File size: 19,456 Bytes
2a50f45
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from functools import partial
import torch
from torch import nn, einsum
import torch.nn.functional as F
from einops import rearrange, repeat
try:
    import xformers
    import xformers.ops
    XFORMERS_IS_AVAILBLE = True
except:
    XFORMERS_IS_AVAILBLE = False
from lvdm.common import (
    checkpoint,
    exists,
    default,
)
from lvdm.basics import (
    zero_module,
)

class RelativePosition(nn.Module):
    """ https://github.com/evelinehong/Transformer_Relative_Position_PyTorch/blob/master/relative_position.py """

    def __init__(self, num_units, max_relative_position):
        super().__init__()
        self.num_units = num_units
        self.max_relative_position = max_relative_position
        self.embeddings_table = nn.Parameter(torch.Tensor(max_relative_position * 2 + 1, num_units))
        nn.init.xavier_uniform_(self.embeddings_table)

    def forward(self, length_q, length_k):
        device = self.embeddings_table.device
        range_vec_q = torch.arange(length_q, device=device)
        range_vec_k = torch.arange(length_k, device=device)
        distance_mat = range_vec_k[None, :] - range_vec_q[:, None]
        distance_mat_clipped = torch.clamp(distance_mat, -self.max_relative_position, self.max_relative_position)
        final_mat = distance_mat_clipped + self.max_relative_position
        final_mat = final_mat.long()
        embeddings = self.embeddings_table[final_mat]
        return embeddings


class CrossAttention(nn.Module):

    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., 
                 relative_position=False, temporal_length=None, img_cross_attention=False):
        super().__init__()
        inner_dim = dim_head * heads
        context_dim = default(context_dim, query_dim)

        self.scale = dim_head**-0.5
        self.heads = heads
        self.dim_head = dim_head
        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))

        self.image_cross_attention_scale = 1.0
        self.text_context_len = 77
        self.img_cross_attention = img_cross_attention
        if self.img_cross_attention:
            self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
            self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False)
        
        self.relative_position = relative_position
        if self.relative_position:
            assert(temporal_length is not None)
            self.relative_position_k = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
            self.relative_position_v = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
        else:
            ## only used for spatial attention, while NOT for temporal attention
            if XFORMERS_IS_AVAILBLE and temporal_length is None:
                self.forward = self.efficient_forward

    def forward(self, x, context=None, mask=None):
        h = self.heads

        q = self.to_q(x)
        context = default(context, x)
        ## considering image token additionally
        if context is not None and self.img_cross_attention:
            context, context_img = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
            k = self.to_k(context)
            v = self.to_v(context)
            k_ip = self.to_k_ip(context_img)
            v_ip = self.to_v_ip(context_img)
        else:
            k = self.to_k(context)
            v = self.to_v(context)

        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
        sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
        if self.relative_position:
            len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1]
            k2 = self.relative_position_k(len_q, len_k)
            sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale # TODO check 
            sim += sim2
        del k

        if exists(mask):
            ## feasible for causal attention mask only
            max_neg_value = -torch.finfo(sim.dtype).max
            mask = repeat(mask, 'b i j -> (b h) i j', h=h)
            sim.masked_fill_(~(mask>0.5), max_neg_value)

        # attention, what we cannot get enough of
        sim = sim.softmax(dim=-1)
        out = torch.einsum('b i j, b j d -> b i d', sim, v)
        if self.relative_position:
            v2 = self.relative_position_v(len_q, len_v)
            out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check
            out += out2
        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)

        ## considering image token additionally
        if context is not None and self.img_cross_attention:
            k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (k_ip, v_ip))
            sim_ip =  torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale
            del k_ip
            sim_ip = sim_ip.softmax(dim=-1)
            out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip)
            out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h)
            out = out + self.image_cross_attention_scale * out_ip
        del q

        return self.to_out(out)
    
    def efficient_forward(self, x, context=None, mask=None):
        q = self.to_q(x)
        context = default(context, x)

        ## considering image token additionally
        if context is not None and self.img_cross_attention:
            context, context_img = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
            k = self.to_k(context)
            v = self.to_v(context)
            k_ip = self.to_k_ip(context_img)
            v_ip = self.to_v_ip(context_img)
        else:
            k = self.to_k(context)
            v = self.to_v(context)

        b, _, _ = q.shape
        q, k, v = map(
            lambda t: t.unsqueeze(3)
            .reshape(b, t.shape[1], self.heads, self.dim_head)
            .permute(0, 2, 1, 3)
            .reshape(b * self.heads, t.shape[1], self.dim_head)
            .contiguous(),
            (q, k, v),
        )
        # actually compute the attention, what we cannot get enough of
        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=None)

