flodl 0.2.1

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

use crate::autograd::{self, Variable};
use crate::tensor::{Device, DType, Result, Tensor, TensorOptions};

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

/// Single-step GRU cell backed by fused ATen `gru_cell` kernel.
///
/// Uses 4 packed parameters (`w_ih`, `w_hh`, `b_ih`, `b_hh`) instead
/// of 6 separate Linear modules, reducing ~25 kernel launches to ~2.
///
/// Takes `(x, h)` where h is optional (None -> zeros).
/// For use with graph forward refs: input = x, ref "hidden" = h.
///
/// Standalone usage: call `forward_step(x, h)` directly.
///
/// ```ignore
/// let gru = GRUCell::new(4, 3)?; // input_size=4, hidden_size=3
/// let x = Variable::new(Tensor::randn(&[1, 4], opts)?, false);
/// let h1 = gru.forward_step(&x, None)?;     // [1, 3]
/// let h2 = gru.forward_step(&x, Some(&h1))?; // [1, 3]
/// ```
pub struct GRUCell {
    w_ih: Parameter,
    w_hh: Parameter,
    b_ih: Parameter,
    b_hh: Parameter,
    hidden_size: i64,
}

impl GRUCell {
    /// Create a GRU cell with given input and hidden dimensions.
    /// All parameters initialized with PyTorch-style uniform(-1/sqrt(hs), 1/sqrt(hs)).
    pub fn new(input_size: i64, hidden_size: i64) -> Result<Self> {
        Self::on_device(input_size, hidden_size, Device::CPU)
    }

    /// Create a GRU cell on a specific device.
    pub fn on_device(input_size: i64, hidden_size: i64, device: Device) -> Result<Self> {
        let bound = 1.0 / (hidden_size as f64).sqrt();
        let opts = TensorOptions { dtype: DType::Float32, device };

        let w_ih = Tensor::rand(&[3 * hidden_size, input_size], opts)?
            .mul_scalar(2.0 * bound)?.add_scalar(-bound)?;
        let w_hh = Tensor::rand(&[3 * hidden_size, hidden_size], opts)?
            .mul_scalar(2.0 * bound)?.add_scalar(-bound)?;
        let b_ih = Tensor::rand(&[3 * hidden_size], opts)?
            .mul_scalar(2.0 * bound)?.add_scalar(-bound)?;
        let b_hh = Tensor::rand(&[3 * hidden_size], opts)?
            .mul_scalar(2.0 * bound)?.add_scalar(-bound)?;

        Ok(GRUCell {
            w_ih: Parameter::new(w_ih, "w_ih"),
            w_hh: Parameter::new(w_hh, "w_hh"),
            b_ih: Parameter::new(b_ih, "b_ih"),
            b_hh: Parameter::new(b_hh, "b_hh"),
            hidden_size,
        })
    }

    /// Single GRU step: returns new hidden state `[batch, hidden_size]`.
    /// Pass `None` for `h` to zero-initialize the hidden state.
    pub fn forward_step(&self, x: &Variable, h: Option<&Variable>) -> Result<Variable> {
        let batch = x.shape()[0];

        let h = match h {
            Some(h) => h.clone(),
            None => {
                let opts = TensorOptions { dtype: DType::Float32, device: self.w_ih.variable.device() };
                Variable::new(
                    Tensor::zeros(&[batch, self.hidden_size], opts)?,
                    false,
                )
            }
        };

        autograd::gru_cell(
            x, &h,
            &self.w_ih.variable, &self.w_hh.variable,
            &self.b_ih.variable, &self.b_hh.variable,
        )
    }
}

impl Module for GRUCell {
    fn name(&self) -> &str { "grucell" }

    /// Module trait forward: runs a single step with zero-initialized hidden state.
    /// Does not accept or return hidden state -- use [`forward_step`](GRUCell::forward_step)
    /// for explicit hidden state control in recurrent loops.
    fn forward(&self, input: &Variable) -> Result<Variable> {
        self.forward_step(input, None)
    }

    fn parameters(&self) -> Vec<Parameter> {
        vec![
            self.w_ih.clone(),
            self.w_hh.clone(),
            self.b_ih.clone(),
            self.b_hh.clone(),
        ]
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    fn from_f32(data: &[f32], shape: &[i64]) -> Tensor {
        Tensor::from_f32(data, shape, crate::tensor::test_device()).unwrap()
    }

