briny_ai 0.5.0

A tiny & efficient AI inference engine
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

use super::{exp, ln};
use crate::nn::{
    tensors::{Tensor, WithGrad},
    TensorFloat,
};
use tensor_optim::TensorOps;

#[cfg(feature = "alloc")]
use alloc::boxed::Box;
#[cfg(feature = "dyntensor")]
use alloc::vec;
#[cfg(not(feature = "alloc"))]
use box_closure::{Align8, OpaqueFn};

/// Cross-entropy loss for arbitrary-rank tensors along last axis, ignoring zeroed timesteps.
#[must_use]
#[cfg(feature = "dyntensor")]
pub fn cross_entropy_nonzero_loss<'a>(
    prediction: &'a WithGrad<Tensor<TensorFloat>>,
    target: &'a Tensor<TensorFloat>,
) -> (
    TensorFloat,
    Box<dyn Fn(TensorFloat) -> Tensor<TensorFloat> + 'a>,
) {
    let shape = prediction.get_value().shape();
    let rank = shape.len();
    let last_dim = shape[rank - 1];
    let outer_size: usize = shape[..rank - 1].iter().product();

    let pred_data = prediction.get_value().data();
    let target_data = target.data();

    // Compute softmax and loss
    let mut softmax = vec![0.0; pred_data.len()];
    let mut loss_sum = 0.0;

    let mut nonzero_mask = alloc::vec![true; outer_size];
    let mut valid_timesteps = 0_usize;

    for (i, val) in nonzero_mask.iter_mut().enumerate().take(outer_size) {
        let offset = i * last_dim;
        let slice = &pred_data[offset..offset + last_dim];
        let t_slice = &target_data[offset..offset + last_dim];

        let is_padding = t_slice.iter().all(|&x| x == 0.0);
        if is_padding {
            *val = false;
            continue;
        }

        let max_val = slice
            .iter()
            .copied()
            .fold(TensorFloat::NEG_INFINITY, TensorFloat::max);
        let exp_sum: TensorFloat = slice.iter().map(|&x| exp(x - max_val)).sum();

        for j in 0..last_dim {
            let s = exp(slice[j] - max_val) / exp_sum;
            softmax[offset + j] = s;
            loss_sum -= t_slice[j] * ln(s);
        }

        valid_timesteps += 1;
    }

    let n_samples = valid_timesteps.max(1) as TensorFloat;
    let loss = loss_sum / n_samples;

    let back = move |grad_output: TensorFloat| {
        let mut grad = vec![0.0; pred_data.len()];
        for (i, val) in nonzero_mask.iter().enumerate().take(outer_size) {
            if !*val {
                continue;
            }

            let offset = i * last_dim;
            let t_slice = &target_data[offset..offset + last_dim];
            let s_slice = &softmax[offset..offset + last_dim];

            for j in 0..last_dim {
                grad[offset + j] = (s_slice[j] - t_slice[j]) * grad_output / n_samples;
            }
        }
        Tensor::new(shape, &grad)
    };

    (loss, Box::new(back))
}

/// Cross-entropy loss for arbitrary-rank tensors along last axis, ignoring zeroed timesteps.
#[cfg(all(feature = "alloc", not(feature = "dyntensor")))]
#[must_use]
pub fn cross_entropy_nonzero_loss<'a, const N: usize, const D: usize>(
    prediction: &'a WithGrad<Tensor<TensorFloat, N, D>>,
    target: &'a Tensor<TensorFloat, N, D>,
) -> (
    TensorFloat,
    Box<dyn Fn(TensorFloat) -> Tensor<TensorFloat, N, D> + 'a>,
) {
    use tensor_optim::ConstTensorOps;

    let shape: &[usize; D] = prediction.get_value().shape_array();
    let pred_data = prediction.get_value().data();
    let target_data = target.data();

    let last_dim = shape[D - 1];
    let outer_size: usize = shape[..D - 1].iter().product();

    let mut softmax = [0.0; N];
    let mut loss_sum = 0.0;

    let mut nonzero_mask = alloc::vec![true; outer_size];
    let mut valid_timesteps = 0_usize;

