Skip to content

Rprop Fix #1421

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

Merged
11 commits merged into from
Dec 23, 2024
Merged

Rprop Fix #1421

Show file tree
Hide file tree
Changes from 7 commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
580dd1f
consistency typo
alinpahontu2912 Dec 6, 2024
bde6db6
print debug
alinpahontu2912 Dec 6, 2024
91079e4
rework step update, remove add max_step and update tests - WIP
alinpahontu2912 Dec 11, 2024
1340fa9
ci test
alinpahontu2912 Dec 12, 2024
1eec3b2
Merge branch 'dotnet:main' into rprop_fix
alinpahontu2912 Dec 12, 2024
ce51507
ci test changes
alinpahontu2912 Dec 16, 2024
732b0d2
update test values
alinpahontu2912 Dec 16, 2024
7d56c40
delete extra spacing
alinpahontu2912 Dec 17, 2024
1ee62e1
change back to pytorch style step size update
alinpahontu2912 Dec 17, 2024
da68f38
remove unused parameter
alinpahontu2912 Dec 18, 2024
a70ac63
Merge branch 'main' into rprop_fix
Dec 20, 2024
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
22 changes: 11 additions & 11 deletions src/TorchSharp/Optimizers/Rprop.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,7 +86,7 @@ public class Rprop : OptimizerHelper
/// <param name="min_step">Minimum allowed step size.</param>
/// <param name="max_step">Maximum allowed step size.</param>
/// <param name="maximize">Maximize the params based on the objective, instead of minimizing.</param>
public Rprop(IEnumerable<Parameter> parameters, double lr = 0.01, double etaminus = 0.5, double etaplus = 1.2, double min_step = 1e-6, double max_step = 50, bool maximize = false)
public Rprop(IEnumerable<Parameter> parameters, double lr = 1e-2, double etaminus = 0.5, double etaplus = 1.2, double min_step = 1e-6, double max_step = 50, bool maximize = false)
Comment thread
alinpahontu2912 marked this conversation as resolved.
: this(new ParamGroup[] { new() { Parameters = parameters } }, lr, etaminus, etaplus, min_step, max_step, maximize)
{
}
Expand Down Expand Up @@ -156,20 +156,20 @@ public override Tensor step(Func<Tensor> closure = null)

state.step += 1;

grad = (max_step != 0)
? grad.add(param, alpha: max_step)
: grad.alias();
var grad_prod = grad.mul(state.prev);
Comment thread
alinpahontu2912 marked this conversation as resolved.
Outdated
Comment thread
alinpahontu2912 marked this conversation as resolved.
Outdated

var sign = grad.mul(state.prev).sign();
sign[sign.gt(0)] = (Tensor)etaplus;
sign[sign.lt(0)] = (Tensor)etaminus;
sign[sign.eq(0)] = (Tensor)1;
var pos_mask = grad_prod.gt(0);
var neg_mask = grad_prod.lt(0);
var zero_mask = grad_prod.eq(0);

state.step_size.mul_(sign).clamp_(min_step, max_step);
var step_size_update = pos_mask.to(torch.float64) * etaplus +
neg_mask.to(torch.float64) * etaminus +
zero_mask.to(torch.float64);

grad = grad.clone();
state.step_size.mul_(step_size_update).clamp_(min_step, max_step);

grad.index_put_(0, sign.eq(etaminus));
grad = grad.clone();
grad.masked_fill_(neg_mask, 0);

param.addcmul_(grad.sign(), state.step_size, -1);

Expand Down
12 changes: 6 additions & 6 deletions test/TorchSharpTest/TestTraining.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1170,7 +1170,7 @@ public void TrainingRprop()

var loss = TrainLoop(seq, x, y, optimizer);

LossIsClose(229.68f, loss);
LossIsClose(77.279f, loss);
}


Expand All @@ -1187,7 +1187,7 @@ public void TrainingRpropMax()

var loss = TrainLoop(seq, x, y, optimizer, maximize:true);

LossIsClose(229.68f, -loss);
LossIsClose(77.279f, -loss);
}

[Fact]
Expand All @@ -1202,8 +1202,8 @@ public void TrainingRpropEtam()
var optimizer = torch.optim.Rprop(seq.parameters(), etaminus: 0.55);

var loss = TrainLoop(seq, x, y, optimizer);

LossIsClose(201.417f, loss);
Comment thread
alinpahontu2912 marked this conversation as resolved.
Outdated
LossIsClose(171.12f, loss);
}

[Fact]
Expand All @@ -1219,7 +1219,7 @@ public void TrainingRpropEtap()

var loss = TrainLoop(seq, x, y, optimizer);

LossIsClose(221.365f, loss);
LossIsClose(65.859f, loss);
}


Expand All @@ -1240,7 +1240,7 @@ public void TrainingRpropParamGroups()

var loss = TrainLoop(seq, x, y, optimizer);

LossIsClose(78.619f, loss);
LossIsClose(66.479f, loss);
}

/// <summary>
Expand Down