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
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
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277

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

use super::lstmcell::LSTMCell;
use super::parameter::Parameter;
use super::Module;

/// Multi-layer LSTM (Long Short-Term Memory) sequence module.
///
/// Wraps multiple [`LSTMCell`] layers and loops over timesteps, matching
/// the PyTorch `nn.LSTM` interface. Each layer feeds its output sequence
/// as input to the next layer.
///
/// ```ignore
/// let lstm = LSTM::new(4, 8, 2)?;       // input=4, hidden=8, 2 layers
/// let x = Variable::new(Tensor::randn(&[10, 1, 4], opts)?, false); // [seq, batch, input]
/// let (output, (h_n, c_n)) = lstm.forward_seq(&x, None)?;
/// // output: [10, 1, 8], h_n: [2, 1, 8], c_n: [2, 1, 8]
/// ```
pub struct LSTM {
    cells: Vec<LSTMCell>,
    hidden_size: i64,
    num_layers: usize,
    batch_first: bool,
}

impl LSTM {
    /// Create a multi-layer LSTM on CPU.
    pub fn new(input_size: i64, hidden_size: i64, num_layers: usize) -> Result<Self> {
        Self::on_device(input_size, hidden_size, num_layers, false, Device::CPU)
    }

    /// Create a multi-layer LSTM on a specific device.
    pub fn on_device(
        input_size: i64,
        hidden_size: i64,
        num_layers: usize,
        batch_first: bool,
        device: Device,
    ) -> Result<Self> {
        assert!(num_layers >= 1, "LSTM requires at least 1 layer");
        let mut cells = Vec::with_capacity(num_layers);
        for layer in 0..num_layers {
            let in_size = if layer == 0 { input_size } else { hidden_size };
            cells.push(LSTMCell::on_device(in_size, hidden_size, device)?);
        }
        Ok(LSTM {
            cells,
            hidden_size,
            num_layers,
            batch_first,
        })
    }

    /// Set batch_first mode. When true, input/output are `[batch, seq, features]`.
    pub fn batch_first(mut self, batch_first: bool) -> Self {
        self.batch_first = batch_first;
        self
    }

    /// Forward pass over a full sequence.
    ///
    /// - `input`: `[seq_len, batch, input_size]` (or `[batch, seq_len, input_size]` if batch_first)
    /// - `state_0`: optional initial state `(h_0, c_0)` each `[num_layers, batch, hidden_size]`,
    ///   or None for zeros
    ///
    /// Returns `(output, (h_n, c_n))`:
    /// - `output`: `[seq_len, batch, hidden_size]` (or `[batch, seq_len, hidden_size]` if batch_first)
    /// - `h_n`: `[num_layers, batch, hidden_size]`
    /// - `c_n`: `[num_layers, batch, hidden_size]`
    pub fn forward_seq(
        &self,
        input: &Variable,
        state_0: Option<(&Variable, &Variable)>,
    ) -> Result<(Variable, (Variable, Variable))> {
        // Normalize to [seq_len, batch, features]
        let x = if self.batch_first {
            input.transpose(0, 1)?
        } else {
            input.clone()
        };

        let seq_len = x.shape()[0];
        let batch = x.shape()[1];
        let hs = self.hidden_size;
        let opts = TensorOptions {
            dtype: DType::Float32,
            device: self.cells[0].parameters()[0].variable.device(),
        };

        // Initialize hidden and cell states per layer
        let (mut h, mut c): (Vec<Variable>, Vec<Variable>) = if let Some((h0, c0)) = state_0 {
            let h_vec = (0..self.num_layers)
                .map(|l| h0.select(0, l as i64))
                .collect::<Result<Vec<_>>>()?;
            let c_vec = (0..self.num_layers)
                .map(|l| c0.select(0, l as i64))
                .collect::<Result<Vec<_>>>()?;
            (h_vec, c_vec)
        } else {
            let h_vec = (0..self.num_layers)
                .map(|_| {
                    Ok(Variable::new(
                        Tensor::zeros(&[batch, hs], opts)?,
                        false,
                    ))
                })
                .collect::<Result<Vec<_>>>()?;
            let c_vec = (0..self.num_layers)
                .map(|_| {
                    Ok(Variable::new(
                        Tensor::zeros(&[batch, hs], opts)?,
                        false,
                    ))
                })
                .collect::<Result<Vec<_>>>()?;
            (h_vec, c_vec)
        };

        // Run through timesteps and layers
        let mut outputs = Vec::with_capacity(seq_len as usize);
        for t in 0..seq_len {
            let mut layer_input = x.select(0, t)?;
            for (l, cell) in self.cells.iter().enumerate() {
                // Pack state for LSTMCell: cat(h, c, dim=1)
                let state = h[l].cat(&c[l], 1)?;
                let new_state = cell.forward_step(&layer_input, Some(&state))?;
                // Unpack: h = state[:, :hs], c = state[:, hs:]
                h[l] = new_state.narrow(1, 0, hs)?;
                c[l] = new_state.narrow(1, hs, hs)?;
                layer_input = h[l].clone();
            }
            outputs.push(layer_input);
        }

        // Stack outputs: [seq_len, batch, hidden_size]
        let output = Variable::stack(&outputs, 0)?;
        let output = if self.batch_first {
            output.transpose(0, 1)?
        } else {
            output
        };

        // Stack final hidden/cell states: [num_layers, batch, hidden_size]
        let h_n = Variable::stack(&h, 0)?;
        let c_n = Variable::stack(&c, 0)?;

