ella_tensor/ops/
shape.rs

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
use crate::{
    shape::{stride_offset, NdimMax},
    Axis, Const, Dyn, IntoShape, RemoveAxis, Shape, Tensor, TensorValue,
};

impl<T, S> Tensor<T, S>
where
    T: TensorValue,
    S: Shape,
{
    pub fn as_dyn(&self) -> Tensor<T, Dyn> {
        Tensor::new(
            self.values().clone(),
            self.shape().as_dyn(),
            self.strides().as_dyn(),
        )
    }

    pub fn reshape<I>(&self, shape: I) -> Tensor<T, I::Shape>
    where
        I: IntoShape,
    {
        let shape = shape.into_shape();
        let strides = shape.default_strides();
        assert_eq!(self.shape().size(), shape.size());

        let values = self.clone().to_standard_layout().into_values();

        Tensor::new(values, shape, strides)
    }

    #[inline]
    pub fn flatten(&self) -> Tensor<T, Const<1>> {
        self.reshape(Const([self.shape().size()]))
    }

    pub fn unsqueeze<A>(&self, axis: A) -> Tensor<T, S::Larger>
    where
        A: Into<Axis>,
    {
        let axis = axis.into();
        Tensor::new(
            self.values().clone(),
            self.shape().insert_axis(axis),
            self.strides().insert_axis(axis),
        )
    }

    pub fn swap_axes<A1, A2>(&self, ax1: A1, ax2: A2) -> Self
    where
        A1: Into<Axis>,
        A2: Into<Axis>,
    {
        let mut shape = self.shape().clone();
        let mut strides = self.strides().clone();
        let ax1 = Axis::index(&ax1.into(), &shape);
        let ax2 = Axis::index(&ax2.into(), &shape);

        shape.slice_mut().swap(ax1, ax2);
        strides.slice_mut().swap(ax1, ax2);
        Tensor::new(self.values().clone(), shape, strides)
    }

    pub fn as_shape<S2>(&self) -> crate::Result<Tensor<T, S2>>
    where
        S2: Shape,
    {
        let shape = S2::from_shape(self.shape())?;
        let strides = S2::from_shape(self.strides())?;
        Ok(Tensor::new(self.values().clone(), shape, strides))
    }

    pub fn broadcast_to<I, O>(&self, shape: I) -> crate::Result<Tensor<T, O>>
    where
        O: Shape,
        I: IntoShape<Shape = O>,
    {
        let to: O = shape.into_shape();
        let from = self.shape();
        if to.ndim() < from.ndim() {
            return Err(crate::ShapeError::broadcast(from.as_ref(), to.as_ref()).into());
        }

        let mut new_stride = to.clone();

        let mut stride_iter = new_stride.slice_mut().iter_mut().rev();
        for ((er, es), dr) in from
            .slice()
            .iter()
            .rev()
            .zip(self.strides().slice().iter().rev())
            .zip(stride_iter.by_ref())
        {
            if *dr == *er {
                *dr = *es;
            } else if *er == 1 {
                *dr = 0;
            } else {
                return Err(crate::ShapeError::broadcast(from.as_ref(), to.as_ref()).into());
            }
        }

        for dr in stride_iter {
            *dr = 0;
        }
        Ok(Tensor::new(self.values().clone(), to, new_stride))
    }

    #[allow(clippy::type_complexity)]
    pub fn broadcast_with<T2, S2>(
        &self,
        other: &Tensor<T2, S2>,
    ) -> crate::Result<(
        Tensor<T, <S as NdimMax<S2>>::Output>,
        Tensor<T2, <S as NdimMax<S2>>::Output>,
    )>
    where
        T2: TensorValue,
        S2: Shape,
        S: NdimMax<S2>,
    {
        let shape = self
            .shape()
            .broadcast::<S2, <S as NdimMax<S2>>::Output>(other.shape())?;
        let t1 = if shape.slice() == self.shape().slice() {
            self.as_shape::<<S as NdimMax<S2>>::Output>()
        } else {
            self.broadcast_to(shape.clone())
        }?;
        let t2 = if shape.slice() == other.shape().slice() {
            other.as_shape::<<S as NdimMax<S2>>::Output>()
        } else {
            other.broadcast_to(shape)
        }?;
        Ok((t1, t2))
    }

    pub fn invert_axis<A: Into<Axis>>(&self, axis: A) -> Self {
        let axis: Axis = axis.into();
        let ax = axis.index(self.shape());
        let stride = self.strides()[ax] as isize;
        let size = self.shape()[ax];
        let values = if size != 0 {
            self.values()
                .offset(stride_offset(size - 1, stride as usize))
        } else {
            self.values().clone()
        };
        let mut strides = self.strides().clone();
        strides[ax] = (-stride) as usize;
        Tensor::new(values, self.shape().clone(), strides)
    }

    pub fn roll(&self, axis: Axis, roll: isize) -> Self
    where
        S: RemoveAxis,
    {
        let len = self.shape().axis(axis);
        let shift = -roll.checked_rem(len as isize).unwrap_or_default();
        if roll == 0 {
            return self.clone();
        }
        Tensor::concat(
            axis,
            &[
                self.slice_axis(axis, shift..),
                self.slice_axis(axis, ..shift),
            ],
        )
        .unwrap()
    }
}

/// > 1-D shape operations
impl<T, S> Tensor<T, S>
where
    T: TensorValue,
    S: Shape + RemoveAxis,
{
    pub fn squeeze(&self, axis: Axis) -> Tensor<T, S::Smaller> {
        debug_assert!(axis.index(self.shape()) < self.shape().ndim());
        debug_assert!(self.shape()[axis.index(self.shape())] <= 1);

        let shape = self.shape().remove_axis(axis);
        let strides = self.strides().remove_axis(axis);
        Tensor::new(self.values().clone(), shape, strides)
    }
}

/// 2-D shape operations
impl<T> Tensor<T, Const<2>>
where
    T: TensorValue,
{
    pub fn t(&self) -> Self {
        self.swap_axes(0, 1)
    }
}

#[cfg(test)]
mod test {
    use crate::Axis;

    #[test]
    fn test_invert_axis() {
        let x = crate::tensor![[1, 2, 3], [4, 5, 6],];

        crate::assert_tensor_eq!(x.invert_axis(0), crate::tensor![[4, 5, 6], [1, 2, 3]]);
        crate::assert_tensor_eq!(x.invert_axis(1), crate::tensor![[3, 2, 1], [6, 5, 4]]);
    }

    #[test]
    fn test_roll() {
        let x = crate::tensor![[1, 2, 3], [4, 5, 6], [7, 8, 9],];

        // check forward roll
        crate::assert_tensor_eq!(
            x.roll(Axis(0), 2),
            crate::tensor![[4, 5, 6], [7, 8, 9], [1, 2, 3]]
        );

        // check backward roll
        crate::assert_tensor_eq!(
            x.roll(Axis(0), -1),
            crate::tensor![[4, 5, 6], [7, 8, 9], [1, 2, 3]]
        );

        // check that rolls cycle back around
        crate::assert_tensor_eq!(
            x.roll(Axis(0), 5),
            crate::tensor![[4, 5, 6], [7, 8, 9], [1, 2, 3]]
        );

        crate::assert_tensor_eq!(
            x.roll(Axis(-1), 1),
            crate::tensor![[3, 1, 2], [6, 4, 5], [9, 7, 8]]
        );
    }
}