tch 0.1.6

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

use tch::nn::layer_norm;
use tch::nn::{Module, OptimizerConfig};
use tch::{kind, nn, Device, Kind, Reduction, Tensor};

#[test]
fn optimizer_test() {
    tch::manual_seed(42);
    // Create some linear data.
    let xs = Tensor::of_slice(&(1..15).collect::<Vec<_>>())
        .to_kind(Kind::Float)
        .view([-1, 1]);
    let ys = &xs * 0.42 + 1.337;

    // Fit a linear model (with deterministic initialization) on the data.
    let vs = nn::VarStore::new(Device::Cpu);
    let mut opt = nn::Sgd::default().build(&vs, 1e-2).unwrap();
    let cfg = nn::LinearConfig {
        ws_init: nn::Init::Const(0.),
        bs_init: Some(nn::Init::Const(0.)),
    };
    let mut linear = nn::linear(vs.root(), 1, 1, cfg);

    let loss = xs.apply(&linear).mse_loss(&ys, Reduction::Mean);
    let initial_loss = f64::from(&loss);
    assert!(initial_loss > 1.0, "initial loss {}", initial_loss);

    // Optimization loop.
    for _idx in 1..50 {
        let loss = xs.apply(&linear).mse_loss(&ys, Reduction::Mean);
        opt.backward_step(&loss);
    }
    let loss = xs.apply(&linear).mse_loss(&ys, Reduction::Mean);
    let final_loss = f64::from(loss);
    assert!(final_loss < 0.25, "final loss {}", final_loss);

    // Reset the weights to their initial values.
    tch::no_grad(|| {
        linear.ws.init(nn::Init::Const(0.));
        linear.bs.init(nn::Init::Const(0.));
    });
    let initial_loss2 = f64::from(xs.apply(&linear).mse_loss(&ys, Reduction::Mean));
    assert_eq!(initial_loss, initial_loss2)
}

fn my_module(p: nn::Path, dim: i64) -> impl nn::Module {
    let x1 = p.zeros("x1", &[dim]);
    let x2 = p.zeros("x2", &[dim]);
    nn::func(move |xs| xs * &x1 + xs.exp() * &x2)
}

#[test]
fn gradient_descent_test() {
    let vs = nn::VarStore::new(Device::Cpu);
    let my_module = my_module(vs.root(), 7);
    let mut opt = nn::Sgd::default().build(&vs, 1e-2).unwrap();
    for _idx in 1..50 {
        let xs = Tensor::zeros(&[7], kind::FLOAT_CPU);
        let ys = Tensor::zeros(&[7], kind::FLOAT_CPU);
        let loss = (my_module.forward(&xs) - ys).pow(2).sum(Kind::Float);
        opt.backward_step(&loss);
    }
}

#[test]
fn bn_test() {
    let opts = (tch::Kind::Float, tch::Device::Cpu);
    let vs = nn::VarStore::new(tch::Device::Cpu);
    let bn = nn::batch_norm1d(vs.root(), 40, Default::default());
    let x = Tensor::randn(&[10, 40], opts);
    let _y = x.apply_t(&bn, true);
}

#[test]
fn layer_norm_test() {
    let opts = (tch::Kind::Float, tch::Device::Cpu);
    let vs = nn::VarStore::new(tch::Device::Cpu);
    let ln = layer_norm(vs.root(), vec![5, 10, 10], Default::default());
    let x = Tensor::randn(&[20, 5, 10, 10], opts);
    let _y = x.apply(&ln);
}

#[test]
fn layer_norm_parameters_test() {
    tch::manual_seed(42);
    // Create some linear data.
    let xs = Tensor::of_slice(&[42.0, 42.0, 42.0, 24.0])
        .to_kind(Kind::Float)
        .view([-1, 2]);
    let ys = &xs * 0.42 + 1.337;

    // Fit a layer normalization layer (with deterministic initialization) on the data.
    let vs = nn::VarStore::new(Device::Cpu);
    let mut opt = nn::Sgd::default().build(&vs, 1.0).unwrap();
    let mut ln = layer_norm(vs.root(), vec![2], Default::default());

    let loss = xs.apply(&ln).mse_loss(&ys, Reduction::Mean);
    let initial_loss = f64::from(&loss);
    assert!(initial_loss > 1.0, "initial loss {}", initial_loss);

    // Optimization loop.
    for _idx in 1..50 {
        let loss = xs.apply(&ln).mse_loss(&ys, Reduction::Mean);
        opt.backward_step(&loss);
    }
    let loss = xs.apply(&ln).mse_loss(&ys, Reduction::Mean);
    let final_loss = f64::from(loss);
    assert!(final_loss < 0.25, "final loss {:?}", final_loss);

