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

use crate::{Shape, Value};
use std::collections::HashMap;

pub trait Func<IN, OUT> {
    fn apply(&self, input: IN) -> OUT;
}

mod tensor_iter;
pub use tensor_iter::TensorIter;

mod eval;
pub use eval::eval;

mod fn_impls;

mod raiexpr;
pub use raiexpr::raiexpr;

pub fn jvp<IN, OUT, F>(func: F, input: IN, tangents: IN::Gradient) -> (OUT, OUT::Gradient)
where
    F: Func<IN, OUT>,
    IN: Value,
    OUT: Value,
{
    let input_tensors = input.tensors();
    let mut grads = HashMap::new();
    IN::grad_map(&input_tensors, tangents, &mut grads);
    let output = func.apply(input);
    let output_tensors = output.tensors();
    let mut jvps = HashMap::new();
    for t in output_tensors.tensor_iter() {
        jvps.insert(t.id(), t.jvp(&mut grads));
    }
    let grad = OUT::grad(&output_tensors, &jvps);
    (output, grad)
}

type VjpFunc<IN, OUT> = Box<dyn Fn(IN) -> OUT>;
pub fn vjp<IN, OUT, F>(func: F, input: IN) -> (OUT, VjpFunc<OUT::Gradient, IN::Gradient>)
where
    F: Func<IN, OUT>,
    IN: Value,
    OUT: Value,
{
    let input_tensors = input.tensors();
    let output = func.apply(input);
    let output_tensors = output.tensors();
    let vjps_fn = move |cotangents: OUT::Gradient| {
        let mut grads = HashMap::new();
        OUT::grad_map(&output_tensors, cotangents, &mut grads);
        for t in output_tensors.tensor_iter() {
            t.vjp(&mut grads);
        }
        let mut vjps = HashMap::new();
        for t in input_tensors.tensor_iter() {
            let id = t.id();
            let c = grads.get(&id).unwrap().clone();
            if c.shape_eq(&t) {
                vjps.insert(id, c);
            } else {
                // reduce sum cotangent to its input shape
                // TODO: will cotangent ndim always > input ndim?
                vjps.insert(id, c.sum(..t.ndim()));
            }
        }
        IN::grad(&input_tensors, &vjps)
    };
    (output, Box::new(vjps_fn))
}

#[derive(Clone)]
pub struct GradFunc<IN, OUT, F>
where
    F: Func<IN, OUT> + Clone,
{
    func: F,
    phantom: std::marker::PhantomData<(IN, OUT)>,
}

impl<IN, OUT, F> GradFunc<IN, OUT, F>
where
    F: Func<IN, OUT> + Clone,
{
    pub fn new(func: F) -> Self {
        Self {
            func,
            phantom: std::marker::PhantomData,
        }
    }
}

impl<IN, OUT, F> Func<IN, IN::Gradient> for GradFunc<IN, OUT, F>
where
    F: Func<IN, OUT> + Clone,
    IN: Value,
    OUT: Value,
{
    fn apply(&self, input: IN) -> IN::Gradient {
        let (output, vjp_fn) = vjp(self.func.clone(), input);
        let mut cotangents = HashMap::new();
        let output_tensors = output.tensors();
        for t in output_tensors.tensor_iter() {
            cotangents.insert(t.id(), t.ones_like());
        }
        vjp_fn(OUT::grad(&output_tensors, &cotangents))
    }
}

pub fn grad<IN, OUT, F>(func: F) -> GradFunc<IN, OUT, F>
where
    F: Func<IN, OUT> + Clone,
    IN: Value,
    OUT: Value,
{
    GradFunc::new(func)
}

#[derive(Clone)]
pub struct ValueAndGradFunc<IN, OUT, F>
where
    F: Func<IN, OUT> + Clone,
{
    func: F,
    phantom: std::marker::PhantomData<(IN, OUT)>,
}

impl<IN, OUT, F> ValueAndGradFunc<IN, OUT, F>
where
    F: Func<IN, OUT> + Clone,
{
    pub fn new(func: F) -> Self {
        Self {
            func,
            phantom: std::marker::PhantomData,
        }
    }
}

impl<IN, OUT, F> Func<IN, (OUT, IN::Gradient)> for ValueAndGradFunc<IN, OUT, F>
where
    F: Func<IN, OUT> + Clone,
    IN: Value,
    OUT: Value,
{
    fn apply(&self, input: IN) -> (OUT, IN::Gradient) {
        let (output, vjp_fn) = vjp(self.func.clone(), input);
        let mut cotangents = HashMap::new();
        let output_tensors = output.tensors();
        for t in output_tensors.tensor_iter() {
            cotangents.insert(t.id(), t.ones_like());
        }
        (output, vjp_fn(OUT::grad(&output_tensors, &cotangents)))
    }
}

pub fn value_and_grad<IN, OUT, F>(func: F) -> ValueAndGradFunc<IN, OUT, F>
where
    F: Func<IN, OUT> + Clone,
    IN: Value,
    OUT: Value,
{
    ValueAndGradFunc::new(func)
}