        ## considering image token additionally
        if context is not None and self.img_cross_attention:
            k_ip, v_ip = map(
                lambda t: t.unsqueeze(3)
                .reshape(b, t.shape[1], self.heads, self.dim_head)
                .permute(0, 2, 1, 3)
                .reshape(b * self.heads, t.shape[1], self.dim_head)
                .contiguous(),
                (k_ip, v_ip),
            )
            out_ip = xformers.ops.memory_efficient_attention(q, k_ip, v_ip, attn_bias=None, op=None)
            out_ip = (
                out_ip.unsqueeze(0)
                .reshape(b, self.heads, out.shape[1], self.dim_head)
                .permute(0, 2, 1, 3)
                .reshape(b, out.shape[1], self.heads * self.dim_head)
            )

        if exists(mask):
            raise NotImplementedError
        out = (
            out.unsqueeze(0)
            .reshape(b, self.heads, out.shape[1], self.dim_head)
            .permute(0, 2, 1, 3)
            .reshape(b, out.shape[1], self.heads * self.dim_head)
        )
        if context is not None and self.img_cross_attention:
            out = out + self.image_cross_attention_scale * out_ip
        return self.to_out(out)


class BasicTransformerBlock(nn.Module):

    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
                disable_self_attn=False, attention_cls=None, img_cross_attention=False):
        super().__init__()
        attn_cls = CrossAttention if attention_cls is None else attention_cls
        self.disable_self_attn = disable_self_attn
        self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
            context_dim=context_dim if self.disable_self_attn else None)
        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
        self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, dropout=dropout,
            img_cross_attention=img_cross_attention)
        self.norm1 = nn.LayerNorm(dim)
        self.norm2 = nn.LayerNorm(dim)
        self.norm3 = nn.LayerNorm(dim)
        self.checkpoint = checkpoint

    def forward(self, x, context=None, mask=None):
        ## implementation tricks: because checkpointing doesn't support non-tensor (e.g. None or scalar) arguments
        input_tuple = (x,)      ## should not be (x), otherwise *input_tuple will decouple x into multiple arguments
        if context is not None:
            input_tuple = (x, context)
        if mask is not None:
            forward_mask = partial(self._forward, mask=mask)
            return checkpoint(forward_mask, (x,), self.parameters(), self.checkpoint)
        if context is not None and mask is not None:
            input_tuple = (x, context, mask)
        return checkpoint(self._forward, input_tuple, self.parameters(), self.checkpoint)

    def _forward(self, x, context=None, mask=None):
        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None, mask=mask) + x
        x = self.attn2(self.norm2(x), context=context, mask=mask) + x
        x = self.ff(self.norm3(x)) + x
        return x


class SpatialTransformer(nn.Module):
    """
    Transformer block for image-like data in spatial axis.
    First, project the input (aka embedding)
    and reshape to b, t, d.
    Then apply standard transformer action.
    Finally, reshape to image
    NEW: use_linear for more efficiency instead of the 1x1 convs
    """

    def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
                 use_checkpoint=True, disable_self_attn=False, use_linear=False, img_cross_attention=False):
        super().__init__()
        self.in_channels = in_channels
        inner_dim = n_heads * d_head
        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
        if not use_linear:
            self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
        else:
            self.proj_in = nn.Linear(in_channels, inner_dim)

        self.transformer_blocks = nn.ModuleList([
            BasicTransformerBlock(
                inner_dim,
                n_heads,
                d_head,
                dropout=dropout,
                context_dim=context_dim,
                img_cross_attention=img_cross_attention,
                disable_self_attn=disable_self_attn,
                checkpoint=use_checkpoint) for d in range(depth)
        ])
        if not use_linear:
            self.proj_out = zero_module(nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
        else:
            self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
        self.use_linear = use_linear


    def forward(self, x, context=None):
        b, c, h, w = x.shape
        x_in = x
        x = self.norm(x)
        if not self.use_linear:
            x = self.proj_in(x)
        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
        if self.use_linear:
            x = self.proj_in(x)
        for i, block in enumerate(self.transformer_blocks):
            x = block(x, context=context)
        if self.use_linear:
            x = self.proj_out(x)
        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
        if not self.use_linear:
            x = self.proj_out(x)
        return x + x_in
    
    
class TemporalTransformer(nn.Module):
    """
    Transformer block for image-like data in temporal axis.
    First, reshape to b, t, d.
    Then apply standard transformer action.
    Finally, reshape to image
    """
    def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
                 use_checkpoint=True, use_linear=False, only_self_att=True, causal_attention=False,
                 relative_position=False, temporal_length=None):
        super().__init__()
        self.only_self_att = only_self_att
        self.relative_position = relative_position
        self.causal_attention = causal_attention
        self.in_channels = in_channels
        inner_dim = n_heads * d_head
        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
        self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
        if not use_linear:
            self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
        else:
            self.proj_in = nn.Linear(in_channels, inner_dim)

