private async Task<bool> Process(DataBatch batch, TcpResponse tcpResponse)
{
IRawClassifierTrainingContext<NetworkParameters> ctx = await GetOrCreateContext(batch);
if (batch.Vectors.Any())
{
var iterations = batch.PropertyOrDefault("Iterations", 1);
foreach (var i in Enumerable.Range(0, iterations))
{
foreach (var inputOutputPair in batch
.Vectors
.Select(v => v.Split(ctx.Parameters.InputVectorSize))
.RandomOrder()
)
{
ctx.Train(inputOutputPair[1], inputOutputPair[0]);
}
}
}
if (!batch.KeepAlive)
{
if (batch.PropertyOrDefault("SaveOutput", false))
{
var network = (MultilayerNetwork)ctx.Output;
foreach (var prop in batch.Properties)
{
network.Properties[prop.Key] = prop.Value;
}
var name = batch.PropertyOrDefault("Name", batch.Id);
await _blobStore.StoreAsync(name, network);
}
ctx.Output.Save(tcpResponse.Content);
_trainingContexts.Remove(batch.Id);
}
else
{
if (batch.SendResponse)
{
WriteSummaryResponse(ctx, tcpResponse.Content);
}
}
return true;
}
private async Task<IRawClassifierTrainingContext<NetworkParameters>> GetOrCreateContext(DataBatch batch)
{
IRawClassifierTrainingContext<NetworkParameters> ctx;
//lock (_trainingContexts)
{
if (!_trainingContexts.TryGetValue(batch.Id, out ctx))
{
var name = batch.PropertyOrDefault("Name", batch.Id);
var layerSizes = batch.Properties["LayerSizes"].Split(',').Select(s => int.Parse(s)).ToArray();
var networkParams = new NetworkParameters(layerSizes);
try
{
var network = new MultilayerNetwork(networkParams);
await _blobStore.RestoreAsync(name, network);
_trainingContexts[batch.Id] = ctx = _trainingContextFactory.Create(network);
}
catch
{
_trainingContexts[batch.Id] = ctx = _trainingContextFactory.Create(networkParams);
}
}
}
return ctx;
}