vit_tch 0.1.0

Vision Transformer models and building blocks for Rust using tch.
Documentation
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
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
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
use tch::{IndexOp, Kind, Tensor, nn, nn::ModuleT};

use crate::{
    config::RegionViTConfig,
    conv_layers::ChannelLayerNorm2D,
    layers::FeedForward,
    tensor::{assert_image_patchable, merge_heads, patch_dim, split_heads},
};

#[derive(Debug)]
struct Downsample {
    conv: nn::Conv2D,
}

impl Downsample {
    fn new(vs: &nn::Path, dim_in: i64, dim_out: i64) -> Self {
        let conv = nn::conv2d(
            vs / "conv",
            dim_in,
            dim_out,
            3,
            nn::ConvConfig {
                stride: 2,
                padding: 1,
                ..Default::default()
            },
        );

        Self { conv }
    }
}

impl nn::Module for Downsample {
    fn forward(&self, xs: &Tensor) -> Tensor {
        xs.apply(&self.conv)
    }
}

#[derive(Debug)]
struct PEG {
    proj: nn::Conv2D,
}

impl PEG {
    fn new(vs: &nn::Path, dim: i64, kernel_size: i64) -> Self {
        assert!(kernel_size > 0, "peg kernel size must be positive");
        let proj = nn::conv2d(
            vs / "proj",
            dim,
            dim,
            kernel_size,
            nn::ConvConfig {
                padding: kernel_size / 2,
                groups: dim,
                ..Default::default()
            },
        );

        Self { proj }
    }
}

impl nn::Module for PEG {
    fn forward(&self, xs: &Tensor) -> Tensor {
        xs + xs.apply(&self.proj)
    }
}

#[derive(Debug)]
enum LocalEncoder {
    Conv(nn::Conv2D),
    ThreeConv {
        conv1: nn::Conv2D,
        norm1: ChannelLayerNorm2D,
        conv2: nn::Conv2D,
        norm2: ChannelLayerNorm2D,
        conv3: nn::Conv2D,
    },
}

impl LocalEncoder {
    fn new(vs: &nn::Path, channels: i64, dim: i64, tokenize_local_3_conv: bool) -> Self {
        if tokenize_local_3_conv {
            let conv1 = nn::conv2d(
                vs / "conv1",
                channels,
                dim,
                3,
                nn::ConvConfig {
                    stride: 2,
                    padding: 1,
                    ..Default::default()
                },
            );
            let norm1 = ChannelLayerNorm2D::new(&(vs / "norm1"), dim, 1e-5);
            let conv2 = nn::conv2d(
                vs / "conv2",
                dim,
                dim,
                3,
                nn::ConvConfig {
                    stride: 2,
                    padding: 1,
                    ..Default::default()
                },
            );
            let norm2 = ChannelLayerNorm2D::new(&(vs / "norm2"), dim, 1e-5);
            let conv3 = nn::conv2d(
                vs / "conv3",
                dim,
                dim,
                3,
                nn::ConvConfig {
                    padding: 1,
                    ..Default::default()
                },
            );

            Self::ThreeConv {
                conv1,
                norm1,
                conv2,
                norm2,
                conv3,
            }
        } else {
            Self::Conv(nn::conv2d(
                vs / "conv",
                channels,
                dim,
                8,
                nn::ConvConfig {
                    stride: 4,
                    padding: 3,
                    ..Default::default()
                },
            ))
        }
    }
}

impl nn::Module for LocalEncoder {
    fn forward(&self, xs: &Tensor) -> Tensor {
        match self {
            Self::Conv(conv) => xs.apply(conv),
            Self::ThreeConv {
                conv1,
                norm1,
                conv2,
                norm2,
                conv3,
            } => xs
                .apply(conv1)
                .apply(norm1)
                .gelu("none")
                .apply(conv2)
                .apply(norm2)
                .gelu("none")
                .apply(conv3),
        }
    }
}

#[derive(Debug)]
struct RegionEncoder {
    patch_size: i64,
    conv: nn::Conv2D,
}

impl RegionEncoder {
    fn new(vs: &nn::Path, channels: i64, dim: i64, patch_size: i64) -> Self {
        assert!(patch_size > 0, "region patch size must be positive");
        let conv = nn::conv2d(
            vs / "conv",
            patch_dim(channels, patch_size, patch_size),
            dim,
            1,
            Default::default(),
        );

