bunsen 0.0.1

burn neural network extension library
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
//! # Residual Block Wrapper
//!
//! [`ResidualBlock`] is a abstraction wrapper around either:
//! * [`BasicBlock`] - the basic `ResNet` conv block, or
//! * [`BottleneckBlock`] - the bottleneck variant `ResNet` conv block.
//!
//! [`ResidualBlockMeta`] defines a common meta api shared by:
//! * [`ResidualBlock`], and
//! * [`ResidualBlockStructureConfig`]
//!
//! [`ResidualBlockStructureConfig`] implements [`Config`], and provides an
//! [`ResidualBlockStructureConfig::init`] constructor pathway to [`ResidualBlock`].
//!
//! [`ResidualBlock`] implements [`Module`],
//! and provides [`ResidualBlock::forward`].
//!
//! [`ResidualBlock`] can also be constructed via:
//! * [`From<BasicBlock<B>>`](`BasicBlock`),
//! * [`From<BottleneckBlock<B>>`](`BottleneckBlock`).

use crate::layers::drop::drop_block::DropBlockOptions;
use crate::models::resnet::basic_block::{BasicBlock, BasicBlockConfig, BasicBlockMeta};
use crate::models::resnet::bottleneck_block::{
    BottleneckBlock, BottleneckBlockConfig, BottleneckBlockMeta, BottleneckPolicyConfig,
};
use crate::models::resnet::util::stride_div_output_resolution;
use crate::utility::probability::expect_probability;
use burn::nn::BatchNormConfig;
use burn::nn::activation::ActivationConfig;
use burn::nn::norm::NormalizationConfig;
use burn::prelude::{Backend, Config, Module, Tensor};

/// Abstract [`ResidualBlock`] Config.
#[derive(Config, Debug)]
pub struct ResidualBlockContractConfig {
    /// The number of input feature planes.
    pub in_planes: usize,

    /// The number of output feature planes.
    pub out_planes: usize,

    /// Dilation rate for conv layers.
    #[config(default = 1)]
    pub dilation: usize,

    /// If set, override the first dilation rate.
    #[config(default = "None")]
    pub first_dilation: Option<usize>,

    /// Downsample the input by 2x?
    #[config(default = "false")]
    pub downsample_input: bool,

    /// Select between [`BasicBlock`] and [`BottleneckBlock`].
    #[config(default = "None")]
    pub bottleneck_policy: Option<BottleneckPolicyConfig>,

    /// [`crate::compat::normalization_wrapper::Normalization`] config.
    ///
    /// The feature size of this config will be replaced
    /// with the appropriate feature size for the input layer.
    #[config(default = "NormalizationConfig::Batch(BatchNormConfig::new(0))")]
    pub normalization: NormalizationConfig,

    /// [`crate::compat::activation_wrapper::Activation`] config.
    #[config(default = "ActivationConfig::Relu")]
    pub activation: ActivationConfig,
}

impl ResidualBlockContractConfig {
    /// Convert to [`ResidualBlockStructureConfig`].
    pub fn to_structure(self) -> ResidualBlockStructureConfig {
        let stride = if self.downsample_input { 2 } else { 1 };

        match self.bottleneck_policy {
            None => BasicBlockConfig::new(self.in_planes, self.out_planes)
                .with_stride(stride)
                .with_dilation(self.dilation)
                .with_first_dilation(self.first_dilation)
                .with_normalization(self.normalization)
                .with_activation(self.activation)
                .into(),
            Some(policy) => BottleneckBlockConfig::new(self.in_planes, self.out_planes)
                .with_stride(stride)
                .with_dilation(self.dilation)
                .with_first_dilation(self.first_dilation)
                .with_normalization(self.normalization)
                .with_activation(self.activation)
                .with_policy(policy)
                .into(),
        }
    }
}

impl From<ResidualBlockContractConfig> for ResidualBlockStructureConfig {
    fn from(config: ResidualBlockContractConfig) -> Self {
        config.to_structure()
    }
}

/// [`ResidualBlock`] Meta API.
///
/// Defines a shared API for [`ResidualBlock`] and [`ResidualBlockStructureConfig`].
pub trait ResidualBlockMeta {
    /// The number of input feature planes.
    fn in_planes(&self) -> usize;

    /// The number of outpu feature planes.
    fn out_planes(&self) -> usize;

    /// The stride of convolution.
    ///
    /// Affects downsample behavior.
    fn stride(&self) -> usize;

    /// Get the output resolution for a given input resolution.
    ///
    /// The input must be a multiple of the stride.
    ///
    /// # Arguments
    ///
    /// - `input_resolution`: \
    ///   ``[in_height=out_height*stride, in_width=out_width*stride]``.
    ///
    /// # Returns
    ///
    /// ``[out_height, out_width]``
    ///
    /// # Panics
    ///
    /// If the input resolution is not a multiple of the stride.
    fn output_resolution(
        &self,
        input_resolution: [usize; 2],
    ) -> [usize; 2] {
        stride_div_output_resolution(input_resolution, self.stride())
    }
}

