public string TrainModel(IpfsJob job)
{
var tmpInput = Ipfs.Get <JToken>(job.input);
var tmpTarget = Ipfs.Get <JToken>(job.target);
var seq = CreateSequential(job.Model);
var inputData = tmpInput.SelectToken("data").ToObject <float[]>();
var inputShape = tmpInput.SelectToken("shape").ToObject <int[]>();
var inputTensor = controller.floatTensorFactory.Create(_data: inputData, _shape: inputShape, _autograd: true);
var targetData = tmpTarget.SelectToken("data").ToObject <float[]>();
var targetShape = tmpTarget.SelectToken("shape").ToObject <int[]>();
var targetTensor = controller.floatTensorFactory.Create(_data: targetData, _shape: targetShape, _autograd: true);
var grad = controller.floatTensorFactory.Create(_data: new float[] { 1, 1, 1, 1 },
_shape: new int[] { 4, 1 });
Loss loss;
switch (job.config.criterion)
{
case "mseloss":
loss = new MSELoss(this.controller);
break;
case "categorical_crossentropy":
loss = new CategoricalCrossEntropyLoss(this.controller);
break;
case "cross_entropy_loss":
loss = new CrossEntropyLoss(this.controller, 1); // TODO -- real value
break;
case "nll_loss":
loss = new NLLLoss(this.controller);
break;
default:
loss = new MSELoss(this.controller);
break;
}
var optimizer = new SGD(this.controller, seq.getParameters(), job.config.lr, 0, 0);
for (var i = 0; i < job.config.iters; ++i)
{
var pred = seq.Forward(inputTensor);
var l = loss.Forward(pred, targetTensor);
l.Backward();
// TODO -- better batch size
optimizer.Step(100, i);
}
var resultJob = new Ipfs();
var response = resultJob.Write(new IpfsJob(job.input, job.target, seq.GetConfig(), job.config));
return(response.Hash);
}
public void TrainModel(MonoBehaviour owner, string input, string target, IpfsJob job, int modelId)
{
var tmpInput = Ipfs.Get <JToken>(input);
var tmpTarget = Ipfs.Get <JToken>(target);
var seq = CreateSequential(job.Model);
var inputData = tmpInput.SelectToken("data").ToObject <float[]>();
var inputShape = tmpInput.SelectToken("shape").ToObject <int[]>();
var inputTensor = controller.floatTensorFactory.Create(_data: inputData, _shape: inputShape, _autograd: true);
var targetData = tmpTarget.SelectToken("data").ToObject <float[]>();
var targetShape = tmpTarget.SelectToken("shape").ToObject <int[]>();
var targetTensor = controller.floatTensorFactory.Create(_data: targetData, _shape: targetShape, _autograd: true);
var grad = controller.floatTensorFactory.Create(_data: new float[] { 1, 1, 1, 1 },
_shape: new int[] { 4, 1 });
// 10 epochs .. make configurable
for (var i = 0; i < 10; ++i)
{
var pred = seq.Forward(inputTensor);
var loss = pred.Sub(targetTensor).Pow(2);
loss.Backward(grad);
foreach (var p in seq.getParameters())
{
var pTensor = controller.floatTensorFactory.Get(p);
pTensor.Sub(pTensor.Grad, inline: true);
}
}
var resultJob = new Ipfs();
var config = new IpfsJobConfig(job.config.lr);
var response = resultJob.Write(new IpfsJob(seq.GetConfig(), config));
var req = new Request();
owner.StartCoroutine(req.AddWeights(owner, modelId, response.Hash));
}