    #[test]
    fn test_grucell_forward() {
        let gru = GRUCell::on_device(4, 3, crate::tensor::test_device()).unwrap();
        let x = Variable::new(Tensor::randn(&[2, 4], crate::tensor::test_opts()).unwrap(), false);
        let h = gru.forward_step(&x, None).unwrap();
        assert_eq!(h.shape(), vec![2, 3]);

        // Second step with hidden state
        let h2 = gru.forward_step(&x, Some(&h)).unwrap();
        assert_eq!(h2.shape(), vec![2, 3]);
    }

    #[test]
    fn test_grucell_gradient() {
        let gru = GRUCell::on_device(4, 3, crate::tensor::test_device()).unwrap();
        let x = Variable::new(
            Tensor::randn(&[2, 4], crate::tensor::test_opts()).unwrap(), true);
        let h = gru.forward_step(&x, None).unwrap();
        let loss = h.sum().unwrap();
        loss.backward().unwrap();

        // All parameters should have gradients
        for p in gru.parameters() {
            assert!(p.variable.grad().is_some(), "missing grad for {}", p.name);
        }
        // Input should have gradient
        assert!(x.grad().is_some());
    }

    #[test]
    fn test_grucell_multi_step() {
        let gru = GRUCell::on_device(4, 3, crate::tensor::test_device()).unwrap();
        let x = Variable::new(
            Tensor::randn(&[2, 4], crate::tensor::test_opts()).unwrap(), false);

        // Run 5 steps, each feeding previous hidden state
        let mut h: Option<Variable> = None;
        for _ in 0..5 {
            let h_new = gru.forward_step(&x, h.as_ref()).unwrap();
            h = Some(h_new);
        }
        assert_eq!(h.unwrap().shape(), vec![2, 3]);
    }

    #[test]
    fn test_grucell_finite_difference() {
        let eps = 1e-3;
        let gru = GRUCell::on_device(2, 2, crate::tensor::test_device()).unwrap();
        let x_data = vec![0.5_f32, -0.3, 0.1, 0.8];
        let x = Variable::new(from_f32(&x_data, &[2, 2]), true);

        let h = gru.forward_step(&x, None).unwrap();
        let loss = h.sum().unwrap();
        loss.backward().unwrap();
        let analytical = x.grad().unwrap().to_f32_vec().unwrap();

        for i in 0..4 {
            let mut xp = x_data.clone();
            let mut xm = x_data.clone();
            xp[i] += eps as f32;
            xm[i] -= eps as f32;

            let xp_var = Variable::new(from_f32(&xp, &[2, 2]), false);
            let xm_var = Variable::new(from_f32(&xm, &[2, 2]), false);
            let fp: f64 = gru.forward_step(&xp_var, None).unwrap()
                .sum().unwrap().item().unwrap();
            let fm: f64 = gru.forward_step(&xm_var, None).unwrap()
                .sum().unwrap().item().unwrap();
            let numerical = (fp - fm) / (2.0 * eps);
            assert!(
                (analytical[i] as f64 - numerical).abs() < 0.05,
                "grad mismatch at {}: analytical={}, numerical={}",
                i, analytical[i], numerical
            );
        }
    }

    #[test]
    fn test_grucell_module_forward() {
        use crate::nn::Module;
        let gru = GRUCell::on_device(4, 3, crate::tensor::test_device()).unwrap();
        let x = Variable::new(Tensor::randn(&[2, 4], crate::tensor::test_opts()).unwrap(), false);
        let y = gru.forward(&x).unwrap();
        assert_eq!(y.shape(), vec![2, 3]);
    }
}