/// [`ResidualBlock`] Config.
///
/// Implements [`ResidualBlockMeta`].
#[derive(Config, Debug)]
pub enum ResidualBlockStructureConfig {
    /// A `ResNet` [`BasicBlock`].
    Basic(BasicBlockConfig),

    /// A `ResNet` [`BottleneckBlock`].
    Bottleneck(BottleneckBlockConfig),
}

impl ResidualBlockMeta for ResidualBlockStructureConfig {
    fn in_planes(&self) -> usize {
        match self {
            Self::Basic(config) => config.in_planes(),
            Self::Bottleneck(config) => config.in_planes(),
        }
    }

    fn out_planes(&self) -> usize {
        match self {
            Self::Basic(config) => config.out_planes(),
            Self::Bottleneck(config) => config.out_planes(),
        }
    }

    fn stride(&self) -> usize {
        match self {
            Self::Basic(config) => config.stride(),
            Self::Bottleneck(config) => config.stride(),
        }
    }

    fn output_resolution(
        &self,
        input_resolution: [usize; 2],
    ) -> [usize; 2] {
        match self {
            Self::Basic(config) => config.output_resolution(input_resolution),
            Self::Bottleneck(config) => config.output_resolution(input_resolution),
        }
    }
}

impl From<BasicBlockConfig> for ResidualBlockStructureConfig {
    fn from(config: BasicBlockConfig) -> Self {
        Self::Basic(config)
    }
}

impl From<BottleneckBlockConfig> for ResidualBlockStructureConfig {
    fn from(config: BottleneckBlockConfig) -> Self {
        Self::Bottleneck(config)
    }
}

impl ResidualBlockStructureConfig {
    /// Initialize a [`ResidualBlock`].
    pub fn init<B: Backend>(
        self,
        device: &B::Device,
    ) -> ResidualBlock<B> {
        match self {
            Self::Basic(config) => config.init(device).into(),
            Self::Bottleneck(config) => config.init(device).into(),
        }
    }

    /// Set drop block options.
    pub fn with_drop_block(
        self,
        options: Option<DropBlockOptions>,
    ) -> Self {
        match self {
            Self::Basic(config) => config.with_drop_block(options).into(),
            Self::Bottleneck(config) => config.with_drop_block(options).into(),
        }
    }

    /// Set the drop path probability.
    pub fn with_drop_path_prob(
        self,
        drop_path_prob: f64,
    ) -> Self {
        let drop_path_prob = expect_probability(drop_path_prob);
        match self {
            Self::Basic(config) => config.with_drop_path_prob(drop_path_prob).into(),
            Self::Bottleneck(config) => config.with_drop_path_prob(drop_path_prob).into(),
        }
    }
}

/// A `ResNet` [`BasicBlock`] or [`BottleneckBlock`] wrapper.
///
/// Implements [`ResidualBlockMeta`].
#[derive(Module, Debug)]
#[allow(clippy::large_enum_variant)]
pub enum ResidualBlock<B: Backend> {
    /// A `ResNet` [`BasicBlock`].
    Basic(BasicBlock<B>),

    /// A `ResNet` [`BottleneckBlock`].
    Bottleneck(BottleneckBlock<B>),
}

impl<B: Backend> From<BasicBlock<B>> for ResidualBlock<B> {
    fn from(block: BasicBlock<B>) -> Self {
        Self::Basic(block)
    }
}

impl<B: Backend> From<BottleneckBlock<B>> for ResidualBlock<B> {
    fn from(block: BottleneckBlock<B>) -> Self {
        Self::Bottleneck(block)
    }
}

impl<B: Backend> ResidualBlockMeta for ResidualBlock<B> {
    fn in_planes(&self) -> usize {
        match self {
            Self::Basic(block) => block.in_planes(),
            Self::Bottleneck(block) => block.in_planes(),
        }
    }

    fn out_planes(&self) -> usize {
        match self {
            Self::Basic(block) => block.out_planes(),
            Self::Bottleneck(block) => block.out_planes(),
        }
    }

    fn stride(&self) -> usize {
        match self {
            Self::Basic(block) => block.stride(),
            Self::Bottleneck(block) => block.stride(),
        }
    }
}

impl<B: Backend> ResidualBlock<B> {
    /// Debug print.
    pub fn debug_print(&self) {
        match self {
            Self::Basic(block) => block.debug_print(),
            Self::Bottleneck(block) => block.debug_print(),
        }
    }

