tensorlogic_train/optimizers/
lamb.rs

1//! LAMB optimizer (Layer-wise Adaptive Moments optimizer for Batch training).
2//!
3//! LAMB is designed for large batch training and uses layer-wise adaptation
4//! of the learning rate based on the ratio of parameter and update norms.
5//!
6//! Reference: You et al., "Large Batch Optimization for Deep Learning:
7//! Training BERT in 76 minutes", ICLR 2020
8
9use super::common::{compute_gradient_norm, GradClipMode, Optimizer, OptimizerConfig};
10use crate::{TrainError, TrainResult};
11use scirs2_core::ndarray::{Array, Ix2};
12use std::collections::HashMap;
13
14/// LAMB optimizer (Layer-wise Adaptive Moments optimizer for Batch training).
15/// Designed for large batch training, uses layer-wise adaptation.
16#[derive(Debug)]
17pub struct LambOptimizer {
18    config: OptimizerConfig,
19    /// First moment estimates.
20    m: HashMap<String, Array<f64, Ix2>>,
21    /// Second moment estimates.
22    v: HashMap<String, Array<f64, Ix2>>,
23    /// Timestep counter.
24    t: usize,
25}
26
27impl LambOptimizer {
28    /// Create a new LAMB optimizer.
29    pub fn new(config: OptimizerConfig) -> Self {
30        Self {
31            config,
32            m: HashMap::new(),
33            v: HashMap::new(),
34            t: 0,
35        }
36    }
37
38    /// Apply gradient clipping if configured.
39    fn clip_gradients(&self, gradients: &mut HashMap<String, Array<f64, Ix2>>) {
40        if let Some(clip_value) = self.config.grad_clip {
41            match self.config.grad_clip_mode {
42                GradClipMode::Value => {
43                    for grad in gradients.values_mut() {
44                        grad.mapv_inplace(|g| g.max(-clip_value).min(clip_value));
45                    }
46                }
47                GradClipMode::Norm => {
48                    let total_norm = compute_gradient_norm(gradients);
49                    if total_norm > clip_value {
50                        let scale = clip_value / total_norm;
51                        for grad in gradients.values_mut() {
52                            grad.mapv_inplace(|g| g * scale);
53                        }
54                    }
55                }
56            }
57        }
58    }
59
60    /// Compute L2 norm of an array.
61    fn compute_norm(arr: &Array<f64, Ix2>) -> f64 {
62        arr.iter().map(|&x| x * x).sum::<f64>().sqrt()
63    }
64}
65
66impl Optimizer for LambOptimizer {
67    fn step(
68        &mut self,
69        parameters: &mut HashMap<String, Array<f64, Ix2>>,
70        gradients: &HashMap<String, Array<f64, Ix2>>,
71    ) -> TrainResult<()> {
72        let mut clipped_gradients = gradients.clone();
73        self.clip_gradients(&mut clipped_gradients);
74        self.t += 1;
75        let lr = self.config.learning_rate;
76        let beta1 = self.config.beta1;
77        let beta2 = self.config.beta2;
78        let eps = self.config.epsilon;
79        let weight_decay = self.config.weight_decay;
80        for (name, param) in parameters.iter_mut() {
81            let grad = clipped_gradients.get(name).ok_or_else(|| {
82                TrainError::OptimizerError(format!("Missing gradient for parameter: {}", name))
83            })?;
84            if !self.m.contains_key(name) {
85                self.m.insert(name.clone(), Array::zeros(param.raw_dim()));
86                self.v.insert(name.clone(), Array::zeros(param.raw_dim()));
87            }
88            let m = self.m.get_mut(name).unwrap();
89            let v = self.v.get_mut(name).unwrap();
90            *m = &*m * beta1 + &(grad * (1.0 - beta1));
91            let grad_squared = grad.mapv(|g| g * g);
92            *v = &*v * beta2 + &(grad_squared * (1.0 - beta2));
93            let m_hat = &*m / (1.0 - beta1.powi(self.t as i32));
94            let v_hat = &*v / (1.0 - beta2.powi(self.t as i32));
95            let adam_step = &m_hat / &v_hat.mapv(|v_val| v_val.sqrt() + eps);
96            let update = &adam_step + &param.mapv(|p| p * weight_decay);
97            let param_norm = Self::compute_norm(param);
98            let update_norm = Self::compute_norm(&update);
99            let trust_ratio = if param_norm > 0.0 && update_norm > 0.0 {
100                param_norm / update_norm
101            } else {
102                1.0
103            };
104            *param = &*param - &(update * (lr * trust_ratio));
105        }
106        Ok(())
107    }
108
109    fn zero_grad(&mut self) {}
110
111    fn get_lr(&self) -> f64 {
112        self.config.learning_rate
113    }
114
115    fn set_lr(&mut self, lr: f64) {
116        self.config.learning_rate = lr;
117    }
118
119    fn state_dict(&self) -> HashMap<String, Vec<f64>> {
120        let mut state = HashMap::new();
121        state.insert("t".to_string(), vec![self.t as f64]);
122        for (name, m_val) in &self.m {
123            state.insert(format!("m_{}", name), m_val.iter().copied().collect());
124        }
125        for (name, v_val) in &self.v {
126            state.insert(format!("v_{}", name), v_val.iter().copied().collect());
127        }
128        state
129    }
130
131    fn load_state_dict(&mut self, state: HashMap<String, Vec<f64>>) {
132        if let Some(t_vals) = state.get("t") {
133            self.t = t_vals[0] as usize;
134        }
135        for (key, values) in state {
136            if let Some(name) = key.strip_prefix("m_") {
137                if let Some(m) = self.m.get(name) {
138                    let shape = m.raw_dim();
139                    if let Ok(arr) = Array::from_shape_vec(shape, values) {
140                        self.m.insert(name.to_string(), arr);
141                    }
142                }
143            } else if let Some(name) = key.strip_prefix("v_") {
144                if let Some(v) = self.v.get(name) {
145                    let shape = v.raw_dim();
146                    if let Ok(arr) = Array::from_shape_vec(shape, values) {
147                        self.v.insert(name.to_string(), arr);
148                    }
149                }
150            }
151        }
152    }
153}
154
155#[cfg(test)]
156mod tests {
157    use super::*;
158    use scirs2_core::ndarray::array;
159
160    #[test]
161    fn test_lamb_optimizer() {
162        let config = OptimizerConfig {
163            learning_rate: 0.001,
164            weight_decay: 0.01,
165            ..Default::default()
166        };
167        let mut optimizer = LambOptimizer::new(config);
168        let mut params = HashMap::new();
169        params.insert("w".to_string(), array![[1.0, 2.0], [3.0, 4.0]]);
170        let mut grads = HashMap::new();
171        grads.insert("w".to_string(), array![[0.1, 0.1], [0.1, 0.1]]);
172        optimizer.step(&mut params, &grads).unwrap();
173        let w = params.get("w").unwrap();
174        assert!(w[[0, 0]] < 1.0);
175    }
176}
tensorlogic_train/optimizers/lamb.rs

tensorlogic_train/optimizers/
lamb.rs
