openaimodel.py 26.7 KB
Newer Older
comfyanonymous's avatar
comfyanonymous committed
1
2
3
4
5
6
7
8
from abc import abstractmethod
import math

import numpy as np
import torch as th
import torch.nn as nn
import torch.nn.functional as F

comfyanonymous's avatar
comfyanonymous committed
9
from .util import (
comfyanonymous's avatar
comfyanonymous committed
10
11
12
13
14
15
    checkpoint,
    avg_pool_nd,
    zero_module,
    normalization,
    timestep_embedding,
)
comfyanonymous's avatar
comfyanonymous committed
16
17
from ..attention import SpatialTransformer
from comfy.ldm.util import exists
comfyanonymous's avatar
comfyanonymous committed
18
import comfy.ops
comfyanonymous's avatar
comfyanonymous committed
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37

class TimestepBlock(nn.Module):
    """
    Any module where forward() takes timestep embeddings as a second argument.
    """

    @abstractmethod
    def forward(self, x, emb):
        """
        Apply the module to `x` given `emb` timestep embeddings.
        """


class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
    """
    A sequential module that passes timestep embeddings to the children that
    support it as an extra input.
    """

38
    def forward(self, x, emb, context=None, transformer_options={}, output_shape=None):
comfyanonymous's avatar
comfyanonymous committed
39
40
41
42
        for layer in self:
            if isinstance(layer, TimestepBlock):
                x = layer(x, emb)
            elif isinstance(layer, SpatialTransformer):
43
                x = layer(x, context, transformer_options)
44
45
            elif isinstance(layer, Upsample):
                x = layer(x, output_shape=output_shape)
comfyanonymous's avatar
comfyanonymous committed
46
47
48
49
            else:
                x = layer(x)
        return x

50
51
52
53
54
55
56
57
58
59
60
61
62
#This is needed because accelerate makes a copy of transformer_options which breaks "current_index"
def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, output_shape=None):
    for layer in ts:
        if isinstance(layer, TimestepBlock):
            x = layer(x, emb)
        elif isinstance(layer, SpatialTransformer):
            x = layer(x, context, transformer_options)
            transformer_options["current_index"] += 1
        elif isinstance(layer, Upsample):
            x = layer(x, output_shape=output_shape)
        else:
            x = layer(x)
    return x
comfyanonymous's avatar
comfyanonymous committed
63
64
65
66
67
68
69
70
71
72

class Upsample(nn.Module):
    """
    An upsampling layer with an optional convolution.
    :param channels: channels in the inputs and outputs.
    :param use_conv: a bool determining if a convolution is applied.
    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
                 upsampling occurs in the inner-two dimensions.
    """

comfyanonymous's avatar
comfyanonymous committed
73
    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=comfy.ops):
comfyanonymous's avatar
comfyanonymous committed
74
75
76
77
78
79
        super().__init__()
        self.channels = channels
        self.out_channels = out_channels or channels
        self.use_conv = use_conv
        self.dims = dims
        if use_conv:
comfyanonymous's avatar
comfyanonymous committed
80
            self.conv = operations.conv_nd(dims, self.channels, self.out_channels, 3, padding=padding, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
81

82
    def forward(self, x, output_shape=None):
comfyanonymous's avatar
comfyanonymous committed
83
84
        assert x.shape[1] == self.channels
        if self.dims == 3:
85
86
87
88
            shape = [x.shape[2], x.shape[3] * 2, x.shape[4] * 2]
            if output_shape is not None:
                shape[1] = output_shape[3]
                shape[2] = output_shape[4]
comfyanonymous's avatar
comfyanonymous committed
89
        else:
90
91
92
93
94
95
            shape = [x.shape[2] * 2, x.shape[3] * 2]
            if output_shape is not None:
                shape[0] = output_shape[2]
                shape[1] = output_shape[3]

        x = F.interpolate(x, size=shape, mode="nearest")
comfyanonymous's avatar
comfyanonymous committed
96
97
98
99
100
101
102
103
104
105
106
107
108
        if self.use_conv:
            x = self.conv(x)
        return x

class Downsample(nn.Module):
    """
    A downsampling layer with an optional convolution.
    :param channels: channels in the inputs and outputs.
    :param use_conv: a bool determining if a convolution is applied.
    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
                 downsampling occurs in the inner-two dimensions.
    """

