tch 0.24.0

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

use tch::{Device, Kind, Tensor};
use tch::{IndexOp, NewAxis};

mod test_utils;
use test_utils::*;

#[test]
fn integer_index() {
    let opt = (Kind::Float, Device::Cpu);

    let tensor = Tensor::arange_start(0, 2 * 3, opt).view([2, 3]);
    let result = tensor.i(1);
    assert_eq!(result.size(), &[3]);
    assert_eq!(vec_i64_from(&result), &[3, 4, 5]);

    let tensor = Tensor::arange_start(0, 2 * 3, opt).view([2, 3]);
    let result = tensor.i((.., 2));
    assert_eq!(result.size(), &[2]);
    assert_eq!(vec_i64_from(&result), &[2, 5]);

    let result = tensor.i((.., -2));
    assert_eq!(result.size(), &[2]);
    assert_eq!(vec_i64_from(&result), &[1, 4]);
}

#[test]
fn range_index() {
    let opt = (Kind::Float, Device::Cpu);

    // Range
    let tensor = Tensor::arange_start(0, 4 * 3, opt).view([4, 3]);
    let result = tensor.i(1..3);
    assert_eq!(result.size(), &[2, 3]);
    assert_eq!(vec_i64_from(&result), &[3, 4, 5, 6, 7, 8]);

    // RangeFull
    let tensor = Tensor::arange_start(0, 2 * 3, opt).view([2, 3]);
    let result = tensor.i(..);
    assert_eq!(result.size(), &[2, 3]);
    assert_eq!(vec_i64_from(&result), &[0, 1, 2, 3, 4, 5]);

    // RangeFrom
    let tensor = Tensor::arange_start(0, 4 * 3, opt).view([4, 3]);
    let result = tensor.i(2..);
    assert_eq!(result.size(), &[2, 3]);
    assert_eq!(vec_i64_from(&result), &[6, 7, 8, 9, 10, 11]);

    // RangeTo
    let tensor = Tensor::arange_start(0, 4 * 3, opt).view([4, 3]);
    let result = tensor.i(..2);
    assert_eq!(result.size(), &[2, 3]);
    assert_eq!(vec_i64_from(&result), &[0, 1, 2, 3, 4, 5]);

    // RangeInclusive
    let tensor = Tensor::arange_start(0, 4 * 3, opt).view([4, 3]);
    let result = tensor.i(1..=2);
    assert_eq!(result.size(), &[2, 3]);
    assert_eq!(vec_i64_from(&result), &[3, 4, 5, 6, 7, 8]);

    // RangeTo
    let tensor = Tensor::arange_start(0, 4 * 3, opt).view([4, 3]);
    let result = tensor.i(..1);
    assert_eq!(result.size(), &[1, 3]);
    assert_eq!(vec_i64_from(&result), &[0, 1, 2]);

    // RangeToInclusive
    let tensor = Tensor::arange_start(0, 4 * 3, opt).view([4, 3]);
    let result = tensor.i(..=1);
    assert_eq!(result.size(), &[2, 3]);
    assert_eq!(vec_i64_from(&result), &[0, 1, 2, 3, 4, 5]);

    // Negative Cases
    //
    // Rust considers both `0..0` and `1..0` to be valid, empty ranges.
    let tensor = Tensor::arange_start(0, 4 * 3, opt).view([4, 3]);
    let result = tensor.i(0..0);
    assert_eq!(result.size(), &[0, 3]);
    assert_eq!(vec_i64_from(&result), &[] as &[i64]);

    // Clippy doesn't like explicitly reversed ranges, so let's pretend the bounds were computed.
    let (start, end) = (1, 0);
    let tensor = Tensor::arange_start(0, 4 * 3, opt).view([4, 3]);
    let result = tensor.i(start..end);
    assert_eq!(result.size(), &[0, 3]);
    assert_eq!(vec_i64_from(&result), &[] as &[i64]);
}

#[test]
fn slice_index() {
    let opt = (Kind::Float, Device::Cpu);

    let tensor = Tensor::arange_start(0, 6 * 2, opt).view([6, 2]);
    let index: &[_] = &[1, 3, 5];
    let result = tensor.i(index);
    assert_eq!(result.size(), &[3, 2]);
    assert_eq!(vec_i64_from(&result), &[2, 3, 6, 7, 10, 11]);

    let tensor = Tensor::arange_start(0, 3 * 4, opt).view([3, 4]);
    let index: &[_] = &[3, 0];
    let result = tensor.i((.., index));
    assert_eq!(result.size(), &[3, 2]);
    assert_eq!(vec_i64_from(&result), &[3, 0, 7, 4, 11, 8]);
}

#[test]
fn new_index() {
    let opt = (Kind::Float, Device::Cpu);

    let tensor = Tensor::arange_start(0, 2 * 3, opt).view([2, 3]);
    let result = tensor.i((NewAxis,));
    assert_eq!(result.size(), &[1, 2, 3]);
    assert_eq!(vec_i64_from(&result), &[0, 1, 2, 3, 4, 5]);

    let tensor = Tensor::arange_start(0, 2 * 3, opt).view([2, 3]);
    let result = tensor.i((.., NewAxis));
    assert_eq!(result.size(), &[2, 1, 3]);
    assert_eq!(vec_i64_from(&result), &[0, 1, 2, 3, 4, 5]);

    let tensor = Tensor::arange_start(0, 2 * 3, opt).view([2, 3]);
    let result = tensor.i((.., .., NewAxis));
    assert_eq!(result.size(), &[2, 3, 1]);
    assert_eq!(vec_i64_from(&result), &[0, 1, 2, 3, 4, 5]);
}

