tensorlogic_train/optimizers/
adam.rs

1//! Adam optimizer (Adaptive Moment Estimation).
2//!
3//! Adam combines the benefits of AdaGrad and RMSProp by maintaining both
4//! first-order (momentum) and second-order moment estimates of gradients.
5//!
6//! Reference: Kingma & Ba, "Adam: A Method for Stochastic Optimization", ICLR 2015
7
8use super::common::{compute_gradient_norm, GradClipMode, Optimizer, OptimizerConfig};
9use crate::{TrainError, TrainResult};
10use scirs2_core::ndarray::{Array, Ix2};
11use std::collections::HashMap;
12
13/// Adam optimizer.
14#[derive(Debug)]
15pub struct AdamOptimizer {
16    config: OptimizerConfig,
17    /// First moment estimates (exponential moving average of gradients).
18    m: HashMap<String, Array<f64, Ix2>>,
19    /// Second moment estimates (exponential moving average of squared gradients).
20    v: HashMap<String, Array<f64, Ix2>>,
21    /// Timestep counter.
22    t: usize,
23}
24
25impl AdamOptimizer {
26    /// Create a new Adam optimizer.
27    pub fn new(config: OptimizerConfig) -> Self {
28        Self {
29            config,
30            m: HashMap::new(),
31            v: HashMap::new(),
32            t: 0,
33        }
34    }
35
36    /// Apply gradient clipping if configured.
37    fn clip_gradients(&self, gradients: &mut HashMap<String, Array<f64, Ix2>>) {
38        if let Some(clip_value) = self.config.grad_clip {
39            match self.config.grad_clip_mode {
40                GradClipMode::Value => {
41                    for grad in gradients.values_mut() {
42                        grad.mapv_inplace(|g| g.max(-clip_value).min(clip_value));
43                    }
44                }
45                GradClipMode::Norm => {
46                    let total_norm = compute_gradient_norm(gradients);
47                    if total_norm > clip_value {
48                        let scale = clip_value / total_norm;
49                        for grad in gradients.values_mut() {
50                            grad.mapv_inplace(|g| g * scale);
51                        }
52                    }
53                }
54            }
55        }
56    }
57}
58
59impl Optimizer for AdamOptimizer {
60    fn step(
61        &mut self,
62        parameters: &mut HashMap<String, Array<f64, Ix2>>,
63        gradients: &HashMap<String, Array<f64, Ix2>>,
64    ) -> TrainResult<()> {
65        let mut clipped_gradients = gradients.clone();
66        self.clip_gradients(&mut clipped_gradients);
67        self.t += 1;
68        let lr = self.config.learning_rate;
69        let beta1 = self.config.beta1;
70        let beta2 = self.config.beta2;
71        let eps = self.config.epsilon;
72        let lr_t =
73            lr * ((1.0 - beta2.powi(self.t as i32)).sqrt()) / (1.0 - beta1.powi(self.t as i32));
74        for (name, param) in parameters.iter_mut() {
75            let grad = clipped_gradients.get(name).ok_or_else(|| {
76                TrainError::OptimizerError(format!("Missing gradient for parameter: {}", name))
77            })?;
78            if !self.m.contains_key(name) {
79                self.m.insert(name.clone(), Array::zeros(param.raw_dim()));
80                self.v.insert(name.clone(), Array::zeros(param.raw_dim()));
81            }
82            let m = self.m.get_mut(name).unwrap();
83            let v = self.v.get_mut(name).unwrap();
84            *m = &*m * beta1 + &(grad * (1.0 - beta1));
85            let grad_squared = grad.mapv(|g| g * g);
86            *v = &*v * beta2 + &(grad_squared * (1.0 - beta2));
87            let update = m.mapv(|m_val| m_val * lr_t) / &v.mapv(|v_val| v_val.sqrt() + eps);
88            *param = &*param - &update;
89        }
90        Ok(())
91    }
92
93    fn zero_grad(&mut self) {}
94
95    fn get_lr(&self) -> f64 {
96        self.config.learning_rate
97    }
98
99    fn set_lr(&mut self, lr: f64) {
100        self.config.learning_rate = lr;
101    }
102
103    fn state_dict(&self) -> HashMap<String, Vec<f64>> {
104        let mut state = HashMap::new();
105        state.insert("t".to_string(), vec![self.t as f64]);
106        for (name, m_val) in &self.m {
107            state.insert(format!("m_{}", name), m_val.iter().copied().collect());
108        }
109        for (name, v_val) in &self.v {
110            state.insert(format!("v_{}", name), v_val.iter().copied().collect());
111        }
112        state
113    }
114
115    fn load_state_dict(&mut self, state: HashMap<String, Vec<f64>>) {
116        if let Some(t_vals) = state.get("t") {
117            self.t = t_vals[0] as usize;
118        }
119        for (key, values) in state {
120            if let Some(name) = key.strip_prefix("m_") {
121                if let Some(m) = self.m.get(name) {
122                    let shape = m.raw_dim();
123                    if let Ok(arr) = Array::from_shape_vec(shape, values) {
124                        self.m.insert(name.to_string(), arr);
125                    }
126                }
127            } else if let Some(name) = key.strip_prefix("v_") {
128                if let Some(v) = self.v.get(name) {
129                    let shape = v.raw_dim();
130                    if let Ok(arr) = Array::from_shape_vec(shape, values) {
131                        self.v.insert(name.to_string(), arr);
132                    }
133                }
134            }
135        }
136    }
137}
138
139#[cfg(test)]
140mod tests {
141    use super::*;
142    use scirs2_core::ndarray::array;
143
144    #[test]
145    fn test_adam_optimizer() {
146        let config = OptimizerConfig {
147            learning_rate: 0.001,
148            ..Default::default()
149        };
150        let mut optimizer = AdamOptimizer::new(config);
151        let mut params = HashMap::new();
152        params.insert("w".to_string(), array![[1.0, 2.0], [3.0, 4.0]]);
153        let mut grads = HashMap::new();
154        grads.insert("w".to_string(), array![[0.1, 0.1], [0.1, 0.1]]);
155        optimizer.step(&mut params, &grads).unwrap();
156        let w = params.get("w").unwrap();
157        assert!(w[[0, 0]] < 1.0);
158    }
159}
tensorlogic_train/optimizers/adam.rs

tensorlogic_train/optimizers/
adam.rs