        Self { patch_size, conv }
    }
}

impl nn::Module for RegionEncoder {
    fn forward(&self, xs: &Tensor) -> Tensor {
        let size = xs.size();
        assert_eq!(
            size.len(),
            4,
            "expected image tensor with shape [batch, channels, height, width]"
        );
        let batch = size[0];
        let channels = size[1];
        let height = size[2];
        let width = size[3];
        assert_image_patchable(height, width, self.patch_size, self.patch_size);
        let grid_h = height / self.patch_size;
        let grid_w = width / self.patch_size;

        xs.view([
            batch,
            channels,
            grid_h,
            self.patch_size,
            grid_w,
            self.patch_size,
        ])
        .permute([0, 1, 3, 5, 2, 4])
        .contiguous()
        .view([
            batch,
            channels * self.patch_size * self.patch_size,
            grid_h,
            grid_w,
        ])
        .apply(&self.conv)
    }
}

#[derive(Debug)]
struct Attention {
    norm: nn::LayerNorm,
    to_qkv: nn::Linear,
    to_out: nn::Linear,
    heads: i64,
    scale: f64,
    dropout: f64,
}

impl Attention {
    fn new(vs: &nn::Path, dim: i64, heads: i64, dim_head: i64, dropout: f64) -> Self {
        assert!(heads > 0, "heads must be positive");
        assert!(dim_head > 0, "dim head must be positive");
        let inner_dim = heads * dim_head;
        let norm = nn::layer_norm(vs / "norm", vec![dim], Default::default());
        let to_qkv = nn::linear(
            vs / "to_qkv",
            dim,
            inner_dim * 3,
            nn::LinearConfig {
                bias: false,
                ..Default::default()
            },
        );
        let to_out = nn::linear(vs / "to_out", inner_dim, dim, Default::default());

        Self {
            norm,
            to_qkv,
            to_out,
            heads,
            scale: (dim_head as f64).powf(-0.5),
            dropout,
        }
    }

    fn forward_t_with_bias(&self, xs: &Tensor, bias: Option<&Tensor>, train: bool) -> Tensor {
        let qkv = xs.apply(&self.norm).apply(&self.to_qkv).chunk(3, -1);
        let q = split_heads(&qkv[0], self.heads) * self.scale;
        let k = split_heads(&qkv[1], self.heads);
        let v = split_heads(&qkv[2], self.heads);
        let mut dots = q.matmul(&k.transpose(-1, -2));

        if let Some(bias) = bias {
            dots += bias;
        }

        let attn = dots.softmax(-1, Kind::Float).dropout(self.dropout, train);

        merge_heads(&attn.matmul(&v))
            .apply(&self.to_out)
            .dropout(self.dropout, train)
    }
}

impl nn::ModuleT for Attention {
    fn forward_t(&self, xs: &Tensor, train: bool) -> Tensor {
        self.forward_t_with_bias(xs, None, train)
    }
}

#[derive(Debug)]
struct R2LLayer {
    attention: Attention,
    feed_forward: FeedForward,
}

impl R2LLayer {
    fn new(
        vs: &nn::Path,
        dim: i64,
        heads: i64,
        dim_head: i64,
        ff_dropout: f64,
        attn_dropout: f64,
    ) -> Self {
        let attention = Attention::new(&(vs / "attention"), dim, heads, dim_head, attn_dropout);
        let feed_forward = FeedForward::new(&(vs / "feed_forward"), dim, dim * 4, ff_dropout);

        Self {
            attention,
            feed_forward,
        }
    }
}

#[derive(Debug)]
struct R2LTransformer {
    layers: Vec<R2LLayer>,
    local_rel_pos_bias: Tensor,
    window_size: i64,
    heads: i64,
}

impl R2LTransformer {
    fn new(vs: &nn::Path, dim: i64, config: &RegionViTConfig, depth: usize) -> Self {
        let rel_positions = 2 * config.window_size - 1;
        let local_rel_pos_bias = vs.var(
            "local_rel_pos_bias",
            &[rel_positions * rel_positions, config.heads],
            nn::Init::Randn {
                mean: 0.0,
                stdev: 1.0,
            },
        );
        let layers = (0..depth)
            .map(|index| {
                let name = format!("layer_{index}");
                R2LLayer::new(
                    &(vs / name.as_str()),
                    dim,
                    config.heads,
                    config.dim_head,
                    config.ff_dropout,
                    config.attn_dropout,
                )
            })
            .collect();

