tch 0.4.0

Rust wrappers for the PyTorch C++ api (libtorch).
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
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273

//! Recurrent Neural Networks
use crate::{Device, Kind, Tensor};

/// Trait for Recurrent Neural Networks.
pub trait RNN {
    type State;

    /// A zero state from which the recurrent network is usually initialized.
    fn zero_state(&self, batch_dim: i64) -> Self::State;

    /// Applies a single step of the recurrent network.
    ///
    /// The input should have dimensions [batch_size, features].
    fn step(&self, input: &Tensor, state: &Self::State) -> Self::State;

    /// Applies multiple steps of the recurrent network.
    ///
    /// The input should have dimensions [batch_size, seq_len, features].
    /// The initial state is the result of applying zero_state.
    fn seq(&self, input: &Tensor) -> (Tensor, Self::State) {
        let batch_dim = input.size()[0];
        let state = self.zero_state(batch_dim);
        self.seq_init(input, &state)
    }

    /// Applies multiple steps of the recurrent network.
    ///
    /// The input should have dimensions [batch_size, seq_len, features].
    fn seq_init(&self, input: &Tensor, state: &Self::State) -> (Tensor, Self::State);
}

/// The state for a LSTM network, this contains two tensors.
#[derive(Debug)]
pub struct LSTMState(pub (Tensor, Tensor));

impl LSTMState {
    /// The hidden state vector, which is also the output of the LSTM.
    pub fn h(&self) -> Tensor {
        (self.0).0.shallow_clone()
    }

    /// The cell state vector.
    pub fn c(&self) -> Tensor {
        (self.0).1.shallow_clone()
    }
}

// The GRU and LSTM layers share the same config.
/// Configuration for the GRU and LSTM layers.
#[derive(Debug, Clone, Copy)]
pub struct RNNConfig {
    pub has_biases: bool,
    pub num_layers: i64,
    pub dropout: f64,
    pub train: bool,
    pub bidirectional: bool,
    pub batch_first: bool,
}

impl Default for RNNConfig {
    fn default() -> Self {
        RNNConfig {
            has_biases: true,
            num_layers: 1,
            dropout: 0.,
            train: true,
            bidirectional: false,
            batch_first: true,
        }
    }
}

/// A Long Short-Term Memory (LSTM) layer.
///
/// https://en.wikipedia.org/wiki/Long_short-term_memory
#[derive(Debug)]
pub struct LSTM {
    flat_weights: Vec<Tensor>,
    hidden_dim: i64,
    config: RNNConfig,
    device: Device,
}

/// Creates a LSTM layer.
pub fn lstm(vs: &super::var_store::Path, in_dim: i64, hidden_dim: i64, c: RNNConfig) -> LSTM {
    let num_directions = if c.bidirectional { 2 } else { 1 };
    let gate_dim = 4 * hidden_dim;
    let mut flat_weights = vec![];
    for layer_idx in 0..c.num_layers {
        for direction_idx in 0..num_directions {
            let in_dim = if layer_idx == 0 {
                in_dim
            } else {
                hidden_dim * num_directions
            };
            let suffix = if direction_idx == 1 { "_reverse" } else { "" };
            let w_ih = vs.kaiming_uniform(
                &format!("weight_ih_l{}{}", layer_idx, suffix),
                &[gate_dim, in_dim],
            );
            let w_hh = vs.kaiming_uniform(
                &format!("weight_hh_l{}{}", layer_idx, suffix),
                &[gate_dim, hidden_dim],
            );
            flat_weights.push(w_ih);
            flat_weights.push(w_hh);
            if c.has_biases {
                let b_ih = vs.zeros(&format!("bias_ih_l{}{}", layer_idx, suffix), &[gate_dim]);
                let b_hh = vs.zeros(&format!("bias_hh_l{}{}", layer_idx, suffix), &[gate_dim]);
                flat_weights.push(b_ih);
                flat_weights.push(b_hh);
            }
        }
    }
    if vs.device().is_cuda() && crate::Cuda::cudnn_is_available() {
        let _ = Tensor::internal_cudnn_rnn_flatten_weight(
            &flat_weights,
            4,
            in_dim,
            2, /* 2 for LSTM see rnn.cpp in pytorch */
            hidden_dim,
            0, /* disables projections */
            c.num_layers,
            c.batch_first,
            c.bidirectional,
        );
    }
    LSTM {
        flat_weights,
        hidden_dim,
        config: c,
        device: vs.device(),
    }
}

impl RNN for LSTM {
    type State = LSTMState;

