burn_core/optim/
adagrad.rs

1use crate::{
2    self as burn, LearningRate, grad_clipping::GradientClippingConfig, module::AutodiffModule,
3    record::Record,
4};
5
6use super::{
7    SimpleOptimizer,
8    decay::{WeightDecay, WeightDecayConfig},
9};
10use crate::config::Config;
11use crate::optim::adaptor::OptimizerAdaptor;
12use crate::tensor::{Tensor, backend::AutodiffBackend};
13use burn_tensor::{backend::Backend, ops::Device};
14
15/// AdaGrad configuration.
16#[derive(Config)]
17pub struct AdaGradConfig {
18    #[config(default = 0.)]
19    lr_decay: f64,
20    #[config(default = 1e-5)]
21    epsilon: f32,
22    /// [Weight decay](WeightDecayConfig) config.
23    weight_decay: Option<WeightDecayConfig>,
24    /// [Gradient Clipping](GradientClippingConfig) config.
25    grad_clipping: Option<GradientClippingConfig>,
26}
27
28/// AdaGrad optimizer
29#[derive(Clone)]
30pub struct AdaGrad {
31    lr_decay: LrDecay,
32    weight_decay: Option<WeightDecay>,
33}
34
35/// AdaGrad state.
36#[derive(Record, Clone, new)]
37pub struct AdaGradState<B: Backend, const D: usize> {
38    lr_decay: LrDecayState<B, D>,
39}
40
41impl<B: Backend> SimpleOptimizer<B> for AdaGrad {
42    type State<const D: usize> = AdaGradState<B, D>;
43
44    fn step<const D: usize>(
45        &self,
46        lr: LearningRate,
47        tensor: Tensor<B, D>,
48        mut grad: Tensor<B, D>,
49        state: Option<Self::State<D>>,
50    ) -> (Tensor<B, D>, Option<Self::State<D>>) {
51        let mut state_lr_decay = None;
52
53        if let Some(state) = state {
54            state_lr_decay = Some(state.lr_decay);
55        }
56
57        if let Some(weight_decay) = &self.weight_decay {
58            grad = weight_decay.transform(grad, tensor.clone());
59        }
60
61        let (grad, state_lr_decay) = self.lr_decay.transform(grad, lr, state_lr_decay);
62
63        let state = AdaGradState::new(state_lr_decay);
64
65        (tensor - grad, Some(state))
66    }
67
68    fn to_device<const D: usize>(mut state: Self::State<D>, device: &Device<B>) -> Self::State<D> {
69        state.lr_decay = state.lr_decay.to_device(device);
70        state
71    }
72}
73
74impl AdaGradConfig {
75    /// Initialize AdaGrad optimizer.
76    ///
77    /// # Returns
78    ///
79    /// Returns an optimizer that can be used to optimize a module.
80    pub fn init<B: AutodiffBackend, M: AutodiffModule<B>>(
81        &self,
82    ) -> OptimizerAdaptor<AdaGrad, M, B> {
83        let optim = AdaGrad {
84            lr_decay: LrDecay {
85                lr_decay: self.lr_decay,
86                epsilon: self.epsilon,
87            },
88            weight_decay: self.weight_decay.as_ref().map(WeightDecay::new),
89        };
90
91        let mut optim = OptimizerAdaptor::from(optim);
92        if let Some(config) = &self.grad_clipping {
93            optim = optim.with_grad_clipping(config.init());
94        }
95        optim
96    }
97}
98
99/// Learning rate decay state (also includes sum state).