    //     Reset the weights to their initial values.
    tch::no_grad(|| {
        if let Some(ws) = &mut ln.ws {
            ws.init(nn::Init::Const(1.));
        }
        if let Some(bs) = &mut ln.bs {
            bs.init(nn::Init::Const(0.));
        }
    });
    let initial_loss2 = f64::from(xs.apply(&ln).mse_loss(&ys, Reduction::Mean));
    assert_eq!(initial_loss, initial_loss2)
}

fn gru_test(rnn_config: nn::RNNConfig) {
    use nn::RNN;
    let batch_dim = 5;
    let seq_len = 3;
    let input_dim = 2;
    let output_dim = 4;
    let vs = nn::VarStore::new(tch::Device::Cpu);
    let gru = nn::gru(&vs.root(), input_dim, output_dim, rnn_config);

    let num_directions = if rnn_config.bidirectional { 2 } else { 1 };
    let layer_dim = rnn_config.num_layers * num_directions;
    //
    // step test
    let input = Tensor::randn(&[batch_dim, input_dim], kind::FLOAT_CPU);
    let nn::GRUState(output) = gru.step(&input, &gru.zero_state(batch_dim));
    assert_eq!(output.size(), [layer_dim, batch_dim, output_dim]);

    // seq test
    let input = Tensor::randn(&[batch_dim, seq_len, input_dim], kind::FLOAT_CPU);
    let (output, _) = gru.seq(&input);
    assert_eq!(
        output.size(),
        [batch_dim, seq_len, output_dim * num_directions]
    );
}

#[test]
fn gru() {
    gru_test(Default::default());
    gru_test(nn::RNNConfig {
        bidirectional: true,
        ..Default::default()
    });
    gru_test(nn::RNNConfig {
        num_layers: 2,
        ..Default::default()
    });
    gru_test(nn::RNNConfig {
        num_layers: 2,
        bidirectional: true,
        ..Default::default()
    });
}

fn lstm_test(rnn_config: nn::RNNConfig) {
    use nn::RNN;
    let batch_dim = 5;
    let seq_len = 3;
    let input_dim = 2;
    let output_dim = 4;
    let vs = nn::VarStore::new(tch::Device::Cpu);
    let lstm = nn::lstm(&vs.root(), input_dim, output_dim, rnn_config);

    let num_directions = if rnn_config.bidirectional { 2 } else { 1 };
    let layer_dim = rnn_config.num_layers * num_directions;
    //
    // step test
    let input = Tensor::randn(&[batch_dim, input_dim], kind::FLOAT_CPU);
    let nn::LSTMState((h, c)) = lstm.step(&input, &lstm.zero_state(batch_dim));
    assert_eq!(h.size(), [layer_dim, batch_dim, output_dim]);
    assert_eq!(c.size(), [layer_dim, batch_dim, output_dim]);

    // seq test
    let input = Tensor::randn(&[batch_dim, seq_len, input_dim], kind::FLOAT_CPU);
    let (output, _) = lstm.seq(&input);
    assert_eq!(
        output.size(),
        [batch_dim, seq_len, output_dim * num_directions]
    );
}

#[test]
fn lstm() {
    lstm_test(Default::default());
    lstm_test(nn::RNNConfig {
        bidirectional: true,
        ..Default::default()
    });
    lstm_test(nn::RNNConfig {
        num_layers: 2,
        ..Default::default()
    });
    lstm_test(nn::RNNConfig {
        num_layers: 2,
        bidirectional: true,
        ..Default::default()
    });
}

fn embedding_test(embedding_config: nn::EmbeddingConfig) {
    let batch_dim = 5;
    let seq_len = 7;
    let input_dim = 10;
    let output_dim = 4;
    let vs = nn::VarStore::new(tch::Device::Cpu);
    let embeddings = nn::embedding(&vs.root(), input_dim, output_dim, embedding_config);

    // forward test
    let input = Tensor::randint(10, &[batch_dim, seq_len], kind::INT64_CPU);
    let output = embeddings.forward(&input);
    assert_eq!(output.size(), [batch_dim, seq_len, output_dim]);

    // padding test
    let padding_idx = if embedding_config.padding_idx < 0 {
        input_dim + embedding_config.padding_idx
    } else {
        embedding_config.padding_idx
    };
    let input = Tensor::of_slice(&vec![padding_idx; 1]);
    let output = embeddings.forward(&input);
    assert_eq!(output.size(), [1, output_dim]);
    assert_eq!(output.get(0), embeddings.ws.get(padding_idx));
}

#[test]
fn embedding() {
    embedding_test(Default::default());
    embedding_test(nn::EmbeddingConfig {
        padding_idx: -1,
        ..Default::default()
    });
    embedding_test(nn::EmbeddingConfig {
        padding_idx: 0,
        ..Default::default()
    });
}