topos 0.13.0

An autodiff compiler stack in Rust.
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

use crate::{
    Activation, Linear, Module, Segment, Sequential, Symbol, Tape, Tensor, Value, Visitor,
};

use super::{named_restore, named_snapshot, restore, snapshot};

/// Builds the same two-stage topology on `tape` with `fill`-valued
/// parameters: the "same code" whose recording order both positional
/// identities rely on.
fn model(tape: &Tape<f64>, fill: f64) -> Sequential<f64> {
    Sequential::new()
        .then(Linear::new(
            tape,
            Tensor::filled([2, 3], fill),
            Tensor::filled([3], fill),
        ))
        .then(Activation::Tanh)
        .then(Linear::new(
            tape,
            Tensor::filled([3, 1], fill),
            Tensor::filled([1], fill),
        ))
}

#[test]
fn positional_checkpoints_round_trip() {
    let input_shape = [1, 2];
    let trained_tape = Tape::new();
    let trained = model(&trained_tape, 0.5);
    let trained_input = trained_tape.leaf(Tensor::filled(input_shape, 1.0_f64));
    let trained_output = trained.express(trained_input).symbol();
    let trained_network = trained_tape.into_network();
    let trained_parameters = trained_network.parameters();
    let payloads = snapshot(&trained_parameters, &trained);
    assert_eq!(payloads.len(), 4);

    // A fresh process: same code, different initialization.
    let fresh_tape = Tape::new();
    let fresh = model(&fresh_tape, 0.0);
    let fresh_input = fresh_tape.leaf(Tensor::filled(input_shape, 1.0_f64));
    let fresh_output = fresh.express(fresh_input).symbol();
    let fresh_network = fresh_tape.into_network();
    let restored = restore(&fresh_network.parameters(), &fresh, payloads);

    assert_eq!(
        trained_network
            .forward(&trained_parameters, [])
            .of(trained_output)
            .to_vec(),
        fresh_network
            .forward(&restored, [])
            .of(fresh_output)
            .to_vec(),
    );
}

#[test]
#[should_panic(expected = "payloads but the module has")]
fn positional_restore_rejects_a_count_mismatch() {
    let tape = Tape::new();
    let module = model(&tape, 0.0);
    let parameters = tape.into_network().parameters();
    let _ = restore(&parameters, &module, vec![Tensor::filled([2, 3], 1.0_f64)]);
}

#[test]
fn named_checkpoints_round_trip() {
    let trained_tape = Tape::new();
    let trained = model(&trained_tape, 0.25);
    let trained_parameters = trained_tape.into_network().parameters();
    let entries = named_snapshot(&trained_parameters, &trained);
    let rendered: Vec<String> = entries.iter().map(|(path, _)| path.to_string()).collect();
    assert_eq!(rendered, ["0.weights", "0.bias", "2.weights", "2.bias"]);

    let fresh_tape = Tape::new();
    let fresh = model(&fresh_tape, 0.0);
    let fresh_parameters = fresh_tape.into_network().parameters();
    let restored = named_restore(&fresh_parameters, &fresh, entries);
    let payloads = snapshot(&restored, &fresh);
    assert_eq!(payloads[0].to_vec(), vec![0.25; 6]);
}

#[test]
#[should_panic(expected = "missing entries for: 2.weights")]
fn named_restore_rejects_missing_entries() {
    let tape = Tape::new();
    let module = model(&tape, 0.5);
    let parameters = tape.into_network().parameters();
    let mut entries = named_snapshot(&parameters, &module);
    entries.remove(2);
    let _ = named_restore(&parameters, &module, entries);
}

#[test]
#[should_panic(expected = "no parameter matches")]
fn named_restore_rejects_unexpected_entries() {
    let tape = Tape::new();
    let first = model(&tape, 0.5);
    let second = Sequential::new().then(Linear::new(
        &tape,
        Tensor::filled([2, 3], 0.5_f64),
        Tensor::filled([3], 0.5),
    ));
    let parameters = tape.into_network().parameters();
    // Entries snapshotted from the larger model cannot all match the
    // smaller one.
    let entries = named_snapshot(&parameters, &first);
    let _ = named_restore(&parameters, &second, entries);
}

#[test]
fn tied_parameters_restore_once() {
    /// One table announced under two paths, the way a tied head
    /// shares an embedding's weights.
    struct Tied(Symbol);
    impl Module<f64> for Tied {
        fn express<'tape>(&self, input: Value<'tape, f64>) -> Value<'tape, f64> {
            input.tape().resolve(self.0)
        }
        fn visit(&self, visitor: &mut dyn Visitor) {
            visitor.enter(Segment::Name("embedding"));
            visitor.parameter("weights", self.0);
            visitor.leave();
            visitor.enter(Segment::Name("head"));
            visitor.parameter("weights", self.0);
            visitor.leave();
        }
    }

    let tape = Tape::new();
    let table = tape.parameter(Tensor::filled([2, 2], 0.5_f64)).symbol();
    let model = Tied(table);
    let parameters = tape.into_network().parameters();

    let entries = named_snapshot(&parameters, &model);
    // One symbol under two paths: both entries are present, and the
    // restore takes the later one in visit order.
    assert_eq!(entries.len(), 2);
    let restored = named_restore(&parameters, &model, entries);
    assert_eq!(restored.of(table).to_vec(), vec![0.5; 4]);
}