ruda-optim 0.21.27

Ruda optimizer updates, gradient state, clipping and learning-rate schedules.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180

use super::TrainingRecord;
use crate::{Adam, AdamConfig, GradientsAccumulator, GradientsParams, Optimizer, adaptor::OptimizerAdaptor,
    SgdConfig, TestAutodiffBackend as B, lr_scheduler::{LrScheduler, step::{StepLrScheduler, StepLrSchedulerConfig}}};
use ruda_model::{module::Param, record::{BinBytesRecorder, FullPrecisionSettings}, tensor::{Tensor, TensorData, Tolerance}};
use ruda_nn::Linear;

type Model = Linear<B>;
type AdamOptimizer = OptimizerAdaptor<Adam, Model, B>;
type Snapshot = TrainingRecord<B, Model, AdamOptimizer, StepLrScheduler, (usize, usize)>;

type MasterOptimizer = OptimizerAdaptor<crate::Fp32MasterOptimizer<crate::AdamW>, Model, B>;
type MixedSnapshot = TrainingRecord<B, Model, MasterOptimizer, StepLrScheduler, (usize, usize)>;
type StoredMixedSnapshot = TrainingRecord<B, Model, MasterOptimizer, StepLrScheduler,
    (ruda_model::module::ModuleDTypeRecord, (usize, usize))>;

fn mixed_model(dtype: ruda_model::tensor::FloatDType) -> Model {
    let device = Default::default();
    Linear {
        weight: Param::from_tensor(Tensor::<B, 2>::from_floats([[0.25], [-0.5]], &device).cast(dtype)),
        bias: Some(Param::from_tensor(Tensor::<B, 1>::from_floats([0.1], &device))),
    }
}

fn accumulate_mixed(model: &Model, accumulator: &mut GradientsAccumulator<Model>, batch: usize) {
    use ruda_model::tensor::{DType, FloatDType};
    let device = Default::default();
    let input = if batch % 2 == 0 { [[1., 0.], [0., 1.]] } else { [[1., 1.], [-1., 2.]] };
    let input = Tensor::<B, 2>::from_floats(input, &device).cast(model.weight.val().dtype());
    let prediction = input.matmul(model.weight.val()).cast(DType::F32)
        + model.bias.as_ref().unwrap().val().unsqueeze::<2>();
    let loss = prediction.square().mean() * 8.;
    accumulator.accumulate_with_dtype(model,
        GradientsParams::from_grads(loss.backward(), model), FloatDType::F32);
}

fn assert_fp32_gradients<const D: usize>(expected: &GradientsParams, actual: &GradientsParams,
    id: ruda_model::module::ParamId) {
    let expected = expected.get::<crate::TestBackend, D>(id).unwrap();
    let actual = actual.get::<crate::TestBackend, D>(id).unwrap();
    assert_eq!(expected.dtype(), ruda_model::tensor::DType::F32);
    assert_eq!(actual.dtype(), ruda_model::tensor::DType::F32);
    expected.to_data().assert_eq(&actual.to_data(), false);
}

#[test]
fn mixed_storage_resume_preserves_fp32_pending_gradients_and_master_updates() {
    use crate::{AdamWConfig, Fp32MasterOptimizer};
    use ruda_model::tensor::{DType, FloatDType};
    let device = Default::default();
    for dtype in [FloatDType::F16, FloatDType::BF16] {
        let config = AdamWConfig::new().with_amsgrad(true);
        let schedule = StepLrSchedulerConfig::new(0.001, 2).with_gamma(0.8);
        let mut model = mixed_model(dtype);
        let mut optimizer: MasterOptimizer = Fp32MasterOptimizer::new(config.build())
            .with_gradient_scale(16.).init();
        let mut scheduler = schedule.init().unwrap();
        let mut accumulator = GradientsAccumulator::new();
        for step in 0..3 {
            accumulate_mixed(&model, &mut accumulator, step * 2);
            accumulate_mixed(&model, &mut accumulator, step * 2 + 1);
            model = optimizer.step(scheduler.step(), model, accumulator.grads());
        }
        accumulate_mixed(&model, &mut accumulator, 6);
        let recorder = BinBytesRecorder::<FullPrecisionSettings>::default();
        let bytes = MixedSnapshot::capture_with_dtypes(
            &model, &optimizer, &scheduler, &accumulator, (3, 1))
            .unwrap().save(&recorder, ()).unwrap();
        let snapshot = StoredMixedSnapshot::load(&recorder, bytes, &device).unwrap();
        let mut restored = snapshot.restore_with_dtypes(self::model(),
            Fp32MasterOptimizer::new(config.build()).with_gradient_scale(16.).init(),
            schedule.init().unwrap(), &device).unwrap();
        assert_eq!(restored.state, (3, 1));
        assert_eq!(restored.model.weight.val().dtype(), dtype.into());
        assert_eq!(restored.model.bias.as_ref().unwrap().val().dtype(), DType::F32);
        assert_eq!(restored.model.weight.id, model.weight.id);
        assert!(restored.model.weight.val().is_require_grad());
        for step in 3..6 {
            if step != 3 {
                accumulate_mixed(&model, &mut accumulator, step * 2);
                accumulate_mixed(&restored.model, &mut restored.accumulator, step * 2);
            }
            accumulate_mixed(&model, &mut accumulator, step * 2 + 1);
            accumulate_mixed(&restored.model, &mut restored.accumulator, step * 2 + 1);
            let lr = scheduler.step();
            assert_eq!(lr, restored.scheduler.step());
            let gradients = accumulator.grads();
            let restored_gradients = restored.accumulator.grads();
            assert_fp32_gradients::<2>(&gradients, &restored_gradients, model.weight.id);
            assert_fp32_gradients::<1>(&gradients, &restored_gradients,
                model.bias.as_ref().unwrap().id);
            model = optimizer.step(lr, model, gradients);
            restored.model = restored.optimizer.step(lr, restored.model, restored_gradients);
            assert_eq!(restored.model.weight.val().dtype(), dtype.into());
            model.weight.val().to_data().assert_eq(&restored.model.weight.val().to_data(), false);
            model.bias.as_ref().unwrap().val().to_data().assert_eq(
                &restored.model.bias.as_ref().unwrap().val().to_data(), false);
        }
    }
}