comfyanonymous's avatar
comfyanonymous committed
109
    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=comfy.ops):
comfyanonymous's avatar
comfyanonymous committed
110
111
112
113
114
115
116
        super().__init__()
        self.channels = channels
        self.out_channels = out_channels or channels
        self.use_conv = use_conv
        self.dims = dims
        stride = 2 if dims != 3 else (1, 2, 2)
        if use_conv:
comfyanonymous's avatar
comfyanonymous committed
117
            self.op = operations.conv_nd(
118
                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding, dtype=dtype, device=device
comfyanonymous's avatar
comfyanonymous committed
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
            )
        else:
            assert self.channels == self.out_channels
            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)

    def forward(self, x):
        assert x.shape[1] == self.channels
        return self.op(x)


class ResBlock(TimestepBlock):
    """
    A residual block that can optionally change the number of channels.
    :param channels: the number of input channels.
    :param emb_channels: the number of timestep embedding channels.
    :param dropout: the rate of dropout.
    :param out_channels: if specified, the number of out channels.
    :param use_conv: if True and out_channels is specified, use a spatial
        convolution instead of a smaller 1x1 convolution to change the
        channels in the skip connection.
    :param dims: determines if the signal is 1D, 2D, or 3D.
    :param use_checkpoint: if True, use gradient checkpointing on this module.
    :param up: if True, use this block for upsampling.
    :param down: if True, use this block for downsampling.
    """

    def __init__(
        self,
        channels,
        emb_channels,
        dropout,
        out_channels=None,
        use_conv=False,
        use_scale_shift_norm=False,
        dims=2,
        use_checkpoint=False,
        up=False,
        down=False,
157
158
        dtype=None,
        device=None,
comfyanonymous's avatar
comfyanonymous committed
159
        operations=comfy.ops
comfyanonymous's avatar
comfyanonymous committed
160
161
162
163
164
165
166
167
168
169
170
    ):
        super().__init__()
        self.channels = channels
        self.emb_channels = emb_channels
        self.dropout = dropout
        self.out_channels = out_channels or channels
        self.use_conv = use_conv
        self.use_checkpoint = use_checkpoint
        self.use_scale_shift_norm = use_scale_shift_norm

        self.in_layers = nn.Sequential(
171
            nn.GroupNorm(32, channels, dtype=dtype, device=device),
comfyanonymous's avatar
comfyanonymous committed
172
            nn.SiLU(),
comfyanonymous's avatar
comfyanonymous committed
173
            operations.conv_nd(dims, channels, self.out_channels, 3, padding=1, dtype=dtype, device=device),
comfyanonymous's avatar
comfyanonymous committed
174
175
176
177
178
        )

        self.updown = up or down

        if up:
179
180
            self.h_upd = Upsample(channels, False, dims, dtype=dtype, device=device)
            self.x_upd = Upsample(channels, False, dims, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
181
        elif down:
182
183
            self.h_upd = Downsample(channels, False, dims, dtype=dtype, device=device)
            self.x_upd = Downsample(channels, False, dims, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
184
185
186
187
188
        else:
            self.h_upd = self.x_upd = nn.Identity()

        self.emb_layers = nn.Sequential(
            nn.SiLU(),
comfyanonymous's avatar
comfyanonymous committed
189
            operations.Linear(
comfyanonymous's avatar
comfyanonymous committed
190
                emb_channels,
191
                2 * self.out_channels if use_scale_shift_norm else self.out_channels, dtype=dtype, device=device
comfyanonymous's avatar
comfyanonymous committed
192
193
194
            ),
        )
        self.out_layers = nn.Sequential(
195
            nn.GroupNorm(32, self.out_channels, dtype=dtype, device=device),
comfyanonymous's avatar
comfyanonymous committed
196
197
198
            nn.SiLU(),
            nn.Dropout(p=dropout),
            zero_module(
comfyanonymous's avatar
comfyanonymous committed
199
                operations.conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
200
201
202
203
204
205
            ),
        )

