model.py 31.3 KB
Newer Older
1
2
# pytorch_diffusion + derived encoder decoder
import math
Fazzie's avatar
Fazzie committed
3
4
5
from typing import Any, Optional

import numpy as np
6
7
8
9
import torch
import torch.nn as nn
from einops import rearrange

Fazzie's avatar
Fazzie committed
10
11
12
13
14
15
16
17
18
19
20
21
22
23
try:
    from lightning.pytorch.utilities import rank_zero_info
except:
    from pytorch_lightning.utilities import rank_zero_info

from ldm.modules.attention import MemoryEfficientCrossAttention

try:
    import xformers
    import xformers.ops
    XFORMERS_IS_AVAILBLE = True
except:
    XFORMERS_IS_AVAILBLE = False
    print("No module 'xformers'. Proceeding without it.")
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41


def get_timestep_embedding(timesteps, embedding_dim):
    """
    This matches the implementation in Denoising Diffusion Probabilistic Models:
    From Fairseq.
    Build sinusoidal embeddings.
    This matches the implementation in tensor2tensor, but differs slightly
    from the description in Section 3.5 of "Attention Is All You Need".
    """
    assert len(timesteps.shape) == 1

    half_dim = embedding_dim // 2
    emb = math.log(10000) / (half_dim - 1)
    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
    emb = emb.to(device=timesteps.device)
    emb = timesteps.float()[:, None] * emb[None, :]
    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
Fazzie's avatar
Fazzie committed
42
43
    if embedding_dim % 2 == 1:    # zero pad
        emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
44
45
46
47
48
    return emb


def nonlinearity(x):
    # swish
Fazzie's avatar
Fazzie committed
49
    return x * torch.sigmoid(x)
50
51
52
53
54
55
56


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


class Upsample(nn.Module):
Fazzie's avatar
Fazzie committed
57

58
59
60
61
    def __init__(self, in_channels, with_conv):
        super().__init__()
        self.with_conv = with_conv
        if self.with_conv:
Fazzie's avatar
Fazzie committed
62
            self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1)
63
64
65
66
67
68
69
70
71

    def forward(self, x):
        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
        if self.with_conv:
            x = self.conv(x)
        return x


class Downsample(nn.Module):
Fazzie's avatar
Fazzie committed
72

73
74
75
76
77
    def __init__(self, in_channels, with_conv):
        super().__init__()
        self.with_conv = with_conv
        if self.with_conv:
            # no asymmetric padding in torch conv, must do it ourselves
Fazzie's avatar
Fazzie committed
78
            self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
79
80
81

    def forward(self, x):
        if self.with_conv:
Fazzie's avatar
Fazzie committed
82
            pad = (0, 1, 0, 1)
83
84
85
86
87
88
89
90
            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
            x = self.conv(x)
        else:
            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
        return x


class ResnetBlock(nn.Module):
Fazzie's avatar
Fazzie committed
91
92

    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, dropout, temb_channels=512):
93
94
95
96
97
98
99
        super().__init__()
        self.in_channels = in_channels
        out_channels = in_channels if out_channels is None else out_channels
        self.out_channels = out_channels
        self.use_conv_shortcut = conv_shortcut

        self.norm1 = Normalize(in_channels)
Fazzie's avatar
Fazzie committed
100
        self.conv1 = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
101
        if temb_channels > 0:
Fazzie's avatar
Fazzie committed
102
            self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
103
104
        self.norm2 = Normalize(out_channels)
        self.dropout = torch.nn.Dropout(dropout)
Fazzie's avatar
Fazzie committed
105
        self.conv2 = torch.nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
106
107
        if self.in_channels != self.out_channels:
            if self.use_conv_shortcut:
Fazzie's avatar
Fazzie committed
108
                self.conv_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
109
            else:
Fazzie's avatar
Fazzie committed
110
                self.nin_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
111
112
113
114
115
116
117
118

    def forward(self, x, temb):
        h = x
        h = self.norm1(h)
        h = nonlinearity(h)
        h = self.conv1(h)

        if temb is not None:
Fazzie's avatar
Fazzie committed
119
            h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
120
121
122
123
124
125
126
127
128
129
130
131

        h = self.norm2(h)
        h = nonlinearity(h)
        h = self.dropout(h)
        h = self.conv2(h)

        if self.in_channels != self.out_channels:
            if self.use_conv_shortcut:
                x = self.conv_shortcut(x)
            else:
                x = self.nin_shortcut(x)

