AdamW implementation (#566)

burn-core/Cargo.toml

@@ -63,6 +63,7 @@ rmp-serde = {workspace = true, optional = true}
 serde_json = {workspace = true, features = ["alloc"]}#Default enables std 
 
 [dev-dependencies]
+tempfile = {workspace = true}
 burn-dataset = {path = "../burn-dataset", version = "0.9.0", features = [
   "fake",
 ]}

burn-core/src/optim/adam.rs

@@ -54,11 +54,11 @@ impl<B: Backend> SimpleOptimizer<B> for Adam<B> {
         state: Option<Self::State<D>>,
     ) -> (Tensor<B, D>, Option<Self::State<D>>) {
         let mut state_weight_decay = None;
-        let mut state_momemtum = None;
+        let mut state_momentum = None;
 
         if let Some(state) = state {
             state_weight_decay = state.weight_decay;
-            state_momemtum = Some(state.momentum);
+            state_momentum = Some(state.momentum);
         }
 
         if let Some(weight_decay) = &self.weight_decay {
@@ -67,9 +67,9 @@ impl<B: Backend> SimpleOptimizer<B> for Adam<B> {
             grad = grad_out;
         }
 
-        let (grad, state_momemtum) = self.momentum.transform(grad, state_momemtum);
+        let (grad, state_momentum) = self.momentum.transform(grad, state_momentum);
 
-        let state = AdamState::new(state_weight_decay, state_momemtum);
+        let state = AdamState::new(state_weight_decay, state_momentum);
         let delta = grad.mul_scalar(lr);
 
         (tensor - delta, Some(state))
@@ -127,9 +127,9 @@ impl AdaptiveMomentum {
     pub fn transform<B: Backend, const D: usize>(
         &self,
         grad: Tensor<B, D>,
-        state: Option<AdaptiveMomentumState<B, D>>,
+        momentum_state: Option<AdaptiveMomentumState<B, D>>,
     ) -> (Tensor<B, D>, AdaptiveMomentumState<B, D>) {
-        let state = if let Some(mut state) = state {
+        let state = if let Some(mut state) = momentum_state {
             let factor = 1.0 - self.beta_1;
             state.moment_1 = state
                 .moment_1

