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
|
//! Linear layer
//!
//! This layer applies a linear transformation to the incoming data, `y = x@w.t() + b`.
//! The bias is optional. The `forward` method can be used to apply the layer, it supports input
//! with a batch dimension (so of shape `(b_sz, in_c)`) or without (of shape `(in_c,)`), the
//! output has shape `(b_sz, out_c)` and `(out_c,)` respectively.
//!
//! ```rust
//! use candle::{Tensor, Device::Cpu};
//! use candle_nn::{Linear, Module};
//! # fn main() -> candle::Result<()> {
//!
//! let w = Tensor::new(&[[1f32, 2.], [3., 4.], [5., 6.]], &Cpu)?;
//! let layer = Linear::new(w, None); // Use no bias.
//! let xs = Tensor::new(&[[10f32, 100.]], &Cpu)?;
//! let ys = layer.forward(&xs)?;
//! assert_eq!(ys.to_vec2::<f32>()?, &[[210.0, 430.0, 650.0]]);
//! # Ok(()) }
//! ```
use candle::{Result, Tensor};
#[derive(Debug)]
pub struct Linear {
weight: Tensor,
bias: Option<Tensor>,
}
impl Linear {
pub fn new(weight: Tensor, bias: Option<Tensor>) -> Self {
Self { weight, bias }
}
}
impl super::Module for Linear {
fn forward(&self, x: &Tensor) -> candle::Result<Tensor> {
let w = match x.dims() {
&[bsize, _, _] => self.weight.broadcast_left(bsize)?.t()?,
_ => self.weight.t()?,
};
let x = x.matmul(&w)?;
match &self.bias {
None => Ok(x),
Some(bias) => x.broadcast_add(bias),
}
}
}
/// Create or initialize a new linear layer.
///
/// This uses some default names for weight and biases, namely `"weight"` and `"bias"`.
pub fn linear(in_dim: usize, out_dim: usize, vs: crate::VarBuilder) -> Result<Linear> {
let init_ws = crate::init::DEFAULT_KAIMING_NORMAL;
let ws = vs.get_or_init((out_dim, in_dim), "weight", init_ws)?;
let bound = 1. / (in_dim as f64).sqrt();
let init_bs = crate::Init::Uniform {
lo: -bound,
up: bound,
};
let bs = vs.get_or_init(out_dim, "bias", init_bs)?;
Ok(Linear::new(ws, Some(bs)))
}
pub fn linear_no_bias(in_dim: usize, out_dim: usize, vs: crate::VarBuilder) -> Result<Linear> {
let init_ws = crate::init::DEFAULT_KAIMING_NORMAL;
let ws = vs.get_or_init((out_dim, in_dim), "weight", init_ws)?;
Ok(Linear::new(ws, None))
}
|
