unet_ldm.py 23.1 KB
Newer Older
patil-suraj's avatar
patil-suraj committed
1
import math
Patrick von Platen's avatar
Patrick von Platen committed
2
from inspect import isfunction
patil-suraj's avatar
patil-suraj committed
3
4
5
6
7

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
Patrick von Platen's avatar
Patrick von Platen committed
8

patil-suraj's avatar
patil-suraj committed
9
10
from ..configuration_utils import ConfigMixin
from ..modeling_utils import ModelMixin
Patrick von Platen's avatar
Patrick von Platen committed
11
from .attention import AttentionBlock
12
from .embeddings import get_timestep_embedding
13
from .resnet import Downsample2D, ResnetBlock2D, Upsample2D
Patrick von Platen's avatar
Patrick von Platen committed
14
from .unet_new import UNetMidBlock2D
Patrick von Platen's avatar
up  
Patrick von Platen committed
15
16
17


# from .resnet import ResBlock
patil-suraj's avatar
patil-suraj committed
18

Patrick von Platen's avatar
Patrick von Platen committed
19

patil-suraj's avatar
patil-suraj committed
20
21
22
23
24
def exists(val):
    return val is not None


def uniq(arr):
Patrick von Platen's avatar
Patrick von Platen committed
25
    return {el: True for el in arr}.keys()
patil-suraj's avatar
patil-suraj committed
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


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


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


def init_(tensor):
    dim = tensor.shape[-1]
    std = 1 / math.sqrt(dim)
    tensor.uniform_(-std, std)
    return tensor


# 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):
Patrick von Platen's avatar
Patrick von Platen committed
57
    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
patil-suraj's avatar
patil-suraj committed
58
59
60
        super().__init__()
        inner_dim = int(dim * mult)
        dim_out = default(dim_out, dim)
Patrick von Platen's avatar
Patrick von Platen committed
61
62
63
        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))
patil-suraj's avatar
patil-suraj committed
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

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


def zero_module(module):
    """
    Zero out the parameters of a module and return it.
    """
    for p in module.parameters():
        p.detach().zero_()
    return module


def Normalize(in_channels):
    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)


def convert_module_to_f16(l):
    """
    Convert primitive modules to float16.
    """
    if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Conv3d)):
        l.weight.data = l.weight.data.half()
        if l.bias is not None:
            l.bias.data = l.bias.data.half()


def convert_module_to_f32(l):
    """
    Convert primitive modules to float32, undoing convert_module_to_f16().
    """
    if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Conv3d)):
        l.weight.data = l.weight.data.float()
        if l.bias is not None:
            l.bias.data = l.bias.data.float()


def conv_nd(dims, *args, **kwargs):
    """
    Create a 1D, 2D, or 3D convolution module.
    """
    if dims == 1:
        return nn.Conv1d(*args, **kwargs)
    elif dims == 2:
        return nn.Conv2d(*args, **kwargs)
    elif dims == 3:
        return nn.Conv3d(*args, **kwargs)
    raise ValueError(f"unsupported dimensions: {dims}")


def linear(*args, **kwargs):
    """
    Create a linear module.
    """
    return nn.Linear(*args, **kwargs)


class GroupNorm32(nn.GroupNorm):
    def __init__(self, num_groups, num_channels, swish, eps=1e-5):
        super().__init__(num_groups=num_groups, num_channels=num_channels, eps=eps)
        self.swish = swish

    def forward(self, x):
        y = super().forward(x.float()).to(x.dtype)
        if self.swish == 1.0:
            y = F.silu(y)
        elif self.swish:
            y = y * F.sigmoid(y * float(self.swish))
        return y


def normalization(channels, swish=0.0):
    """
    Make a standard normalization layer, with an optional swish activation.

Patrick von Platen's avatar
Patrick von Platen committed
140
    :param channels: number of input channels. :return: an nn.Module for normalization.
patil-suraj's avatar
patil-suraj committed
141
142
143
144
    """
    return GroupNorm32(num_channels=channels, num_groups=32, swish=swish)