burn-core/src/optim/adamw.rs

@@ -0,0 +1,357 @@
+use crate::{
+    self as burn, grad_clipping::GradientClippingConfig, module::ADModule, record::Record,
+    LearningRate,
+};
+use std::marker::PhantomData;
+
+use super::{Optimizer, SimpleOptimizer};
+use crate::config::Config;
+use crate::optim::adaptor::OptimizerAdaptor;
+use crate::tensor::{backend::ADBackend, Tensor};
+use burn_tensor::{backend::Backend, ElementConversion};
+
+/// AdamW configuration.
+#[derive(Config)]
+pub struct AdamWConfig {
+    /// Parameter for AdamW.
+    #[config(default = 0.9)]
+    beta_1: f32,
+    /// Parameter for AdamW.
+    #[config(default = 0.999)]
+    beta_2: f32,
+    /// A value required for numerical stability.
+    #[config(default = 1e-5)]
+    epsilon: f32,
+    /// [Weight decay](WeightDecayConfig) config.
+    #[config(default = 1e-4)]
+    weight_decay: f32,
+    /// [Gradient Clipping](GradientClippingConfig) config.
+    grad_clipping: Option<GradientClippingConfig>,
+}
+
+/// AdamW optimizer as described in the paper [Decoupled Weight Decay Regularization, Loshchilov and Hutter, 2019](https://arxiv.org/abs/1711.05101).
+pub struct AdamW<B: Backend> {
+    momentum: AdaptiveMomentumW,
+    weight_decay: f32,
+    _phantom: PhantomData<B>,
+}
+
+/// AdamW state.
+#[derive(Record, Clone, new)]
+pub struct AdamWState<B: Backend, const D: usize> {
+    momentum: AdaptiveMomentumWState<B, D>,
+}
+
+impl<B: Backend> SimpleOptimizer<B> for AdamW<B> {
+    type State<const D: usize> = AdamWState<B, D>;
+
+    /// A single optimization step for any tensor that represents the parameters of a model.
+    fn step<const D: usize>(
+        &self,
+        // Learning rate.
+        lr: LearningRate,
+        // Any tensor that represents the parameters of a model.
+        tensor: Tensor<B, D>,
+        // Gradient of the loss w.r.t. the parameters.
+        grad: Tensor<B, D>,
+        // State of the optimizer.
+        state: Option<Self::State<D>>,
+    ) -> (Tensor<B, D>, Option<Self::State<D>>) {
+        let tensor_updated = tensor.clone() - tensor.mul_scalar(lr).mul_scalar(self.weight_decay);
+
+        let (raw_delta, momentum_state) = self.momentum.transform(grad, state.map(|s| s.momentum));
+
+        let state = AdamWState {
+            momentum: momentum_state,
+        };
+
+        (tensor_updated - raw_delta.mul_scalar(lr), Some(state))
+    }
+
+    fn to_device<const D: usize>(
+        mut state: Self::State<D>,
+        device: &<B as Backend>::Device,
+    ) -> Self::State<D> {
+        state.momentum = state.momentum.to_device(device);
+        state
+    }
+}
+
+impl AdamWConfig {
+    /// Initialize AdamW optimizer.
+    ///
+    /// # Returns
+    ///
+    /// Returns an optimizer that can be used to optimize a module.
+    pub fn init<B: ADBackend, M: ADModule<B>>(&self) -> impl Optimizer<M, B> {
+        let optim = AdamW {
+            momentum: AdaptiveMomentumW {
+                beta_1: self.beta_1,
+                beta_2: self.beta_2,
+                epsilon: self.epsilon,
+            },
+            weight_decay: self.weight_decay,
+            _phantom: Default::default(),
+        };
+
+        let mut optim = OptimizerAdaptor::from(optim);
+        if let Some(config) = &self.grad_clipping {
+            optim = optim.with_grad_clipping(config.init());
+        }
+        optim
+    }
+}
+
+/// Adaptive momentum state.
+#[derive(Record, new, Clone)]
+pub struct AdaptiveMomentumWState<B: Backend, const D: usize> {
+    time: usize,
+    moment_1: Tensor<B, D>,
+    moment_2: Tensor<B, D>,
+}
+
+struct AdaptiveMomentumW {
+    beta_1: f32,
+    beta_2: f32,
+    epsilon: f32,
+}
+
+impl AdaptiveMomentumW {
+    pub fn transform<B: Backend, const D: usize>(
+        &self,
+        grad: Tensor<B, D>,
+        state: Option<AdaptiveMomentumWState<B, D>>,
+    ) -> (Tensor<B, D>, AdaptiveMomentumWState<B, D>) {
+        let state = if let Some(mut state) = state {
+            // Update first moment estimate.
+            let factor = 1.0 - self.beta_1;
+            state.moment_1 = state
+                .moment_1
+                .mul_scalar(self.beta_1)
+                .add(grad.clone().mul_scalar(factor));
+
+            // Update second moment estimate.
+            let factor = 1.0 - self.beta_2;
+            state.moment_2 = state
+                .moment_2
+                .mul_scalar(self.beta_2)
+                .add(grad.powf(2.0).mul_scalar(factor));
+
+            // Update time.
+            state.time += 1;
+
+            state
+        } else {
+            // Initialize first moment estimate.
+            let factor = 1.0 - self.beta_1;
+            let moment_1 = grad.clone().mul_scalar(factor);
+
+            // Initialize second moment estimate.
+            let factor = 1.0 - self.beta_2;
+            let moment_2 = grad.powf(2.0).mul_scalar(factor);
+
+            AdaptiveMomentumWState::new(0, moment_1, moment_2)
+        };
+
+        let time: i32 = (state.time as i32).elem();
+
+        // Compute bias-corrected first and second moment estimates.
+        let moment_1_corrected = state
+            .moment_1
+            .clone()
+            .div_scalar(1f32 - self.beta_1.powi(time));
+
+        let moment_2_corrected = state
+            .moment_2
+            .clone()
+            .div_scalar(1f32 - self.beta_2.powi(time));
+
+        // Compute update delta. This still needs to be scaled by the learning rate.
+        let update_delta = moment_1_corrected
+            .clone()
+            .div(moment_2_corrected.clone().sqrt().add_scalar(self.epsilon));
+
+        (
+            update_delta,
+            AdaptiveMomentumWState::new(state.time, state.moment_1, state.moment_2),
+        )
+    }
+}
+
+impl<B: Backend, const D: usize> AdaptiveMomentumWState<B, D> {
+    /// Move state to device.
+    ///
+    /// # Arguments
+    ///
+    /// * `device` - Device to move state to.
+    ///
+    /// # Returns
+    ///
+    /// Returns state moved to device.
+    pub fn to_device(mut self, device: &B::Device) -> Self {
+        self.moment_1 = self.moment_1.to_device(device);
+        self.moment_2 = self.moment_2.to_device(device);
+        self
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::module::{Module, Param};
