rustyml 0.11.0

A high-performance machine learning & deep learning library in pure Rust, offering ML algorithms and neural network support
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

/// Common implementation for `output_shape` method in normalization layers
macro_rules! normalization_layer_output_shape {
    ($self:expr) => {
        if !$self.input_shape.is_empty() {
            format!(
                "({})",
                $self
                    .input_shape
                    .iter()
                    .map(|x| x.to_string())
                    .collect::<Vec<_>>()
                    .join(", ")
            )
        } else {
            String::from("Unknown")
        }
    };
}

/// Common implementation for `update_parameters_sgd` method in normalization layers
macro_rules! normalization_layer_update_parameters_sgd {
    ($self:expr, $lr:expr) => {{
        if let Some(grad_gamma) = &$self.grad_gamma {
            $self.gamma = &$self.gamma - &(grad_gamma * $lr);
        }
        if let Some(grad_beta) = &$self.grad_beta {
            $self.beta = &$self.beta - &(grad_beta * $lr);
        }
    }};
}

/// Common implementation for `update_parameters_adam` method in normalization layers
macro_rules! normalization_layer_update_parameters_adam {
    ($self:expr, $lr:expr, $beta1:expr, $beta2:expr, $epsilon:expr, $t:expr) => {{
        // Initialize moment estimates if not already done
        if $self.optimizer_cache.adam_states.is_none() {
            $self.optimizer_cache.adam_states = Some(AdamStatesNormalizationLayer {
                m_gamma: Tensor::zeros($self.gamma.raw_dim()),
                v_gamma: Tensor::zeros($self.gamma.raw_dim()),
                m_beta: Tensor::zeros($self.beta.raw_dim()),
                v_beta: Tensor::zeros($self.beta.raw_dim()),
            });
        }

        if let Some(grad_gamma) = &$self.grad_gamma {
            let adam_states = $self.optimizer_cache.adam_states.as_mut().unwrap();
            let m_gamma = &mut adam_states.m_gamma;
            let v_gamma = &mut adam_states.v_gamma;

            // Update biased first moment estimate
            *m_gamma = m_gamma.clone() * $beta1 + grad_gamma * (1.0 - $beta1);

            // Update biased second raw moment estimate
            *v_gamma = v_gamma.clone() * $beta2 + &(grad_gamma * grad_gamma) * (1.0 - $beta2);

            // Compute bias-corrected first moment estimate
            let m_hat = m_gamma.clone() / (1.0 - $beta1.powi($t as i32));

            // Compute bias-corrected second raw moment estimate
            let v_hat = v_gamma.clone() / (1.0 - $beta2.powi($t as i32));

            // Update parameters
            $self.gamma = &$self.gamma - &(&m_hat / &(v_hat.mapv(|x| x.sqrt()) + $epsilon) * $lr);
        }

        if let Some(grad_beta) = &$self.grad_beta {
            let adam_states = $self.optimizer_cache.adam_states.as_mut().unwrap();
            let m_beta = &mut adam_states.m_beta;
            let v_beta = &mut adam_states.v_beta;

            *m_beta = m_beta.clone() * $beta1 + grad_beta * (1.0 - $beta1);
            *v_beta = v_beta.clone() * $beta2 + &(grad_beta * grad_beta) * (1.0 - $beta2);

            let m_hat = m_beta.clone() / (1.0 - $beta1.powi($t as i32));
            let v_hat = v_beta.clone() / (1.0 - $beta2.powi($t as i32));

            $self.beta = &$self.beta - &(&m_hat / &(v_hat.mapv(|x| x.sqrt()) + $epsilon) * $lr);
        }
    }};
}

/// Common implementation for `update_parameters_rmsprop` method in normalization layers
macro_rules! normalization_layer_update_parameters_rmsprop {
    ($self:expr, $lr:expr, $rho:expr, $epsilon:expr) => {{
        // Initialize cache if not already done
        if $self.optimizer_cache.rmsprop_cache.is_none() {
            $self.optimizer_cache.rmsprop_cache = Some(RMSpropCacheNormalizationLayer {
                cache_gamma: Tensor::zeros($self.gamma.raw_dim()),
                cache_beta: Tensor::zeros($self.beta.raw_dim()),
            });
        }

        if let Some(grad_gamma) = &$self.grad_gamma {
            let rmsprop_cache = $self.optimizer_cache.rmsprop_cache.as_mut().unwrap();
            let cache_gamma = &mut rmsprop_cache.cache_gamma;

            // Update cache
            *cache_gamma = cache_gamma.clone() * $rho + &(grad_gamma * grad_gamma) * (1.0 - $rho);

            // Update parameters
            $self.gamma =
                &$self.gamma - &(grad_gamma / &(cache_gamma.mapv(|x| x.sqrt()) + $epsilon) * $lr);
        }

        if let Some(grad_beta) = &$self.grad_beta {
            let rmsprop_cache = $self.optimizer_cache.rmsprop_cache.as_mut().unwrap();
            let cache_beta = &mut rmsprop_cache.cache_beta;

            *cache_beta = cache_beta.clone() * $rho + &(grad_beta * grad_beta) * (1.0 - $rho);

            $self.beta =
                &$self.beta - &(grad_beta / &(cache_beta.mapv(|x| x.sqrt()) + $epsilon) * $lr);
        }
    }};
}

