fastnn 0.3.3

A robust deep learning library in Rust with CUDA kernels
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
188
189
190
191
192
193
194
195
196
197
198
199
200
201

//! Single-layer, unidirectional LSTM and GRU built from ordinary tensor ops, so
//! one graph node per gate per timestep. Stack them in a [`Sequential`](super::Sequential).

use crate::tensor::Tensor;

use super::module::Module;
use super::param::Param;

/// Weights shared by both cells.
struct Gates {
    input_weight: Param,
    hidden_weight: Param,
    input_bias: Param,
    hidden_bias: Param,
}

impl Gates {
    /// `count` gates of `hidden` units each, over an `input`-wide input.
    fn new(input: usize, hidden: usize, count: usize) -> Gates {
        let width = count * hidden;
        Gates {
            input_weight: Param::new(Tensor::xavier_uniform(&[width, input], input, hidden)),
            hidden_weight: Param::new(Tensor::xavier_uniform(&[width, hidden], hidden, hidden)),
            input_bias: Param::new(Tensor::zeros(&[width])),
            hidden_bias: Param::new(Tensor::zeros(&[width])),
        }
    }

    /// `x · W_ihᵀ + b_ih` and `h · W_hhᵀ + b_hh`, each `[batch, count·hidden]`.
    fn project(&self, x: &Tensor, h: &Tensor) -> (Tensor, Tensor) {
        let from_input = x
            .matmul_nt(&self.input_weight.tensor())
            .add(&self.input_bias.tensor());
        let from_hidden = h
            .matmul_nt(&self.hidden_weight.tensor())
            .add(&self.hidden_bias.tensor());
        (from_input, from_hidden)
    }

    fn named_parameters(&self) -> Vec<(String, Param)> {
        vec![
            ("input_weight".into(), self.input_weight.clone()),
            ("hidden_weight".into(), self.hidden_weight.clone()),
            ("input_bias".into(), self.input_bias.clone()),
            ("hidden_bias".into(), self.hidden_bias.clone()),
        ]
    }
}

/// Long short-term memory.
pub struct LSTM {
    gates: Gates,
    input_size: usize,
    hidden_size: usize,
}

impl LSTM {
    pub fn new(input_size: usize, hidden_size: usize) -> LSTM {
        let gates = Gates::new(input_size, hidden_size, 4);
        // forget gate starts open, at zero bias it halves the cell state every step
        let mut bias = gates.input_bias.value().to_vec();
        bias[hidden_size..2 * hidden_size].fill(1.0);
        gates
            .input_bias
            .set_value(Tensor::from_vec(bias, &[4 * hidden_size]));

        LSTM {
            gates,
            input_size,
            hidden_size,
        }
    }

    /// Run over `[batch, sequence, input_size]`, returning outputs and the final `(h, c)`.
    pub fn run(
        &self,
        input: &Tensor,
        initial: Option<(&Tensor, &Tensor)>,
    ) -> (Tensor, Tensor, Tensor) {
        let (batch, steps) = check_sequence(input, self.input_size, "LSTM");
        let zeros = || Tensor::zeros(&[batch, self.hidden_size]).to(input.device());
        let (mut h, mut c) = match initial {
            Some((h0, c0)) => (h0.clone(), c0.clone()),
            None => (zeros(), zeros()),
        };

        let mut outputs = Vec::with_capacity(steps);
        for t in 0..steps {
            let x = step_input(input, t, batch, self.input_size);
            let (from_input, from_hidden) = self.gates.project(&x, &h);
            let gates = from_input.add(&from_hidden);

            let slice = |index: usize| gates.narrow(1, index * self.hidden_size, self.hidden_size);
            let input_gate = slice(0).sigmoid();
            let forget_gate = slice(1).sigmoid();
            let candidate = slice(2).tanh();
            let output_gate = slice(3).sigmoid();

            c = forget_gate.mul(&c).add(&input_gate.mul(&candidate));
            h = output_gate.mul(&c.tanh());
            outputs.push(h.clone());
        }

        (stack_steps(&outputs), h, c)
    }
}

impl Module for LSTM {
    fn forward(&self, input: &Tensor) -> Tensor {
        self.run(input, None).0
    }

    fn named_parameters(&self) -> Vec<(String, Param)> {
        self.gates.named_parameters()
    }
}

/// Gated recurrent unit, LSTM's gating with no separate cell state.
pub struct GRU {
    gates: Gates,
    input_size: usize,
    hidden_size: usize,
}

impl GRU {
    pub fn new(input_size: usize, hidden_size: usize) -> GRU {
        GRU {
            gates: Gates::new(input_size, hidden_size, 3),
            input_size,
            hidden_size,
        }
    }

    /// Run over `[batch, sequence, input_size]`, returning outputs and the final hidden state.
    pub fn run(&self, input: &Tensor, initial: Option<&Tensor>) -> (Tensor, Tensor) {
        let (batch, steps) = check_sequence(input, self.input_size, "GRU");
        let mut h = initial
            .cloned()
            .unwrap_or_else(|| Tensor::zeros(&[batch, self.hidden_size]).to(input.device()));

        let mut outputs = Vec::with_capacity(steps);
        for t in 0..steps {
            let x = step_input(input, t, batch, self.input_size);
            let (from_input, from_hidden) = self.gates.project(&x, &h);

            let part = |source: &Tensor, index: usize| {
                source.narrow(1, index * self.hidden_size, self.hidden_size)
            };
            let reset = part(&from_input, 0).add(&part(&from_hidden, 0)).sigmoid();
            let update = part(&from_input, 1).add(&part(&from_hidden, 1)).sigmoid();
            // reset scales the hidden contribution only
            let candidate = part(&from_input, 2)
                .add(&reset.mul(&part(&from_hidden, 2)))
                .tanh();

            let keep = update.neg().add_scalar(1.0);
            h = keep.mul(&candidate).add(&update.mul(&h));
            outputs.push(h.clone());
        }

        (stack_steps(&outputs), h)
    }
}

impl Module for GRU {
    fn forward(&self, input: &Tensor) -> Tensor {
        self.run(input, None).0
    }

    fn named_parameters(&self) -> Vec<(String, Param)> {
        self.gates.named_parameters()
    }
}

fn check_sequence(input: &Tensor, input_size: usize, layer: &str) -> (usize, usize) {
    assert_eq!(
        input.ndim(),
        3,
        "{layer} expects [batch, sequence, features], got {:?}",
        input.shape()
    );
    assert_eq!(
        input.dim(2),
        input_size,
        "{layer} expects {input_size} features, got {:?}",
        input.shape()
    );
    (input.dim(0), input.dim(1))
}

/// Timestep `t` as `[batch, features]`.
fn step_input(input: &Tensor, t: usize, batch: usize, features: usize) -> Tensor {
    input
        .narrow(1, t, 1)
        .reshape(&[batch as i64, features as i64])
}

/// Per-step `[batch, hidden]` outputs into `[batch, sequence, hidden]`.
fn stack_steps(steps: &[Tensor]) -> Tensor {
    Tensor::stack(&steps.iter().collect::<Vec<_>>(), 1)
}