+    use crate::optim::{GradientsParams, Optimizer};
+    use crate::record::{BinFileRecorder, FullPrecisionSettings, Recorder};
+    use crate::tensor::{Data, Distribution, Tensor};
+    use crate::{nn, TestADBackend, TestBackend};
+    use tempfile::TempDir;
+
+    const LEARNING_RATE: LearningRate = 0.01;
+
+    #[test]
+    fn test_adamw_optimizer_save_load_state() {
+        let linear = nn::LinearConfig::new(6, 6).init();
+        let x = Tensor::<TestADBackend, 2>::random([2, 6], Distribution::Default);
+        let mut optimizer = create_adamw();
+        let grads = linear.forward(x).backward();
+        let grads = GradientsParams::from_grads(grads, &linear);
+        let _linear = optimizer.step(LEARNING_RATE, linear, grads);
+        let temp_dir = TempDir::new().unwrap();
+        BinFileRecorder::<FullPrecisionSettings>::default()
+            .record(optimizer.to_record(), temp_dir.path().join("test_optim"))
+            .unwrap();
+
+        let state_optim_before = optimizer.to_record();
+        let state_optim_before_copy = optimizer.to_record();
+        let optimizer = create_adamw();
+        let optimizer = optimizer.load_record(state_optim_before_copy);
+        let state_optim_after = optimizer.to_record();
+
+        assert_eq!(state_optim_before.len(), state_optim_after.len());
+    }
+
+    const ASSERT_PRECISION: usize = 6;
+
+    #[test]
+    fn test_adamw_optimizer_with_numbers() {
+        let linear = given_linear_layer(
+            Data::from([
+                [-0.3206, 0.1374, 0.4043, 0.3200, 0.0859, 0.0671],
+                [0.0777, -0.0185, -0.3667, 0.2550, 0.1955, -0.2922],
+                [-0.0190, 0.0346, -0.2962, 0.2484, -0.2780, 0.3130],
+                [-0.2980, -0.2214, -0.3715, -0.2981, -0.0761, 0.1626],
+                [0.3300, -0.2182, 0.3717, -0.1729, 0.3796, -0.0304],
+                [-0.0159, -0.0120, 0.1258, 0.1921, 0.0293, 0.3833],
+            ]),
+            Data::from([-0.3905, 0.0884, -0.0970, 0.1176, 0.1366, 0.0130]),
+        );
+        let x_1 = Tensor::from_floats([
+            [0.6294, 0.0940, 0.8176, 0.8824, 0.5228, 0.4310],
+            [0.7152, 0.9559, 0.7893, 0.5684, 0.5939, 0.8883],
+        ])
+        .require_grad();
+        let x_2 = Tensor::from_floats([
+            [0.8491, 0.2108, 0.8939, 0.4433, 0.5527, 0.2528],
+            [0.3270, 0.0412, 0.5538, 0.9605, 0.3195, 0.9085],
+        ])
+        .require_grad();
+
+        let mut optimizer = AdamWConfig::new()
+            .with_epsilon(1e-8)
+            .with_beta_1(0.9)
+            .with_beta_2(0.999)
+            .with_weight_decay(0.5)
+            .init();
+
+        let grads = linear.forward(x_1).backward();
+        let grads = GradientsParams::from_grads(grads, &linear);
+        let linear = optimizer.step(LEARNING_RATE, linear, grads);
+
+        let grads = linear.forward(x_2).backward();
+        let grads = GradientsParams::from_grads(grads, &linear);
+        let linear = optimizer.step(LEARNING_RATE, linear, grads);
+
+        let state_updated = linear.into_record();
+        let weights_expected = Data::from([
+            [-0.337295, 0.117827, 0.380358, 0.296868, 0.065232, 0.046534],
+            [
+                0.057032, -0.036518, -0.382951, 0.232516, 0.173738, -0.309182,
+            ],
+            [
+                -0.038703, 0.016052, -0.313155, 0.225982, -0.295039, 0.289981,
+            ],
+            [
+                -0.314920, -0.237394, -0.387704, -0.315067, -0.095153, 0.141081,
+            ],
+            [
+                0.306815, -0.234226, 0.348083, -0.191115, 0.356002, -0.049993,
+            ],
+            [-0.035634, -0.030083, 0.104636, 0.170244, 0.009196, 0.359580],
+        ]);
+        let bias_expected = Data::from([
+            -0.406555, 0.067568, -0.115982, 0.096477, 0.115287, -0.007080,
+        ]);
+
+        let t_state_updated: Tensor<TestADBackend, 2> =
+            Tensor::from_data(state_updated.weight.to_data());
+        let t_state_expected: Tensor<TestADBackend, 2> =
+            Tensor::from_data(weights_expected.clone());
+
+        let t_actual_difference = t_state_updated.sub(t_state_expected);
+        let expected_difference: Tensor<TestADBackend, 2> = Tensor::from_floats([
+            [
+                -0.016695, -0.019573, -0.023942, -0.023132, -0.020668, -0.020566,
+            ],
+            [
+                -0.020668, -0.018018, -0.016251, -0.022484, -0.021762, -0.016982,
+            ],
+            [
+                -0.019703, -0.018548, -0.016955, -0.022418, -0.017039, -0.023019,
+            ],
+            [
+                -0.016920, -0.015994, -0.016204, -0.016967, -0.019053, -0.021519,
+            ],
+            [
+                -0.023185, -0.016026, -0.023617, -0.018215, -0.023598, -0.019593,
+            ],
+            [
+                -0.019734, -0.018083, -0.021164, -0.021856, -0.020104, -0.023720,
+            ],
+        ]);
+
+        t_actual_difference
+            .into_data()
+            .assert_approx_eq(&expected_difference.into_data(), ASSERT_PRECISION);
+
+        let (weight_updated, bias_updated) = (
+            state_updated.weight.to_data(),
+            state_updated.bias.unwrap().to_data(),
+        );
+
+        bias_updated.assert_approx_eq(&bias_expected, ASSERT_PRECISION);
+        weight_updated.assert_approx_eq(&weights_expected, ASSERT_PRECISION);
+    }
+
+    fn given_linear_layer(weight: Data<f32, 2>, bias: Data<f32, 1>) -> nn::Linear<TestADBackend> {
+        let record = nn::LinearRecord {
+            weight: Param::from(Tensor::from_data(weight)),
+            bias: Some(Param::from(Tensor::from_data(bias))),
+        };
+
+        nn::LinearConfig::new(6, 6).init_with(record)
+    }
+
+    fn create_adamw(
+    ) -> OptimizerAdaptor<AdamW<TestBackend>, nn::Linear<TestADBackend>, TestADBackend> {
+        let config = AdamWConfig::new();
+        AdamW {
+            momentum: AdaptiveMomentumW {
+                beta_1: config.beta_1,
+                beta_2: config.beta_2,
+                epsilon: config.epsilon,
+            },
+            weight_decay: config.weight_decay,
+            _phantom: Default::default(),
+        }
+        .into()
+    }
+}