Fazzie's avatar
Fazzie committed
132
        return x + h
133
134
135


class AttnBlock(nn.Module):
Fazzie's avatar
Fazzie committed
136

137
138
139
140
141
    def __init__(self, in_channels):
        super().__init__()
        self.in_channels = in_channels

        self.norm = Normalize(in_channels)
Fazzie's avatar
Fazzie committed
142
143
144
145
        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)
146
147
148
149
150
151
152
153
154

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

        # compute attention
Fazzie's avatar
Fazzie committed
155
156
157
158
159
        b, c, h, w = q.shape
        q = q.reshape(b, c, h * w)
        q = q.permute(0, 2, 1)    # b,hw,c
        k = k.reshape(b, c, h * w)    # b,c,hw
        w_ = torch.bmm(q, k)    # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
160
161
162
163
        w_ = w_ * (int(c)**(-0.5))
        w_ = torch.nn.functional.softmax(w_, dim=2)

        # attend to values
Fazzie's avatar
Fazzie committed
164
165
166
167
        v = v.reshape(b, c, h * w)
        w_ = w_.permute(0, 2, 1)    # b,hw,hw (first hw of k, second of q)
        h_ = torch.bmm(v, w_)    # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
        h_ = h_.reshape(b, c, h, w)
168
169
170

        h_ = self.proj_out(h_)

Fazzie's avatar
Fazzie committed
171
172
        return x + h_

173

Fazzie's avatar
Fazzie committed
174
175
176
177
178
179
class MemoryEfficientAttnBlock(nn.Module):
    """
        Uses xformers efficient implementation,
        see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
        Note: this is a single-head self-attention operation
    """
Fazzie's avatar
Fazzie committed
180

Fazzie's avatar
Fazzie committed
181
182
183
184
185
186
    #
    def __init__(self, in_channels):
        super().__init__()
        self.in_channels = in_channels

        self.norm = Normalize(in_channels)
Fazzie's avatar
Fazzie committed
187
188
189
190
        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)
Fazzie's avatar
Fazzie committed
191
        self.attention_op: Optional[Any] = None
192

Fazzie's avatar
Fazzie committed
193
194
195
196
197
198
199
200
201
202
203
204
    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, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v))

        q, k, v = map(
Fazzie's avatar
Fazzie committed
205
206
            lambda t: t.unsqueeze(3).reshape(B, t.shape[1], 1, C).permute(0, 2, 1, 3).reshape(B * 1, t.shape[1], C).
            contiguous(),
Fazzie's avatar
Fazzie committed
207
208
209
210
            (q, k, v),
        )
        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)

Fazzie's avatar
Fazzie committed
211
        out = (out.unsqueeze(0).reshape(B, 1, out.shape[1], C).permute(0, 2, 1, 3).reshape(B, out.shape[1], C))
Fazzie's avatar
Fazzie committed
212
213
        out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C)
        out = self.proj_out(out)
Fazzie's avatar
Fazzie committed
214
        return x + out
Fazzie's avatar
Fazzie committed
215
216
217


class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention):
Fazzie's avatar
Fazzie committed
218

Fazzie's avatar
Fazzie committed
219
220
221
222
223
224
225
226
227
    def forward(self, x, context=None, mask=None):
        b, c, h, w = x.shape
        x = rearrange(x, 'b c h w -> b (h w) c')
        out = super().forward(x, context=context, mask=mask)
        out = rearrange(out, 'b (h w) c -> b c h w', h=h, w=w, c=c)
        return x + out


