tensorlogic_train/optimizers/
adabelief.rs

1//! AdaBelief optimizer (NeurIPS 2020).
2//!
3//! AdaBelief adapts the step size according to the "belief" in the gradient direction.
4//! It uses the variance of gradients (belief) to adapt the learning rate, which can
5//! achieve faster convergence and better generalization than Adam/AdamW.
6//!
7//! Reference: Zhuang et al. "AdaBelief Optimizer: Adapting Stepsizes by the Belief
8//! in Observed Gradients" (NeurIPS 2020)
9
10use super::common::{compute_gradient_norm, GradClipMode, Optimizer, OptimizerConfig};
11use crate::{TrainError, TrainResult};
12use scirs2_core::ndarray::{Array, Ix2};
13use std::collections::HashMap;
14
15/// AdaBelief optimizer (NeurIPS 2020).
16///
17/// AdaBelief adapts the step size according to the "belief" in the gradient direction.
18/// It uses the variance of gradients (belief) to adapt the learning rate, which can
19/// achieve faster convergence and better generalization than Adam/AdamW.
20///
21/// Reference: Zhuang et al. "AdaBelief Optimizer: Adapting Stepsizes by the Belief
22/// in Observed Gradients" (NeurIPS 2020)
23#[derive(Debug)]
24pub struct AdaBeliefOptimizer {
25    config: OptimizerConfig,
26    /// First moment estimates (exponential moving average of gradients).
27    m: HashMap<String, Array<f64, Ix2>>,
28    /// Second moment estimates (variance of gradients).
29    s: HashMap<String, Array<f64, Ix2>>,
30    /// Timestep counter.
31    t: usize,
32}
33
34impl AdaBeliefOptimizer {
35    /// Create a new AdaBelief optimizer.
36    pub fn new(config: OptimizerConfig) -> Self {
37        Self {
38            config,
39            m: HashMap::new(),
40            s: HashMap::new(),
41            t: 0,
42        }
43    }
44
45    /// Apply gradient clipping if configured.
46    fn clip_gradients(&self, gradients: &mut HashMap<String, Array<f64, Ix2>>) {
47        if let Some(clip_value) = self.config.grad_clip {
48            match self.config.grad_clip_mode {
49                GradClipMode::Value => {
50                    for grad in gradients.values_mut() {
51                        grad.mapv_inplace(|g| g.max(-clip_value).min(clip_value));
52                    }
53                }
54                GradClipMode::Norm => {
55                    let total_norm = compute_gradient_norm(gradients);
56                    if total_norm > clip_value {
57                        let scale = clip_value / total_norm;
58                        for grad in gradients.values_mut() {
59                            grad.mapv_inplace(|g| g * scale);
60                        }
61                    }
62                }
63            }
64        }
65    }
66}
67
68impl Optimizer for AdaBeliefOptimizer {
69    fn step(
70        &mut self,
71        parameters: &mut HashMap<String, Array<f64, Ix2>>,
72        gradients: &HashMap<String, Array<f64, Ix2>>,
73    ) -> TrainResult<()> {
74        let mut clipped_gradients = gradients.clone();
75        self.clip_gradients(&mut clipped_gradients);
76        self.t += 1;
77        let lr = self.config.learning_rate;
78        let beta1 = self.config.beta1;
79        let beta2 = self.config.beta2;
80        let eps = self.config.epsilon;
81        let weight_decay = self.config.weight_decay;
82        let bias_correction1 = 1.0 - beta1.powi(self.t as i32);
83        let bias_correction2 = 1.0 - beta2.powi(self.t as i32);
84        for (name, param) in parameters.iter_mut() {
85            let grad = clipped_gradients.get(name).ok_or_else(|| {
86                TrainError::OptimizerError(format!("Missing gradient for parameter: {}", name))
87            })?;
88            if !self.m.contains_key(name) {
89                self.m.insert(name.clone(), Array::zeros(param.raw_dim()));
90                self.s.insert(name.clone(), Array::zeros(param.raw_dim()));
91            }
92            let m = self.m.get_mut(name).unwrap();
93            let s = self.s.get_mut(name).unwrap();
94            *m = &*m * beta1 + &(grad * (1.0 - beta1));
95            let grad_diff = grad - &*m;
96            let grad_diff_squared = grad_diff.mapv(|g| g * g);
97            *s = &*s * beta2 + &(grad_diff_squared * (1.0 - beta2));
98            let m_hat = &*m / bias_correction1;
99            let s_hat = &*s / bias_correction2;
100            if weight_decay > 0.0 {
101                param.mapv_inplace(|p| p * (1.0 - lr * weight_decay));
102            }
103            let update = m_hat / (s_hat.mapv(|v| v.sqrt()) + eps);
104            *param = &*param - &(update * lr);
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, s_val) in &self.s {
126            state.insert(format!("s_{}", name), s_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_val) = state.get("t") {
133            self.t = t_val[0] as usize;
134        }
135        for (key, values) in state {
136            if let Some(name) = key.strip_prefix("m_") {
137                if let Some(m_array) = self.m.get(name) {
138                    let shape = m_array.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("s_") {
144                if let Some(s_array) = self.s.get(name) {
145                    let shape = s_array.raw_dim();
146                    if let Ok(arr) = Array::from_shape_vec(shape, values) {
147                        self.s.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_adabelief_optimizer() {
162        let config = OptimizerConfig {
163            learning_rate: 0.001,
164            weight_decay: 0.01,
165            ..Default::default()
166        };
167        let mut optimizer = AdaBeliefOptimizer::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, 0.2], [0.3, 0.4]]);
172        for _ in 0..5 {
173            optimizer.step(&mut params, &grads).unwrap();
174        }
175        let w = params.get("w").unwrap();
176        assert!(w[[0, 0]] < 1.0);
177        assert!(w[[1, 1]] < 4.0);
178        let state = optimizer.state_dict();
179        assert!(state.contains_key("t"));
180        assert!(state.contains_key("m_w"));
181        assert!(state.contains_key("s_w"));
182    }
183}
tensorlogic_train/optimizers/adabelief.rs

tensorlogic_train/optimizers/
adabelief.rs