#[cfg(target_os = "linux")]
#[test]
fn complex_index() {
    let opt = (Kind::Float, Device::Cpu);

    let tensor = Tensor::arange_start(0, 2 * 3 * 5 * 7, opt).view([2, 3, 5, 7]);
    let result = tensor.i((1, 1..2, vec![2, 3, 0].as_slice(), NewAxis, 3..));
    assert_eq!(result.size(), &[1, 3, 1, 4]);
    assert_eq!(
        vec_i64_from(&result),
        &[157, 158, 159, 160, 164, 165, 166, 167, 143, 144, 145, 146]
    );
}

#[test]
fn index_3d() {
    let values: Vec<i64> = (0..24).collect();
    let tensor = tch::Tensor::from_slice(&values).view((2, 3, 4));
    assert_eq!(vec_i64_from(&tensor.i((0, 0, 0))), &[0]);
    assert_eq!(vec_i64_from(&tensor.i((1, 0, 0))), &[12]);
    assert_eq!(vec_i64_from(&tensor.i((0..2, 0, 0))), &[0, 12]);
}

#[test]
fn tensor_index() {
    let t = Tensor::arange(6, (Kind::Int64, Device::Cpu)).view((2, 3));
    let rows_select = Tensor::from_slice(&[0i64, 1, 0]);
    let column_select = Tensor::from_slice(&[1i64, 2, 2]);

    let selected = t.index(&[Some(rows_select), Some(column_select)]);
    assert_eq!(selected.size(), &[3]);
    assert_eq!(vec_i64_from(&selected), &[1, 5, 2]);
}

#[test]
fn tensor_index2() {
    let t = Tensor::arange(400, (Kind::Int64, Device::Cpu)).view((2, 2, 10, 10));

    let selected = t.index(&[
        Some(Tensor::from(0i64)),
        Some(Tensor::from(1i64)),
        Some(Tensor::from_slice(&[1i64, 2, 3])),
        Some(Tensor::from_slice(&[5i64, 6, 7])),
    ]);
    assert_eq!(selected.size(), &[3]);
    // 115 = 0 * 200 + 1 * 100 + 1 * 10 + 5
    // 126 = 0 * 200 + 1 * 100 + 2 * 10 + 6
    // 137 = 0 * 200 + 1 * 100 + 3 * 10 + 7
    assert_eq!(vec_i64_from(&selected), &[115, 126, 137]);
}

#[test]
fn tensor_multi_index() {
    let t = Tensor::arange(6, (Kind::Int64, Device::Cpu)).view((2, 3));

    let select_1 = Tensor::from_slice(&[0i64, 1, 0]);
    let select_2 = Tensor::from_slice(&[1i64, 0, 0]);
    let select_final = Tensor::stack(&[select_1, select_2], 0);
    assert_eq!(select_final.size(), &[2, 3]);

    let selected = t.index(&[Some(select_final)]); // index only rows
    assert_eq!(selected.size(), &[2, 3, 3]);
    assert_eq!(vec_i64_from(&selected), &[0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 0, 1, 2]);
    // after flattening
}

#[test]
fn tensor_put() {
    let t = Tensor::arange(6, (Kind::Int64, Device::Cpu)).view((2, 3));
    let rows_select = Tensor::from_slice(&[0i64, 1, 0]);
    let column_select = Tensor::from_slice(&[1i64, 2, 2]);
    let values = Tensor::from_slice(&[10i64, 12, 24]);

    let updated = t.index_put(&[Some(rows_select), Some(column_select)], &values, false);
    assert_eq!(vec_i64_from(&updated), &[0i64, 10, 24, 3, 4, 12]); // after flattening
}

#[test]
fn indexing_doc() {
    let tensor = Tensor::from_slice(&[1, 2, 3, 4, 5, 6]).view((2, 3));
    let t = tensor.i(1);
    assert_eq!(vec_i64_from(&t), [4, 5, 6]);
    let t = tensor.i((.., -2));
    assert_eq!(vec_i64_from(&t), [2, 5]);

    let tensor = Tensor::from_slice(&[1, 2, 3, 4, 5, 6]).view((2, 3));
    let t = tensor.i((.., 1..));
    assert_eq!(t.size(), [2, 2]);
    assert_eq!(vec_i64_from(&t.contiguous().view(-1)), [2, 3, 5, 6]);
    let t = tensor.i((..1, ..));
    assert_eq!(t.size(), [1, 3]);
    assert_eq!(vec_i64_from(&t.contiguous().view(-1)), [1, 2, 3]);
    let t = tensor.i((.., 1..2));
    assert_eq!(t.size(), [2, 1]);
    assert_eq!(vec_i64_from(&t.contiguous().view(-1)), [2, 5]);
    let t = tensor.i((.., 1..=2));
    assert_eq!(t.size(), [2, 2]);
    assert_eq!(vec_i64_from(&t.contiguous().view(-1)), [2, 3, 5, 6]);

    let tensor = Tensor::from_slice(&[1, 2, 3, 4, 5, 6]).view((2, 3));
    let t = tensor.i((NewAxis,));
    assert_eq!(t.size(), &[1, 2, 3]);
    let t = tensor.i((.., .., NewAxis));
    assert_eq!(t.size(), &[2, 3, 1]);
}