def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
Fazzie's avatar
Fazzie committed
228
229
    assert attn_type in ["vanilla", "vanilla-xformers", "memory-efficient-cross-attn", "linear",
                         "none"], f'attn_type {attn_type} unknown'
Fazzie's avatar
Fazzie committed
230
231
    if XFORMERS_IS_AVAILBLE and attn_type == "vanilla":
        attn_type = "vanilla-xformers"
232
    if attn_type == "vanilla":
Fazzie's avatar
Fazzie committed
233
        assert attn_kwargs is None
234
        return AttnBlock(in_channels)
Fazzie's avatar
Fazzie committed
235
236
237
238
239
240
    elif attn_type == "vanilla-xformers":
        rank_zero_info(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...")
        return MemoryEfficientAttnBlock(in_channels)
    elif type == "memory-efficient-cross-attn":
        attn_kwargs["query_dim"] = in_channels
        return MemoryEfficientCrossAttentionWrapper(**attn_kwargs)
241
242
243
    elif attn_type == "none":
        return nn.Identity(in_channels)
    else:
Fazzie's avatar
Fazzie committed
244
        raise NotImplementedError()
245
246
247


class Model(nn.Module):
Fazzie's avatar
Fazzie committed
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262

    def __init__(self,
                 *,
                 ch,
                 out_ch,
                 ch_mult=(1, 2, 4, 8),
                 num_res_blocks,
                 attn_resolutions,
                 dropout=0.0,
                 resamp_with_conv=True,
                 in_channels,
                 resolution,
                 use_timestep=True,
                 use_linear_attn=False,
                 attn_type="vanilla"):
263
        super().__init__()
Fazzie's avatar
Fazzie committed
264
265
        if use_linear_attn:
            attn_type = "linear"
266
        self.ch = ch
Fazzie's avatar
Fazzie committed
267
        self.temb_ch = self.ch * 4
268
269
270
271
272
273
274
275
        self.num_resolutions = len(ch_mult)
        self.num_res_blocks = num_res_blocks
        self.resolution = resolution
        self.in_channels = in_channels

        self.use_timestep = use_timestep
        if self.use_timestep:
            # timestep embedding
Fazzie's avatar
Fazzie committed
276
            self.temb = nn.Module()
277
            self.temb.dense = nn.ModuleList([
Fazzie's avatar
Fazzie committed
278
279
                torch.nn.Linear(self.ch, self.temb_ch),
                torch.nn.Linear(self.temb_ch, self.temb_ch),
280
281
282
            ])

        # downsampling
Fazzie's avatar
Fazzie committed
283
        self.conv_in = torch.nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
284
285

        curr_res = resolution
Fazzie's avatar
Fazzie committed
286
        in_ch_mult = (1,) + tuple(ch_mult)
287
288
289
290
        self.down = nn.ModuleList()
        for i_level in range(self.num_resolutions):
            block = nn.ModuleList()
            attn = nn.ModuleList()
Fazzie's avatar
Fazzie committed
291
292
            block_in = ch * in_ch_mult[i_level]
            block_out = ch * ch_mult[i_level]
293
            for i_block in range(self.num_res_blocks):
Fazzie's avatar
Fazzie committed
294
295
296
297
298
                block.append(
                    ResnetBlock(in_channels=block_in,
                                out_channels=block_out,
                                temb_channels=self.temb_ch,
                                dropout=dropout))
299
300
301
                block_in = block_out
                if curr_res in attn_resolutions:
                    attn.append(make_attn(block_in, attn_type=attn_type))
Fazzie's avatar
Fazzie committed
302
            down = nn.Module()
303
304
            down.block = block
            down.attn = attn
Fazzie's avatar
Fazzie committed
305
            if i_level != self.num_resolutions - 1:
306
307
308
309
310
                down.downsample = Downsample(block_in, resamp_with_conv)
                curr_res = curr_res // 2
            self.down.append(down)

        # middle
Fazzie's avatar
Fazzie committed
311
        self.mid = nn.Module()
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
        self.mid.block_1 = ResnetBlock(in_channels=block_in,
                                       out_channels=block_in,
                                       temb_channels=self.temb_ch,
                                       dropout=dropout)
        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
        self.mid.block_2 = ResnetBlock(in_channels=block_in,
                                       out_channels=block_in,
                                       temb_channels=self.temb_ch,
                                       dropout=dropout)