burn-core/src/optim/mod.rs

@@ -5,6 +5,7 @@ pub mod decay;
 pub mod momentum;
 
 mod adam;
+mod adamw;
 mod base;
 mod grad_accum;
 mod grads;
@@ -13,6 +14,7 @@ mod simple;
 mod visitor;
 
 pub use adam::*;
+pub use adamw::*;
 pub use base::*;
 pub use grad_accum::*;
 pub use grads::*;

burn-core/src/optim/momentum.rs

@@ -22,7 +22,7 @@ pub struct MomentumConfig {
 
 /// State of [Momentum](Momentum).
 #[derive(Record, Clone, new)]
-pub struct MomemtumState<B: Backend, const D: usize> {
+pub struct MomentumState<B: Backend, const D: usize> {
     velocity: Tensor<B, D>,
 }
 
@@ -57,8 +57,8 @@ impl<B: Backend> Momentum<B> {
     pub fn transform<const D: usize>(
         &self,
         grad: Tensor<B, D>,
-        state: Option<MomemtumState<B, D>>,
+        state: Option<MomentumState<B, D>>,
-    ) -> (Tensor<B, D>, MomemtumState<B, D>) {
+    ) -> (Tensor<B, D>, MomentumState<B, D>) {
         let velocity = if let Some(state) = state {
             grad.clone()
                 .mul_scalar(1.0 - self.dampening)
@@ -72,11 +72,11 @@ impl<B: Backend> Momentum<B> {
             false => velocity.clone(),
         };
 
-        (grad, MomemtumState::new(velocity))
+        (grad, MomentumState::new(velocity))
     }
 }
 
-impl<B: Backend, const D: usize> MomemtumState<B, D> {
+impl<B: Backend, const D: usize> MomentumState<B, D> {
     /// Moves the state to a device.
     ///
     /// # Arguments

burn-core/src/optim/sgd.rs

@@ -3,7 +3,7 @@ use crate::module::ADModule;
 use crate::{self as burn, LearningRate};
 
 use super::decay::{WeightDecay, WeightDecayConfig, WeightDecayState};
-use super::momentum::{MomemtumState, Momentum, MomentumConfig};
+use super::momentum::{Momentum, MomentumConfig, MomentumState};
 use super::SimpleOptimizer;
 use crate::config::Config;
 use crate::optim::adaptor::OptimizerAdaptor;
@@ -34,7 +34,7 @@ pub struct Sgd<B: Backend> {
 #[derive(Record, Clone, new)]
 pub struct SgdState<B: Backend, const D: usize> {
     weight_decay: Option<WeightDecayState<B, D>>,
-    momentum: Option<MomemtumState<B, D>>,
+    momentum: Option<MomentumState<B, D>>,
 }
 
 impl SgdConfig {