Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Refactor/optim #272

Merged
merged 13 commits into from
Apr 5, 2023
Merged

Refactor/optim #272

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
46b74e6
Wip
nathanielsimard Apr 4, 2023
ec7986f
Merge branch 'main' into refactor/optim
nathanielsimard Apr 4, 2023
1437256
Update burn-core
nathanielsimard Apr 4, 2023
dd55207
Adapt burn train
nathanielsimard Apr 4, 2023
cb90875
some fixes
nathanielsimard Apr 4, 2023
382c83e
Fix old checkpoints were not removed
nathanielsimard Apr 4, 2023
8405ac8
Refactoring
nathanielsimard Apr 5, 2023
083ee95
Remove state
nathanielsimard Apr 5, 2023
f4fee4b
Update inference test
nathanielsimard Apr 5, 2023
b6f6e2b
Run clippy
nathanielsimard Apr 5, 2023
b30278a
Cleanup
nathanielsimard Apr 5, 2023
defa717
Fix no std
nathanielsimard Apr 5, 2023
605a0b3
Add comment
nathanielsimard Apr 5, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
some fixes
  • Loading branch information
@nathanielsimard
nathanielsimard committed Apr 4, 2023
commit cb90875d65b0cee209b911691b6225f9c92032da
6 changes: 6 additions & 0 deletions burn-core/src/module/param/id.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,12 @@ impl From<&str> for ParamId {
}
}

impl From<String> for ParamId {
fn from(value: String) -> Self {
Self { value }
}
}

impl Default for ParamId {
fn default() -> Self {
Self::new()
Expand Down
214 changes: 110 additions & 104 deletions burn-core/src/optim/adam.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,107 +153,113 @@ impl AdaptiveMomentum {
}
}

// #[cfg(test)]
// mod tests {
// use super::*;
// use crate::module::{Module, Param};
// use crate::tensor::{Data, Distribution, Tensor};
// use crate::{nn, TestADBackend};
//
// #[test]
// fn test_adam_optimizer_save_load_state() {
// let linear = nn::LinearConfig::new(6, 6).init();
// let x = Tensor::<TestADBackend, 2>::random([2, 6], Distribution::Standard);
// let mut optimizer = Adam::new(&AdamConfig::new(0.01));
// let grads = linear.forward(x).backward();
// let grads = GradientsParams::from_grads(grads, &linear);
// let linear = optimizer.update_module(linear, grads);
//
// let state_optim_before = optimizer.state(&linear);
// let mut optimizer = Adam::new(&AdamConfig::new(0.01));
// optimizer.load(&linear, &state_optim_before).unwrap();
// let state_optim_after = optimizer.state(&linear);
//
// assert_eq!(state_optim_before, state_optim_after);
// }
//
// #[test]
// fn test_adam_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 = Adam::new(
// &AdamConfig::new(0.01)
// .with_epsilon(1e-8)
// .with_beta_1(0.9)
// .with_beta_2(0.999),
// );
//
// let grads = linear.forward(x_1).backward();
// let grads = GradientsParams::from_grads(grads, &linear);
// let linear = optimizer.update_module(linear, grads);
//
// let grads = linear.forward(x_2).backward();
// let grads = GradientsParams::from_grads(grads, &linear);
// let linear = optimizer.update_module(linear, grads);
//
// let state_updated = linear.into_record();
// let state_expected = given_linear_record(
// Data::from([
// [-0.3405, 0.1191, 0.3843, 0.3000, 0.0661, 0.0471],
// [0.0577, -0.0367, -0.3846, 0.2360, 0.1756, -0.3122],
// [-0.0389, 0.0150, -0.3161, 0.2284, -0.2978, 0.2930],
// [-0.3180, -0.2396, -0.3915, -0.3181, -0.0960, 0.1427],
// [0.3100, -0.2365, 0.3517, -0.1929, 0.3597, -0.0504],
// [-0.0358, -0.0303, 0.1059, 0.1721, 0.0095, 0.3634],
// ]),
// Data::from([-0.4105, 0.0684, -0.1170, 0.0976, 0.1166, -0.0070]),
// );
// let (weight_updated, bias_updated) = (
// state_updated.weight.to_data(),
// state_updated.bias.unwrap().to_data(),
// );
// let (weight_expected, bias_expected) = (
// state_expected.weight.to_data(),
// state_expected.bias.unwrap().to_data(),
// );
//
// bias_updated.assert_approx_eq(&bias_expected, 2);
// weight_updated.assert_approx_eq(&weight_expected, 2);
// }
//
// fn given_linear_layer(weight: Data<f32, 2>, bias: Data<f32, 1>) -> nn::Linear<TestADBackend> {
// let linear = nn::LinearConfig::new(6, 6).init();
// let record = given_linear_record(weight, bias);
//
// linear.load_record(record)
// }
//
// fn given_linear_record(
// weight: Data<f32, 2>,
// bias: Data<f32, 1>,
// ) -> nn::LinearRecord<TestADBackend> {
// nn::LinearRecord {
// weight: Param::from(Tensor::from_data(weight)),
// bias: Some(Param::from(Tensor::from_data(bias))),
// }
// }
// }
#[cfg(test)]
mod tests {
use super::*;
use crate::module::{Module, Param};
use crate::optim::{GradientsParams, Optimizer};
use crate::record::DebugRecordSettings;
use crate::tensor::{Data, Distribution, Tensor};
use crate::{nn, TestADBackend};

