moose 0.2.2

Encrypted learning and data processing framework
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

//! Abstraction layer for fixed-point values.

use crate::computation::*;
use crate::error::Result;
#[cfg(feature = "compile")]
use crate::execution::symbolic::Symbolic;
use crate::execution::Session;
use crate::kernels::*;
use crate::replicated::*;
use serde::{Deserialize, Serialize};

mod ops;

/// Abstract fixed-point tensor.
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub enum FixedTensor<HostFixedT, MirFixedT, RepFixedT> {
    Host(HostFixedT),
    Mirrored3(MirFixedT),
    Replicated(RepFixedT),
}

impl<HostFixedT, MirFixedT, RepFixedT> Placed for FixedTensor<HostFixedT, MirFixedT, RepFixedT>
where
    HostFixedT: Placed,
    HostFixedT::Placement: Into<Placement>,
    MirFixedT: Placed,
    MirFixedT::Placement: Into<Placement>,
    RepFixedT: Placed,
    RepFixedT::Placement: Into<Placement>,
{
    type Placement = Placement;

    fn placement(&self) -> Result<Self::Placement> {
        match self {
            FixedTensor::Host(x) => Ok(x.placement()?.into()),
            FixedTensor::Mirrored3(x) => Ok(x.placement()?.into()),
            FixedTensor::Replicated(x) => Ok(x.placement()?.into()),
        }
    }
}

pub trait FixedpointTensor {
    fn fractional_precision(&self) -> u32;
    fn integral_precision(&self) -> u32;
}

impl<RepRingT> FixedpointTensor for RepFixedTensor<RepRingT> {
    fn fractional_precision(&self) -> u32 {
        self.fractional_precision
    }

    fn integral_precision(&self) -> u32 {
        self.integral_precision
    }
}

#[cfg(feature = "compile")]
impl<RepRingT: Placed> FixedpointTensor for Symbolic<RepFixedTensor<RepRingT>> {
    fn fractional_precision(&self) -> u32 {
        match self {
            Symbolic::Symbolic(_) => unimplemented!(), // TODO(Dragos) extract from underlying op signature
            Symbolic::Concrete(x) => x.fractional_precision,
        }
    }

    fn integral_precision(&self) -> u32 {
        match self {
            Symbolic::Symbolic(_) => unimplemented!(), // TODO(Dragos) extract from underlying op signature
            Symbolic::Concrete(x) => x.integral_precision,
        }
    }
}

pub(crate) trait PrefixMul<S: Session, RepFixedT> {
    fn prefix_mul(&self, sess: &S, x: Vec<RepFixedT>) -> Vec<RepFixedT>;
}

impl<S: Session, RepFixedT> PrefixMul<S, RepFixedT> for ReplicatedPlacement
where
    RepFixedT: FixedpointTensor,
    ReplicatedPlacement: PlacementMul<S, RepFixedT, RepFixedT, RepFixedT>,
    ReplicatedPlacement: PlacementTruncPr<S, RepFixedT, RepFixedT>,
{
    fn prefix_mul(&self, sess: &S, x: Vec<RepFixedT>) -> Vec<RepFixedT> {
        let elementwise_mul =
            |rep: &ReplicatedPlacement, sess: &S, x: &RepFixedT, y: &RepFixedT| -> RepFixedT {
                assert_eq!(x.fractional_precision(), y.fractional_precision());
                rep.trunc_pr(sess, x.fractional_precision(), &rep.mul(sess, x, y))
            };

        self.prefix_op(sess, x, elementwise_mul)
    }
}

pub(crate) trait PolynomialEval<S: Session, RepFixedTensorT> {
    fn polynomial_eval(&self, sess: &S, coeffs: Vec<f64>, x: RepFixedTensorT) -> RepFixedTensorT;
}

impl<S: Session, RepFixedTensorT, MirFixedT> PolynomialEval<S, RepFixedTensorT>
    for ReplicatedPlacement
where
    RepFixedTensorT: FixedpointTensor,
    RepFixedTensorT: Clone,
    ReplicatedPlacement: PlacementMul<S, MirFixedT, RepFixedTensorT, RepFixedTensorT>,
    ReplicatedPlacement: PlacementTruncPr<S, RepFixedTensorT, RepFixedTensorT>,
    ReplicatedPlacement: PlacementAddN<S, RepFixedTensorT, RepFixedTensorT>,
    ReplicatedPlacement: PlacementAdd<S, RepFixedTensorT, MirFixedT, RepFixedTensorT>,
    ReplicatedPlacement: ShapeFill<S, RepFixedTensorT, Result = MirFixedT>,
    ReplicatedPlacement: PrefixMul<S, RepFixedTensorT>,
{
    fn polynomial_eval(&self, sess: &S, coeffs: Vec<f64>, x: RepFixedTensorT) -> RepFixedTensorT {
        assert!(!coeffs.is_empty());
        let mut degree = coeffs.len() - 1;

        // Exclude coefficients under precision
        for coeff in coeffs.iter().rev() {
            if *coeff < 2f64.powi(-(x.fractional_precision() as i32 + 1)) as f64 {
                degree -= 1
            } else {
                break;
            }
        }

        let coeffs_mir: Vec<_> = coeffs[0..degree + 1]
            .iter()
            .map(|coeff| {
                self.shape_fill(
                    sess,
                    coeff.as_fixedpoint(x.fractional_precision() as usize),
                    &x,
                )
            })
            .collect();

        let x_n: Vec<RepFixedTensorT> = (0..degree).map(|_| x.clone()).collect();

        let x_pre_mul = self.prefix_mul(sess, x_n);

        // TODO [Yann]
        // If x_pre_mul could be concatenated in one tensor, we could use a single
        // multiplication instead of doing a for loop.
        let x_mul_coeffs: Vec<RepFixedTensorT> = (0..x_pre_mul.len())
            .map(|i| self.mul(sess, &coeffs_mir[i + 1], &x_pre_mul[i]))
            .collect();

        let x_mul_coeffs_added = self.add_n(sess, &x_mul_coeffs);
        let x_mul_coeffs_added_fixed_trunc =
            self.trunc_pr(sess, x.fractional_precision(), &x_mul_coeffs_added);

        self.add(sess, &x_mul_coeffs_added_fixed_trunc, &coeffs_mir[0])
    }
}