    for (i, val) in nonzero_mask.iter_mut().enumerate().take(outer_size) {
        let offset = i * last_dim;
        let slice = &pred_data[offset..offset + last_dim];
        let t_slice = &target_data[offset..offset + last_dim];

        let is_padding = t_slice.iter().all(|&x| x == 0.0);
        if is_padding {
            *val = false;
            continue;
        }

        let max_val = slice
            .iter()
            .copied()
            .fold(TensorFloat::NEG_INFINITY, TensorFloat::max);
        let exp_sum: TensorFloat = slice.iter().map(|&x| exp(x - max_val)).sum();

        for j in 0..last_dim {
            let s = exp(slice[j] - max_val) / exp_sum;
            softmax[offset + j] = s;
            loss_sum -= t_slice[j] * ln(s);
        }

        valid_timesteps += 1;
    }

    let n_samples = valid_timesteps.max(1) as TensorFloat;
    let loss = loss_sum / n_samples;

    let back = move |grad_output: TensorFloat| {
        let mut grad = [0.0; N];
        for (i, val) in nonzero_mask.iter().enumerate().take(outer_size) {
            if !*val {
                continue;
            }

            let offset = i * last_dim;
            let t_slice = &target_data[offset..offset + last_dim];
            let s_slice = &softmax[offset..offset + last_dim];

            for j in 0..last_dim {
                grad[offset + j] = (s_slice[j] - t_slice[j]) * grad_output / n_samples;
            }
        }
        Tensor::new(shape, &grad)
    };

    (loss, Box::new(back))
}

/// Cross-entropy loss for arbitrary-rank tensors along last axis, ignoring zeroed timesteps.
#[cfg(not(feature = "alloc"))]
#[must_use]
pub fn cross_entropy_nonzero_loss<'a, const N: usize, const D: usize>(
    prediction: &'a WithGrad<Tensor<TensorFloat, N, D>>,
    target: &'a Tensor<TensorFloat, N, D>,
) -> (
    TensorFloat,
    OpaqueFn<'a, TensorFloat, Tensor<TensorFloat, N, D>, Align8<128>>,
) {
    use tensor_optim::ConstTensorOps;

    let shape: &[usize; D] = prediction.get_value().shape_array();
    let pred_data = prediction.get_value().data();
    let target_data = target.data();

    let last_dim = shape[D - 1];
    let outer_size: usize = shape[..D - 1].iter().product();

    let mut softmax = [0.0; N];
    let mut loss_sum = 0.0;

    let mut nonzero_mask = [true; N]; // outer_size <= N
    let mut valid_timesteps = 0_usize;

    for (i, val) in nonzero_mask.iter_mut().enumerate().take(outer_size) {
        let offset = i * last_dim;
        let slice = &pred_data[offset..offset + last_dim];
        let t_slice = &target_data[offset..offset + last_dim];

        let is_padding = t_slice.iter().all(|&x| x == 0.0);
        if is_padding {
            *val = false;
            continue;
        }

        let max_val = slice
            .iter()
            .copied()
            .fold(TensorFloat::NEG_INFINITY, TensorFloat::max);
        let exp_sum: TensorFloat = slice.iter().map(|&x| exp(x - max_val)).sum();

        for j in 0..last_dim {
            let s = exp(slice[j] - max_val) / exp_sum;
            softmax[offset + j] = s;
            loss_sum -= t_slice[j] * ln(s);
        }

        valid_timesteps += 1;
    }

    let n_samples = valid_timesteps.max(1) as TensorFloat;
    let loss = loss_sum / n_samples;

    let back = move |grad_output: TensorFloat| {
        let mut grad = [0.0; N];
        for (i, val) in nonzero_mask.iter().enumerate().take(outer_size) {
            if !*val {
                continue;
            }

            let offset = i * last_dim;
            let t_slice = &target_data[offset..offset + last_dim];
            let s_slice = &softmax[offset..offset + last_dim];

            for j in 0..last_dim {
                grad[offset + j] = (s_slice[j] - t_slice[j]) * grad_output / n_samples;
            }
        }
        Tensor::new(shape, &grad)
    };

    (loss, OpaqueFn::new(back))
}