fn model() -> Model {
    let device = Default::default();
    Linear {
        weight: Param::from_tensor(Tensor::from_floats([[0.25], [-0.5]], &device)),
        bias: Some(Param::from_tensor(Tensor::from_floats([0.1], &device))),
    }
}

fn loss(model: &Model, batch: usize) -> Tensor<B, 1> {
    let device = Default::default();
    let (x, y) = if batch % 2 == 0 {
        ([[1.0, 0.0], [0.0, 1.0]], [[2.5], [-2.5]])
    } else {
        ([[1.0, 1.0], [-1.0, 2.0]], [[-0.5], [-7.5]])
    };
    let difference = model.forward::<2>(Tensor::from_floats(x, &device)) - Tensor::from_floats(y, &device);
    difference.clone().mul(difference).mean()
}

fn accumulate(model: &Model, accumulator: &mut GradientsAccumulator<Model>, batch: usize) {
    accumulator.accumulate(model, GradientsParams::from_grads(loss(model, batch).backward(), model));
}

#[test]
fn linear_mse_sgd_matches_analytic_update() {
    let device = Default::default();
    let model = Linear::<B> { weight: Param::from_tensor(Tensor::zeros([2, 1], &device)), bias: None };
    let x = Tensor::<B, 2>::from_floats([[1.0, 0.0], [0.0, 1.0]], &device);
    let y = Tensor::from_floats([[2.0], [-3.0]], &device);
    let residual = model.forward(x) - y;
    let loss = residual.clone().mul(residual).mean();
    assert!((loss.clone().into_scalar() - 6.5).abs() < 1e-6);
    let gradients = GradientsParams::from_grads(loss.backward(), &model);
    let mut optimizer = SgdConfig::new().init();
    let model = optimizer.step(0.1, model, gradients);
    model.weight.val().to_data().assert_approx_eq::<f32>(&TensorData::from([[0.2], [-0.3]]), Tolerance::absolute(1e-6));
}

#[test]
fn adam_resume_preserves_pending_gradients_and_next_updates() {
    let device = Default::default();
    let mut model = model();
    let initial_loss = loss(&model, 0).into_scalar() + loss(&model, 1).into_scalar();
    let config = AdamConfig::new().with_amsgrad(true);
    let schedule = StepLrSchedulerConfig::new(0.05, 2).with_gamma(0.8);
    let mut optimizer: AdamOptimizer = config.init();
    let mut scheduler = schedule.init().unwrap();
    let mut accumulator = GradientsAccumulator::new();
    for step in 0..3 {
        accumulate(&model, &mut accumulator, step * 2);
        accumulate(&model, &mut accumulator, step * 2 + 1);
        model = optimizer.step(scheduler.step(), model, accumulator.grads());
    }
    accumulate(&model, &mut accumulator, 6);
    let recorder = BinBytesRecorder::<FullPrecisionSettings>::default();
    let bytes = Snapshot::capture(&model, &optimizer, &scheduler, &accumulator, (3, 1))
        .unwrap().save(&recorder, ()).unwrap();
    let snapshot = Snapshot::load(&recorder, bytes, &device).unwrap();
    let mut restored = snapshot.restore(self::model(), config.init(), schedule.init().unwrap(), &device).unwrap();
    assert_eq!(restored.state, (3, 1));
    assert_eq!(model.weight.id, restored.model.weight.id);
    for step in 3..8 {
        if step != 3 {
            accumulate(&model, &mut accumulator, step * 2);
            accumulate(&restored.model, &mut restored.accumulator, step * 2);
        }
        accumulate(&model, &mut accumulator, step * 2 + 1);
        accumulate(&restored.model, &mut restored.accumulator, step * 2 + 1);
        let lr = scheduler.step();
        assert_eq!(lr, restored.scheduler.step());
        model = optimizer.step(lr, model, accumulator.grads());
        restored.model = restored.optimizer.step(lr, restored.model, restored.accumulator.grads());
        model.weight.val().to_data().assert_eq(&restored.model.weight.val().to_data(), false);
        model.bias.as_ref().unwrap().val().to_data()
            .assert_eq(&restored.model.bias.as_ref().unwrap().val().to_data(), false);
        loss(&model, 0).to_data().assert_eq(&loss(&restored.model, 0).to_data(), false);
    }
    let final_loss = loss(&model, 0).into_scalar() + loss(&model, 1).into_scalar();
    assert!(final_loss < initial_loss, "loss did not decrease: {initial_loss} -> {final_loss}");
}