145
class TimestepEmbedSequential(nn.Sequential):
patil-suraj's avatar
patil-suraj committed
146
    """
Patrick von Platen's avatar
Patrick von Platen committed
147
    A sequential module that passes timestep embeddings to the children that support it as an extra input.
patil-suraj's avatar
patil-suraj committed
148
149
150
151
    """

    def forward(self, x, emb, context=None):
        for layer in self:
152
            if isinstance(layer, ResnetBlock2D) or isinstance(layer, TimestepEmbedSequential):
patil-suraj's avatar
patil-suraj committed
153
154
155
156
157
158
159
160
161
162
                x = layer(x, emb)
            elif isinstance(layer, SpatialTransformer):
                x = layer(x, context)
            else:
                x = layer(x)
        return x


def count_flops_attn(model, _x, y):
    """
Patrick von Platen's avatar
Patrick von Platen committed
163
    A counter for the `thop` package to count the operations in an attention operation. Meant to be used like:
patil-suraj's avatar
patil-suraj committed
164
        macs, params = thop.profile(
Patrick von Platen's avatar
Patrick von Platen committed
165
            model, inputs=(inputs, timestamps), custom_ops={QKVAttention: QKVAttention.count_flops},
patil-suraj's avatar
patil-suraj committed
166
167
168
169
170
171
172
        )
    """
    b, c, *spatial = y[0].shape
    num_spatial = int(np.prod(spatial))
    # We perform two matmuls with the same number of ops.
    # The first computes the weight matrix, the second computes
    # the combination of the value vectors.
Patrick von Platen's avatar
Patrick von Platen committed
173
    matmul_ops = 2 * b * (num_spatial**2) * c
patil-suraj's avatar
patil-suraj committed
174
175
176
    model.total_ops += torch.DoubleTensor([matmul_ops])


patil-suraj's avatar
patil-suraj committed
177
class UNetLDMModel(ModelMixin, ConfigMixin):
patil-suraj's avatar
patil-suraj committed
178
    """
Patrick von Platen's avatar
Patrick von Platen committed
179
180
181
182
183
184
185
186
    The full UNet model with attention and timestep embedding. :param in_channels: channels in the input Tensor. :param
    model_channels: base channel count for the model. :param out_channels: channels in the output Tensor. :param
    num_res_blocks: number of residual blocks per downsample. :param attention_resolutions: a collection of downsample
    rates at which
        attention will take place. May be a set, list, or tuple. For example, if this contains 4, then at 4x
        downsampling, attention will be used.
    :param dropout: the dropout probability. :param channel_mult: channel multiplier for each level of the UNet. :param
    conv_resample: if True, use learned convolutions for upsampling and
patil-suraj's avatar
patil-suraj committed
187
        downsampling.
Patrick von Platen's avatar
Patrick von Platen committed
188
189
    :param dims: determines if the signal is 1D, 2D, or 3D. :param num_classes: if specified (as an int), then this
    model will be
patil-suraj's avatar
patil-suraj committed
190
        class-conditional with `num_classes` classes.
Patrick von Platen's avatar
Patrick von Platen committed
191
192
    :param use_checkpoint: use gradient checkpointing to reduce memory usage. :param num_heads: the number of attention
    heads in each attention layer. :param num_heads_channels: if specified, ignore num_heads and instead use
patil-suraj's avatar
patil-suraj committed
193
194
195
                               a fixed channel width per attention head.
    :param num_heads_upsample: works with num_heads to set a different number
                               of heads for upsampling. Deprecated.
Patrick von Platen's avatar
Patrick von Platen committed
196
197
    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism. :param resblock_updown: use residual blocks
    for up/downsampling. :param use_new_attention_order: use a different attention pattern for potentially
patil-suraj's avatar
patil-suraj committed
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
                                    increased efficiency.
    """

    def __init__(
        self,
        image_size,
        in_channels,
        model_channels,
        out_channels,
        num_res_blocks,
        attention_resolutions,
        dropout=0,
        channel_mult=(1, 2, 4, 8),
        conv_resample=True,
        dims=2,
        num_classes=None,
        use_checkpoint=False,
        use_fp16=False,
        num_heads=-1,
        num_head_channels=-1,
        num_heads_upsample=-1,
        use_scale_shift_norm=False,
        resblock_updown=False,
        use_new_attention_order=False,
Patrick von Platen's avatar
Patrick von Platen committed
222
223
224
225
        use_spatial_transformer=False,  # custom transformer support
        transformer_depth=1,  # custom transformer support
        context_dim=None,  # custom transformer support
        n_embed=None,  # custom support for prediction of discrete ids into codebook of first stage vq model
patil-suraj's avatar
patil-suraj committed
226
227
228
        legacy=True,
    ):
        super().__init__()
Patrick von Platen's avatar
Patrick von Platen committed
229

