diffusion_rs_core/models/vaes/
vae.rs

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
#![allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]

use diffusion_rs_common::core::{Result, Tensor, D};
use diffusion_rs_common::nn::{Activation, Conv2d, Conv2dConfig, GroupNorm};
use diffusion_rs_common::{conv2d, group_norm, linear, VarBuilder};
use serde::Deserialize;
use tracing::{span, Span};

fn default_act() -> Activation {
    Activation::Silu
}

#[derive(Debug, Clone, Deserialize)]
pub struct VAEConfig {
    pub in_channels: usize,
    pub out_channels: usize,
    pub block_out_channels: Vec<usize>,
    pub layers_per_block: usize,
    #[serde(default = "default_act")]
    pub act_fn: Activation,
    pub latent_channels: usize,
    pub norm_num_groups: usize,
    pub mid_block_add_attention: bool,
    pub down_block_types: Vec<String>,
    pub up_block_types: Vec<String>,
}

fn scaled_dot_product_attention(q: &Tensor, k: &Tensor, v: &Tensor) -> Result<Tensor> {
    let dim = q.dim(D::Minus1)?;
    let scale_factor = 1.0 / (dim as f64).sqrt();
    let attn_weights = (q.matmul(&k.t()?)? * scale_factor)?;
    diffusion_rs_common::nn::ops::softmax_last_dim(&attn_weights)?.matmul(v)
}

#[derive(Debug, Clone)]
struct AttnBlock {
    q: Conv2d,
    k: Conv2d,
    v: Conv2d,
    out: Conv2d,
    norm: GroupNorm,
    attn: Span,
}

impl AttnBlock {
    fn new(in_c: usize, vb: VarBuilder, cfg: &VAEConfig) -> Result<Self> {
        let q = linear(in_c, in_c, vb.pp("to_q"))?;
        let q = Conv2d::new(
            q.weight()
                .clone()
                .unsqueeze(D::Minus1)?
                .unsqueeze(D::Minus1)?,
            q.bias().cloned(),
            Conv2dConfig::default(),
        );
        let k = linear(in_c, in_c, vb.pp("to_k"))?;
        let k = Conv2d::new(
            k.weight()
                .clone()
                .unsqueeze(D::Minus1)?
                .unsqueeze(D::Minus1)?,
            k.bias().cloned(),
            Conv2dConfig::default(),
        );
        let v = linear(in_c, in_c, vb.pp("to_v"))?;
        let v = Conv2d::new(
            v.weight()
                .clone()
                .unsqueeze(D::Minus1)?
                .unsqueeze(D::Minus1)?,
            v.bias().cloned(),
            Conv2dConfig::default(),
        );
        let out = linear(in_c, in_c, vb.pp("to_out.0"))?;
        let out = Conv2d::new(
            out.weight()
                .clone()
                .unsqueeze(D::Minus1)?
                .unsqueeze(D::Minus1)?,
            out.bias().cloned(),
            Conv2dConfig::default(),
        );
        let norm = group_norm(cfg.norm_num_groups, in_c, 1e-6, vb.pp("group_norm"))?;
        Ok(Self {
            q,
            k,
            v,
            out,
            norm,
            attn: span!(tracing::Level::TRACE, "vae-attn"),
        })
    }
}

impl diffusion_rs_common::core::Module for AttnBlock {
    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        let _span = self.attn.enter();
        let init_xs = xs;
        let xs = xs.apply(&self.norm)?;
        let q = xs.apply(&self.q)?;
        let k = xs.apply(&self.k)?;
        let v = xs.apply(&self.v)?;
        let (b, c, h, w) = q.dims4()?;
        let q = q.flatten_from(2)?.t()?.unsqueeze(1)?;
        let k = k.flatten_from(2)?.t()?.unsqueeze(1)?;
        let v = v.flatten_from(2)?.t()?.unsqueeze(1)?;
        let xs = scaled_dot_product_attention(&q, &k, &v)?;
        let xs = xs.squeeze(1)?.t()?.reshape((b, c, h, w))?;
        xs.apply(&self.out)? + init_xs
    }
}