    /// Apply the wrapped block to the input.
    ///
    /// # Arguments
    ///
    /// - `input`: ``[batch, in_planes, in_height=out_height*stride, in_width=out_width*stride]``.
    ///
    /// # Returns
    ///
    /// A ``[batch, out_planes, out_height, out_width]`` tensor;
    pub fn forward(
        &self,
        input: Tensor<B, 4>,
    ) -> Tensor<B, 4> {
        match self {
            Self::Basic(block) => block.forward(input),
            Self::Bottleneck(block) => block.forward(input),
        }
    }

    /// Set the drop path probability.
    pub fn with_drop_path_prob(
        self,
        drop_path_prob: f64,
    ) -> Self {
        let drop_path_prob = expect_probability(drop_path_prob);
        match self {
            Self::Basic(block) => block.with_drop_path_prob(drop_path_prob).into(),
            Self::Bottleneck(block) => block.with_drop_path_prob(drop_path_prob).into(),
        }
    }

    /// Set drop block options.
    pub fn with_drop_block(
        self,
        options: Option<DropBlockOptions>,
    ) -> Self {
        match self {
            Self::Basic(config) => config.with_drop_block(options).into(),
            Self::Bottleneck(config) => config.with_drop_block(options).into(),
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use bimm_contracts::assert_shape_contract;
    use burn::backend::NdArray;

    #[test]
    fn test_residual_block_config() {
        let in_planes = 16;
        let out_planes = 32;

        {
            let inner_cfg = BasicBlockConfig::new(in_planes, out_planes).with_stride(2);

            let cfg: ResidualBlockStructureConfig = inner_cfg.clone().into();
            assert!(matches!(cfg, ResidualBlockStructureConfig::Basic(_)));
            assert_eq!(cfg.in_planes(), in_planes);
            assert_eq!(cfg.out_planes(), out_planes);
            assert_eq!(cfg.stride(), 2);
            assert_eq!(cfg.output_resolution([20, 20]), [10, 10]);
        }

        {
            let inner_cfg = BottleneckBlockConfig::new(in_planes, out_planes).with_stride(2);

            let cfg: ResidualBlockStructureConfig = inner_cfg.clone().into();
            assert!(matches!(cfg, ResidualBlockStructureConfig::Bottleneck(_)));
            assert_eq!(cfg.in_planes(), in_planes);
            assert_eq!(cfg.out_planes(), out_planes);
            assert_eq!(cfg.stride(), 2);
            assert_eq!(cfg.output_resolution([20, 20]), [10, 10]);
        }
    }

    #[test]
    fn test_residual_block_basic_block() {
        type B = NdArray;
        let device = Default::default();

        let batch_size = 2;
        let in_planes = 16;
        let planes = 32;
        let in_height = 8;
        let in_width = 8;
        let out_height = 4;
        let out_width = 4;

        let cfg: ResidualBlockStructureConfig = BasicBlockConfig::new(in_planes, planes)
            .with_stride(2)
            .into();

        let block: ResidualBlock<B> = cfg.init(&device);
        assert!(matches!(block, ResidualBlock::Basic(_)));
        assert_eq!(block.in_planes(), in_planes);
        assert_eq!(block.out_planes(), planes);
        assert_eq!(block.stride(), 2);
        assert_eq!(block.output_resolution([20, 20]), [10, 10]);

        let input = Tensor::ones([batch_size, in_planes, in_height, in_width], &device);
        let output = block.forward(input);

        assert_shape_contract!(
            ["batch", "out_channels", "out_height", "out_width"],
            &output.dims(),
            &[
                ("batch", batch_size),
                ("out_channels", planes),
                ("out_height", out_height),
                ("out_width", out_width)
            ],
        );
    }

    #[test]
    fn test_residual_block_bottleneck_block() {
        use burn::backend::Wgpu;
        type B = Wgpu;
        let device = Default::default();

        let batch_size = 2;
        let in_planes = 16;
        let planes = 32;
        let in_height = 8;
        let in_width = 8;
        let out_height = 4;
        let out_width = 4;

        let cfg: ResidualBlockStructureConfig = BottleneckBlockConfig::new(in_planes, planes)
            .with_stride(2)
            .into();

        let block: ResidualBlock<B> = cfg.init(&device);
        assert!(matches!(block, ResidualBlock::Bottleneck(_)));
        assert_eq!(block.in_planes(), in_planes);
        assert_eq!(block.out_planes(), planes);
        assert_eq!(block.stride(), 2);
        assert_eq!(block.output_resolution([20, 20]), [10, 10]);

        let input = Tensor::ones([batch_size, in_planes, in_height, in_width], &device);
        let output = block.forward(input);

        assert_shape_contract!(
            ["batch", "out_planes", "out_height", "out_width"],
            &output.dims(),
            &[
                ("batch", batch_size),
                ("out_planes", planes),
                ("out_height", out_height),
                ("out_width", out_width)
            ],
        );
    }
}