        # upsampling
        self.up = nn.ModuleList()
        for i_level in reversed(range(self.num_resolutions)):
            block = nn.ModuleList()
            attn = nn.ModuleList()
Fazzie's avatar
Fazzie committed
327
328
329
            block_out = ch * ch_mult[i_level]
            skip_in = ch * ch_mult[i_level]
            for i_block in range(self.num_res_blocks + 1):
330
                if i_block == self.num_res_blocks:
Fazzie's avatar
Fazzie committed
331
332
333
334
335
336
                    skip_in = ch * in_ch_mult[i_level]
                block.append(
                    ResnetBlock(in_channels=block_in + skip_in,
                                out_channels=block_out,
                                temb_channels=self.temb_ch,
                                dropout=dropout))
337
338
339
                block_in = block_out
                if curr_res in attn_resolutions:
                    attn.append(make_attn(block_in, attn_type=attn_type))
Fazzie's avatar
Fazzie committed
340
            up = nn.Module()
341
342
343
344
345
            up.block = block
            up.attn = attn
            if i_level != 0:
                up.upsample = Upsample(block_in, resamp_with_conv)
                curr_res = curr_res * 2
Fazzie's avatar
Fazzie committed
346
            self.up.insert(0, up)    # prepend to get consistent order
347
348
349

        # end
        self.norm_out = Normalize(block_in)
Fazzie's avatar
Fazzie committed
350
        self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374

    def forward(self, x, t=None, context=None):
        #assert x.shape[2] == x.shape[3] == self.resolution
        if context is not None:
            # assume aligned context, cat along channel axis
            x = torch.cat((x, context), dim=1)
        if self.use_timestep:
            # timestep embedding
            assert t is not None
            temb = get_timestep_embedding(t, self.ch)
            temb = self.temb.dense[0](temb)
            temb = nonlinearity(temb)
            temb = self.temb.dense[1](temb)
        else:
            temb = None

        # downsampling
        hs = [self.conv_in(x)]
        for i_level in range(self.num_resolutions):
            for i_block in range(self.num_res_blocks):
                h = self.down[i_level].block[i_block](hs[-1], temb)
                if len(self.down[i_level].attn) > 0:
                    h = self.down[i_level].attn[i_block](h)
                hs.append(h)
Fazzie's avatar
Fazzie committed
375
            if i_level != self.num_resolutions - 1:
376
377
378
379
380
381
382
383
384
385
                hs.append(self.down[i_level].downsample(hs[-1]))

        # middle
        h = hs[-1]
        h = self.mid.block_1(h, temb)
        h = self.mid.attn_1(h)
        h = self.mid.block_2(h, temb)

        # upsampling
        for i_level in reversed(range(self.num_resolutions)):
Fazzie's avatar
Fazzie committed
386
387
            for i_block in range(self.num_res_blocks + 1):
                h = self.up[i_level].block[i_block](torch.cat([h, hs.pop()], dim=1), temb)
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
                if len(self.up[i_level].attn) > 0:
                    h = self.up[i_level].attn[i_block](h)
            if i_level != 0:
                h = self.up[i_level].upsample(h)

        # end
        h = self.norm_out(h)
        h = nonlinearity(h)
        h = self.conv_out(h)
        return h

    def get_last_layer(self):
        return self.conv_out.weight


class Encoder(nn.Module):
Fazzie's avatar
Fazzie committed
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419

    def __init__(self,
                 *,
                 ch,
                 out_ch,
                 ch_mult=(1, 2, 4, 8),
                 num_res_blocks,
                 attn_resolutions,
                 dropout=0.0,
                 resamp_with_conv=True,
                 in_channels,
                 resolution,
                 z_channels,
                 double_z=True,
                 use_linear_attn=False,
                 attn_type="vanilla",
420
421
                 **ignore_kwargs):
        super().__init__()
Fazzie's avatar
Fazzie committed
422
423
        if use_linear_attn:
            attn_type = "linear"
424
425
426
427
428
429
430
431
        self.ch = ch
        self.temb_ch = 0
        self.num_resolutions = len(ch_mult)
        self.num_res_blocks = num_res_blocks
        self.resolution = resolution
        self.in_channels = in_channels