#[derive(Debug, Clone)]
struct ResnetBlock {
    norm1: GroupNorm,
    conv1: Conv2d,
    norm2: GroupNorm,
    conv2: Conv2d,
    nin_shortcut: Option<Conv2d>,
    act_fn: Activation,
    resnet: Span,
}

impl ResnetBlock {
    fn new(in_c: usize, out_c: usize, vb: VarBuilder, cfg: &VAEConfig) -> Result<Self> {
        let conv_cfg = diffusion_rs_common::nn::Conv2dConfig {
            padding: 1,
            ..Default::default()
        };
        let norm1 = group_norm(cfg.norm_num_groups, in_c, 1e-6, vb.pp("norm1"))?;
        let conv1 = conv2d(in_c, out_c, 3, conv_cfg, vb.pp("conv1"))?;
        let norm2 = group_norm(cfg.norm_num_groups, out_c, 1e-6, vb.pp("norm2"))?;
        let conv2 = conv2d(out_c, out_c, 3, conv_cfg, vb.pp("conv2"))?;
        let nin_shortcut = if in_c == out_c {
            None
        } else {
            Some(conv2d(
                in_c,
                out_c,
                1,
                Default::default(),
                vb.pp("conv_shortcut"),
            )?)
        };
        Ok(Self {
            norm1,
            conv1,
            norm2,
            conv2,
            nin_shortcut,
            act_fn: cfg.act_fn,
            resnet: span!(tracing::Level::TRACE, "vae-resnet"),
        })
    }
}

impl diffusion_rs_common::core::Module for ResnetBlock {
    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        let _span = self.resnet.enter();
        let h = xs
            .apply(&self.norm1)?
            .apply(&self.act_fn)?
            .apply(&self.conv1)?
            .apply(&self.norm2)?
            .apply(&self.act_fn)?
            .apply(&self.conv2)?;
        match self.nin_shortcut.as_ref() {
            None => xs + h,
            Some(c) => xs.apply(c)? + h,
        }
    }
}

#[derive(Debug, Clone)]
struct Downsample {
    conv: Conv2d,
    downsample: Span,
}

impl Downsample {
    fn new(in_c: usize, vb: VarBuilder) -> Result<Self> {
        let conv_cfg = diffusion_rs_common::nn::Conv2dConfig {
            stride: 2,
            ..Default::default()
        };
        let conv = conv2d(in_c, in_c, 3, conv_cfg, vb.pp("conv"))?;
        Ok(Self {
            conv,
            downsample: span!(tracing::Level::TRACE, "vae-downsample"),
        })
    }
}

impl diffusion_rs_common::core::Module for Downsample {
    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        let _span = self.downsample.enter();
        let xs = xs.pad_with_zeros(D::Minus1, 0, 1)?;
        let xs = xs.pad_with_zeros(D::Minus2, 0, 1)?;
        xs.apply(&self.conv)
    }
}

#[derive(Debug, Clone)]
struct Upsample {
    conv: Conv2d,
    upsample: Span,
}

impl Upsample {
    fn new(in_c: usize, vb: VarBuilder) -> Result<Self> {
        let conv_cfg = diffusion_rs_common::nn::Conv2dConfig {
            padding: 1,
            ..Default::default()
        };
        let conv = conv2d(in_c, in_c, 3, conv_cfg, vb.pp("conv"))?;
        Ok(Self {
            conv,
            upsample: span!(tracing::Level::TRACE, "vae-upsample"),
        })
    }
}

impl diffusion_rs_common::core::Module for Upsample {
    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        let _ = self.upsample.enter();
        let (_, _, h, w) = xs.dims4()?;
        xs.upsample_nearest2d(h * 2, w * 2)?.apply(&self.conv)
    }
}

#[derive(Debug, Clone)]
struct DownBlock {
    block: Vec<ResnetBlock>,
    downsample: Option<Downsample>,
}

#[derive(Debug, Clone)]
pub struct Encoder {
    conv_in: Conv2d,
    mid_block_1: ResnetBlock,
    mid_attn_1: Option<AttnBlock>,
    mid_block_2: ResnetBlock,
    norm_out: GroupNorm,
    conv_out: Conv2d,
    down: Vec<DownBlock>,
    act_fn: Activation,
}