        Self {
            layers,
            local_rel_pos_bias,
            window_size: config.window_size,
            heads: config.heads,
        }
    }

    fn relative_position_bias(&self, window_h: i64, window_w: i64) -> Tensor {
        let positions = (0..window_h)
            .flat_map(|h| (0..window_w).map(move |w| (h, w)))
            .collect::<Vec<_>>();
        let mut indices = Vec::with_capacity(positions.len() * positions.len());
        let span = self.window_size * 2 - 1;

        for &(h1, w1) in &positions {
            for &(h2, w2) in &positions {
                let rel_h = h1 - h2 + self.window_size - 1;
                let rel_w = w1 - w2 + self.window_size - 1;
                indices.push(rel_h * span + rel_w);
            }
        }

        let tokens = window_h * window_w;
        let index = Tensor::from_slice(&indices)
            .to_kind(Kind::Int64)
            .to_device(self.local_rel_pos_bias.device());
        let bias = self
            .local_rel_pos_bias
            .index_select(0, &index)
            .view([tokens, tokens, self.heads])
            .permute([2, 0, 1])
            .unsqueeze(0);
        let zeros = Tensor::zeros(
            [1, self.heads, tokens + 1, tokens + 1],
            (bias.kind(), bias.device()),
        );

        zeros.i((.., .., 1.., 1..)).copy_(&bias);
        zeros
    }

    fn forward_pair_t(
        &self,
        local_tokens: &Tensor,
        region_tokens: &Tensor,
        train: bool,
    ) -> (Tensor, Tensor) {
        let local_size = local_tokens.size();
        let region_size = region_tokens.size();
        let batch = local_size[0];
        let dim = local_size[1];
        let local_h = local_size[2];
        let local_w = local_size[3];
        let region_h = region_size[2];
        let region_w = region_size[3];
        assert_image_patchable(local_h, local_w, region_h, region_w);
        let window_h = local_h / region_h;
        let window_w = local_w / region_w;
        let local_window_tokens = window_h * window_w;
        let rel_pos_bias = self.relative_position_bias(window_h, window_w);
        let mut local_tokens =
            local_tokens
                .permute([0, 2, 3, 1])
                .contiguous()
                .view([batch, local_h * local_w, dim]);
        let mut region_tokens = region_tokens.permute([0, 2, 3, 1]).contiguous().view([
            batch,
            region_h * region_w,
            dim,
        ]);

        for layer in &self.layers {
            let region_attn = layer.attention.forward_t(&region_tokens, train);
            region_tokens = &region_tokens + region_attn;
            let local_windows = local_tokens
                .view([batch, local_h, local_w, dim])
                .view([batch, region_h, window_h, region_w, window_w, dim])
                .permute([0, 1, 3, 2, 4, 5])
                .contiguous()
                .view([batch * region_h * region_w, local_window_tokens, dim]);
            let region_windows = region_tokens.view([batch * region_h * region_w, 1, dim]);
            let mut tokens = Tensor::cat(&[region_windows, local_windows], 1);
            let attn = layer
                .attention
                .forward_t_with_bias(&tokens, Some(&rel_pos_bias), train);
            tokens += attn;
            let ff = layer.feed_forward.forward_t(&tokens, train);
            tokens += ff;
            region_tokens = tokens
                .i((.., 0..1, ..))
                .view([batch, region_h * region_w, dim]);
            local_tokens = tokens
                .i((.., 1.., ..))
                .view([batch, region_h, region_w, window_h, window_w, dim])
                .permute([0, 1, 3, 2, 4, 5])
                .contiguous()
                .view([batch, local_h * local_w, dim]);
        }

        let local_tokens = local_tokens
            .view([batch, local_h, local_w, dim])
            .permute([0, 3, 1, 2]);
        let region_tokens = region_tokens
            .view([batch, region_h, region_w, dim])
            .permute([0, 3, 1, 2]);