patil-suraj's avatar
patil-suraj committed
230
        # register all __init__ params with self.register
231
        self.register_to_config(
patil-suraj's avatar
patil-suraj committed
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
            image_size=image_size,
            in_channels=in_channels,
            model_channels=model_channels,
            out_channels=out_channels,
            num_res_blocks=num_res_blocks,
            attention_resolutions=attention_resolutions,
            dropout=dropout,
            channel_mult=channel_mult,
            conv_resample=conv_resample,
            dims=dims,
            num_classes=num_classes,
            use_fp16=use_fp16,
            num_heads=num_heads,
            num_head_channels=num_head_channels,
            num_heads_upsample=num_heads_upsample,
            use_scale_shift_norm=use_scale_shift_norm,
            resblock_updown=resblock_updown,
            use_spatial_transformer=use_spatial_transformer,
            transformer_depth=transformer_depth,
            context_dim=context_dim,
            n_embed=n_embed,
            legacy=legacy,
        )

        if use_spatial_transformer:
Patrick von Platen's avatar
Patrick von Platen committed
257
258
259
            assert (
                context_dim is not None
            ), "Fool!! You forgot to include the dimension of your cross-attention conditioning..."
patil-suraj's avatar
patil-suraj committed
260
261

        if context_dim is not None:
Patrick von Platen's avatar
Patrick von Platen committed
262
263
264
            assert (
                use_spatial_transformer
            ), "Fool!! You forgot to use the spatial transformer for your cross-attention conditioning..."
patil-suraj's avatar
patil-suraj committed
265
266
267
268
269

        if num_heads_upsample == -1:
            num_heads_upsample = num_heads

        if num_heads == -1:
Patrick von Platen's avatar
Patrick von Platen committed
270
            assert num_head_channels != -1, "Either num_heads or num_head_channels has to be set"
patil-suraj's avatar
patil-suraj committed
271
272

        if num_head_channels == -1:
Patrick von Platen's avatar
Patrick von Platen committed
273
            assert num_heads != -1, "Either num_heads or num_head_channels has to be set"
patil-suraj's avatar
patil-suraj committed
274
275
276
277
278
279
280
281
282
283
284

        self.image_size = image_size
        self.in_channels = in_channels
        self.model_channels = model_channels
        self.out_channels = out_channels
        self.num_res_blocks = num_res_blocks
        self.attention_resolutions = attention_resolutions
        self.dropout = dropout
        self.channel_mult = channel_mult
        self.conv_resample = conv_resample
        self.num_classes = num_classes
patil-suraj's avatar
patil-suraj committed
285
        self.dtype_ = torch.float16 if use_fp16 else torch.float32
patil-suraj's avatar
patil-suraj committed
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
        self.num_heads = num_heads
        self.num_head_channels = num_head_channels
        self.num_heads_upsample = num_heads_upsample
        self.predict_codebook_ids = n_embed is not None

        time_embed_dim = model_channels * 4
        self.time_embed = nn.Sequential(
            linear(model_channels, time_embed_dim),
            nn.SiLU(),
            linear(time_embed_dim, time_embed_dim),
        )

        if self.num_classes is not None:
            self.label_emb = nn.Embedding(num_classes, time_embed_dim)