        if self.out_channels == channels:
            self.skip_connection = nn.Identity()
        elif use_conv:
comfyanonymous's avatar
comfyanonymous committed
206
            self.skip_connection = operations.conv_nd(
207
                dims, channels, self.out_channels, 3, padding=1, dtype=dtype, device=device
comfyanonymous's avatar
comfyanonymous committed
208
209
            )
        else:
comfyanonymous's avatar
comfyanonymous committed
210
            self.skip_connection = operations.conv_nd(dims, channels, self.out_channels, 1, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
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

    def forward(self, x, emb):
        """
        Apply the block to a Tensor, conditioned on a timestep embedding.
        :param x: an [N x C x ...] Tensor of features.
        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
        :return: an [N x C x ...] Tensor of outputs.
        """
        return checkpoint(
            self._forward, (x, emb), self.parameters(), self.use_checkpoint
        )


    def _forward(self, x, emb):
        if self.updown:
            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
            h = in_rest(x)
            h = self.h_upd(h)
            x = self.x_upd(x)
            h = in_conv(h)
        else:
            h = self.in_layers(x)
        emb_out = self.emb_layers(emb).type(h.dtype)
        while len(emb_out.shape) < len(h.shape):
            emb_out = emb_out[..., None]
        if self.use_scale_shift_norm:
            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
            scale, shift = th.chunk(emb_out, 2, dim=1)
            h = out_norm(h) * (1 + scale) + shift
            h = out_rest(h)
        else:
            h = h + emb_out
            h = self.out_layers(h)
        return self.skip_connection(x) + h

246
247
248
249
250
251
252
253
class Timestep(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.dim = dim

    def forward(self, t):
        return timestep_embedding(t, self.dim)

comfyanonymous's avatar
comfyanonymous committed
254
255
256
257
def apply_control(h, control, name):
    if control is not None and name in control and len(control[name]) > 0:
        ctrl = control[name].pop()
        if ctrl is not None:
258
259
260
261
            try:
                h += ctrl
            except:
                print("warning control could not be applied", h.shape, ctrl.shape)
comfyanonymous's avatar
comfyanonymous committed
262
    return h
263

comfyanonymous's avatar
comfyanonymous committed
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
class UNetModel(nn.Module):
    """
    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 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
        downsampling.
    :param dims: determines if the signal is 1D, 2D, or 3D.
    :param num_classes: if specified (as an int), then this model will be
        class-conditional with `num_classes` classes.
    :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
                               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.
    :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
                                    increased efficiency.
    """

    def __init__(
        self,
        image_size,
        in_channels,
        model_channels,
        out_channels,
        num_res_blocks,
        dropout=0,
        channel_mult=(1, 2, 4, 8),
        conv_resample=True,
        dims=2,
        num_classes=None,
        use_checkpoint=False,
303
        dtype=th.float32,
comfyanonymous's avatar
comfyanonymous committed
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
        num_heads=-1,
        num_head_channels=-1,
        num_heads_upsample=-1,
        use_scale_shift_norm=False,
        resblock_updown=False,
        use_new_attention_order=False,
        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
        legacy=True,
        disable_self_attentions=None,
        num_attention_blocks=None,
        disable_middle_self_attn=False,
        use_linear_in_transformer=False,
319
        adm_in_channels=None,
320
        transformer_depth_middle=None,
321
        transformer_depth_output=None,
322
        device=None,
comfyanonymous's avatar
comfyanonymous committed
323
        operations=comfy.ops,
comfyanonymous's avatar
comfyanonymous committed
324
325
    ):
        super().__init__()
comfyanonymous's avatar
comfyanonymous committed
326
        assert use_spatial_transformer == True, "use_spatial_transformer has to be true"
comfyanonymous's avatar
comfyanonymous committed
327
328
329
330
331
        if use_spatial_transformer:
            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'

        if context_dim is not None:
            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
332
333
334
            # from omegaconf.listconfig import ListConfig
            # if type(context_dim) == ListConfig:
            #     context_dim = list(context_dim)
comfyanonymous's avatar
comfyanonymous committed
335
336
337
338
339
340
341
342
343
344
345
346
347
348

        if num_heads_upsample == -1:
            num_heads_upsample = num_heads

        if num_heads == -1:
            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'

        if num_head_channels == -1:
            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'

        self.image_size = image_size
        self.in_channels = in_channels
        self.model_channels = model_channels
        self.out_channels = out_channels
349

comfyanonymous's avatar
comfyanonymous committed
350
351
352
353
354
355
356
        if isinstance(num_res_blocks, int):
            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
        else:
            if len(num_res_blocks) != len(channel_mult):
                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
                                 "as a list/tuple (per-level) with the same length as channel_mult")
            self.num_res_blocks = num_res_blocks
357

comfyanonymous's avatar
comfyanonymous committed
358
359
360
361
362
363
        if disable_self_attentions is not None:
            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
            assert len(disable_self_attentions) == len(channel_mult)
        if num_attention_blocks is not None:
            assert len(num_attention_blocks) == len(self.num_res_blocks)

