burn_optim/optim/
adagrad.rs

1use burn_core as burn;
2
3use crate::RecordState;
4
5use burn::config::Config;
6use burn::tensor::Device;
7use burn::tensor::Tensor;
8
9use super::{
10    Optimizer,
11    decay::{WeightDecay, WeightDecayConfig},
12    module_optimizer::ModuleOptimizer,
13};
14use crate::{LearningRate, grad_clipping::GradientClippingConfig};
15
16/// AdaGrad configuration.
17#[derive(Config, Debug)]
18pub struct AdaGradConfig {
19    #[config(default = 0.)]
20    lr_decay: f64,
21    #[config(default = 1e-5)]
22    epsilon: f32,
23    /// [Weight decay](WeightDecayConfig) config.
24    weight_decay: Option<WeightDecayConfig>,
25    /// [Gradient Clipping](GradientClippingConfig) config.
26    grad_clipping: Option<GradientClippingConfig>,
27}
28
29/// AdaGrad optimizer
30#[derive(Clone)]
31pub struct AdaGrad {
32    lr_decay: LrDecay,
33    weight_decay: Option<WeightDecay>,
34}
35
36/// AdaGrad state.
37#[derive(RecordState, Clone, new)]
38pub struct AdaGradState<const D: usize> {
39    lr_decay: LrDecayState<D>,
40}
41
42impl Optimizer for AdaGrad {
43    type State<const D: usize> = AdaGradState<D>;
44
45    fn step<const D: usize>(
46        &self,
47        lr: LearningRate,
48        tensor: Tensor<D>,
49        mut grad: Tensor<D>,
50        state: Option<Self::State<D>>,
51    ) -> (Tensor<D>, Option<Self::State<D>>) {
52        let mut state_lr_decay = None;
53
54        if let Some(state) = state {
55            state_lr_decay = Some(state.lr_decay);
56        }
57
58        if let Some(weight_decay) = &self.weight_decay {
59            grad = weight_decay.transform(grad, tensor.clone());
60        }
61
62        let (grad, state_lr_decay) = self.lr_decay.transform(grad, lr, state_lr_decay);
63
64        let state = AdaGradState::new(state_lr_decay);
65
66        (tensor - grad, Some(state))
67    }
68
69    fn to_device<const D: usize>(mut state: Self::State<D>, device: &Device) -> Self::State<D> {
70        state.lr_decay = state.lr_decay.to_device(device);
71        state
72    }
73}
74
75impl AdaGradConfig {
76    /// Build an [`AdaGrad`] from the config.
77    pub(crate) fn build(&self) -> AdaGrad {
78        AdaGrad {
79            lr_decay: LrDecay {
80                lr_decay: self.lr_decay,
81                epsilon: self.epsilon,
82            },
83            weight_decay: self.weight_decay.as_ref().map(WeightDecay::new),
84        }
85    }
86
87    /// Initialize AdaGrad optimizer.
88    ///
89    /// # Returns
90    ///
91    /// Returns an optimizer that can be used to optimize a module.
92    pub fn init(&self) -> ModuleOptimizer {
93        let mut optim = ModuleOptimizer::from(self.build());
94        if let Some(config) = &self.grad_clipping {
95            optim = optim.with_grad_clipping(config.init());
96        }
97        optim
98    }
99}
100
101/// Learning rate decay state (also includes sum state).
102#[derive(RecordState, new, Clone)]
103pub struct LrDecayState<const D: usize> {
104    time: usize,
105    sum: Tensor<D>,
106}
107
108#[derive(Clone)]
109struct LrDecay {
110    lr_decay: f64,
111    epsilon: f32,
112}
113
114impl LrDecay {
115    pub fn transform<const D: usize>(
116        &self,
117        grad: Tensor<D>,
118        lr: LearningRate,
119        lr_decay_state: Option<LrDecayState<D>>,
120    ) -> (Tensor<D>, LrDecayState<D>) {
121        let state = if let Some(mut state) = lr_decay_state {
122            state.sum = state.sum.add(grad.clone().square());
123            state.time += 1;
124            state
125        } else {
126            LrDecayState::new(1, grad.clone().square())
127        };
128
129        let new_lr = lr / (1. + (state.time as f64 - 1.) * self.lr_decay);
130
131        let grad = grad
132            .div(state.sum.clone().sqrt().add_scalar(self.epsilon))
133            .mul_scalar(new_lr);
134
135        (grad, state)
136    }
137}
138
139impl<const D: usize> LrDecayState<D> {
140    /// Move state to device.
141    ///
142    /// # Arguments
143    ///
144    /// * `device` - Device to move state to.
145    ///
146    /// # Returns
147    ///
148    /// Returns state moved to device.
