burn_core/optim/

adam.rs

use crate::{
    self as burn, LearningRate, grad_clipping::GradientClippingConfig, module::AutodiffModule,
    record::Record,
};

use super::{
    SimpleOptimizer,
    decay::{WeightDecay, WeightDecayConfig},
};
use crate::config::Config;
use crate::optim::adaptor::OptimizerAdaptor;
use crate::tensor::{Tensor, backend::AutodiffBackend};
use burn_tensor::{backend::Backend, ops::Device};

#[cfg(not(feature = "std"))]
use num_traits::Float;

/// Adam configuration.
#[derive(Config)]
pub struct AdamConfig {
    /// Parameter for Adam.
    #[config(default = 0.9)]
    beta_1: f32,
    /// Parameter for Adam.
    #[config(default = 0.999)]
    beta_2: f32,
    /// A value required for numerical stability.
    #[config(default = 1e-5)]
    epsilon: f32,
    /// [Weight decay](WeightDecayConfig) config.
    weight_decay: Option<WeightDecayConfig>,
    /// [Gradient Clipping](GradientClippingConfig) config.
    grad_clipping: Option<GradientClippingConfig>,
}

/// Adam optimizer as described in the paper [Adam: A Method for Stochastic Optimization](https://arxiv.org/pdf/1412.6980.pdf).
#[derive(Clone)]
pub struct Adam {
    momentum: AdaptiveMomentum,
    weight_decay: Option<WeightDecay>,
}

/// Adam state.
#[derive(Record, Clone, new)]
pub struct AdamState<B: Backend, const D: usize> {
    /// The current adaptive momentum.
    pub momentum: AdaptiveMomentumState<B, D>,
}

impl<B: Backend> SimpleOptimizer<B> for Adam {
    type State<const D: usize> = AdamState<B, D>;

    fn step<const D: usize>(
        &self,
        lr: LearningRate,
        tensor: Tensor<B, D>,
        mut grad: Tensor<B, D>,
        state: Option<Self::State<D>>,
    ) -> (Tensor<B, D>, Option<Self::State<D>>) {
        let mut state_momentum = None;

        if let Some(state) = state {
            state_momentum = Some(state.momentum);
        }

        if let Some(weight_decay) = &self.weight_decay {
            grad = weight_decay.transform(grad, tensor.clone());
        }

        let (grad, state_momentum) = self.momentum.transform(grad, state_momentum);

        let state = AdamState::new(state_momentum);
        let delta = grad.mul_scalar(lr);

        (tensor - delta, Some(state))
    }

    fn to_device<const D: usize>(mut state: Self::State<D>, device: &Device<B>) -> Self::State<D> {
        state.momentum = state.momentum.to_device(device);
        state
    }
}

impl AdamConfig {
    /// Initialize Adam optimizer.
    ///
    /// # Returns
    ///
    /// Returns an optimizer that can be used to optimize a module.
    pub fn init<B: AutodiffBackend, M: AutodiffModule<B>>(&self) -> OptimizerAdaptor<Adam, M, B> {
        let optim = Adam {
            momentum: AdaptiveMomentum {
                beta_1: self.beta_1,
                beta_2: self.beta_2,
                epsilon: self.epsilon,
            },
            weight_decay: self.weight_decay.as_ref().map(WeightDecay::new),
        };

        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 AdaptiveMomentumState<B: Backend, const D: usize> {
    /// The number of iterations aggregated.
    pub time: usize,
    /// The first order momentum.
    pub moment_1: Tensor<B, D>,
    /// The second order momentum.
    pub moment_2: Tensor<B, D>,
}

#[derive(Clone)]
struct AdaptiveMomentum {
    beta_1: f32,
    beta_2: f32,
    epsilon: f32,
}

impl AdaptiveMomentum {
    pub fn transform<B: Backend, const D: usize>(
        &self,
        grad: Tensor<B, D>,
        momentum_state: Option<AdaptiveMomentumState<B, D>>,
    ) -> (Tensor<B, D>, AdaptiveMomentumState<B, D>) {
        let state = if let Some(mut state) = momentum_state {
            let factor = 1.0 - self.beta_1;
            state.moment_1 = state
                .moment_1
                .mul_scalar(self.beta_1)
                .add(grad.clone().mul_scalar(factor));

            let factor = 1.0 - self.beta_2;
            state.moment_2 = state
                .moment_2
                .mul_scalar(self.beta_2)
                .add(grad.powi_scalar(2).mul_scalar(factor));

            state.time += 1;

            state
        } else {
            let factor = 1.0 - self.beta_1;
            let moment_1 = grad.clone().mul_scalar(factor);

            let factor = 1.0 - self.beta_2;
            let moment_2 = grad.powi_scalar(2).mul_scalar(factor);

            AdaptiveMomentumState::new(1, moment_1, moment_2)
        };

        let time = state.time as i32;
        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));

        let grad = moment_1_corrected.div(moment_2_corrected.sqrt().add_scalar(self.epsilon));

        (grad, state)
    }
}

impl<B: Backend, const D: usize> AdaptiveMomentumState<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 burn_tensor::Tolerance;
    use burn_tensor::ops::FloatElem;

    use super::*;
    use crate::module::{Module, Param};
    use crate::optim::{GradientsParams, Optimizer};
    use crate::tensor::{Distribution, Tensor, TensorData};
    use crate::{TestAutodiffBackend, nn};

    const LEARNING_RATE: LearningRate = 0.01;

