self.token_emb = nn.Embedding(num_tokens, emb_dim)
        self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if (
                    use_pos_emb and not attn_layers.has_pos_emb) else always(0)
        self.emb_dropout = nn.Dropout(emb_dropout)

        self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity()
        self.attn_layers = attn_layers
        self.norm = nn.LayerNorm(dim)

        self.init_()

        self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t()

        # memory tokens (like [cls]) from Memory Transformers paper
        num_memory_tokens = default(num_memory_tokens, 0)
        self.num_memory_tokens = num_memory_tokens
        if num_memory_tokens > 0:
            self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim))

            # let funnel encoder know number of memory tokens, if specified
            if hasattr(attn_layers, 'num_memory_tokens'):
                attn_layers.num_memory_tokens = num_memory_tokens

    def init_(self):
        nn.init.normal_(self.token_emb.weight, std=0.02)

    def forward(
            self,
            x,
            return_embeddings=False,
            mask=None,
            return_mems=False,
            return_attn=False,
            mems=None,
            **kwargs
    ):
        b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens
        x = self.token_emb(x)
        x += self.pos_emb(x)
        x = self.emb_dropout(x)

        x = self.project_emb(x)

        if num_mem > 0:
            mem = repeat(self.memory_tokens, 'n d -> b n d', b=b)
            x = torch.cat((mem, x), dim=1)

            # auto-handle masking after appending memory tokens
            if exists(mask):
                mask = F.pad(mask, (num_mem, 0), value=True)

        x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs)
        x = self.norm(x)

        mem, x = x[:, :num_mem], x[:, num_mem:]

        out = self.to_logits(x) if not return_embeddings else x

        if return_mems:
            hiddens = intermediates.hiddens
            new_mems = list(map(lambda pair: torch.cat(pair, dim=-2), zip(mems, hiddens))) if exists(mems) else hiddens
            new_mems = list(map(lambda t: t[..., -self.max_mem_len:, :].detach(), new_mems))
            return out, new_mems

        if return_attn:
            attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates))
            return out, attn_maps

        return out