        (local_tokens, region_tokens)
    }
}

#[derive(Debug)]
struct RegionViTStage {
    downsample: Option<Downsample>,
    peg: Option<PEG>,
    transformer: R2LTransformer,
}

impl RegionViTStage {
    fn new(
        vs: &nn::Path,
        dim_in: i64,
        dim: i64,
        depth: usize,
        config: &RegionViTConfig,
        needs_downsample: bool,
        needs_peg: bool,
    ) -> Self {
        let downsample =
            needs_downsample.then(|| Downsample::new(&(vs / "downsample"), dim_in, dim));
        let peg = needs_peg.then(|| PEG::new(&(vs / "peg"), dim, 3));
        let transformer = R2LTransformer::new(&(vs / "transformer"), dim, config, depth);

        Self {
            downsample,
            peg,
            transformer,
        }
    }

    fn forward_pair_t(
        &self,
        local_tokens: &Tensor,
        region_tokens: &Tensor,
        train: bool,
    ) -> (Tensor, Tensor) {
        let mut local_tokens = local_tokens.shallow_clone();
        let mut region_tokens = region_tokens.shallow_clone();

        if let Some(downsample) = &self.downsample {
            local_tokens = local_tokens.apply(downsample);
            region_tokens = region_tokens.apply(downsample);
        }

        if let Some(peg) = &self.peg {
            local_tokens = local_tokens.apply(peg);
        }

        self.transformer
            .forward_pair_t(&local_tokens, &region_tokens, train)
    }
}

#[derive(Debug)]
pub struct RegionViT {
    local_patch_size: i64,
    region_patch_size: i64,
    local_encoder: LocalEncoder,
    region_encoder: RegionEncoder,
    stages: Vec<RegionViTStage>,
    norm: nn::LayerNorm,
    head: nn::Linear,
}

impl RegionViT {
    pub fn new(vs: &nn::Path, config: RegionViTConfig) -> Self {
        assert!(!config.dim.is_empty(), "dim must not be empty");
        assert_eq!(
            config.depth.len(),
            config.dim.len(),
            "depth must match dim length"
        );
        assert!(
            config.local_patch_size > 0,
            "local patch size must be positive"
        );
        assert!(config.window_size > 0, "window size must be positive");
        let init_dim = config.dim[0];
        let region_patch_size = config.local_patch_size * config.window_size;
        let local_encoder = LocalEncoder::new(
            &(vs / "local_encoder"),
            config.channels,
            init_dim,
            config.tokenize_local_3_conv,
        );
        let region_encoder = RegionEncoder::new(
            &(vs / "region_encoder"),
            config.channels,
            init_dim,
            region_patch_size,
        );
        let mut dim_in = init_dim;
        let stages = config
            .dim
            .iter()
            .zip(&config.depth)
            .enumerate()
            .map(|(index, (&dim, &depth))| {
                let name = format!("stage_{index}");
                let not_first = index != 0;
                let stage = RegionViTStage::new(
                    &(vs / name.as_str()),
                    dim_in,
                    dim,
                    depth,
                    &config,
                    not_first,
                    not_first && config.use_peg,
                );
                dim_in = dim;
                stage
            })
            .collect::<Vec<_>>();
        let last_dim = *config.dim.last().expect("dim must not be empty");
        let norm = nn::layer_norm(vs / "norm", vec![last_dim], Default::default());
        let head = nn::linear(
            vs / "head",
            last_dim,
            config.num_classes,
            Default::default(),
        );

        Self {
            local_patch_size: config.local_patch_size,
            region_patch_size,
            local_encoder,
            region_encoder,
            stages,
            norm,
            head,
        }
    }
}

impl nn::ModuleT for RegionViT {
    fn forward_t(&self, xs: &Tensor, train: bool) -> Tensor {
        let size = xs.size();
        assert_eq!(
            size.len(),
            4,
            "expected image tensor with shape [batch, channels, height, width]"
        );
        assert_image_patchable(
            size[2],
            size[3],
            self.region_patch_size,
            self.region_patch_size,
        );
        assert_image_patchable(
            size[2],
            size[3],
            self.local_patch_size,
            self.local_patch_size,
        );
        let mut local_tokens = xs.apply(&self.local_encoder);
        let mut region_tokens = xs.apply(&self.region_encoder);

        for stage in &self.stages {
            (local_tokens, region_tokens) =
                stage.forward_pair_t(&local_tokens, &region_tokens, train);
        }

        region_tokens
            .mean_dim(&[2_i64, 3][..], false, Kind::Float)
            .apply(&self.norm)
            .apply(&self.head)
    }
}