#[test]
fn test_adam_optimizer_save_load_state() {
let linear = nn::LinearConfig::new(6, 6).init();
let x = Tensor::<TestADBackend, 2>::random([2, 6], Distribution::Standard);
let mut optimizer = AdamConfig::new(0.01).init();
let grads = linear.forward(x).backward();
let grads = GradientsParams::from_grads(grads, &linear);
let _linear = optimizer.step(linear, grads);
optimizer
.to_record()
.record::<DebugRecordSettings>("/tmp/test_optim".into())
.unwrap();

let state_optim_before = optimizer.to_record();
let state_optim_before_copy = optimizer.to_record();
let optimizer = AdamConfig::new(0.01).init::<TestADBackend, nn::Linear<TestADBackend>>();
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 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 = AdamConfig::new(0.01)
.with_epsilon(1e-8)
.with_beta_1(0.9)
.with_beta_2(0.999)
.init();

let grads = linear.forward(x_1).backward();
let grads = GradientsParams::from_grads(grads, &linear);
let linear = optimizer.step(linear, grads);

let grads = linear.forward(x_2).backward();
let grads = GradientsParams::from_grads(grads, &linear);
let linear = optimizer.step(linear, grads);

let state_updated = linear.into_record();
let state_expected = given_linear_record(
Data::from([
[-0.3405, 0.1191, 0.3843, 0.3000, 0.0661, 0.0471],
[0.0577, -0.0367, -0.3846, 0.2360, 0.1756, -0.3122],
[-0.0389, 0.0150, -0.3161, 0.2284, -0.2978, 0.2930],
[-0.3180, -0.2396, -0.3915, -0.3181, -0.0960, 0.1427],
[0.3100, -0.2365, 0.3517, -0.1929, 0.3597, -0.0504],
[-0.0358, -0.0303, 0.1059, 0.1721, 0.0095, 0.3634],
]),
Data::from([-0.4105, 0.0684, -0.1170, 0.0976, 0.1166, -0.0070]),
);
let (weight_updated, bias_updated) = (
state_updated.weight.to_data(),
state_updated.bias.unwrap().to_data(),
);
let (weight_expected, bias_expected) = (
state_expected.weight.to_data(),
state_expected.bias.unwrap().to_data(),
);

bias_updated.assert_approx_eq(&bias_expected, 2);
weight_updated.assert_approx_eq(&weight_expected, 2);
}

fn given_linear_layer(weight: Data<f32, 2>, bias: Data<f32, 1>) -> nn::Linear<TestADBackend> {
let linear = nn::LinearConfig::new(6, 6).init();
let record = given_linear_record(weight, bias);

linear.load_record(record)
}

fn given_linear_record(
weight: Data<f32, 2>,
bias: Data<f32, 1>,
) -> nn::LinearRecord<TestADBackend> {
nn::LinearRecord {
weight: Param::from(Tensor::from_data(weight)),
bias: Some(Param::from(Tensor::from_data(bias))),
}
}
}
Loading