        # downsampling
Fazzie's avatar
Fazzie committed
432
        self.conv_in = torch.nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
433
434

        curr_res = resolution
Fazzie's avatar
Fazzie committed
435
        in_ch_mult = (1,) + tuple(ch_mult)
436
437
438
439
440
        self.in_ch_mult = in_ch_mult
        self.down = nn.ModuleList()
        for i_level in range(self.num_resolutions):
            block = nn.ModuleList()
            attn = nn.ModuleList()
Fazzie's avatar
Fazzie committed
441
442
            block_in = ch * in_ch_mult[i_level]
            block_out = ch * ch_mult[i_level]
443
            for i_block in range(self.num_res_blocks):
Fazzie's avatar
Fazzie committed
444
445
446
447
448
                block.append(
                    ResnetBlock(in_channels=block_in,
                                out_channels=block_out,
                                temb_channels=self.temb_ch,
                                dropout=dropout))
449
450
451
                block_in = block_out
                if curr_res in attn_resolutions:
                    attn.append(make_attn(block_in, attn_type=attn_type))
Fazzie's avatar
Fazzie committed
452
            down = nn.Module()
453
454
            down.block = block
            down.attn = attn
Fazzie's avatar
Fazzie committed
455
            if i_level != self.num_resolutions - 1:
456
457
458
459
460
                down.downsample = Downsample(block_in, resamp_with_conv)
                curr_res = curr_res // 2
            self.down.append(down)

        # middle
Fazzie's avatar
Fazzie committed
461
        self.mid = nn.Module()
462
463
464
465
466
467
468
469
470
471
472
473
474
        self.mid.block_1 = ResnetBlock(in_channels=block_in,
                                       out_channels=block_in,
                                       temb_channels=self.temb_ch,
                                       dropout=dropout)
        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
        self.mid.block_2 = ResnetBlock(in_channels=block_in,
                                       out_channels=block_in,
                                       temb_channels=self.temb_ch,
                                       dropout=dropout)

        # end
        self.norm_out = Normalize(block_in)
        self.conv_out = torch.nn.Conv2d(block_in,
Fazzie's avatar
Fazzie committed
475
                                        2 * z_channels if double_z else z_channels,
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
                                        kernel_size=3,
                                        stride=1,
                                        padding=1)

    def forward(self, x):
        # timestep embedding
        temb = None

        # downsampling
        hs = [self.conv_in(x)]
        for i_level in range(self.num_resolutions):
            for i_block in range(self.num_res_blocks):
                h = self.down[i_level].block[i_block](hs[-1], temb)
                if len(self.down[i_level].attn) > 0:
                    h = self.down[i_level].attn[i_block](h)
                hs.append(h)
Fazzie's avatar
Fazzie committed
492
            if i_level != self.num_resolutions - 1:
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
                hs.append(self.down[i_level].downsample(hs[-1]))

        # middle
        h = hs[-1]
        h = self.mid.block_1(h, temb)
        h = self.mid.attn_1(h)
        h = self.mid.block_2(h, temb)

        # end
        h = self.norm_out(h)
        h = nonlinearity(h)
        h = self.conv_out(h)
        return h


class Decoder(nn.Module):
Fazzie's avatar
Fazzie committed
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526

    def __init__(self,
                 *,
                 ch,
                 out_ch,
                 ch_mult=(1, 2, 4, 8),
                 num_res_blocks,
                 attn_resolutions,
                 dropout=0.0,
                 resamp_with_conv=True,
                 in_channels,
                 resolution,
                 z_channels,
                 give_pre_end=False,
                 tanh_out=False,
                 use_linear_attn=False,
                 attn_type="vanilla",
                 **ignorekwargs):
527
        super().__init__()
Fazzie's avatar
Fazzie committed
528
529
        if use_linear_attn:
            attn_type = "linear"
530
531
532
533
534
535
536
537
538
539
        self.ch = ch
        self.temb_ch = 0
        self.num_resolutions = len(ch_mult)
        self.num_res_blocks = num_res_blocks
        self.resolution = resolution
        self.in_channels = in_channels
        self.give_pre_end = give_pre_end
        self.tanh_out = tanh_out

        # compute in_ch_mult, block_in and curr_res at lowest res
Fazzie's avatar
Fazzie committed
540
541
542
543
544
        in_ch_mult = (1,) + tuple(ch_mult)
        block_in = ch * ch_mult[self.num_resolutions - 1]
        curr_res = resolution // 2**(self.num_resolutions - 1)
        self.z_shape = (1, z_channels, curr_res, curr_res)
        rank_zero_info("Working with z of shape {} = {} dimensions.".format(self.z_shape, np.prod(self.z_shape)))
545
546