impl Encoder {
    pub fn new(cfg: &VAEConfig, vb: VarBuilder) -> Result<Self> {
        if !cfg
            .down_block_types
            .iter()
            .all(|x| x == "DownEncoderBlock2D")
        {
            diffusion_rs_common::bail!(
                "All down (encoder) block types must be `DownEncoderBlock2D`"
            );
        }
        let conv_cfg = diffusion_rs_common::nn::Conv2dConfig {
            padding: 1,
            ..Default::default()
        };
        let base_ch = cfg.block_out_channels[0];
        let mut block_in = base_ch;
        let conv_in = conv2d(cfg.in_channels, block_in, 3, conv_cfg, vb.pp("conv_in"))?;

        let mut down = Vec::with_capacity(cfg.block_out_channels.len());
        let vb_d = vb.pp("down_blocks");
        for (i_level, out_channels) in cfg.block_out_channels.iter().enumerate() {
            let mut block = Vec::with_capacity(cfg.layers_per_block);
            let vb_d = vb_d.pp(i_level);
            let vb_resnets = vb_d.pp("resnets");
            block_in = if i_level == 0 {
                base_ch
            } else {
                cfg.block_out_channels[i_level - 1]
            };
            let block_out = *out_channels;
            for i_block in 0..cfg.layers_per_block {
                let b = ResnetBlock::new(block_in, block_out, vb_resnets.pp(i_block), cfg)?;
                block.push(b);
                block_in = block_out;
            }
            let downsample = if i_level != cfg.block_out_channels.len() - 1 {
                Some(Downsample::new(block_in, vb_d.pp("downsamplers.0"))?)
            } else {
                None
            };
            let block = DownBlock { block, downsample };
            down.push(block)
        }

        // TODO: this is technically not general enough. Should always start with 1 resnet, then unet num_layers (defaults to 1 so this is OK)
        // repeats of attention and resnet!
        // https://github.com/huggingface/diffusion_rs/blob/243d9a49864ebb4562de6304a5fb9b9ebb496c6e/src/diffusion_rs/models/unets/unet_2d_blocks.py#L644-L729
        // https://github.com/huggingface/diffusion_rs/blob/243d9a49864ebb4562de6304a5fb9b9ebb496c6e/src/diffusion_rs/models/unets/unet_2d_blocks.py#L625
        let mid_block_1 = ResnetBlock::new(block_in, block_in, vb.pp("mid_block.resnets.0"), cfg)?;
        let mid_attn_1 = if cfg.mid_block_add_attention {
            Some(AttnBlock::new(
                block_in,
                vb.pp("mid_block.attentions.0"),
                cfg,
            )?)
        } else {
            None
        };
        let mid_block_2 = ResnetBlock::new(block_in, block_in, vb.pp("mid_block.resnets.1"), cfg)?;
        let conv_out = conv2d(
            block_in,
            2 * cfg.latent_channels,
            3,
            conv_cfg,
            vb.pp("conv_out"),
        )?;
        let norm_out = group_norm(cfg.norm_num_groups, block_in, 1e-6, vb.pp("conv_norm_out"))?;
        Ok(Self {
            conv_in,
            mid_block_1,
            mid_attn_1,
            mid_block_2,
            norm_out,
            conv_out,
            down,
            act_fn: cfg.act_fn,
        })
    }
}

impl diffusion_rs_common::nn::Module for Encoder {
    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        let mut h = xs.apply(&self.conv_in)?;
        for block in self.down.iter() {
            for b in block.block.iter() {
                h = h.apply(b)?
            }
            if let Some(ds) = block.downsample.as_ref() {
                h = h.apply(ds)?
            }
        }
        h = h.apply(&self.mid_block_1)?;
        if let Some(attn) = &self.mid_attn_1 {
            h = h.apply(attn)?;
        }
        h.apply(&self.mid_block_2)?
            .apply(&self.norm_out)?
            .apply(&self.act_fn)?
            .apply(&self.conv_out)
    }
}

#[derive(Debug, Clone)]
struct UpBlock {
    block: Vec<ResnetBlock>,
    upsample: Option<Upsample>,
}

#[derive(Debug, Clone)]
pub struct Decoder {
    conv_in: Conv2d,
    mid_block_1: ResnetBlock,
    mid_attn_1: Option<AttnBlock>,
    mid_block_2: ResnetBlock,
    norm_out: GroupNorm,
    conv_out: Conv2d,
    up: Vec<UpBlock>,
    act_fn: Activation,
}

impl Decoder {
    pub fn new(cfg: &VAEConfig, vb: VarBuilder) -> Result<Self> {
        if !cfg.up_block_types.iter().all(|x| x == "UpDecoderBlock2D") {
            diffusion_rs_common::bail!("All up (decoder) block types must be `UpDecoderBlock2D`");
        }
        let conv_cfg = diffusion_rs_common::nn::Conv2dConfig {
            padding: 1,
            ..Default::default()
        };
        let base_ch = cfg.block_out_channels[0];
        let mut block_in = cfg.block_out_channels.last().copied().unwrap_or(base_ch);