        Ok((output, (h_n, c_n)))
    }
}

impl Module for LSTM {
    fn name(&self) -> &str { "lstm" }

    /// Module trait forward: runs the full sequence with zero-initialized state.
    /// Returns only the output sequence (not h_n/c_n). Use [`forward_seq`](LSTM::forward_seq)
    /// for explicit state access.
    fn forward(&self, input: &Variable) -> Result<Variable> {
        let (output, _) = self.forward_seq(input, None)?;
        Ok(output)
    }

    fn parameters(&self) -> Vec<Parameter> {
        self.cells.iter().flat_map(|c| c.parameters()).collect()
    }
}

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

    #[test]
    fn test_lstm_shapes() {
        let dev = crate::tensor::test_device();
        let opts = crate::tensor::test_opts();
        let lstm = LSTM::on_device(4, 8, 2, false, dev).unwrap();

        let x = Variable::new(Tensor::randn(&[5, 3, 4], opts).unwrap(), false);
        let (output, (h_n, c_n)) = lstm.forward_seq(&x, None).unwrap();

        assert_eq!(output.shape(), vec![5, 3, 8]); // [seq, batch, hidden]
        assert_eq!(h_n.shape(), vec![2, 3, 8]);    // [layers, batch, hidden]
        assert_eq!(c_n.shape(), vec![2, 3, 8]);    // [layers, batch, hidden]
    }

    #[test]
    fn test_lstm_batch_first() {
        let dev = crate::tensor::test_device();
        let opts = crate::tensor::test_opts();
        let lstm = LSTM::on_device(4, 8, 2, true, dev).unwrap();

        let x = Variable::new(Tensor::randn(&[3, 5, 4], opts).unwrap(), false);
        let (output, (h_n, c_n)) = lstm.forward_seq(&x, None).unwrap();

        assert_eq!(output.shape(), vec![3, 5, 8]); // [batch, seq, hidden]
        assert_eq!(h_n.shape(), vec![2, 3, 8]);
        assert_eq!(c_n.shape(), vec![2, 3, 8]);
    }

    #[test]
    fn test_lstm_with_initial_state() {
        let dev = crate::tensor::test_device();
        let opts = crate::tensor::test_opts();
        let lstm = LSTM::on_device(4, 8, 2, false, dev).unwrap();

        let x = Variable::new(Tensor::randn(&[5, 3, 4], opts).unwrap(), false);
        let h0 = Variable::new(Tensor::randn(&[2, 3, 8], opts).unwrap(), false);
        let c0 = Variable::new(Tensor::randn(&[2, 3, 8], opts).unwrap(), false);
        let (output, (h_n, c_n)) = lstm.forward_seq(&x, Some((&h0, &c0))).unwrap();

        assert_eq!(output.shape(), vec![5, 3, 8]);
        assert_eq!(h_n.shape(), vec![2, 3, 8]);
        assert_eq!(c_n.shape(), vec![2, 3, 8]);
    }

    #[test]
    fn test_lstm_single_layer() {
        let dev = crate::tensor::test_device();
        let opts = crate::tensor::test_opts();
        let lstm = LSTM::on_device(4, 8, 1, false, dev).unwrap();

        let x = Variable::new(Tensor::randn(&[5, 3, 4], opts).unwrap(), false);
        let (output, (h_n, c_n)) = lstm.forward_seq(&x, None).unwrap();

        assert_eq!(output.shape(), vec![5, 3, 8]);
        assert_eq!(h_n.shape(), vec![1, 3, 8]);
        assert_eq!(c_n.shape(), vec![1, 3, 8]);
    }

    #[test]
    fn test_lstm_gradient() {
        let dev = crate::tensor::test_device();
        let opts = crate::tensor::test_opts();
        let lstm = LSTM::on_device(4, 8, 2, false, dev).unwrap();

        let x = Variable::new(Tensor::randn(&[5, 3, 4], opts).unwrap(), true);
        let (output, _) = lstm.forward_seq(&x, None).unwrap();
        let loss = output.sum().unwrap();
        loss.backward().unwrap();

        for p in lstm.parameters() {
            assert!(p.variable.grad().is_some(), "missing grad for {}", p.name);
        }
        assert!(x.grad().is_some());
    }

    #[test]
    fn test_lstm_module_forward() {
        let dev = crate::tensor::test_device();
        let opts = crate::tensor::test_opts();
        let lstm = LSTM::on_device(4, 8, 2, false, dev).unwrap();

        let x = Variable::new(Tensor::randn(&[5, 3, 4], opts).unwrap(), false);
        let y = lstm.forward(&x).unwrap();
        assert_eq!(y.shape(), vec![5, 3, 8]);
    }

    #[test]
    fn test_lstm_parameters_count() {
        let dev = crate::tensor::test_device();
        let lstm = LSTM::on_device(4, 8, 2, false, dev).unwrap();

        // Layer 0: 4 params (w_ih, w_hh, b_ih, b_hh)
        // Layer 1: 4 params
        assert_eq!(lstm.parameters().len(), 8);
    }

    #[test]
    fn test_lstm_builder_pattern() {
        let dev = crate::tensor::test_device();
        let lstm = LSTM::on_device(4, 8, 1, false, dev).unwrap().batch_first(true);
        let opts = crate::tensor::test_opts();
        let x = Variable::new(Tensor::randn(&[3, 5, 4], opts).unwrap(), false);
        let (output, _) = lstm.forward_seq(&x, None).unwrap();
        assert_eq!(output.shape(), vec![3, 5, 8]); // [batch, seq, hidden]
    }
}