        self.input_blocks = nn.ModuleList(
Patrick von Platen's avatar
Patrick von Platen committed
302
            [TimestepEmbedSequential(conv_nd(dims, in_channels, model_channels, 3, padding=1))]
patil-suraj's avatar
patil-suraj committed
303
304
305
306
307
308
309
310
        )
        self._feature_size = model_channels
        input_block_chans = [model_channels]
        ch = model_channels
        ds = 1
        for level, mult in enumerate(channel_mult):
            for _ in range(num_res_blocks):
                layers = [
Patrick von Platen's avatar
Patrick von Platen committed
311
                    ResnetBlock2D(
Patrick von Platen's avatar
up  
Patrick von Platen committed
312
313
314
315
316
317
318
                        in_channels=ch,
                        out_channels=mult * model_channels,
                        dropout=dropout,
                        temb_channels=time_embed_dim,
                        eps=1e-5,
                        non_linearity="silu",
                        overwrite_for_ldm=True,
patil-suraj's avatar
patil-suraj committed
319
320
321
322
323
324
325
326
327
328
                    )
                ]
                ch = mult * model_channels
                if ds in attention_resolutions:
                    if num_head_channels == -1:
                        dim_head = ch // num_heads
                    else:
                        num_heads = ch // num_head_channels
                        dim_head = num_head_channels
                    if legacy:
Patrick von Platen's avatar
Patrick von Platen committed
329
                        # num_heads = 1
patil-suraj's avatar
patil-suraj committed
330
331
332
333
334
335
                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
                    layers.append(
                        AttentionBlock(
                            ch,
                            num_heads=num_heads,
                            num_head_channels=dim_head,
Patrick von Platen's avatar
Patrick von Platen committed
336
337
338
                        )
                        if not use_spatial_transformer
                        else SpatialTransformer(
patil-suraj's avatar
patil-suraj committed
339
340
341
342
343
344
345
346
347
348
                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
                        )
                    )
                self.input_blocks.append(TimestepEmbedSequential(*layers))
                self._feature_size += ch
                input_block_chans.append(ch)
            if level != len(channel_mult) - 1:
                out_ch = ch
                self.input_blocks.append(
                    TimestepEmbedSequential(
349
                        Downsample2D(ch, use_conv=conv_resample, out_channels=out_ch, padding=1, name="op")
patil-suraj's avatar
patil-suraj committed
350
351
352
353
354
355
356
357
358
359
360
361
362
                    )
                )
                ch = out_ch
                input_block_chans.append(ch)
                ds *= 2
                self._feature_size += ch

        if num_head_channels == -1:
            dim_head = ch // num_heads
        else:
            num_heads = ch // num_head_channels
            dim_head = num_head_channels
        if legacy:
Patrick von Platen's avatar
Patrick von Platen committed
363
            # num_heads = 1
patil-suraj's avatar
patil-suraj committed
364
            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
Patrick von Platen's avatar
Patrick von Platen committed
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383

        if dim_head < 0:
            dim_head = None
        self.mid = UNetMidBlock2D(
            in_channels=ch,
            dropout=dropout,
            temb_channels=time_embed_dim,
            resnet_eps=1e-5,
            resnet_act_fn="silu",
            resnet_time_scale_shift="scale_shift" if use_scale_shift_norm else "default",
            attention_layer_type="self" if not use_spatial_transformer else "spatial",
            attn_num_heads=num_heads,
            attn_num_head_channels=dim_head,
            attn_depth=transformer_depth,
            attn_encoder_channels=context_dim,
        )

        # TODO(Patrick) - delete after weight conversion
        # init to be able to overwrite `self.mid`
patil-suraj's avatar
patil-suraj committed
384
        self.middle_block = TimestepEmbedSequential(
Patrick von Platen's avatar
Patrick von Platen committed
385
            ResnetBlock2D(
Patrick von Platen's avatar
up  
Patrick von Platen committed
386
387
388
389
390
391
392
                in_channels=ch,
                out_channels=None,
                dropout=dropout,
                temb_channels=time_embed_dim,
                eps=1e-5,
                non_linearity="silu",
                overwrite_for_ldm=True,
patil-suraj's avatar
patil-suraj committed
393
394
395
396
397
            ),
            AttentionBlock(
                ch,
                num_heads=num_heads,
                num_head_channels=dim_head,
Patrick von Platen's avatar
Patrick von Platen committed
398
399
400
            )
            if not use_spatial_transformer
            else SpatialTransformer(ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim),
Patrick von Platen's avatar
Patrick von Platen committed
401
            ResnetBlock2D(
Patrick von Platen's avatar
up  
Patrick von Platen committed
402
403
404
405
406
407
408
                in_channels=ch,
                out_channels=None,
                dropout=dropout,
                temb_channels=time_embed_dim,
                eps=1e-5,
                non_linearity="silu",
                overwrite_for_ldm=True,
patil-suraj's avatar
patil-suraj committed
409
410
            ),
        )
411
412
413
        self.mid.resnets[0] = self.middle_block[0]
        self.mid.attentions[0] = self.middle_block[1]
        self.mid.resnets[1] = self.middle_block[2]
Patrick von Platen's avatar
Patrick von Platen committed
414

patil-suraj's avatar
patil-suraj committed
415
416
417
418
419
420
421
        self._feature_size += ch

