fastnn 0.3.3

A robust deep learning library in Rust with CUDA kernels
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

//! `try_*` variants of the ops that can disagree about shape. Each validates and
//! then calls the panicking op, so there is only one implementation of each.

use crate::error::{Error, Result};

use super::Tensor;
use super::shape;

impl Tensor {
    /// [`matmul`](Tensor::matmul), reporting a mismatch instead of panicking.
    pub fn try_matmul(&self, other: &Tensor) -> Result<Tensor> {
        self.check_same_device(other, "matmul")?;
        // Rank first: `ndim() - 2` underflows on a 1-D operand.
        self.check_matmul_rank(other)?;
        self.check_matmul_dims(other, self.last_dim(), other.dim(other.ndim() - 2))?;
        Ok(self.matmul(other))
    }

    /// [`matmul_nt`](Tensor::matmul_nt), reporting a mismatch instead of panicking.
    pub fn try_matmul_nt(&self, other: &Tensor) -> Result<Tensor> {
        self.check_same_device(other, "matmul_nt")?;
        self.check_matmul_rank(other)?;
        self.check_matmul_dims(other, self.last_dim(), other.last_dim())?;
        Ok(self.matmul_nt(other))
    }

    /// [`matmul_tn`](Tensor::matmul_tn), reporting a mismatch instead of panicking.
    pub fn try_matmul_tn(&self, other: &Tensor) -> Result<Tensor> {
        self.check_same_device(other, "matmul_tn")?;
        self.check_matmul_rank(other)?;
        self.check_matmul_dims(
            other,
            self.dim(self.ndim() - 2),
            other.dim(other.ndim() - 2),
        )?;
        Ok(self.matmul_tn(other))
    }

    /// [`add`](Tensor::add), reporting a broadcast failure instead of panicking.
    pub fn try_add(&self, other: &Tensor) -> Result<Tensor> {
        self.check_broadcast(other, "add")?;
        Ok(self.add(other))
    }

    /// [`sub`](Tensor::sub), reporting a broadcast failure instead of panicking.
    pub fn try_sub(&self, other: &Tensor) -> Result<Tensor> {
        self.check_broadcast(other, "sub")?;
        Ok(self.sub(other))
    }

    /// [`mul`](Tensor::mul), reporting a broadcast failure instead of panicking.
    pub fn try_mul(&self, other: &Tensor) -> Result<Tensor> {
        self.check_broadcast(other, "mul")?;
        Ok(self.mul(other))
    }

    /// [`div`](Tensor::div), reporting a broadcast failure instead of panicking.
    pub fn try_div(&self, other: &Tensor) -> Result<Tensor> {
        self.check_broadcast(other, "div")?;
        Ok(self.div(other))
    }

    /// [`reshape`](Tensor::reshape), reporting an impossible shape instead of panicking.
    pub fn try_reshape(&self, dims: &[i64]) -> Result<Tensor> {
        let placeholders = dims.iter().filter(|&&d| d == -1).count();
        if placeholders > 1 {
            return Err(Error::Shape(format!(
                "reshape: more than one -1 in {dims:?}"
            )));
        }
        if dims.iter().any(|&d| d == 0 || d < -1) {
            return Err(Error::Shape(format!(
                "reshape: dimensions must be positive or -1, got {dims:?}"
            )));
        }

        let known: usize = dims
            .iter()
            .filter(|&&d| d != -1)
            .map(|&d| d as usize)
            .product();
        let fits = if placeholders == 1 {
            known > 0 && self.numel().is_multiple_of(known)
        } else {
            known == self.numel()
        };
        if !fits {
            return Err(Error::Shape(format!(
                "reshape: {} elements do not fit shape {dims:?}",
                self.numel()
            )));
        }
        Ok(self.reshape(dims))
    }