/// Common implementation for `update_parameters_ada_grad` method in normalization layers
macro_rules! normalization_layer_update_parameters_ada_grad {
    ($self:expr, $lr:expr, $epsilon:expr) => {{
        // Initialize accumulated gradient if not already done
        if $self.optimizer_cache.ada_grad_cache.is_none() {
            $self.optimizer_cache.ada_grad_cache = Some(AdaGradStatesNormalizationLayer {
                acc_grad_gamma: Tensor::zeros($self.gamma.raw_dim()),
                acc_grad_beta: Tensor::zeros($self.beta.raw_dim()),
            });
        }

        if let Some(grad_gamma) = &$self.grad_gamma {
            let ada_grad_cache = $self.optimizer_cache.ada_grad_cache.as_mut().unwrap();
            let acc_grad_gamma = &mut ada_grad_cache.acc_grad_gamma;

            // Accumulate squared gradients
            *acc_grad_gamma = acc_grad_gamma.clone() + &(grad_gamma * grad_gamma);

            // Update parameters
            $self.gamma = &$self.gamma
                - &(grad_gamma / &(acc_grad_gamma.mapv(|x| x.sqrt()) + $epsilon) * $lr);
        }

        if let Some(grad_beta) = &$self.grad_beta {
            let ada_grad_cache = $self.optimizer_cache.ada_grad_cache.as_mut().unwrap();
            let acc_grad_beta = &mut ada_grad_cache.acc_grad_beta;

            *acc_grad_beta = acc_grad_beta.clone() + &(grad_beta * grad_beta);

            $self.beta =
                &$self.beta - &(grad_beta / &(acc_grad_beta.mapv(|x| x.sqrt()) + $epsilon) * $lr);
        }
    }};
}

/// Common implementation for computing gamma and beta gradients in normalization layers
///
/// This macro computes the gradients for gamma and beta parameters by summing over
/// all spatial dimensions and batch dimensions for each channel. It handles both
/// parallel and sequential computation based on the total number of elements.
///
/// # Parameters
///
/// - `$grad_gamma:expr` - Mutable tensor to store gamma gradients
/// - `$grad_beta:expr` - Mutable tensor to store beta gradients
/// - `$grad_output:expr` - Gradient from the next layer
/// - `$x_normalized:expr` - Normalized input from forward pass
/// - `$batch_size:expr` - Number of samples in the batch
/// - `$num_channels:expr` - Number of channels
/// - `$spatial_size:expr` - Size of spatial dimensions (product of all spatial dims)
/// - `$total_elements:expr` - Total number of elements in the tensor
/// - `$parallel_threshold:expr` - Threshold for switching to parallel computation
macro_rules! compute_normalization_layer_parameter_gradients {
    (
        $grad_gamma:expr,
        $grad_beta:expr,
        $grad_output:expr,
        $x_normalized:expr,
        $batch_size:expr,
        $num_channels:expr,
        $spatial_size:expr,
        $total_elements:expr,
        $parallel_threshold:expr
    ) => {{
        if $total_elements >= $parallel_threshold {
            // Parallel computation of parameter gradients
            // Compute both gamma and beta gradients in a single pass
            let (grad_gamma_vec, grad_beta_vec): (Vec<f32>, Vec<f32>) = (0..$num_channels)
                .into_par_iter()
                .map(|channel_idx| {
                    let mut gamma_sum = 0.0f32;
                    let mut beta_sum = 0.0f32;

                    for batch_idx in 0..$batch_size {
                        let start = (batch_idx * $num_channels + channel_idx) * $spatial_size;
                        let end = start + $spatial_size;

                        for i in start..end {
                            let grad_out = $grad_output.as_slice().unwrap()[i];
                            gamma_sum += grad_out * $x_normalized.as_slice().unwrap()[i];
                            beta_sum += grad_out;
                        }
                    }

                    (gamma_sum, beta_sum)
                })
                .unzip();

            $grad_gamma
                .as_slice_mut()
                .unwrap()
                .copy_from_slice(&grad_gamma_vec);
            $grad_beta
                .as_slice_mut()
                .unwrap()
                .copy_from_slice(&grad_beta_vec);
        } else {
            // Sequential computation
            for channel_idx in 0..$num_channels {
                let mut gamma_sum = 0.0f32;
                let mut beta_sum = 0.0f32;

                for batch_idx in 0..$batch_size {
                    let start = (batch_idx * $num_channels + channel_idx) * $spatial_size;
                    let end = start + $spatial_size;

                    for i in start..end {
                        gamma_sum += $grad_output.as_slice().unwrap()[i]
                            * $x_normalized.as_slice().unwrap()[i];
                        beta_sum += $grad_output.as_slice().unwrap()[i];
                    }
                }

                $grad_gamma.as_slice_mut().unwrap()[channel_idx] = gamma_sum;
                $grad_beta.as_slice_mut().unwrap()[channel_idx] = beta_sum;
            }
        }
    }};
}

/// Batch Normalization layer for neural networks
pub mod batch_normalization;
/// Group Normalization layer for neural networks
pub mod group_normalization;
/// Instance Normalization layer for neural networks
pub mod instance_normalization;
/// Layer Normalization layer for neural networks
pub mod layer_normalization;

pub use batch_normalization::BatchNormalization;
pub use group_normalization::GroupNormalization;
pub use instance_normalization::InstanceNormalization;
pub use layer_normalization::{LayerNormalization, LayerNormalizationAxis};