rustyml 0.16.0

A high-performance machine learning & deep learning library in pure Rust, offering ML algorithms and neural network support
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

//! The Reverse layer, and the bidirectional graph it makes correct
//!
//! A recurrent layer with `go_backwards` returns its states in processing order, so slot 0
//! holds the state that came from the last input timestep. Keras puts that branch back into
//! input order before it merges. These tests pin the layer, its 3 refusals, and the invariant
//! that ties it to the recurrent family.

use ndarray::{Array, Axis, IxDyn};
use rustyml::neural_network::graph::GraphBuilder;
use rustyml::neural_network::layers::{Activation, Concatenate, Dense, LSTM, Reverse};
use rustyml::neural_network::losses::mean_squared_error::MeanSquaredError;
use rustyml::neural_network::optimizers::Adam;
use rustyml::neural_network::sequential::SequentialBuilder;
use rustyml::neural_network::traits::UnaryLayer;
use rustyml::neural_network::{Ctx, Shape, Tensor};

/// A tensor whose every value names its own position
fn ramp(shape: &[usize]) -> Tensor {
    let count: usize = shape.iter().product();
    let values: Vec<f32> = (0..count).map(|i| i as f32).collect();
    Array::from_shape_vec(IxDyn(shape), values).expect("the formula fills the shape")
}

/// Reverses 1 tensor through a built layer
fn reversed(axis: i32, input: &Tensor) -> Tensor {
    let mut layer = Reverse::new(axis);
    let mut ctx = Ctx::inference();
    layer
        .forward_mut(input, &mut ctx)
        .expect("the layer serves this rank and axis")
}

/// The layer flips the named axis and leaves every other axis as it was
#[test]
fn reverse_flips_the_named_axis_alone() {
    let x = ramp(&[2, 3, 2]);
    let got = reversed(1, &x);
    assert_eq!(got.shape(), &[2, 3, 2]);
    for batch in 0..2 {
        for step in 0..3 {
            for feature in 0..2 {
                assert_eq!(
                    got[[batch, step, feature]],
                    x[[batch, 2 - step, feature]],
                    "slot {step} must hold what slot {} held",
                    2 - step
                );
            }
        }
    }
}

/// A negative axis counts back from the end of the full rank
#[test]
fn reverse_resolves_a_negative_axis_against_the_rank() {
    let x = ramp(&[2, 3, 4]);
    assert_eq!(reversed(-1, &x), reversed(2, &x));
    assert_eq!(reversed(-2, &x), reversed(1, &x));
}

/// The output holds the standard memory order, so the next layer reads it as 1 slice
#[test]
fn reverse_emits_the_standard_memory_order() {
    let got = reversed(1, &ramp(&[2, 3, 4]));
    assert!(
        got.as_slice().is_some(),
        "a reversed view carries a negative stride, and the output must not"
    );
}

/// A rank-2 input is refused, because its axis 1 holds features and not a sequence
///
/// This is the shape a recurrent branch gives when `return_sequences` stays unset. Without the
/// refusal the layer reorders the features of every sample and reports nothing.
#[test]
fn reverse_refuses_a_rank_2_input() {
    let Err(error) = SequentialBuilder::new()
        .add(Reverse::new(1))
        .build(&Shape::known(&[2, 5]))
    else {
        panic!("a rank-2 input holds no axis whose order carries meaning");
    };
    let text = format!("{error}");
    assert!(text.contains("rank 3 or more"), "got {text}");
    assert!(text.contains("with_return_sequences"), "got {text}");
}

/// The batch axis is refused, because reversing it would move a sample against its target
#[test]
fn reverse_refuses_the_batch_axis() {
    for axis in [0_i32, -3] {
        let Err(error) = SequentialBuilder::new()
            .add(Reverse::new(axis))
            .build(&Shape::known(&[2, 3, 4]))
        else {
            panic!("axis 0 is the batch axis");
        };
        assert!(
            format!("{error}").contains("batch axis"),
            "the message must name the batch axis"
        );
    }
}

/// An axis the rank does not hold is refused, and the message names both
#[test]
fn reverse_refuses_an_axis_the_rank_does_not_hold() {
    let Err(error) = SequentialBuilder::new()
        .add(Reverse::new(7))
        .build(&Shape::known(&[2, 3, 4]))
    else {
        panic!("rank 3 holds no axis 7");
    };
    let text = format!("{error}");
    assert!(text.contains("axis 7"), "got {text}");
    assert!(text.contains("rank 3"), "got {text}");
}