364
365
366
        transformer_depth = transformer_depth[:]
        transformer_depth_output = transformer_depth_output[:]

comfyanonymous's avatar
comfyanonymous committed
367
368
369
370
371
        self.dropout = dropout
        self.channel_mult = channel_mult
        self.conv_resample = conv_resample
        self.num_classes = num_classes
        self.use_checkpoint = use_checkpoint
372
        self.dtype = dtype
comfyanonymous's avatar
comfyanonymous committed
373
374
375
376
377
378
379
        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(
comfyanonymous's avatar
comfyanonymous committed
380
            operations.Linear(model_channels, time_embed_dim, dtype=self.dtype, device=device),
comfyanonymous's avatar
comfyanonymous committed
381
            nn.SiLU(),
comfyanonymous's avatar
comfyanonymous committed
382
            operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
comfyanonymous's avatar
comfyanonymous committed
383
384
385
386
387
388
389
390
        )

        if self.num_classes is not None:
            if isinstance(self.num_classes, int):
                self.label_emb = nn.Embedding(num_classes, time_embed_dim)
            elif self.num_classes == "continuous":
                print("setting up linear c_adm embedding layer")
                self.label_emb = nn.Linear(1, time_embed_dim)
391
392
393
394
            elif self.num_classes == "sequential":
                assert adm_in_channels is not None
                self.label_emb = nn.Sequential(
                    nn.Sequential(
comfyanonymous's avatar
comfyanonymous committed
395
                        operations.Linear(adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
396
                        nn.SiLU(),
comfyanonymous's avatar
comfyanonymous committed
397
                        operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
398
399
                    )
                )
comfyanonymous's avatar
comfyanonymous committed
400
401
402
403
404
405
            else:
                raise ValueError()

        self.input_blocks = nn.ModuleList(
            [
                TimestepEmbedSequential(
comfyanonymous's avatar
comfyanonymous committed
406
                    operations.conv_nd(dims, in_channels, model_channels, 3, padding=1, dtype=self.dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
                )
            ]
        )
        self._feature_size = model_channels
        input_block_chans = [model_channels]
        ch = model_channels
        ds = 1
        for level, mult in enumerate(channel_mult):
            for nr in range(self.num_res_blocks[level]):
                layers = [
                    ResBlock(
                        ch,
                        time_embed_dim,
                        dropout,
                        out_channels=mult * model_channels,
                        dims=dims,
                        use_checkpoint=use_checkpoint,
                        use_scale_shift_norm=use_scale_shift_norm,
425
426
                        dtype=self.dtype,
                        device=device,
comfyanonymous's avatar
comfyanonymous committed
427
                        operations=operations,
comfyanonymous's avatar
comfyanonymous committed
428
429
430
                    )
                ]
                ch = mult * model_channels
431
432
                num_transformers = transformer_depth.pop(0)
                if num_transformers > 0:
comfyanonymous's avatar
comfyanonymous committed
433
434
435
436
437
438
439
440
441
442
443
444
445
446
                    if num_head_channels == -1:
                        dim_head = ch // num_heads
                    else:
                        num_heads = ch // num_head_channels
                        dim_head = num_head_channels
                    if legacy:
                        #num_heads = 1
                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
                    if exists(disable_self_attentions):
                        disabled_sa = disable_self_attentions[level]
                    else:
                        disabled_sa = False

