flodl 0.3.0

floDl — a flow-graph deep learning framework built on libtorch
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

use crate::autograd::{Variable, instance_norm};
use crate::tensor::{Result, Device, Tensor};

use super::parameter::Parameter;
use super::Module;

/// Instance normalization layer.
///
/// Normalizes each instance (per-channel, per-sample) independently.
/// Equivalent to PyTorch's `nn.InstanceNorm1d` / `nn.InstanceNorm2d` / `nn.InstanceNorm3d`.
///
/// Input: `[N, C, ...]` (any number of spatial dims).
/// Output: same shape, normalized per (N, C) slice.
pub struct InstanceNorm {
    weight: Option<Parameter>,
    bias: Option<Parameter>,
    /// Number of channels (C) expected in the input.
    pub num_features: i64,
    eps: f64,
    momentum: f64,
    /// Whether this layer has learnable affine parameters.
    pub affine: bool,
}

impl InstanceNorm {
    /// Create an InstanceNorm layer. If `affine`, learns weight and bias.
    pub fn new(num_features: i64, affine: bool) -> Result<Self> {
        Self::build(num_features, affine, 1e-5, 0.1, Device::CPU)
    }

    /// Create on a specific device.
    pub fn on_device(num_features: i64, affine: bool, device: Device) -> Result<Self> {
        Self::build(num_features, affine, 1e-5, 0.1, device)
    }

    fn build(
        num_features: i64, affine: bool, eps: f64, momentum: f64, device: Device,
    ) -> Result<Self> {
        let (weight, bias) = if affine {
            let w = Tensor::ones(&[num_features], crate::tensor::TensorOptions {
                dtype: crate::tensor::DType::Float32,
                device,
            })?;
            let b = Tensor::zeros(&[num_features], crate::tensor::TensorOptions {
                dtype: crate::tensor::DType::Float32,
                device,
            })?;
            (
                Some(Parameter::new(w, "weight")),
                Some(Parameter::new(b, "bias")),
            )
        } else {
            (None, None)
        };

        Ok(Self { weight, bias, num_features, eps, momentum, affine })
    }
}

impl Module for InstanceNorm {
    fn name(&self) -> &str { "instance_norm" }

    fn forward(&self, input: &Variable) -> Result<Variable> {
        instance_norm(
            input,
            self.weight.as_ref().map(|p| &p.variable),
            self.bias.as_ref().map(|p| &p.variable),
            None, None,
            true, self.momentum, self.eps,
        )
    }

    fn parameters(&self) -> Vec<Parameter> {
        let mut params = Vec::new();
        if let Some(ref w) = self.weight {
            params.push(w.clone());
        }
        if let Some(ref b) = self.bias {
            params.push(b.clone());
        }
        params
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::tensor::{Tensor, test_device, test_opts};

    #[test]
    fn test_instance_norm_forward() {
        let norm = InstanceNorm::on_device(3, true, test_device()).unwrap();
        let x = Variable::new(
            Tensor::randn(&[2, 3, 4, 4], test_opts()).unwrap(),
            false,
        );
        let y = norm.forward(&x).unwrap();
        assert_eq!(y.shape(), vec![2, 3, 4, 4]);
    }

    #[test]
    fn test_instance_norm_no_affine() {
        let norm = InstanceNorm::on_device(3, false, test_device()).unwrap();
        assert_eq!(norm.parameters().len(), 0);
        let x = Variable::new(
            Tensor::randn(&[1, 3, 8, 8], test_opts()).unwrap(),
            false,
        );
        let y = norm.forward(&x).unwrap();
        assert_eq!(y.shape(), vec![1, 3, 8, 8]);
    }

    #[test]
    fn test_instance_norm_gradient() {
        let norm = InstanceNorm::on_device(3, true, test_device()).unwrap();
        let x = Variable::new(
            Tensor::randn(&[2, 3, 4, 4], test_opts()).unwrap(), true,
        );
        let y = norm.forward(&x).unwrap().sum().unwrap();
        y.backward().unwrap();
        assert!(x.grad().is_some());
    }

    #[test]
    fn test_instance_norm_3d_input() {
        let norm = InstanceNorm::on_device(4, true, test_device()).unwrap();
        let x = Variable::new(
            Tensor::randn(&[2, 4, 16], test_opts()).unwrap(), false,
        );
        let y = norm.forward(&x).unwrap();
        assert_eq!(y.shape(), vec![2, 4, 16]);
    }

    #[test]
    fn test_instance_norm_affine_parameters() {
        let norm = InstanceNorm::on_device(8, true, test_device()).unwrap();
        assert_eq!(norm.parameters().len(), 2);
    }
}