/// A checkpoint refuses a model whose Reverse names another axis
///
/// The layer changes no shape and holds no array, so the recorded build shapes and the empty
/// weight roster are identical for any 2 axes. The reported type carries the axis, and it is
/// the only field a strict load can separate them by.
#[test]
fn a_checkpoint_refuses_a_different_reverse_axis() {
    let x = ramp(&[2, 3, 4]);
    let build = |axis: i32| {
        SequentialBuilder::new()
            .add(Reverse::new(axis))
            .add(
                Dense::new(2, Activation::Linear)
                    .unwrap()
                    .with_random_state(4),
            )
            .build(&Shape::known(&[2, 3, 4]))
            .expect("the stack builds")
    };

    let saved = build(1);
    let dir = std::env::temp_dir().join(format!("rustyml_reverse_axis_{}", std::process::id()));
    std::fs::create_dir_all(&dir).expect("the temporary directory is writable");
    let path = dir.join("m.rustyml");
    saved.save_to_path(&path).expect("the model saves");

    let mut other = build(2);
    let outcome = other.load_from_path(&path);
    let _ = std::fs::remove_dir_all(&dir);

    let error = outcome.expect_err("a load must not accept another axis");
    assert!(
        format!("{error}").contains("Reverse"),
        "the message names the layer: {error}"
    );
    let _ = x;
}

/// A reversed backward branch starts at the first input timestep
///
/// This is the invariant that ties the layer to the recurrent family, and it is what Keras
/// defines. Processing step k of a `go_backwards` layer consumes input timestep `t - 1 - k`, so
/// its last processing state consumed input timestep 0. After the reversal that state sits at
/// slot 0, which is where a forward branch holds the state of input timestep 0. The 2 branches
/// therefore agree timestep for timestep, which is what a merge layer needs.
#[test]
fn a_reversed_backward_branch_starts_at_the_first_input_timestep() {
    let x = ramp(&[2, 5, 3]);
    let mut ctx = Ctx::inference();

    let mut sequence_layer = LSTM::new(2, Activation::Tanh)
        .unwrap()
        .with_go_backwards(true)
        .with_return_sequences(true)
        .with_random_state(11);
    let sequence = sequence_layer.forward_mut(&x, &mut ctx).unwrap();

    let mut last_layer = LSTM::new(2, Activation::Tanh)
        .unwrap()
        .with_go_backwards(true)
        .with_random_state(11);
    let last = last_layer.forward_mut(&x, &mut ctx).unwrap();

    let aligned = reversed(1, &sequence);
    let first_slot = aligned.index_axis(Axis(1), 0);
    assert_eq!(
        first_slot.shape(),
        last.shape(),
        "the returned last state and 1 slot of the sequence hold the same shape"
    );
    for (slot, state) in first_slot.iter().zip(last.iter()) {
        assert_eq!(
            slot.to_bits(),
            state.to_bits(),
            "slot 0 of the reversed sequence is the state the layer returns without \
             `return_sequences`, bit for bit"
        );
    }
}

/// A bidirectional graph builds, aligns its 2 branches, and trains both of them
#[test]
fn a_bidirectional_graph_trains_both_directions() {
    let x = ramp(&[3, 4, 2]);
    let y = ramp(&[3, 4, 4]);

    let mut builder = GraphBuilder::new();
    let input = builder.input(Shape::known(&[3, 4, 2]));
    let forward = builder.add(
        LSTM::new(2, Activation::Tanh)
            .unwrap()
            .with_return_sequences(true)
            .with_random_state(1),
        &[input],
    );
    let backward = builder.add(
        LSTM::new(2, Activation::Tanh)
            .unwrap()
            .with_return_sequences(true)
            .with_go_backwards(true)
            .with_random_state(2),
        &[input],
    );
    let aligned = builder.add(Reverse::new(1), &[backward]);
    let merged = builder.add(Concatenate::new(-1), &[forward, aligned]);
    let mut model = builder.build(&[merged]).expect("the graph builds");

    // 2 sibling layers are 2 arena entries, so their names never collide
    let paths = model.weight_paths();
    assert_eq!(
        paths.len(),
        6,
        "2 recurrent layers hold 3 arrays each: {paths:?}"
    );
    assert!(paths.contains(&"0.kernel".to_string()), "{paths:?}");
    assert!(paths.contains(&"1.kernel".to_string()), "{paths:?}");

    model.compile(
        Adam::new(0.01, 0.9, 0.999, 1e-8, 0.0).unwrap(),
        MeanSquaredError::new(),
    );
    let before_forward = model.weight("0.kernel").unwrap().sum();
    let before_backward = model.weight("1.kernel").unwrap().sum();

    let history = model.fit(&[&x], &[&y], 3).expect("the model trains");
    assert_eq!(history.loss().len(), 3);
    assert!(
        history.loss()[2] < history.loss()[0],
        "the loss must fall: {:?}",
        history.loss()
    );
    assert_ne!(
        model.weight("0.kernel").unwrap().sum().to_bits(),
        before_forward.to_bits(),
        "the forward branch must train"
    );
    assert_ne!(
        model.weight("1.kernel").unwrap().sum().to_bits(),
        before_backward.to_bits(),
        "the backward branch must train"
    );
}