autograd 0.8.0

Tensors and differentiable operations in Rust
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

extern crate ndarray;

use ndarray_ext::NdArray;
use op;
use ops;
use std::f32;
use tensor::Tensor;

pub struct ELU {
    pub alpha: f32,
}

pub struct ELUGrad {
    pub alpha: f32,
}

pub struct Identity;

pub struct ReLU;

pub struct Sigmoid;

pub struct Softplus;

pub struct Softmax {
    pub axis: isize,
}

#[inline]
pub fn softmax_forward(x: &NdArray, axis: isize) -> NdArray {
    let axis = if axis < 0 {
        (x.ndim() as isize + axis) as usize
    } else {
        axis as usize
    };

    let mut a = x.shape().to_vec();
    a[axis] = 1;
    let reduced_shape = a.as_slice();
    let max_fn = f32::max;
    // unwrap is safe
    let ref max = x.fold_axis(ndarray::Axis(axis), f32::MIN, move |&a, &b| max_fn(a, b))
        .into_shape(ndarray::IxDyn(reduced_shape))
        .unwrap();
    // subtract `max` to prevent overflow
    let mut tmp = x - max;
    tmp.mapv_inplace(|a| a.exp());
    // unwrap is safe
    let sum = tmp.sum_axis(ndarray::Axis(axis))
        .into_shape(ndarray::IxDyn(reduced_shape))
        .unwrap();
    tmp /= &sum;
    tmp
}

impl op::Op for Softmax {
    fn name(&self) -> &str {
        "Softmax"
    }

    fn compute(&self, ctx: ::runtime::OpComputeContext) -> op::ComputeResult {
        vec![Ok(softmax_forward(ctx.grab_inputs()[0], self.axis))]
    }

    fn grad(&self, gy: &Tensor, _: &[&Tensor], output: &Tensor) -> Vec<Option<Tensor>> {
        let sum = ops::reduce_sum(&(output * gy), &[self.axis], true);
        vec![Some((gy - sum) * output)]
    }
}

impl op::Op for Softplus {
    fn name(&self) -> &str {
        "Softplus"
    }

    fn compute(&self, ctx: ::runtime::OpComputeContext) -> op::ComputeResult {
        let xs = ctx.grab_inputs();
        let e = f32::consts::E;
        vec![Ok(xs[0].mapv(move |a| (a.exp() + 1.).log(e)))]
    }

    fn grad(&self, gy: &Tensor, xs: &[&Tensor], _: &Tensor) -> Vec<Option<Tensor>> {
        let a = &ops::exp(xs[0]);
        let b = a + 1;
        let gx = gy * (a / b);
        vec![Some(gx)]
    }
}

impl op::Op for Sigmoid {
    fn name(&self) -> &str {
        "Sigmoid"
    }

    fn compute(&self, ctx: ::runtime::OpComputeContext) -> op::ComputeResult {
        let x = ctx.grab_inputs()[0];
        vec![Ok(x.mapv(|a| ((a * 0.5).tanh() * 0.5) + 0.5))]
    }

    fn grad(&self, gy: &Tensor, _: &[&Tensor], y: &Tensor) -> Vec<Option<Tensor>> {
        vec![Some(gy * (y - ops::square(y)))]
    }
}

impl op::Op for ReLU {
    fn name(&self) -> &str {
        "ReLU"
    }

    fn compute(&self, ctx: ::runtime::OpComputeContext) -> op::ComputeResult {
        let x = ctx.grab_inputs()[0];
        vec![Ok(x.map(|a| a.max(0.)))]
    }

    fn grad(&self, gy: &Tensor, inputs: &[&Tensor], _: &Tensor) -> Vec<Option<Tensor>> {
        let bin = ops::greater(inputs[0], &ops::scalar(0.));
        // inplace is ok because second derivative of relu is 0.
        // (`mul_inplace` returns `None` as input gradient.)
        vec![Some(ops::mul_inplace(bin, gy))]
    }
}

impl op::Op for Identity {
    fn name(&self) -> &str {
        "Identity"
    }

    fn compute(&self, _: ::runtime::OpComputeContext) -> op::ComputeResult {
        // do nothing
        vec![Err(::op::ComputeError::Delegate { to: 0 })]
    }

    fn grad(&self, gy: &Tensor, _: &[&Tensor], _: &Tensor) -> Vec<Option<Tensor>> {
        // use gy's array with rc increment.
        vec![Some(gy.clone())]
    }
}

impl op::Op for ELU {
    fn name(&self) -> &str {
        "ELU"
    }

    fn compute(&self, ctx: ::runtime::OpComputeContext) -> op::ComputeResult {
        let x = ctx.grab_inputs()[0];
        let ret = x.mapv(move |a| {
            if a > 0. {
                a
            } else {
                self.alpha * (a.exp() - 1.)
            }
        });
        vec![Ok(ret)]
    }

    fn grad(&self, gy: &Tensor, inputs: &[&Tensor], _: &Tensor) -> Vec<Option<Tensor>> {
        let gx = Tensor::builder()
            .set_inputs(vec![inputs[0], gy])
            .set_shape(gy.shape())
            .build(ELUGrad { alpha: self.alpha });
        vec![Some(gx)]
    }
}

impl op::Op for ELUGrad {
    fn name(&self) -> &str {
        "ELUGrad"
    }

    fn compute(&self, ctx: ::runtime::OpComputeContext) -> op::ComputeResult {
        let xs = ctx.grab_inputs();
        let x = xs[0];
        let gy = xs[1];
        let a = x.mapv(move |a| {
            if a > 0. {
                1.
            } else {
                self.alpha * (a.exp() - 1.) + self.alpha
            }
        });
        vec![Ok(a * gy)]
    }

    fn grad(&self, _: &Tensor, _: &[&Tensor], _: &Tensor) -> Vec<Option<Tensor>> {
        vec![None, None]
    }
}