acyclib 0.3.0

ML library for directed acyclic tensor graphs.
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

use std::{collections::HashMap, sync::Arc};

use crate::{
    device::{
        Device, OperationError,
        tensor::{DenseMatrix, Matrix, Shape, SparseMatrix},
    },
    graph::{Graph, GraphNodeId, GraphNodeIdTy},
    trainer::TrainerError,
};

pub trait DataLoader: Send + Sync + 'static {
    type Error: Send + Sync;

    fn map_batches<F: FnMut(PreparedBatchHost) -> bool>(self, batch_size: usize, f: F) -> Result<(), Self::Error>;
}

pub struct PreparedBatchHost {
    pub batch_size: usize,
    pub inputs: HashMap<String, HostMatrix>,
}

pub enum HostMatrix {
    Sparse(HostSparseMatrix),
    Dense(HostDenseMatrix),
}

pub struct HostSparseMatrix {
    vals: Vec<i32>,
    shape: Shape,
    nnz: usize,
    batch_size: Option<usize>,
}

impl HostSparseMatrix {
    /// # Safety
    /// All values must be in the range -1..shape.size()
    pub unsafe fn new(vals: Vec<i32>, batch_size: Option<usize>, shape: Shape, nnz: usize) -> Self {
        assert_eq!(batch_size.unwrap_or(1) * nnz, vals.len());

        Self { vals, shape, nnz, batch_size }
    }
}

pub struct HostDenseMatrix {
    vals: Vec<f32>,
    shape: Shape,
    batch_size: Option<usize>,
}

impl HostDenseMatrix {
    pub fn new(vals: Vec<f32>, batch_size: Option<usize>, shape: Shape) -> Self {
        assert_eq!(batch_size.unwrap_or(1) * shape.size(), vals.len());

        Self { vals, shape, batch_size }
    }
}

pub struct PreparedBatchDevice<D: Device> {
    pub device: Arc<D>,
    pub inputs: HashMap<String, Matrix<D>>,
    pub batch_size: usize,
}

impl<D: Device> PreparedBatchDevice<D> {
    pub fn new(device: Arc<D>, data: &PreparedBatchHost) -> Result<Self, D::DeviceError> {
        let mut inputs = HashMap::new();

        let mut max_batch_size = 1;

        for (id, matrix) in &data.inputs {
            let matrix = match matrix {
                HostMatrix::Sparse(HostSparseMatrix { vals, shape, nnz, batch_size }) => {
                    max_batch_size = batch_size.unwrap_or(1).max(max_batch_size);

                    let mut sparse = SparseMatrix::zeroed(device.clone(), shape.size(), *nnz, *batch_size)?;

                    // # Safety
                    // HostMatrix::Sparse is verified on construction.
                    unsafe {
                        sparse.load_from_slice(*nnz, *batch_size, vals)?;
                    }

                    Matrix::Sparse(sparse)
                }
                HostMatrix::Dense(HostDenseMatrix { vals, shape, batch_size }) => {
                    max_batch_size = batch_size.unwrap_or(1).max(max_batch_size);
                    let mut dense = DenseMatrix::zeroed(device.clone(), shape.size(), *batch_size)?;
                    dense.load_from_slice(*batch_size, vals)?;
                    Matrix::Dense(dense)
                }
            };

            inputs.insert(id.clone(), matrix);
        }

        device.synchronise()?;

        Ok(Self { device, inputs, batch_size: max_batch_size })
    }

    pub fn load_new_data(&mut self, data: &PreparedBatchHost) -> Result<(), OperationError<D::DeviceError>> {
        for (id, matrix) in &data.inputs {
            match matrix {
                HostMatrix::Sparse(HostSparseMatrix { vals, shape, nnz, batch_size }) => {
                    let sparse = self.inputs.get_mut(id).unwrap().sparse_mut()?;

                    if shape.size() != sparse.single_size || *nnz != sparse.nnz {
                        return Err(OperationError::InvalidTensorFormat);
                    }

                    // # Safety
                    // HostMatrix::Sparse is verified on construction.
                    unsafe {
                        sparse.load_non_blocking_from_host(*nnz, *batch_size, vals)?;
                    }
                }
                HostMatrix::Dense(HostDenseMatrix { vals, shape, batch_size }) => {
                    let dense = self.inputs.get_mut(id).unwrap().dense_mut()?;

                    if shape.size() != dense.single_size {
                        return Err(OperationError::InvalidTensorFormat);
                    }

                    unsafe {
                        dense.load_non_blocking_from_host(*batch_size, vals)?;
                    }
                }
            }
        }

        self.device.synchronise()?;

        Ok(())
    }

    pub fn load_into_graph(&mut self, graph: &mut Graph<D>) -> Result<(), TrainerError<D>> {
        for (id, matrix) in &mut self.inputs {
            if let Some(idx) = graph.input_idx(id) {
                let tensor = graph.get(GraphNodeId::new(idx, GraphNodeIdTy::Values)).unwrap();
                tensor.swap_with(matrix).map_err(TrainerError::Unexpected)?;
            }
        }

        Ok(())
    }
}