    /// [`index_select`](Tensor::index_select), reporting an out-of-range id.
    pub fn try_index_select(&self, ids: &[usize]) -> Result<Tensor> {
        if self.ndim() != 2 {
            return Err(Error::Shape(format!(
                "index_select: needs a 2-D table, got {:?}",
                self.shape()
            )));
        }
        let rows = self.dim(0);
        if let Some(&bad) = ids.iter().find(|&&id| id >= rows) {
            return Err(Error::Shape(format!(
                "index_select: row {bad} outside 0..{rows}"
            )));
        }
        Ok(self.index_select(ids))
    }

    fn check_same_device(&self, other: &Tensor, op: &str) -> Result<()> {
        if self.device() != other.device() {
            return Err(Error::Shape(format!(
                "{op}: {} and {} are on different devices",
                self.device(),
                other.device()
            )));
        }
        Ok(())
    }

    fn check_matmul_rank(&self, other: &Tensor) -> Result<()> {
        if self.ndim() < 2 || other.ndim() < 2 {
            return Err(Error::Shape(format!(
                "matmul: needs 2+ dimensions, got {:?} and {:?}",
                self.shape(),
                other.shape()
            )));
        }
        Ok(())
    }

    /// Assumes the rank check already passed.
    fn check_matmul_dims(&self, other: &Tensor, inner: usize, other_inner: usize) -> Result<()> {
        if inner != other_inner {
            return Err(Error::Shape(format!(
                "matmul: inner dimensions {inner} and {other_inner} disagree for {:?} and {:?}",
                self.shape(),
                other.shape()
            )));
        }

        let batch = shape::numel(&self.shape()[..self.ndim() - 2]);
        let other_batch = shape::numel(&other.shape()[..other.ndim() - 2]);
        if batch != other_batch && batch != 1 && other_batch != 1 {
            return Err(Error::Shape(format!(
                "matmul: batch counts {batch} and {other_batch} are not broadcastable"
            )));
        }
        Ok(())
    }

    fn check_broadcast(&self, other: &Tensor, op: &str) -> Result<()> {
        self.check_same_device(other, op)?;

        let ndim = self.ndim().max(other.ndim());
        let a = shape::pad_left(self.shape(), ndim);
        let b = shape::pad_left(other.shape(), ndim);
        if a.iter().zip(&b).any(|(&x, &y)| x != y && x != 1 && y != 1) {
            return Err(Error::Shape(format!(
                "{op}: cannot broadcast {:?} and {:?}",
                self.shape(),
                other.shape()
            )));
        }
        Ok(())
    }
}

#[cfg(test)]
mod tests {
    use crate::tensor::Tensor;

    #[test]
    fn matmul_reports_instead_of_panicking() {
        let x = Tensor::zeros(&[4, 8]);
        assert!(x.try_matmul(&Tensor::zeros(&[8, 2])).is_ok());
        assert!(x.try_matmul(&Tensor::zeros(&[3, 3])).is_err());
        assert!(x.try_matmul(&Tensor::zeros(&[8])).is_err());
        assert!(Tensor::zeros(&[8]).try_matmul(&x).is_err());
        assert!(x.try_matmul_tn(&Tensor::zeros(&[4])).is_err());
        assert!(x.try_matmul_nt(&Tensor::zeros(&[8])).is_err());
    }

    #[test]
    fn broadcast_rules_match_the_panicking_ops() {
        let x = Tensor::zeros(&[4, 3]);
        assert!(x.try_add(&Tensor::zeros(&[1, 3])).is_ok());
        assert!(x.try_add(&Tensor::zeros(&[3])).is_ok());
        assert!(x.try_add(&Tensor::zeros(&[2, 3])).is_err());
    }

    #[test]
    fn reshape_rejects_what_cannot_fit() {
        let x = Tensor::zeros(&[4, 3]);
        assert!(x.try_reshape(&[2, 6]).is_ok());
        assert!(x.try_reshape(&[2, -1]).is_ok());
        assert!(x.try_reshape(&[5, 5]).is_err());
        assert!(x.try_reshape(&[-1, -1]).is_err());
        assert!(x.try_reshape(&[0, 12]).is_err());
    }

    #[test]
    fn index_select_rejects_out_of_range_rows() {
        let table = Tensor::zeros(&[5, 3]);
        assert!(table.try_index_select(&[0, 4]).is_ok());
        assert!(table.try_index_select(&[0, 5]).is_err());
    }
}