        # z to block_in
Fazzie's avatar
Fazzie committed
547
        self.conv_in = torch.nn.Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1)
548
549

        # middle
Fazzie's avatar
Fazzie committed
550
        self.mid = nn.Module()
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
        self.mid.block_1 = ResnetBlock(in_channels=block_in,
                                       out_channels=block_in,
                                       temb_channels=self.temb_ch,
                                       dropout=dropout)
        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
        self.mid.block_2 = ResnetBlock(in_channels=block_in,
                                       out_channels=block_in,
                                       temb_channels=self.temb_ch,
                                       dropout=dropout)

        # upsampling
        self.up = nn.ModuleList()
        for i_level in reversed(range(self.num_resolutions)):
            block = nn.ModuleList()
            attn = nn.ModuleList()
Fazzie's avatar
Fazzie committed
566
567
568
569
570
571
572
            block_out = ch * ch_mult[i_level]
            for i_block in range(self.num_res_blocks + 1):
                block.append(
                    ResnetBlock(in_channels=block_in,
                                out_channels=block_out,
                                temb_channels=self.temb_ch,
                                dropout=dropout))
573
574
575
                block_in = block_out
                if curr_res in attn_resolutions:
                    attn.append(make_attn(block_in, attn_type=attn_type))
Fazzie's avatar
Fazzie committed
576
            up = nn.Module()
577
578
579
580
581
            up.block = block
            up.attn = attn
            if i_level != 0:
                up.upsample = Upsample(block_in, resamp_with_conv)
                curr_res = curr_res * 2
Fazzie's avatar
Fazzie committed
582
            self.up.insert(0, up)    # prepend to get consistent order
583
584
585

        # end
        self.norm_out = Normalize(block_in)
Fazzie's avatar
Fazzie committed
586
        self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604

    def forward(self, z):
        #assert z.shape[1:] == self.z_shape[1:]
        self.last_z_shape = z.shape

        # timestep embedding
        temb = None

        # z to block_in
        h = self.conv_in(z)

        # middle
        h = self.mid.block_1(h, temb)
        h = self.mid.attn_1(h)
        h = self.mid.block_2(h, temb)

        # upsampling
        for i_level in reversed(range(self.num_resolutions)):
Fazzie's avatar
Fazzie committed
605
            for i_block in range(self.num_res_blocks + 1):
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
                h = self.up[i_level].block[i_block](h, temb)
                if len(self.up[i_level].attn) > 0:
                    h = self.up[i_level].attn[i_block](h)
            if i_level != 0:
                h = self.up[i_level].upsample(h)

        # end
        if self.give_pre_end:
            return h

        h = self.norm_out(h)
        h = nonlinearity(h)
        h = self.conv_out(h)
        if self.tanh_out:
            h = torch.tanh(h)
        return h


class SimpleDecoder(nn.Module):
Fazzie's avatar
Fazzie committed
625

626
627
    def __init__(self, in_channels, out_channels, *args, **kwargs):
        super().__init__()
Fazzie's avatar
Fazzie committed
628
629
630
631
632
633
634
635
        self.model = nn.ModuleList([
            nn.Conv2d(in_channels, in_channels, 1),
            ResnetBlock(in_channels=in_channels, out_channels=2 * in_channels, temb_channels=0, dropout=0.0),
            ResnetBlock(in_channels=2 * in_channels, out_channels=4 * in_channels, temb_channels=0, dropout=0.0),
            ResnetBlock(in_channels=4 * in_channels, out_channels=2 * in_channels, temb_channels=0, dropout=0.0),
            nn.Conv2d(2 * in_channels, in_channels, 1),
            Upsample(in_channels, with_conv=True)
        ])
636
637
        # end
        self.norm_out = Normalize(in_channels)
Fazzie's avatar
Fazzie committed
638
        self.conv_out = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
639
640
641

    def forward(self, x):
        for i, layer in enumerate(self.model):
Fazzie's avatar
Fazzie committed
642
            if i in [1, 2, 3]:
643
644
645
646
647
648
649
650
651
652
653
                x = layer(x, None)
            else:
                x = layer(x)