        if relative_position:
            assert(temporal_length is not None)
            attention_cls = partial(CrossAttention, relative_position=True, temporal_length=temporal_length)
        else:
            attention_cls = None
        if self.causal_attention:
            assert(temporal_length is not None)
            self.mask = torch.tril(torch.ones([1, temporal_length, temporal_length]))

        if self.only_self_att:
            context_dim = None
        self.transformer_blocks = nn.ModuleList([
            BasicTransformerBlock(
                inner_dim,
                n_heads,
                d_head,
                dropout=dropout,
                context_dim=context_dim,
                attention_cls=attention_cls,
                checkpoint=use_checkpoint) for d in range(depth)
        ])
        if not use_linear:
            self.proj_out = zero_module(nn.Conv1d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
        else:
            self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
        self.use_linear = use_linear

    def forward(self, x, context=None):
        b, c, t, h, w = x.shape
        x_in = x
        x = self.norm(x)
        x = rearrange(x, 'b c t h w -> (b h w) c t').contiguous()
        if not self.use_linear:
            x = self.proj_in(x)
        x = rearrange(x, 'bhw c t -> bhw t c').contiguous()
        if self.use_linear:
            x = self.proj_in(x)

        if self.causal_attention:
            mask = self.mask.to(x.device)
            mask = repeat(mask, 'l i j -> (l bhw) i j', bhw=b*h*w)
        else:
            mask = None

        if self.only_self_att:
            ## note: if no context is given, cross-attention defaults to self-attention
            for i, block in enumerate(self.transformer_blocks):
                x = block(x, mask=mask)
            x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
        else:
            x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
            context = rearrange(context, '(b t) l con -> b t l con', t=t).contiguous()
            for i, block in enumerate(self.transformer_blocks):
                # calculate each batch one by one (since number in shape could not greater then 65,535 for some package)
                for j in range(b):
                    context_j = repeat(
                        context[j],
                        't l con -> (t r) l con', r=(h * w) // t, t=t).contiguous()
                    ## note: causal mask will not applied in cross-attention case
                    x[j] = block(x[j], context=context_j)
        
        if self.use_linear:
            x = self.proj_out(x)
            x = rearrange(x, 'b (h w) t c -> b c t h w', h=h, w=w).contiguous()
        if not self.use_linear:
            x = rearrange(x, 'b hw t c -> (b hw) c t').contiguous()
            x = self.proj_out(x)
            x = rearrange(x, '(b h w) c t -> b c t h w', b=b, h=h, w=w).contiguous()

        return x + x_in
    

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)


class LinearAttention(nn.Module):
    def __init__(self, dim, heads=4, dim_head=32):
        super().__init__()
        self.heads = heads
        hidden_dim = dim_head * heads
        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
        self.to_out = nn.Conv2d(hidden_dim, dim, 1)

    def forward(self, x):
        b, c, h, w = x.shape
        qkv = self.to_qkv(x)
        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
        k = k.softmax(dim=-1)  
        context = torch.einsum('bhdn,bhen->bhde', k, v)
        out = torch.einsum('bhde,bhdn->bhen', context, q)
        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
        return self.to_out(out)


class SpatialSelfAttention(nn.Module):
    def __init__(self, in_channels):
        super().__init__()
        self.in_channels = in_channels

        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
        self.q = torch.nn.Conv2d(in_channels,
                                 in_channels,
                                 kernel_size=1,
                                 stride=1,
                                 padding=0)
        self.k = torch.nn.Conv2d(in_channels,
                                 in_channels,
                                 kernel_size=1,
                                 stride=1,
                                 padding=0)
        self.v = torch.nn.Conv2d(in_channels,
                                 in_channels,
                                 kernel_size=1,
                                 stride=1,
                                 padding=0)
        self.proj_out = torch.nn.Conv2d(in_channels,
                                        in_channels,
                                        kernel_size=1,
                                        stride=1,
                                        padding=0)

    def forward(self, x):
        h_ = x
        h_ = self.norm(h_)
        q = self.q(h_)
        k = self.k(h_)
        v = self.v(h_)

        # compute attention
        b,c,h,w = q.shape
        q = rearrange(q, 'b c h w -> b (h w) c')
        k = rearrange(k, 'b c h w -> b c (h w)')
        w_ = torch.einsum('bij,bjk->bik', q, k)

        w_ = w_ * (int(c)**(-0.5))
        w_ = torch.nn.functional.softmax(w_, dim=2)

        # attend to values
        v = rearrange(v, 'b c h w -> b c (h w)')
        w_ = rearrange(w_, 'b i j -> b j i')
        h_ = torch.einsum('bij,bjk->bik', v, w_)
        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
        h_ = self.proj_out(h_)

        return x+h_