        // TODO: this is technically not general enough. Should always start with 1 resnet, then unet num_layers (defaults to 1 so this is OK)
        // repeats of attention and resnet!
        // https://github.com/huggingface/diffusion_rs/blob/243d9a49864ebb4562de6304a5fb9b9ebb496c6e/src/diffusion_rs/models/unets/unet_2d_blocks.py#L644-L729
        // https://github.com/huggingface/diffusion_rs/blob/243d9a49864ebb4562de6304a5fb9b9ebb496c6e/src/diffusion_rs/models/unets/unet_2d_blocks.py#L625
        let mid_block_1 = ResnetBlock::new(block_in, block_in, vb.pp("mid_block.resnets.0"), cfg)?;
        let mid_attn_1 = if cfg.mid_block_add_attention {
            Some(AttnBlock::new(
                block_in,
                vb.pp("mid_block.attentions.0"),
                cfg,
            )?)
        } else {
            None
        };
        let mid_block_2 = ResnetBlock::new(block_in, block_in, vb.pp("mid_block.resnets.1"), cfg)?;

        let conv_in = conv2d(cfg.latent_channels, block_in, 3, conv_cfg, vb.pp("conv_in"))?;

        let mut up = Vec::with_capacity(cfg.block_out_channels.len());
        let vb_u = vb.pp("up_blocks");
        for (i_level, out_channels) in cfg.block_out_channels.iter().rev().enumerate() {
            let block_out = *out_channels;
            let vb_u = vb_u.pp(i_level);
            let vb_resnets = vb_u.pp("resnets");
            let mut block = Vec::with_capacity(cfg.layers_per_block + 1);
            for i_block in 0..=cfg.layers_per_block {
                let b = ResnetBlock::new(block_in, block_out, vb_resnets.pp(i_block), cfg)?;
                block.push(b);
                block_in = block_out;
            }
            let upsample = if i_level != 3 {
                Some(Upsample::new(block_in, vb_u.pp("upsamplers.0"))?)
            } else {
                None
            };
            let block = UpBlock { block, upsample };
            up.push(block)
        }

        let norm_out = group_norm(cfg.norm_num_groups, base_ch, 1e-6, vb.pp("conv_norm_out"))?;
        let conv_out = conv2d(base_ch, cfg.out_channels, 3, conv_cfg, vb.pp("conv_out"))?;
        Ok(Self {
            conv_in,
            mid_block_1,
            mid_attn_1,
            mid_block_2,
            norm_out,
            conv_out,
            up,
            act_fn: cfg.act_fn,
        })
    }
}

impl diffusion_rs_common::nn::Module for Decoder {
    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        let h = xs.apply(&self.conv_in)?;
        let mut h = h.apply(&self.mid_block_1)?;
        if let Some(attn) = &self.mid_attn_1 {
            h = h.apply(attn)?;
        }
        h = h.apply(&self.mid_block_2)?;
        for block in self.up.iter() {
            for b in block.block.iter() {
                h = h.apply(b)?
            }
            if let Some(us) = block.upsample.as_ref() {
                h = h.apply(us)?
            }
        }
        h.apply(&self.norm_out)?
            .apply(&self.act_fn)?
            .apply(&self.conv_out)
    }
}

#[derive(Debug, Clone)]
pub struct DiagonalGaussian {
    sample: bool,
    chunk_dim: usize,
}

impl DiagonalGaussian {
    pub fn new(sample: bool, chunk_dim: usize) -> Result<Self> {
        Ok(Self { sample, chunk_dim })
    }
}

impl diffusion_rs_common::nn::Module for DiagonalGaussian {
    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        let chunks = xs.chunk(2, self.chunk_dim)?;
        if self.sample {
            let std = (&chunks[1] * 0.5)?.exp()?;
            &chunks[0] + (std * chunks[0].randn_like(0., 1.))?
        } else {
            Ok(chunks[0].clone())
        }
    }
}