100#[derive(Record, new, Clone)]
101pub struct LrDecayState<B: Backend, const D: usize> {
102    time: usize,
103    sum: Tensor<B, D>,
104}
105
106#[derive(Clone)]
107struct LrDecay {
108    lr_decay: f64,
109    epsilon: f32,
110}
111
112impl LrDecay {
113    pub fn transform<B: Backend, const D: usize>(
114        &self,
115        grad: Tensor<B, D>,
116        lr: LearningRate,
117        lr_decay_state: Option<LrDecayState<B, D>>,
118    ) -> (Tensor<B, D>, LrDecayState<B, D>) {
119        let state = if let Some(mut state) = lr_decay_state {
120            state.sum = state.sum.add(grad.clone().powf_scalar(2.));
121            state.time += 1;
122            state
123        } else {
124            LrDecayState::new(1, grad.clone().powf_scalar(2.))
125        };
126
127        let new_lr = lr / (1. + (state.time as f64 - 1.) * self.lr_decay);
128
129        let grad = grad
130            .div(state.sum.clone().sqrt().add_scalar(self.epsilon))
131            .mul_scalar(new_lr);
132
133        (grad, state)
134    }
135}
136
137impl<B: Backend, const D: usize> LrDecayState<B, D> {
138    /// Move state to device.
139    ///
140    /// # Arguments
141    ///
142    /// * `device` - Device to move state to.
143    ///
144    /// # Returns
145    ///
146    /// Returns state moved to device.
147    pub fn to_device(mut self, device: &B::Device) -> Self {
148        self.sum = self.sum.to_device(device);
149        self
150    }
151}
152
153#[cfg(test)]
154mod tests {
155    use burn_tensor::Tolerance;
156    use burn_tensor::ops::FloatElem;
157
158    use super::*;
159    use crate::module::{Module, Param};
160    use crate::optim::{GradientsParams, Optimizer};
161    use crate::tensor::{Distribution, Tensor, TensorData};
162    use crate::{TestAutodiffBackend, nn, nn::Linear};
163
164    const LEARNING_RATE: LearningRate = 0.01;
165
166    #[test]
167    fn test_adagrad_optimizer_save_load_state() {
168        let device = Default::default();
169        let linear = nn::LinearConfig::new(6, 6).init(&device);
170        let x = Tensor::<TestAutodiffBackend, 2>::random([2, 6], Distribution::Default, &device);
171        let mut optimizer = create_adagrad();
172        let grads = linear.forward(x).backward();
173        let grads = GradientsParams::from_grads(grads, &linear);
174        let _linear = optimizer.step(LEARNING_RATE, linear, grads);
175
176        #[cfg(feature = "std")]
177        {
178            use crate::record::{BinFileRecorder, FullPrecisionSettings, Recorder};
179
180            BinFileRecorder::<FullPrecisionSettings>::default()
181                .record(
182                    optimizer.to_record(),
183                    std::env::temp_dir().as_path().join("test_optim_adagrad"),
184                )
185                .unwrap();
186        }
187        #[cfg(not(feature = "std"))]
188        {
189            use crate::record::{BinBytesRecorder, FullPrecisionSettings, Recorder};
190
191            let result = BinBytesRecorder::<FullPrecisionSettings>::default()
192                .record(optimizer.to_record(), ())
193                .unwrap();
194            assert!(!result.is_empty());
195        }
196
197        let state_optim_before = optimizer.to_record();
198        let state_optim_before_copy = optimizer.to_record();
199        let optimizer = create_adagrad();
200        let optimizer = optimizer.load_record(state_optim_before_copy);
201        let state_optim_after = optimizer.to_record();
202
203        assert_eq!(state_optim_before.len(), state_optim_after.len());
204    }
205
206    #[test]
207    fn test_adagrad_optimizer_with_numbers() {
208        let device = Default::default();
209        let linear = given_linear_layer(
210            TensorData::from([
211                [-0.3206, 0.1374, 0.4043, 0.3200, 0.0859, 0.0671],
212                [0.0777, -0.0185, -0.3667, 0.2550, 0.1955, -0.2922],