        h = self.norm_out(x)
        h = nonlinearity(h)
        x = self.conv_out(h)
        return x


class UpsampleDecoder(nn.Module):
Fazzie's avatar
Fazzie committed
654
655

    def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, ch_mult=(2, 2), dropout=0.0):
656
657
658
659
660
661
        super().__init__()
        # upsampling
        self.temb_ch = 0
        self.num_resolutions = len(ch_mult)
        self.num_res_blocks = num_res_blocks
        block_in = in_channels
Fazzie's avatar
Fazzie committed
662
        curr_res = resolution // 2**(self.num_resolutions - 1)
663
664
665
666
667
668
        self.res_blocks = nn.ModuleList()
        self.upsample_blocks = nn.ModuleList()
        for i_level in range(self.num_resolutions):
            res_block = []
            block_out = ch * ch_mult[i_level]
            for i_block in range(self.num_res_blocks + 1):
Fazzie's avatar
Fazzie committed
669
670
671
672
673
                res_block.append(
                    ResnetBlock(in_channels=block_in,
                                out_channels=block_out,
                                temb_channels=self.temb_ch,
                                dropout=dropout))
674
675
676
677
678
679
680
681
                block_in = block_out
            self.res_blocks.append(nn.ModuleList(res_block))
            if i_level != self.num_resolutions - 1:
                self.upsample_blocks.append(Upsample(block_in, True))
                curr_res = curr_res * 2

        # end
        self.norm_out = Normalize(block_in)
Fazzie's avatar
Fazzie committed
682
        self.conv_out = torch.nn.Conv2d(block_in, out_channels, kernel_size=3, stride=1, padding=1)
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698

    def forward(self, x):
        # upsampling
        h = x
        for k, i_level in enumerate(range(self.num_resolutions)):
            for i_block in range(self.num_res_blocks + 1):
                h = self.res_blocks[i_level][i_block](h, None)
            if i_level != self.num_resolutions - 1:
                h = self.upsample_blocks[k](h)
        h = self.norm_out(h)
        h = nonlinearity(h)
        h = self.conv_out(h)
        return h


class LatentRescaler(nn.Module):
Fazzie's avatar
Fazzie committed
699

700
701
702
703
    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
        super().__init__()
        # residual block, interpolate, residual block
        self.factor = factor
Fazzie's avatar
Fazzie committed
704
705
706
707
708
        self.conv_in = nn.Conv2d(in_channels, mid_channels, kernel_size=3, stride=1, padding=1)
        self.res_block1 = nn.ModuleList([
            ResnetBlock(in_channels=mid_channels, out_channels=mid_channels, temb_channels=0, dropout=0.0)
            for _ in range(depth)
        ])
709
        self.attn = AttnBlock(mid_channels)
Fazzie's avatar
Fazzie committed
710
711
712
713
714
715
716
717
718
719
        self.res_block2 = nn.ModuleList([
            ResnetBlock(in_channels=mid_channels, out_channels=mid_channels, temb_channels=0, dropout=0.0)
            for _ in range(depth)
        ])

        self.conv_out = nn.Conv2d(
            mid_channels,
            out_channels,
            kernel_size=1,
        )
720
721
722
723
724

    def forward(self, x):
        x = self.conv_in(x)
        for block in self.res_block1:
            x = block(x, None)
Fazzie's avatar
Fazzie committed
725
726
727
        x = torch.nn.functional.interpolate(x,
                                            size=(int(round(x.shape[2] * self.factor)),
                                                  int(round(x.shape[3] * self.factor))))
728
729
730
731
732
733
734
735
        x = self.attn(x)
        for block in self.res_block2:
            x = block(x, None)
        x = self.conv_out(x)
        return x


class MergedRescaleEncoder(nn.Module):
Fazzie's avatar
Fazzie committed
736
737
738
739
740
741
742
743
744
745
746
747
748

