nuvec-nn 0.4.0

Neural-network modules, layers and optimizers for nuvec
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

//! Loading and saving weights by name ([`crate::Module::load_weights`],
//! [`crate::Module::save_weights`]).

use std::collections::{BTreeMap, HashMap};
use std::path::{Path, PathBuf};

use nuvec::tree::{LeafKind, Leaves, LoadMode, Tree, Visit, Visitor};
use nuvec::{Array, Error, Result};

/// Where [`crate::Module::load_weights`] reads named weights from: a safetensors file (a
/// path), or named arrays (`Vec<(String, Array)>`, a `HashMap` or `BTreeMap`).
pub trait WeightSource {
    /// The named weights.
    fn weights(self) -> Result<Vec<(String, Array)>>;
}

fn from_file(path: &Path) -> Result<Vec<(String, Array)>> {
    Ok(nuvec::io::load_safetensors(path)?
        .tensors
        .into_iter()
        .collect())
}

impl WeightSource for &Path {
    fn weights(self) -> Result<Vec<(String, Array)>> {
        from_file(self)
    }
}

impl WeightSource for &PathBuf {
    fn weights(self) -> Result<Vec<(String, Array)>> {
        from_file(self)
    }
}

impl WeightSource for PathBuf {
    fn weights(self) -> Result<Vec<(String, Array)>> {
        from_file(&self)
    }
}

impl WeightSource for &str {
    fn weights(self) -> Result<Vec<(String, Array)>> {
        from_file(Path::new(self))
    }
}

impl WeightSource for String {
    fn weights(self) -> Result<Vec<(String, Array)>> {
        from_file(Path::new(&self))
    }
}

impl WeightSource for Vec<(String, Array)> {
    fn weights(self) -> Result<Vec<(String, Array)>> {
        Ok(self)
    }
}

impl<const N: usize> WeightSource for [(String, Array); N] {
    fn weights(self) -> Result<Vec<(String, Array)>> {
        Ok(self.into())
    }
}

impl<S> WeightSource for HashMap<String, Array, S> {
    fn weights(self) -> Result<Vec<(String, Array)>> {
        Ok(self.into_iter().collect())
    }
}

impl WeightSource for BTreeMap<String, Array> {
    fn weights(self) -> Result<Vec<(String, Array)>> {
        Ok(self.into_iter().collect())
    }
}

/// `load_flat`'s error, named for `op`.
fn renamed(e: Error, op: &str) -> Error {
    match e {
        Error::InvalidArgument(m) => Error::InvalidArgument(m.replacen("load_flat", op, 1)),
        e => e,
    }
}

pub(crate) fn load<T: Tree>(tree: &mut T, src: impl WeightSource, mode: LoadMode) -> Result<()> {
    let op = match mode {
        LoadMode::Strict => "load_weights",
        _ => "load_weights_partial",
    };
    tree.load_flat(src.weights()?, mode)
        .map_err(|e| renamed(e, op))
}

/// Writes the leaves of `tree` (all, or only the trainable ones) as safetensors, named by
/// path. Leaves that cannot be evaluated (unloaded placeholders) are reported by name.
pub(crate) fn save<T: Leaves + ?Sized>(tree: &T, path: &Path, trainable_only: bool) -> Result<()> {
    struct Named(Vec<(String, Array)>, bool);
    impl Visitor for Named {
        fn leaf(&mut self, path: &str, a: &Array, _: LeafKind, trainable: bool) {
            if trainable || !self.1 {
                self.0.push((path.to_owned(), a.clone()));
            }
        }
    }
    let op = if trainable_only {
        "save_trainable"
    } else {
        "save_weights"
    };
    let mut named = Named(Vec::new(), trainable_only);
    tree.visit(&mut Visit::new(&mut named));
    let named = named.0;
    let mut seen = std::collections::HashSet::new();
    for (p, _) in &named {
        if !seen.insert(p.as_str()) {
            return Err(Error::InvalidArgument(format!(
                "{op}: the path {p:?} appears twice"
            )));
        }
    }
    let arrays: Vec<Array> = named.iter().map(|(_, a)| a.clone()).collect();
    if nuvec::eval(&arrays).is_err() {
        let failed: Vec<&str> = named
            .iter()
            .filter(|(_, a)| a.eval().is_err())
            .map(|(p, _)| p.as_str())
            .collect();
        return Err(Error::InvalidArgument(format!(
            "{op}: cannot evaluate {} (unloaded placeholders?): {}",
            if failed.len() == 1 {
                "a leaf"
            } else {
                "leaves"
            },
            failed.join(", ")
        )));
    }
    let tensors: Vec<(&str, &Array)> = named.iter().map(|(p, a)| (p.as_str(), a)).collect();
    nuvec::io::save_safetensors(path, &tensors, &[("format", "mlx")])
}