public WeightsAndBias RunEpoch(Example[] trainSet, ISGDModel <WeightsAndBias> model, WeightsAndBias curWeights)
{
foreach (var e in trainSet)
{
model.Update(curWeights, e);
}
return(curWeights);
}
public Params RunEpoch(DSet <Example> trainSet, ISGDModel <Params> model, Params curParams)
{
Params accumulatedParams =
trainSet.Fold(
(wb, e) => { return(model.Update(wb, e)); },
model.AddParams,
curParams);
var averagedParams = model.ScaleParams(accumulatedParams, (float)(1.0 / trainSet.NumPartitions));
return(averagedParams);
}
public float GetAccuracy(DSet <Example> dataset, ISGDModel <Params> model, Params curParams)
{
int hits = dataset.Fold((c, ex) => model.Predict(curParams, ex) == ex.Label ? c + 1 : c, (c1, c2) => c1 + c2, 0);
return((float)hits / dataset.Count());
}
/// <summary>
/// Runs the outer loop of an SGD, parameterized by type of dataset (e.g.: Example[] or DSet<Example>)
/// and type of model Params (currently only WeightsAndBias, but could be anything, such as a big neural net).
/// The epochRunner runs each epoch and decides the strategy to run the epoch (distributed, sequential, multi-threaded),
/// and it is also used to evaluates the accuracy after each 10 epochs.
/// The model is passed to the epochRunner to the the individual updates, and so long as the param types match, any model
/// can be used with any epoch runner. This is why we require each model to be able to average its parameters.
/// To do: Use a validation set based criterion to stop, rather than number of epochs.
/// </summary>
private static Params RunSGD <Dataset, Params>(Dataset trainSet, Dataset testSet, ISGDModel <Params> model, Params initialParams, ISGDEpoch <Dataset, Params> epochRunner, int numEpochs)
{
var weightsAndBias = initialParams;
var sw = System.Diagnostics.Stopwatch.StartNew();
for (int epoch = 0; epoch < numEpochs; epoch++)
{
sw.Restart();
Params epochResult = epochRunner.RunEpoch(trainSet, model, weightsAndBias);
Console.WriteLine($"Epoch {epoch}. took: {sw.Elapsed}");
if (epoch % 10 == 0)
{
sw.Restart();
float trainAcc = epochRunner.GetAccuracy(trainSet, model, epochResult);
Console.WriteLine($"Train Accuracy: {trainAcc}");
if (testSet != null)
{
float testAcc = epochRunner.GetAccuracy(testSet, model, epochResult);
Console.WriteLine($"Test Accuracy: {testAcc}");
}
Console.WriteLine($"Evaluating accuracies took: {sw.Elapsed}.");
Console.WriteLine();
}
weightsAndBias = epochResult;
}
return(weightsAndBias);
}
public float GetAccuracy(Example[] dataset, ISGDModel <WeightsAndBias> model, WeightsAndBias curWeights)
{
int hits = dataset.Sum(ex => model.Predict(curWeights, ex) == ex.Label ? 1 : 0);
return((float)hits / dataset.Length);
}
