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
|
use crate::{op::BackpropOp, op::Op, Error, Result, Tensor};
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct ParamsConv1D {
pub(crate) b_size: usize,
// Maybe we should have a version without l_in as this bit depends on the input and not only on
// the weights.
pub(crate) l_in: usize,
pub(crate) c_out: usize,
pub(crate) c_in: usize,
pub(crate) k_size: usize,
pub(crate) padding: usize,
pub(crate) stride: usize,
}
impl ParamsConv1D {
pub(crate) fn l_out(&self) -> usize {
let dilation = 1;
(self.l_in + 2 * self.padding - dilation * (self.k_size - 1) - 1) / self.stride + 1
}
pub(crate) fn out_dims(&self) -> Vec<usize> {
let l_out = self.l_out();
vec![self.b_size, self.c_out, l_out]
}
}
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct ParamsConv2D {
pub(crate) b_size: usize,
pub(crate) i_h: usize,
pub(crate) i_w: usize,
pub(crate) k_h: usize,
pub(crate) k_w: usize,
pub(crate) c_out: usize,
pub(crate) c_in: usize,
pub(crate) padding: usize,
pub(crate) stride: usize,
}
impl ParamsConv2D {
pub(crate) fn out_h(&self) -> usize {
let dilation = 1;
(self.i_h + 2 * self.padding - dilation * (self.k_h - 1) - 1) / self.stride + 1
}
pub(crate) fn out_w(&self) -> usize {
let dilation = 1;
(self.i_w + 2 * self.padding - dilation * (self.k_w - 1) - 1) / self.stride + 1
}
pub(crate) fn out_dims(&self) -> Vec<usize> {
vec![self.b_size, self.c_out, self.out_h(), self.out_w()]
}
}
impl Tensor {
fn conv1d_single_group(&self, kernel: &Self, params: &ParamsConv1D) -> Result<Self> {
let storage =
self.storage()
.conv1d(self.layout(), &kernel.storage(), kernel.layout(), params)?;
let op = BackpropOp::new2(self, kernel, |arg, kernel| Op::Conv1D {
arg,
kernel,
padding: params.padding,
stride: params.stride,
});
let out_dims = params.out_dims();
Ok(crate::tensor::from_storage(storage, out_dims, op, false))
}
/// Applies a 1D convolution over the input tensor.
pub fn conv1d(
&self,
kernel: &Self,
padding: usize,
stride: usize,
groups: usize,
) -> Result<Self> {
let (c_out, c_in_k, k_size) = kernel.dims3()?;
let (b_size, c_in, l_in) = self.dims3()?;
if c_in != c_in_k * groups {
Err(Error::Conv1dInvalidArgs {
inp_shape: self.shape().clone(),
k_shape: kernel.shape().clone(),
padding,
stride,
msg: "the number of in-channels on the input doesn't match the kernel size",
}
.bt())?
}
let params = ParamsConv1D {
b_size,
l_in,
c_out: c_out / groups,
c_in: c_in / groups,
k_size,
padding,
stride,
};
if groups == 1 {
self.conv1d_single_group(kernel, ¶ms)
} else {
let blocks = self.chunk(groups, 1)?;
let kernel = kernel.chunk(groups, 0)?;
let blocks = blocks
.iter()
.zip(&kernel)
.map(|(block, kernel)| block.conv1d_single_group(kernel, ¶ms))
.collect::<Result<Vec<_>>>()?;
Tensor::cat(&blocks, 1)
}
}
fn conv2d_single_group(&self, kernel: &Self, params: &ParamsConv2D) -> Result<Self> {
let storage =
self.storage()
.conv2d(self.layout(), &kernel.storage(), kernel.layout(), params)?;
let op = BackpropOp::new2(self, kernel, |arg, kernel| Op::Conv2D {
arg,
kernel,
padding: params.padding,
stride: params.stride,
});
let out_dims = params.out_dims();
Ok(crate::tensor::from_storage(storage, out_dims, op, false))
}
/// Applies a 2D convolution over the input tensor.
pub fn conv2d(
&self,
kernel: &Self,
padding: usize,
stride: usize,
groups: usize,
) -> Result<Self> {
let (b_size, c_in, i_h, i_w) = self.dims4()?;
let (c_out, c_in_k, k_h, k_w) = kernel.dims4()?;
if c_in != c_in_k * groups {
crate::bail!(
"in_channel mismatch between input ({c_in}, groups {groups}) and kernel ({c_in_k})"
)
}
let params = ParamsConv2D {
b_size,
i_h,
i_w,
k_h,
k_w,
c_out: c_out / groups,
c_in: c_in / groups,
padding,
stride,
};
if groups == 1 {
self.conv2d_single_group(kernel, ¶ms)
} else {
let blocks = self.chunk(groups, 1)?;
let kernel = kernel.chunk(groups, 0)?;
let blocks = blocks
.iter()
.zip(&kernel)
.map(|(block, kernel)| block.conv2d_single_group(kernel, ¶ms))
.collect::<Result<Vec<_>>>()?;
Tensor::cat(&blocks, 1)
}
}
}
|