213                [-0.0190, 0.0346, -0.2962, 0.2484, -0.2780, 0.3130],
214                [-0.2980, -0.2214, -0.3715, -0.2981, -0.0761, 0.1626],
215                [0.3300, -0.2182, 0.3717, -0.1729, 0.3796, -0.0304],
216                [-0.0159, -0.0120, 0.1258, 0.1921, 0.0293, 0.3833],
217            ]),
218            TensorData::from([-0.3905, 0.0884, -0.0970, 0.1176, 0.1366, 0.0130]),
219        );
220        let x_1 = Tensor::<TestAutodiffBackend, 2>::from_floats(
221            [
222                [0.6294, 0.0940, 0.8176, 0.8824, 0.5228, 0.4310],
223                [0.7152, 0.9559, 0.7893, 0.5684, 0.5939, 0.8883],
224            ],
225            &device,
226        )
227        .require_grad();
228        let x_2 = Tensor::<TestAutodiffBackend, 2>::from_floats(
229            [
230                [0.8491, 0.2108, 0.8939, 0.4433, 0.5527, 0.2528],
231                [0.3270, 0.0412, 0.5538, 0.9605, 0.3195, 0.9085],
232            ],
233            &device,
234        )
235        .require_grad();
236
237        let mut optimizer = AdaGradConfig::new()
238            .with_epsilon(1e-8)
239            .with_lr_decay(0.5)
240            .init();
241
242        let grads = linear.forward(x_1).backward();
243        let grads = GradientsParams::from_grads(grads, &linear);
244        let linear = optimizer.step(LEARNING_RATE, linear, grads);
245
246        let grads = linear.forward(x_2).backward();
247        let grads = GradientsParams::from_grads(grads, &linear);
248        let linear = optimizer.step(LEARNING_RATE, linear, grads);
249
250        let state_updated = linear.into_record();
251        let weights_expected = TensorData::from([
252            [-0.334989, 0.123011, 0.389911, 0.305611, 0.071511, 0.052711],
253            [
254                0.066144, -0.030056, -0.378256, 0.243444, 0.183944, -0.303756,
255            ],
256            [
257                -0.033462, 0.020138, -0.310662, 0.233938, -0.292462, 0.298538,
258            ],
259            [
260                -0.312636, -0.236036, -0.386136, -0.312736, -0.090736, 0.147964,
261            ],
262            [
263                0.315896, -0.232304, 0.357596, -0.187004, 0.365496, -0.044504,
264            ],
265            [-0.030305, -0.026405, 0.111395, 0.177695, 0.014895, 0.368895],
266        ]);
267        let bias_expected = TensorData::from([
268            -0.405214, 0.073686, -0.111714, 0.102886, 0.121886, -0.001714,
269        ]);
270
271        let (weight_updated, bias_updated) = (
272            state_updated.weight.val().into_data(),
273            state_updated.bias.unwrap().val().into_data(),
274        );
275
276        type FT = FloatElem<TestAutodiffBackend>;
277        let tolerance = Tolerance::absolute(1e-6);
278        bias_updated.assert_approx_eq::<FT>(&bias_expected, tolerance);
279        weight_updated.assert_approx_eq::<FT>(&weights_expected, tolerance);
280    }
281
282    fn given_linear_layer(weight: TensorData, bias: TensorData) -> nn::Linear<TestAutodiffBackend> {
283        let device = Default::default();
284        let record = nn::LinearRecord {
285            weight: Param::from_data(weight, &device),
286            bias: Some(Param::from_data(bias, &device)),
287        };
288
289        nn::LinearConfig::new(6, 6)
290            .init(&device)
291            .load_record(record)
292    }
293
294    fn create_adagrad()
295    -> OptimizerAdaptor<AdaGrad, Linear<TestAutodiffBackend>, TestAutodiffBackend> {
296        let config = AdaGradConfig::new();
297        AdaGrad {
298            lr_decay: LrDecay {
299                lr_decay: config.lr_decay,
300                epsilon: config.epsilon,
301            },
302            weight_decay: config.weight_decay.as_ref().map(WeightDecay::new),
303        }
304        .into()
305    }
306}
burn_core/optim/adagrad.rs

burn_core/optim/
adagrad.rs