        self.output_blocks = nn.ModuleList([])
        for level, mult in list(enumerate(channel_mult))[::-1]:
            for i in range(num_res_blocks + 1):
                ich = input_block_chans.pop()
                layers = [
Patrick von Platen's avatar
Patrick von Platen committed
422
                    ResnetBlock2D(
Patrick von Platen's avatar
up  
Patrick von Platen committed
423
                        in_channels=ch + ich,
patil-suraj's avatar
patil-suraj committed
424
                        out_channels=model_channels * mult,
Patrick von Platen's avatar
up  
Patrick von Platen committed
425
426
427
428
429
430
                        dropout=dropout,
                        temb_channels=time_embed_dim,
                        eps=1e-5,
                        non_linearity="silu",
                        overwrite_for_ldm=True,
                    ),
patil-suraj's avatar
patil-suraj committed
431
432
433
434
435
436
437
438
439
                ]
                ch = model_channels * mult
                if ds in attention_resolutions:
                    if num_head_channels == -1:
                        dim_head = ch // num_heads
                    else:
                        num_heads = ch // num_head_channels
                        dim_head = num_head_channels
                    if legacy:
Patrick von Platen's avatar
Patrick von Platen committed
440
                        # num_heads = 1
patil-suraj's avatar
patil-suraj committed
441
442
443
444
445
446
                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
                    layers.append(
                        AttentionBlock(
                            ch,
                            num_heads=num_heads_upsample,
                            num_head_channels=dim_head,
Patrick von Platen's avatar
Patrick von Platen committed
447
448
449
                        )
                        if not use_spatial_transformer
                        else SpatialTransformer(
patil-suraj's avatar
patil-suraj committed
450
451
452
453
454
                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
                        )
                    )
                if level and i == num_res_blocks:
                    out_ch = ch
455
                    layers.append(Upsample2D(ch, use_conv=conv_resample, out_channels=out_ch))
patil-suraj's avatar
patil-suraj committed
456
457
458
459
460
461
462
463
464
465
466
                    ds //= 2
                self.output_blocks.append(TimestepEmbedSequential(*layers))
                self._feature_size += ch

        self.out = nn.Sequential(
            normalization(ch),
            nn.SiLU(),
            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
        )
        if self.predict_codebook_ids:
            self.id_predictor = nn.Sequential(
Patrick von Platen's avatar
Patrick von Platen committed
467
468
469
470
                normalization(ch),
                conv_nd(dims, model_channels, n_embed, 1),
                # nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
            )
patil-suraj's avatar
patil-suraj committed
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487

    def convert_to_fp16(self):
        """
        Convert the torso of the model to float16.
        """
        self.input_blocks.apply(convert_module_to_f16)
        self.middle_block.apply(convert_module_to_f16)
        self.output_blocks.apply(convert_module_to_f16)

    def convert_to_fp32(self):
        """
        Convert the torso of the model to float32.
        """
        self.input_blocks.apply(convert_module_to_f32)
        self.middle_block.apply(convert_module_to_f32)
        self.output_blocks.apply(convert_module_to_f32)

Patrick von Platen's avatar
Patrick von Platen committed
488
    def forward(self, x, timesteps=None, context=None, y=None, **kwargs):
patil-suraj's avatar
patil-suraj committed
489
        """
Patrick von Platen's avatar
Patrick von Platen committed
490
491
492
        Apply the model to an input batch. :param x: an [N x C x ...] Tensor of inputs. :param timesteps: a 1-D batch
        of timesteps. :param context: conditioning plugged in via crossattn :param y: an [N] Tensor of labels, if
        class-conditional. :return: an [N x C x ...] Tensor of outputs.
patil-suraj's avatar
patil-suraj committed
493
494
495
496
497
        """
        assert (y is not None) == (
            self.num_classes is not None
        ), "must specify y if and only if the model is class-conditional"
        hs = []
patil-suraj's avatar
patil-suraj committed
498
499
        if not torch.is_tensor(timesteps):
            timesteps = torch.tensor([timesteps], dtype=torch.long, device=x.device)
500
        t_emb = get_timestep_embedding(timesteps, self.model_channels, flip_sin_to_cos=True, downscale_freq_shift=0)
patil-suraj's avatar
patil-suraj committed
501
502
503
504
505
506
        emb = self.time_embed(t_emb)

        if self.num_classes is not None:
            assert y.shape == (x.shape[0],)
            emb = emb + self.label_emb(y)

patil-suraj's avatar
patil-suraj committed
507
        h = x.type(self.dtype_)
patil-suraj's avatar
patil-suraj committed
508
509
510
        for module in self.input_blocks:
            h = module(h, emb, context)
            hs.append(h)
Patrick von Platen's avatar
Patrick von Platen committed
511
        h = self.mid(h, emb, context)
patil-suraj's avatar
patil-suraj committed
512
513
514
515
516
517
518
519
520
521
        for module in self.output_blocks:
            h = torch.cat([h, hs.pop()], dim=1)
            h = module(h, emb, context)
        h = h.type(x.dtype)
        if self.predict_codebook_ids:
            return self.id_predictor(h)
        else:
            return self.out(h)


