rai_core/transforms/
mod.rs

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

pub trait Func<InKind, In, Out> {
    fn invoke(&self, input: In) -> Out;
}

mod tensor_iter;
pub use tensor_iter::TensorIter;

mod eval;
pub use eval::eval;

mod fn_impls;

mod optimize;
pub use optimize::optimize;

mod raiexpr;
pub use raiexpr::raiexpr;

pub fn jvp<K, IN, OUT, F>(func: F, input: IN, tangents: IN::Gradient) -> (OUT, OUT::Gradient)
where
    F: Func<K, 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.invoke(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)
}

pub fn linearize<'a, K, IN, OUT, F>(
    func: F,
    input: IN,
) -> (OUT, impl Fn(IN::Gradient) -> OUT::Gradient + Clone + 'a)
where
    F: Func<K, IN, OUT> + Clone + 'a,
    IN: Value,
    OUT: Value,
{
    let input_tensors = input.tensors();
    let output = func.invoke(input);
    let output_tensors = output.tensors();
    let jvp_fn = move |tangents: IN::Gradient| {
        let mut grads = HashMap::new();
        IN::grad_map(&input_tensors, tangents, &mut grads);
        let mut jvps = HashMap::new();
        for t in output_tensors.tensor_iter() {
            jvps.insert(t.id(), t.jvp(&mut grads));
        }
        OUT::grad(&output_tensors, &jvps)
    };
    (output, jvp_fn)
}

pub fn vjp<'a, K, IN, OUT, F>(
    func: F,
    input: IN,
) -> (OUT, impl Fn(OUT::Gradient) -> IN::Gradient + Clone + 'a)
where
    F: Func<K, IN, OUT> + Clone + 'a,
    IN: Value,
    OUT: Value,
{
    let input_tensors = input.tensors();
    let output = func.invoke(input);
    let output_tensors = output.tensors();
    let vjps_fn = move |cotangents: OUT::Gradient| {
        let mut cotangent_cache = HashMap::new();
        OUT::grad_map(&output_tensors, cotangents, &mut cotangent_cache);
        let mut tape = BTreeSet::new();
        let mut stack = Vec::new();
        // use iterative instead of recursive to avoid stack overflow
        // TODO: use proper topo sort algorithm, now sort by id in BTreeSet
        for output in output_tensors.tensor_iter() {
            stack.push(output.clone());
        }
        while let Some(t) = stack.pop() {
            if tape.contains(&t) || t.inputs().is_empty() {
                continue;
            }
            tape.insert(t.clone());
            for input in t.inputs().iter() {
                stack.push(input.clone());
            }
        }
        // run the tape backwards
        for t in tape.iter().rev() {
            let primals = &*t.inputs();
            let cotangent = cotangent_cache
                .entry(t.id())
                .or_insert_with(|| t.ones_like());
            let cotangents = t.primitive().vjp(t, primals, cotangent);
            for (primal, cotan) in primals.iter().zip(cotangents.into_iter()) {
                let id = primal.id();
                if let Some(sum) = cotangent_cache.get(&id) {
                    cotangent_cache.insert(id, sum + cotan);
                } else {
                    cotangent_cache.insert(id, cotan);
                }
            }
        }
        // collect the final cotangents for inputs
        let mut vjps = HashMap::new();
        for t in input_tensors.tensor_iter() {
            let id = t.id();
            let c = cotangent_cache.get(&id).unwrap().clone();
            vjps.insert(id, c);
        }
        IN::grad(&input_tensors, &vjps)
    };
    (output, vjps_fn)
}

pub fn grad<'a, K, IN, OUT, F>(func: F) -> impl Fn(IN) -> IN::Gradient + Clone + 'a
where
    F: Func<K, IN, OUT> + Clone + 'a,
    IN: Value,
    OUT: Value,
{
    move |input: IN| {
        let (output, vjp_fn) = vjp(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 value_and_grad<'a, K, IN, OUT, F>(func: F) -> impl Fn(IN) -> (OUT, IN::Gradient) + Clone + 'a
where
    F: Func<K, IN, OUT> + Clone + 'a,
    IN: Value,
    OUT: Value,
{
    move |input: IN| {
        let (output, vjp_fn) = vjp(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)))
    }
}