    #[test]
    fn test_adam_optimizer_save_load_state() {
        let device = Default::default();
        let linear = nn::LinearConfig::new(6, 6).init(&device);
        let x = Tensor::<TestAutodiffBackend, 2>::random([2, 6], Distribution::Default, &device);
        let mut optimizer = create_adam();
        let grads = linear.forward(x).backward();
        let grads = GradientsParams::from_grads(grads, &linear);
        let _linear = optimizer.step(LEARNING_RATE, linear, grads);

        #[cfg(feature = "std")]
        {
            use crate::record::{BinFileRecorder, FullPrecisionSettings, Recorder};

            BinFileRecorder::<FullPrecisionSettings>::default()
                .record(
                    optimizer.to_record(),
                    std::env::temp_dir().as_path().join("test_optim_adam"),
                )
                .unwrap();
        }
        #[cfg(not(feature = "std"))]
        {
            use crate::record::{BinBytesRecorder, FullPrecisionSettings, Recorder};

            let result = BinBytesRecorder::<FullPrecisionSettings>::default()
                .record(optimizer.to_record(), ())
                .unwrap();
            assert!(!result.is_empty());
        }

        let state_optim_before = optimizer.to_record();
        let state_optim_before_copy = optimizer.to_record();
        let optimizer = create_adam();
        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());
    }

    #[test]
    fn test_adam_optimizer_with_numbers() {
        let device = Default::default();
        let linear = given_linear_layer(
            TensorData::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],
            ]),
            TensorData::from([-0.3905, 0.0884, -0.0970, 0.1176, 0.1366, 0.0130]),
        );
        let x_1 = Tensor::<TestAutodiffBackend, 2>::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],
            ],
            &device,
        )
        .require_grad();
        let x_2 = Tensor::<TestAutodiffBackend, 2>::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],
            ],
            &device,
        )
        .require_grad();

        let mut optimizer = AdamConfig::new()
            .with_epsilon(1e-8)
            .with_beta_1(0.9)
            .with_beta_2(0.999)
            .with_weight_decay(Some(WeightDecayConfig::new(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 = TensorData::from([
            [-0.340528, 0.118929, 0.384336, 0.300010, 0.066034, 0.047154],
            [
                0.057757, -0.036690, -0.386649, 0.235010, 0.175624, -0.312133,
            ],
            [
                -0.038940, 0.016306, -0.316151, 0.228410, -0.297819, 0.293047,
            ],
            [
                -0.317929, -0.239100, -0.391449, -0.318087, -0.095948, 0.142651,
            ],
            [
                0.310050, -0.235909, 0.351736, -0.192888, 0.359710, -0.050343,
            ],
            [-0.035840, -0.030203, 0.105840, 0.172110, 0.009440, 0.363346],
        ]);
        let bias_expected = TensorData::from([
            -0.410499, 0.068401, -0.116999, 0.097601, 0.116601, -0.006999,
        ]);

        let (weight_updated, bias_updated) = (
            state_updated.weight.to_data(),
            state_updated.bias.unwrap().to_data(),
        );

        type FT = FloatElem<TestAutodiffBackend>;
        let tolerance = Tolerance::absolute(1e-2);
        bias_updated.assert_approx_eq::<FT>(&bias_expected, tolerance);
        weight_updated.assert_approx_eq::<FT>(&weights_expected, tolerance);
    }

    #[test]
    fn test_adam_optimizer_no_nan() {
        let linear = given_linear_layer(
            TensorData::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],
            ]),
            TensorData::from([-0.3905, 0.0884, -0.0970, 0.1176, 0.1366, 0.0130]),
        );

        let x = Tensor::<TestAutodiffBackend, 2>::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],
            ],
            &Default::default(),
        )
        .require_grad();

        let mut optimizer = AdamConfig::new()
            .with_epsilon(1e-8)
            .with_beta_1(0.9)
            .with_beta_2(0.999)
            .with_weight_decay(Some(WeightDecayConfig::new(0.5)))
            .init();

        let grads = linear.forward(x.clone()).backward();
        let grads = GradientsParams::from_grads(grads, &linear);
        let linear = optimizer.step(LEARNING_RATE, linear, grads);

        let grads = linear.forward(x).backward();
        let grads = GradientsParams::from_grads(grads, &linear);
        let linear = optimizer.step(LEARNING_RATE, linear, grads);

        let state_updated = linear.into_record();
        assert!(!state_updated.weight.to_data().as_slice::<f32>().unwrap()[0].is_nan());
    }

    fn given_linear_layer(weight: TensorData, bias: TensorData) -> nn::Linear<TestAutodiffBackend> {
        let device = Default::default();
        let record = nn::LinearRecord {
            weight: Param::from_data(weight, &device),
            bias: Some(Param::from_data(bias, &device)),
        };

        nn::LinearConfig::new(6, 6)
            .init(&device)
            .load_record(record)
    }

    fn create_adam() -> OptimizerAdaptor<Adam, nn::Linear<TestAutodiffBackend>, TestAutodiffBackend>
    {
        let config = AdamConfig::new();
        Adam {
            momentum: AdaptiveMomentum {
                beta_1: config.beta_1,
                beta_2: config.beta_2,
                epsilon: config.epsilon,
            },
            weight_decay: config.weight_decay.as_ref().map(WeightDecay::new),
        }
        .into()
    }
}