    fn zero_state(&self, batch_dim: i64) -> LSTMState {
        let num_directions = if self.config.bidirectional { 2 } else { 1 };
        let layer_dim = self.config.num_layers * num_directions;
        let shape = [layer_dim, batch_dim, self.hidden_dim];
        let zeros = Tensor::zeros(&shape, (Kind::Float, self.device));
        LSTMState((zeros.shallow_clone(), zeros.shallow_clone()))
    }

    fn step(&self, input: &Tensor, in_state: &LSTMState) -> LSTMState {
        let input = input.unsqueeze(1);
        let (_output, state) = self.seq_init(&input, in_state);
        state
    }

    fn seq_init(&self, input: &Tensor, in_state: &LSTMState) -> (Tensor, LSTMState) {
        let LSTMState((h, c)) = in_state;
        let flat_weights = self.flat_weights.iter().collect::<Vec<_>>();
        let (output, h, c) = input.lstm(
            &[h, c],
            &flat_weights,
            self.config.has_biases,
            self.config.num_layers,
            self.config.dropout,
            self.config.train,
            self.config.bidirectional,
            self.config.batch_first,
        );
        (output, LSTMState((h, c)))
    }
}

/// A GRU state, this contains a single tensor.
#[derive(Debug)]
pub struct GRUState(pub Tensor);

impl GRUState {
    pub fn value(&self) -> Tensor {
        self.0.shallow_clone()
    }
}

/// A Gated Recurrent Unit (GRU) layer.
///
/// https://en.wikipedia.org/wiki/Gated_recurrent_unit
#[derive(Debug)]
pub struct GRU {
    flat_weights: Vec<Tensor>,
    hidden_dim: i64,
    config: RNNConfig,
    device: Device,
}

/// Creates a new GRU layer.
pub fn gru(vs: &super::var_store::Path, in_dim: i64, hidden_dim: i64, c: RNNConfig) -> GRU {
    let num_directions = if c.bidirectional { 2 } else { 1 };
    let gate_dim = 3 * hidden_dim;
    let mut flat_weights = vec![];
    for layer_idx in 0..c.num_layers {
        for direction_idx in 0..num_directions {
            let in_dim = if layer_idx == 0 {
                in_dim
            } else {
                hidden_dim * num_directions
            };
            let suffix = if direction_idx == 1 { "_reverse" } else { "" };
            let w_ih = vs.kaiming_uniform(
                &format!("weight_ih_l{}{}", layer_idx, suffix),
                &[gate_dim, in_dim],
            );
            let w_hh = vs.kaiming_uniform(
                &format!("weight_hh_l{}{}", layer_idx, suffix),
                &[gate_dim, hidden_dim],
            );
            flat_weights.push(w_ih);
            flat_weights.push(w_hh);
            if c.has_biases {
                let b_ih = vs.zeros(&format!("bias_ih_l{}{}", layer_idx, suffix), &[gate_dim]);
                let b_hh = vs.zeros(&format!("bias_hh_l{}{}", layer_idx, suffix), &[gate_dim]);
                flat_weights.push(b_ih);
                flat_weights.push(b_hh);
            }
        }
    }
    if vs.device().is_cuda() && crate::Cuda::cudnn_is_available() {
        let _ = Tensor::internal_cudnn_rnn_flatten_weight(
            &flat_weights,
            4,
            in_dim,
            3, /* 3 for GRU see rnn.cpp in pytorch */
            hidden_dim,
            0, /* disables projections */
            c.num_layers,
            c.batch_first,
            c.bidirectional,
        );
    }
    GRU {
        flat_weights,
        hidden_dim,
        config: c,
        device: vs.device(),
    }
}

impl RNN for GRU {
    type State = GRUState;

    fn zero_state(&self, batch_dim: i64) -> GRUState {
        let num_directions = if self.config.bidirectional { 2 } else { 1 };
        let layer_dim = self.config.num_layers * num_directions;
        let shape = [layer_dim, batch_dim, self.hidden_dim];
        GRUState(Tensor::zeros(&shape, (Kind::Float, self.device)))
    }

    fn step(&self, input: &Tensor, in_state: &GRUState) -> GRUState {
        let input = input.unsqueeze(1);
        let (_output, state) = self.seq_init(&input, in_state);
        state
    }

    fn seq_init(&self, input: &Tensor, in_state: &GRUState) -> (Tensor, GRUState) {
        let GRUState(h) = in_state;
        let (output, h) = input.gru(
            h,
            &self.flat_weights,
            self.config.has_biases,
            self.config.num_layers,
            self.config.dropout,
            self.config.train,
            self.config.bidirectional,
            self.config.batch_first,
        );
        (output, GRUState(h))
    }
}