                    if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
comfyanonymous's avatar
comfyanonymous committed
447
                        layers.append(SpatialTransformer(
448
                                ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
comfyanonymous's avatar
comfyanonymous committed
449
                                disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
comfyanonymous's avatar
comfyanonymous committed
450
                                use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
comfyanonymous's avatar
comfyanonymous committed
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
                            )
                        )
                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(
                        ResBlock(
                            ch,
                            time_embed_dim,
                            dropout,
                            out_channels=out_ch,
                            dims=dims,
                            use_checkpoint=use_checkpoint,
                            use_scale_shift_norm=use_scale_shift_norm,
                            down=True,
469
470
                            dtype=self.dtype,
                            device=device,
comfyanonymous's avatar
comfyanonymous committed
471
                            operations=operations
comfyanonymous's avatar
comfyanonymous committed
472
473
474
                        )
                        if resblock_updown
                        else Downsample(
comfyanonymous's avatar
comfyanonymous committed
475
                            ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations
comfyanonymous's avatar
comfyanonymous committed
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
                        )
                    )
                )
                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:
            #num_heads = 1
            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
492
        mid_block = [
comfyanonymous's avatar
comfyanonymous committed
493
494
495
496
497
498
499
            ResBlock(
                ch,
                time_embed_dim,
                dropout,
                dims=dims,
                use_checkpoint=use_checkpoint,
                use_scale_shift_norm=use_scale_shift_norm,
500
501
                dtype=self.dtype,
                device=device,
comfyanonymous's avatar
comfyanonymous committed
502
                operations=operations
503
504
505
            )]
        if transformer_depth_middle >= 0:
            mid_block += [SpatialTransformer(  # always uses a self-attn
506
                            ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim,
comfyanonymous's avatar
comfyanonymous committed
507
                            disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
comfyanonymous's avatar
comfyanonymous committed
508
                            use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
comfyanonymous's avatar
comfyanonymous committed
509
510
511
512
513
514
515
516
                        ),
            ResBlock(
                ch,
                time_embed_dim,
                dropout,
                dims=dims,
                use_checkpoint=use_checkpoint,
                use_scale_shift_norm=use_scale_shift_norm,
517
518
                dtype=self.dtype,
                device=device,
comfyanonymous's avatar
comfyanonymous committed
519
                operations=operations
520
521
            )]
        self.middle_block = TimestepEmbedSequential(*mid_block)
comfyanonymous's avatar
comfyanonymous committed
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
        self._feature_size += ch

        self.output_blocks = nn.ModuleList([])
        for level, mult in list(enumerate(channel_mult))[::-1]:
            for i in range(self.num_res_blocks[level] + 1):
                ich = input_block_chans.pop()
                layers = [
                    ResBlock(
                        ch + ich,
                        time_embed_dim,
                        dropout,
                        out_channels=model_channels * mult,
                        dims=dims,
                        use_checkpoint=use_checkpoint,
                        use_scale_shift_norm=use_scale_shift_norm,
537
538
                        dtype=self.dtype,
                        device=device,
comfyanonymous's avatar
comfyanonymous committed
539
                        operations=operations
comfyanonymous's avatar
comfyanonymous committed
540
541
542
                    )
                ]
                ch = model_channels * mult
543
544
                num_transformers = transformer_depth_output.pop()
                if num_transformers > 0:
comfyanonymous's avatar
comfyanonymous committed
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
                    if num_head_channels == -1:
                        dim_head = ch // num_heads
                    else:
                        num_heads = ch // num_head_channels
                        dim_head = num_head_channels
                    if legacy:
                        #num_heads = 1
                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
                    if exists(disable_self_attentions):
                        disabled_sa = disable_self_attentions[level]
                    else:
                        disabled_sa = False

                    if not exists(num_attention_blocks) or i < num_attention_blocks[level]:
                        layers.append(
comfyanonymous's avatar
comfyanonymous committed
560
                            SpatialTransformer(
561
                                ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
comfyanonymous's avatar
comfyanonymous committed
562
                                disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
comfyanonymous's avatar
comfyanonymous committed
563
                                use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
comfyanonymous's avatar
comfyanonymous committed
564
565
566
567
568
569
570
571
572
573
574
575
576
577
                            )
                        )
                if level and i == self.num_res_blocks[level]:
                    out_ch = ch
                    layers.append(
                        ResBlock(
                            ch,
                            time_embed_dim,
                            dropout,
                            out_channels=out_ch,
                            dims=dims,
                            use_checkpoint=use_checkpoint,
                            use_scale_shift_norm=use_scale_shift_norm,
                            up=True,
578
579
                            dtype=self.dtype,
                            device=device,
comfyanonymous's avatar
comfyanonymous committed
580
                            operations=operations
comfyanonymous's avatar
comfyanonymous committed
581
582
                        )
                        if resblock_updown
comfyanonymous's avatar
comfyanonymous committed
583
                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations)
comfyanonymous's avatar
comfyanonymous committed
584
585
586
587
588
589
                    )
                    ds //= 2
                self.output_blocks.append(TimestepEmbedSequential(*layers))
                self._feature_size += ch

