candle-einops 0.2.0

Compile-time einops and einsum tensor operations for Candle
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

use candle_core::{Device, Result, Shape, Tensor, Var};
use candle_einops::einops;

fn flat_f32(tensor: &Tensor) -> Result<Vec<f32>> {
    tensor.flatten_all()?.to_vec1::<f32>()
}

fn assert_close(actual: &[f32], expected: &[f32], context: &str) {
    assert_eq!(actual.len(), expected.len(), "{context}: length");
    for (index, (&actual, &expected)) in actual.iter().zip(expected).enumerate() {
        let tolerance = 1e-5 * (1. + expected.abs());
        assert!(
            (actual - expected).abs() <= tolerance,
            "{context}[{index}]: actual={actual}, expected={expected}"
        );
    }
}

fn compare_forward_and_gradient(
    data: &[f32],
    shape: &[usize],
    macro_operation: impl FnOnce(&Tensor) -> Result<Tensor>,
    candle_operation: impl FnOnce(&Tensor) -> Result<Tensor>,
) -> Result<()> {
    let device = Device::Cpu;
    let macro_input = Var::from_vec(data.to_vec(), Shape::from_dims(shape), &device)?;
    let candle_input = Var::from_vec(data.to_vec(), Shape::from_dims(shape), &device)?;

    let macro_output = macro_operation(macro_input.as_tensor())?;
    let candle_output = candle_operation(candle_input.as_tensor())?;
    assert_eq!(macro_output.dims(), candle_output.dims());
    assert_close(
        &flat_f32(&macro_output)?,
        &flat_f32(&candle_output)?,
        "forward",
    );

    let weight_values = (1..=macro_output.elem_count())
        .map(|value| value as f32 / 7.)
        .collect::<Vec<_>>();
    let weights = Tensor::from_vec(
        weight_values,
        Shape::from_dims(macro_output.dims()),
        &device,
    )?;
    let macro_loss = macro_output.mul(&weights)?.sum_all()?;
    let candle_loss = candle_output.mul(&weights)?.sum_all()?;
    let macro_gradients = macro_loss.backward()?;
    let candle_gradients = candle_loss.backward()?;
    let Some(macro_gradient) = macro_gradients.get(macro_input.as_tensor()) else {
        candle_core::bail!("einops path did not produce an input gradient")
    };
    let Some(candle_gradient) = candle_gradients.get(candle_input.as_tensor()) else {
        candle_core::bail!("Candle oracle did not produce an input gradient")
    };
    assert_eq!(macro_gradient.dims(), candle_gradient.dims());
    assert_close(
        &flat_f32(macro_gradient)?,
        &flat_f32(candle_gradient)?,
        "gradient",
    );
    Ok(())
}

#[test]
fn transpose_gradient_matches_candle() -> Result<()> {
    let data = (1..=24).map(|value| value as f32).collect::<Vec<_>>();
    compare_forward_and_gradient(
        &data,
        &[2, 3, 4],
        |input| einops!("a b c -> c a b", input),
        |input| input.permute((2, 0, 1)),
    )
}

#[test]
fn composition_and_decomposition_gradients_match_candle() -> Result<()> {
    let data = (1..=24).map(|value| value as f32).collect::<Vec<_>>();
    compare_forward_and_gradient(
        &data,
        &[2, 3, 4],
        |input| einops!("a b c -> (b a) c", input),
        |input| input.permute((1, 0, 2))?.reshape((6, 4)),
    )?;

    compare_forward_and_gradient(
        &data,
        &[6, 4],
        |input| einops!("(a b:2) c -> b a c", input),
        |input| input.reshape((3, 2, 4))?.permute((1, 0, 2)),
    )
}

#[test]
fn repeat_gradient_matches_candle() -> Result<()> {
    let data = (1..=6).map(|value| value as f32).collect::<Vec<_>>();
    compare_forward_and_gradient(
        &data,
        &[2, 3],
        |input| einops!("a b -> copies:3 b a", input),
        |input| input.permute((1, 0))?.unsqueeze(0)?.repeat((3, 1, 1)),
    )
}

#[test]
fn sum_and_mean_gradients_match_candle() -> Result<()> {
    let data = (1..=24).map(|value| value as f32).collect::<Vec<_>>();
    compare_forward_and_gradient(
        &data,
        &[2, 3, 4],
        |input| einops!("a sum(b) c -> c a", input),
        |input| input.sum(1)?.permute((1, 0)),
    )?;

    compare_forward_and_gradient(
        &data,
        &[2, 3, 4],
        |input| einops!("a mean(b) c -> c a", input),
        |input| input.mean(1)?.permute((1, 0)),
    )
}

#[test]
fn product_gradients_match_direct_candle_for_zero_cases() -> Result<()> {
    for (name, data) in [
        ("no zero", [2., 3., 4.]),
        ("one zero", [0., 3., 4.]),
        ("multiple zeros", [0., 0., 4.]),
    ] {
        compare_forward_and_gradient(
            &data,
            &[1, 3],
            |input| einops!("row prod(column) -> row", input),
            |input| {
                input
                    .narrow(1, 0, 1)?
                    .squeeze(1)?
                    .mul(&input.narrow(1, 1, 1)?.squeeze(1)?)?
                    .mul(&input.narrow(1, 2, 1)?.squeeze(1)?)
            },
        )
        .map_err(|error| error.with_path(name))?;
    }
    Ok(())
}

#[test]
fn unique_minimum_and_maximum_gradients_match_candle() -> Result<()> {
    let data = [3., 1., 5., 7., 9., 2.];
    compare_forward_and_gradient(
        &data,
        &[2, 3],
        |input| einops!("row min(column) -> row", input),
        |input| input.min(1),
    )?;
    compare_forward_and_gradient(
        &data,
        &[2, 3],
        |input| einops!("row max(column) -> row", input),
        |input| input.max(1),
    )
}