    def __init__(self,
                 in_channels,
                 ch,
                 resolution,
                 out_ch,
                 num_res_blocks,
                 attn_resolutions,
                 dropout=0.0,
                 resamp_with_conv=True,
                 ch_mult=(1, 2, 4, 8),
                 rescale_factor=1.0,
                 rescale_module_depth=1):
749
750
        super().__init__()
        intermediate_chn = ch * ch_mult[-1]
Fazzie's avatar
Fazzie committed
751
752
753
754
755
756
757
758
759
760
        self.encoder = Encoder(in_channels=in_channels,
                               num_res_blocks=num_res_blocks,
                               ch=ch,
                               ch_mult=ch_mult,
                               z_channels=intermediate_chn,
                               double_z=False,
                               resolution=resolution,
                               attn_resolutions=attn_resolutions,
                               dropout=dropout,
                               resamp_with_conv=resamp_with_conv,
761
                               out_ch=None)
Fazzie's avatar
Fazzie committed
762
763
764
765
766
        self.rescaler = LatentRescaler(factor=rescale_factor,
                                       in_channels=intermediate_chn,
                                       mid_channels=intermediate_chn,
                                       out_channels=out_ch,
                                       depth=rescale_module_depth)
767
768
769
770
771
772
773
774

    def forward(self, x):
        x = self.encoder(x)
        x = self.rescaler(x)
        return x


class MergedRescaleDecoder(nn.Module):
Fazzie's avatar
Fazzie committed
775
776
777
778
779
780
781
782
783
784
785
786
787

    def __init__(self,
                 z_channels,
                 out_ch,
                 resolution,
                 num_res_blocks,
                 attn_resolutions,
                 ch,
                 ch_mult=(1, 2, 4, 8),
                 dropout=0.0,
                 resamp_with_conv=True,
                 rescale_factor=1.0,
                 rescale_module_depth=1):
788
        super().__init__()
Fazzie's avatar
Fazzie committed
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
        tmp_chn = z_channels * ch_mult[-1]
        self.decoder = Decoder(out_ch=out_ch,
                               z_channels=tmp_chn,
                               attn_resolutions=attn_resolutions,
                               dropout=dropout,
                               resamp_with_conv=resamp_with_conv,
                               in_channels=None,
                               num_res_blocks=num_res_blocks,
                               ch_mult=ch_mult,
                               resolution=resolution,
                               ch=ch)
        self.rescaler = LatentRescaler(factor=rescale_factor,
                                       in_channels=z_channels,
                                       mid_channels=tmp_chn,
                                       out_channels=tmp_chn,
                                       depth=rescale_module_depth)
805
806
807
808
809
810
811
812

    def forward(self, x):
        x = self.rescaler(x)
        x = self.decoder(x)
        return x


class Upsampler(nn.Module):
Fazzie's avatar
Fazzie committed
813

814
815
816
    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
        super().__init__()
        assert out_size >= in_size
Fazzie's avatar
Fazzie committed
817
818
819
820
821
822
823
824
        num_blocks = int(np.log2(out_size // in_size)) + 1
        factor_up = 1. + (out_size % in_size)
        rank_zero_info(
            f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}"
        )
        self.rescaler = LatentRescaler(factor=factor_up,
                                       in_channels=in_channels,
                                       mid_channels=2 * in_channels,
825
                                       out_channels=in_channels)
Fazzie's avatar
Fazzie committed
826
827
828
829
830
831
832
        self.decoder = Decoder(out_ch=out_channels,
                               resolution=out_size,
                               z_channels=in_channels,
                               num_res_blocks=2,
                               attn_resolutions=[],
                               in_channels=None,
                               ch=in_channels,
833
834
835
836
837
838
839
840
841
                               ch_mult=[ch_mult for _ in range(num_blocks)])

    def forward(self, x):
        x = self.rescaler(x)
        x = self.decoder(x)
        return x


class Resize(nn.Module):
Fazzie's avatar
Fazzie committed
842

843
844
845
846
847
    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
        super().__init__()
        self.with_conv = learned
        self.mode = mode
        if self.with_conv:
Fazzie's avatar
Fazzie committed
848
849
            rank_zero_info(
                f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
850
851
852
            raise NotImplementedError()
            assert in_channels is not None
            # no asymmetric padding in torch conv, must do it ourselves
Fazzie's avatar
Fazzie committed
853
            self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=4, stride=2, padding=1)
854
855

    def forward(self, x, scale_factor=1.0):
Fazzie's avatar
Fazzie committed
856
        if scale_factor == 1.0:
857
858
859
860
            return x
        else:
            x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
        return x