        self.out = nn.Sequential(
590
            nn.GroupNorm(32, ch, dtype=self.dtype, device=device),
comfyanonymous's avatar
comfyanonymous committed
591
            nn.SiLU(),
comfyanonymous's avatar
comfyanonymous committed
592
            zero_module(operations.conv_nd(dims, model_channels, out_channels, 3, padding=1, dtype=self.dtype, device=device)),
comfyanonymous's avatar
comfyanonymous committed
593
594
595
        )
        if self.predict_codebook_ids:
            self.id_predictor = nn.Sequential(
596
            nn.GroupNorm(32, ch, dtype=self.dtype, device=device),
comfyanonymous's avatar
comfyanonymous committed
597
            operations.conv_nd(dims, model_channels, n_embed, 1, dtype=self.dtype, device=device),
comfyanonymous's avatar
comfyanonymous committed
598
599
600
            #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
        )

601
    def forward(self, x, timesteps=None, context=None, y=None, control=None, transformer_options={}, **kwargs):
comfyanonymous's avatar
comfyanonymous committed
602
603
604
605
606
607
608
609
        """
        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.
        """
610
        transformer_options["original_shape"] = list(x.shape)
611
        transformer_options["current_index"] = 0
612
        transformer_patches = transformer_options.get("patches", {})
613

comfyanonymous's avatar
comfyanonymous committed
614
615
616
617
        assert (y is not None) == (
            self.num_classes is not None
        ), "must specify y if and only if the model is class-conditional"
        hs = []
618
        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(self.dtype)
comfyanonymous's avatar
comfyanonymous committed
619
620
621
622
623
624
625
        emb = self.time_embed(t_emb)

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

        h = x.type(self.dtype)
comfyanonymous's avatar
comfyanonymous committed
626
        for id, module in enumerate(self.input_blocks):
627
            transformer_options["block"] = ("input", id)
628
            h = forward_timestep_embed(module, h, emb, context, transformer_options)
comfyanonymous's avatar
comfyanonymous committed
629
            h = apply_control(h, control, 'input')
630
631
632
633
634
            if "input_block_patch" in transformer_patches:
                patch = transformer_patches["input_block_patch"]
                for p in patch:
                    h = p(h, transformer_options)

comfyanonymous's avatar
comfyanonymous committed
635
            hs.append(h)
comfyanonymous's avatar
comfyanonymous committed
636

637
        transformer_options["block"] = ("middle", 0)
638
        h = forward_timestep_embed(self.middle_block, h, emb, context, transformer_options)
comfyanonymous's avatar
comfyanonymous committed
639
        h = apply_control(h, control, 'middle')
comfyanonymous's avatar
comfyanonymous committed
640

641
642
        for id, module in enumerate(self.output_blocks):
            transformer_options["block"] = ("output", id)
comfyanonymous's avatar
comfyanonymous committed
643
            hsp = hs.pop()
comfyanonymous's avatar
comfyanonymous committed
644
            hsp = apply_control(hsp, control, 'output')
645

646
647
648
649
650
            if "output_block_patch" in transformer_patches:
                patch = transformer_patches["output_block_patch"]
                for p in patch:
                    h, hsp = p(h, hsp, transformer_options)

comfyanonymous's avatar
comfyanonymous committed
651
            h = th.cat([h, hsp], dim=1)
comfyanonymous's avatar
comfyanonymous committed
652
            del hsp
653
654
655
656
            if len(hs) > 0:
                output_shape = hs[-1].shape
            else:
                output_shape = None
657
            h = forward_timestep_embed(module, h, emb, context, transformer_options, output_shape)
comfyanonymous's avatar
comfyanonymous committed
658
659
660
661
662
        h = h.type(x.dtype)
        if self.predict_codebook_ids:
            return self.id_predictor(h)
        else:
            return self.out(h)