149    pub fn to_device(mut self, device: &Device) -> Self {
150        self.sum = self.sum.to_device(device);
151        self
152    }
153}
154
155#[cfg(test)]
156mod tests {
157    use burn::tensor::Tolerance;
158
159    use super::*;
160    use crate::GradientsParams;
161    use burn::module::Param;
162    use burn::tensor::{Distribution, Tensor, TensorData};
163    use burn_nn::{Linear, LinearConfig};
164
165    const LEARNING_RATE: LearningRate = 0.01;
166
167    #[test]
168    fn test_adagrad_optimizer_save_load_state() {
169        let device = Device::default().autodiff();
170        let linear = LinearConfig::new(6, 6).init(&device);
171        let x = Tensor::<2>::random([2, 6], Distribution::Default, &device);
172        let mut optimizer = create_adagrad();
173        let grads = linear.forward(x).backward();
174        let grads = GradientsParams::from_grads(grads, &linear);
175        let _linear = optimizer.step(LEARNING_RATE.into(), linear, grads);
176
177        let bytes = optimizer.into_bytes().unwrap();
178        assert!(!bytes.is_empty());
179
180        #[cfg(feature = "std")]
181        optimizer
182            .save(std::env::temp_dir().as_path().join("test_optim_adagrad"))
183            .unwrap();
184
185        let state_optim_before = optimizer.to_record();
186        let optimizer = create_adagrad().from_bytes(bytes).unwrap();
187        let state_optim_after = optimizer.to_record();
188
189        assert_eq!(state_optim_before.len(), state_optim_after.len());
190    }
191
192    #[test]
193    fn test_adagrad_optimizer_with_numbers() {
194        let device = Device::default().autodiff();
195        let linear = given_linear_layer(
196            TensorData::from([
197                [-0.3206, 0.1374, 0.4043, 0.3200, 0.0859, 0.0671],
198                [0.0777, -0.0185, -0.3667, 0.2550, 0.1955, -0.2922],
199                [-0.0190, 0.0346, -0.2962, 0.2484, -0.2780, 0.3130],
200                [-0.2980, -0.2214, -0.3715, -0.2981, -0.0761, 0.1626],
201                [0.3300, -0.2182, 0.3717, -0.1729, 0.3796, -0.0304],
202                [-0.0159, -0.0120, 0.1258, 0.1921, 0.0293, 0.3833],
203            ]),
204            TensorData::from([-0.3905, 0.0884, -0.0970, 0.1176, 0.1366, 0.0130]),
205            &device,
206        );
207        let x_1 = Tensor::<2>::from_floats(
208            [
209                [0.6294, 0.0940, 0.8176, 0.8824, 0.5228, 0.4310],
210                [0.7152, 0.9559, 0.7893, 0.5684, 0.5939, 0.8883],
211            ],
212            &device,
213        )
214        .require_grad();
215        let x_2 = Tensor::<2>::from_floats(
216            [
217                [0.8491, 0.2108, 0.8939, 0.4433, 0.5527, 0.2528],
218                [0.3270, 0.0412, 0.5538, 0.9605, 0.3195, 0.9085],
219            ],
220            &device,
221        )
222        .require_grad();
223
224        let mut optimizer = AdaGradConfig::new()
225            .with_epsilon(1e-8)
226            .with_lr_decay(0.5)
227            .init();
228
229        let grads = linear.forward(x_1).backward();
230        let grads = GradientsParams::from_grads(grads, &linear);
231        let linear = optimizer.step(LEARNING_RATE.into(), linear, grads);
232
233        let grads = linear.forward(x_2).backward();
234        let grads = GradientsParams::from_grads(grads, &linear);
235        let linear = optimizer.step(LEARNING_RATE.into(), linear, grads);
236
237        let state_updated = linear;
238        let weights_expected = TensorData::from([
239            [-0.334989, 0.123011, 0.389911, 0.305611, 0.071511, 0.052711],
240            [
241                0.066144, -0.030056, -0.378256, 0.243444, 0.183944, -0.303756,
242            ],
243            [
244                -0.033462, 0.020138, -0.310662, 0.233938, -0.292462, 0.298538,
245            ],
246            [
247                -0.312636, -0.236036, -0.386136, -0.312736, -0.090736, 0.147964,
248            ],
249            [
250                0.315896, -0.232304, 0.357596, -0.187004, 0.365496, -0.044504,
251            ],
252            [-0.030305, -0.026405, 0.111395, 0.177695, 0.014895, 0.368895],
253        ]);
254        let bias_expected = TensorData::from([
255            -0.405214, 0.073686, -0.111714, 0.102886, 0.121886, -0.001714,
256        ]);
257
258        let (weight_updated, bias_updated) = (
259            state_updated.weight.val().into_data(),
260            state_updated.bias.unwrap().val().into_data(),
261        );
262
263        let tolerance = Tolerance::absolute(1e-6);
264        bias_updated.assert_approx_eq::<f32>(&bias_expected, tolerance);
265        weight_updated.assert_approx_eq::<f32>(&weights_expected, tolerance);
266    }
267
268    fn given_linear_layer(weight: TensorData, bias: TensorData, device: &Device) -> Linear {
269        Linear {
270            weight: Param::from_data(weight, device),
271            bias: Some(Param::from_data(bias, device)),
272        }
273    }
274
275    fn create_adagrad() -> ModuleOptimizer {
276        let config = AdaGradConfig::new();
277        AdaGrad {
278            lr_decay: LrDecay {
279                lr_decay: config.lr_decay,
280                epsilon: config.epsilon,
281            },
282            weight_decay: config.weight_decay.as_ref().map(WeightDecay::new),
283        }
284        .into()
285    }
286}
burn_optim/optim/adagrad.rs

burn_optim/optim/
adagrad.rs