522
class SpatialTransformer(nn.Module):
patil-suraj's avatar
patil-suraj committed
523
    """
524
525
    Transformer block for image-like data. First, project the input (aka embedding) and reshape to b, t, d. Then apply
    standard transformer action. Finally, reshape to image
patil-suraj's avatar
patil-suraj committed
526
527
    """

528
    def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0.0, context_dim=None):
patil-suraj's avatar
patil-suraj committed
529
530
        super().__init__()
        self.in_channels = in_channels
531
532
        inner_dim = n_heads * d_head
        self.norm = Normalize(in_channels)
patil-suraj's avatar
patil-suraj committed
533

534
        self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
patil-suraj's avatar
patil-suraj committed
535

536
537
538
539
540
        self.transformer_blocks = nn.ModuleList(
            [
                BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim)
                for d in range(depth)
            ]
patil-suraj's avatar
patil-suraj committed
541
542
        )

543
        self.proj_out = zero_module(nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
patil-suraj's avatar
patil-suraj committed
544

545
546
547
548
549
550
551
552
553
554
555
556
    def forward(self, x, context=None):
        # note: if no context is given, cross-attention defaults to self-attention
        b, c, h, w = x.shape
        x_in = x
        x = self.norm(x)
        x = self.proj_in(x)
        x = x.permute(0, 2, 3, 1).reshape(b, h * w, c)
        for block in self.transformer_blocks:
            x = block(x, context=context)
        x = x.reshape(b, h, w, c).permute(0, 3, 1, 2)
        x = self.proj_out(x)
        return x + x_in
patil-suraj's avatar
patil-suraj committed
557
558


559
560
561
562
563
564
565
566
567
568
569
570
571
572
class BasicTransformerBlock(nn.Module):
    def __init__(self, dim, n_heads, d_head, dropout=0.0, context_dim=None, gated_ff=True, checkpoint=True):
        super().__init__()
        self.attn1 = CrossAttention(
            query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout
        )  # is a self-attention
        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
        self.attn2 = CrossAttention(
            query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, dropout=dropout
        )  # is self-attn if context is none
        self.norm1 = nn.LayerNorm(dim)
        self.norm2 = nn.LayerNorm(dim)
        self.norm3 = nn.LayerNorm(dim)
        self.checkpoint = checkpoint
patil-suraj's avatar
patil-suraj committed
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
    def forward(self, x, context=None):
        x = self.attn1(self.norm1(x)) + x
        x = self.attn2(self.norm2(x), context=context) + x
        x = self.ff(self.norm3(x)) + x
        return x


class CrossAttention(nn.Module):
    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
        super().__init__()
        inner_dim = dim_head * heads
        context_dim = default(context_dim, query_dim)

        self.scale = dim_head**-0.5
        self.heads = heads

        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))

    def reshape_heads_to_batch_dim(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        head_size = self.heads
        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
        return tensor

    def reshape_batch_dim_to_heads(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        head_size = self.heads
        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
        return tensor

    def forward(self, x, context=None, mask=None):
        batch_size, sequence_length, dim = x.shape

        h = self.heads

        q = self.to_q(x)
        context = default(context, x)
        k = self.to_k(context)
        v = self.to_v(context)

        q = self.reshape_heads_to_batch_dim(q)
        k = self.reshape_heads_to_batch_dim(k)
        v = self.reshape_heads_to_batch_dim(v)

        sim = torch.einsum("b i d, b j d -> b i j", q, k) * self.scale

        if exists(mask):
            mask = mask.reshape(batch_size, -1)
            max_neg_value = -torch.finfo(sim.dtype).max
            mask = mask[:, None, :].repeat(h, 1, 1)
            sim.masked_fill_(~mask, max_neg_value)

        # attention, what we cannot get enough of
        attn = sim.softmax(dim=-1)

        out = torch.einsum("b i j, b j d -> b i d", attn, v)
        out = self.reshape_batch_dim_to_heads(out)
        return self.to_out(out)