WIP: Generic trainer

• Adds a generic trainer class and trainer script
• Moved the model specific stuff to config files; example config added
• uses .to(device) throughout

bob/learn/pytorch/config/generic/oulu_deep_pixbis.py

@@ -137,33 +137,24 @@ optimizer = optim.Adam(filter(lambda p: p.requires_grad, network.parameters()),l
 
 
 def compute_loss(network,img, labels, device):
-	
-	img = img.to(device)
+	"""
+	Compute the losses, given the network, data and labels and 
+	device in which the computation will be performed. 
+	"""
 
-	if not isinstance(labels,list):
+	imagesv = Variable(img['image'].to(device))
 
-		labels=[labels]
+	labelsv_pixel = Variable(labels['pixel_mask'].to(device))
 
-	labelsv=[]
-	for label in labels:
-		labelsv.append(label.to(device))
-	labels=labelsv.copy()
+	labelsv_binary = Variable(labels['binary_target'].to(device))
 
-
-	imagesv = Variable(img)
-
-	labelsv=[]
-	for label in labels:
-		labelsvt = Variable(label) # needs a list here
-		labelsv.append(labelsvt)
-
-	out= network(imagesv)
+	out = network(imagesv)
 
 	beta=0.5
 
-	loss_pixel = criterion_pixel(out[0].squeeze(1),labelsv[0].float())
+	loss_pixel = criterion_pixel(out[0].squeeze(1),labelsv_pixel.float())
 							
-	loss_bce = criterion_bce(out[1],labelsv[1].unsqueeze(1).float())
+	loss_bce = criterion_bce(out[1],labelsv_binary.unsqueeze(1).float())
 						
 	loss = beta*loss_bce + (1.0-beta)*loss_pixel