add huber loss

src/sparcscore/ml/plmodels.py

@@ -206,9 +206,7 @@ def training_step(self, batch):
  data, target = batch
  target = target.unsqueeze(1)
  output = self.network(data) # Forward pass, only one output
- loss = F.mse_loss(output, target) # L2 loss
-
- # accuracy metrics for regression???
+ loss = F.huber_loss(output, target, delta=1.0, reduction='mean') # consider looking at parameters again
 
  self.log('loss/train', loss, on_step=False, on_epoch=True, prog_bar=True)
  self.log('mse/train', self.mse(output, target), on_epoch=True, prog_bar=True)
@@ -220,9 +218,7 @@ def validation_step(self, batch):
  data, target = batch
  target = target.unsqueeze(1)
  output = self.network(data)
- loss = F.mse_loss(output, target)
-
- # accuracy metrics for regression???
+ loss = F.huber_loss(output, target, delta=1.0, reduction='mean')
 
  self.log('loss/val', loss, on_step=False, on_epoch=True, prog_bar=True)
  self.log('mse/val', self.mse(output, target), on_epoch=True, prog_bar=True)
@@ -234,7 +230,7 @@ def test_step(self, batch):
  data, target = batch
  target = target.unsqueeze(1)
  output = self.network(data)
- loss = F.mse_loss(output, target)
+ loss = F.huber_loss(output, target, delta=1.0, reduction='mean')
 
  # accuracy metrics for regression???