gradients 0.3.4

An OpenCL, CUDA and CPU based Deep Learning Library
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

use crate::Param;
use custos::{Alloc, CDatatype, CPU, GraphReturn};
use custos_math::Matrix;

#[cfg(feature = "opencl")]
use custos::{opencl::enqueue_kernel, CLDevice};

pub struct SGD<'a, T> {
    lr: T,
    weight_momentum: Vec<Matrix<'a, T>>,
    bias_momentum: Vec<Matrix<'a, T>>,
    momentum: T,
}

impl<'a, T: CDatatype> SGD<'a, T> {
    pub fn new(lr: T) -> Self {
        SGD {
            lr,
            weight_momentum: Vec::new(),
            bias_momentum: Vec::new(),
            momentum: T::one() / T::two(),
        }
    }

    pub fn momentum(mut self, momentum: T) -> Self {
        self.momentum = momentum;
        self
    }

    pub fn step<D: Alloc<T> + SGDOp<T> + GraphReturn>(&mut self, device: &'a D, params: Vec<Param<T>>) {
        if self.momentum > T::zero() {
            if self.weight_momentum.len() < params.len() {
                for param in &params {
                    self.weight_momentum
                        .push(Matrix::new(device, param.weights.dims()));
                    self.bias_momentum
                        .push(Matrix::new(device, param.bias.dims()));
                }
            }
            return device.step_momentum(self, params);
        }
        device.step(self, params)
    }
}

pub trait SGDOp<T: CDatatype> {
    fn step(&self, sgd: &mut SGD<T>, params: Vec<Param<T>>) {
        for mut param in params {
            param.weights -= param.dweights * sgd.lr;
            param.bias -= param.dbias * sgd.lr;
        }
    }
    fn step_momentum(&self, sgd: &mut SGD<T>, params: Vec<Param<T>>);
}

impl<T: CDatatype> SGDOp<T> for CPU {
    fn step_momentum(&self, sgd: &mut SGD<T>, mut params: Vec<Param<T>>) {
        for (layer_idx, param) in params.iter_mut().enumerate() {
            for (idx, w) in param.weights.iter_mut().enumerate() {
                let update = sgd.momentum * sgd.weight_momentum[layer_idx][idx]
                    + param.dweights[idx] * sgd.lr;
                *w -= update;
                sgd.weight_momentum[layer_idx][idx] = update;
            }

            for (idx, b) in param.bias.iter_mut().enumerate() {
                let update =
                    sgd.momentum * sgd.bias_momentum[layer_idx][idx] + param.dbias[idx] * sgd.lr;
                *b -= update;
                sgd.bias_momentum[layer_idx][idx] = update;
            }
        }
    }
}

#[cfg(feature = "opencl")]
impl<T: CDatatype> SGDOp<T> for CLDevice {
    fn step_momentum(&self, sgd: &mut SGD<T>, params: Vec<Param<T>>) {
        let src = format!(
            "
            __kernel void sgd_momentum(
                __global {dt}* values, 
                __global const {dt}* dvalues, 
                __global {dt}* value_momentum,
                const {dt} momentum,
                const {dt} lr)
                {{
                    size_t i = get_global_id(0);
                    
                    {dt} value_update = momentum * value_momentum[i] + dvalues[i] * lr;
                    values[i] -= value_update;
                    value_momentum[i] = value_update;
                }}
        ",
            dt = T::as_c_type_str()
        );

        for (idx, param) in params.iter().enumerate() {
            enqueue_kernel(
                self,
                &src,
                [param.weights.size(), 0, 0],
                None,
                &[
                    &param.weights,
                    &param.dweights,
                    &sgd.weight_momentum[idx],
                    &sgd.momentum,
                    &sgd.lr,
                ],
            )
            .unwrap();

            enqueue_kernel(
                self,
                &src,
                [param.bias.size(), 0, 0],
                None,
                &[
                    &param.bias,
                    &param.dbias,
                    &sgd.bias_momentum[idx],
                    &sgd.momentum,
                    &sgd.lr,
                ],
            )
            .unwrap();
            /*
            KernelOptions::new(
                self,
                param.weights.as_buf(),
                [param.weights.size(), 0, 0],
                &src,
            )
            .unwrap()
            .add_arg(&param.dweights)
            .add_arg(&sgd.weight_momentum[idx])
            .add_arg(&sgd.momentum)
            .add_arg(&sgd.lr)
            .run()
            .unwrap();

            KernelOptions::new(self, param.bias.as_buf(), [param.bias.size(), 0, 0], &src)
                .unwrap()
                .add_arg(&param.dbias)
                .add_arg(&sgd.bias_momentum[idx])
                .add_arg(&sgd.momentum)
                .add_arg(&sgd.lr)
                .run()
                .unwrap();
            */
        }
    }
}