boost loss without init of weights

Author: jcapp4

chebai/loss/boost_bce.py

@@ -5,13 +5,13 @@
 import extras.weight_loader as f
 
 
-class BCE_Point_boosting(torch.nn.BCEWithLogitsLoss):
+class BCE_Boosting(torch.nn.BCEWithLogitsLoss):
 
     def __init__(
         self,
         **kwargs
     ):
-        super().__init__(reduction=None,**kwargs)
+        super().__init__(reduction='none',**kwargs)
 
     def forward(
             self,
@@ -22,11 +22,5 @@ def forward(
     )-> torch.Tensor:
         weights = kwargs['weights']
         loss = super().forward(input=input,target=target)
-        weights_tensor = f.create_weight_tensor(weights)
-        loss_scaled = torch.matmul(weights_tensor,loss)
-        return torch.mean(loss_scaled)
-
-        
-        
-        
-        
+        loss_scaled = loss * weights
+        return torch.mean(loss_scaled)

chebai/models/base.py

@@ -7,6 +7,11 @@
 
 from chebai.preprocessing.structures import XYData
 
+import sys
+sys.path.insert(1,'/home/programmer/Bachelorarbeit/python-chebai')
+
+import extras.weight_loader as f
+
 logging.getLogger("pysmiles").setLevel(logging.CRITICAL)
 
 _MODEL_REGISTRY = dict()
@@ -264,6 +269,8 @@ def _execute(
                 loss_kwargs = dict()
                 if self.pass_loss_kwargs:
                     loss_kwargs = loss_kwargs_candidates
+                dict_weights = f.get_weights(data['idents'])
+                loss_kwargs['weights'] = f.create_data_weights(len(data['idents']),data['labels'].size(dim=1),dict_weights,data['idents'])
                 loss_kwargs["current_epoch"] = self.trainer.current_epoch
                 loss = self.criterion(loss_data, loss_labels, **loss_kwargs)
                 if isinstance(loss, tuple):

configs/loss/boost_bce.yml

@@ -0,0 +1 @@
+class_path: chebai.loss.boost_bce.BCE_Boosting

extras/weight_loader.py

@@ -31,7 +31,7 @@ def mock_init_weights(path="/home/programmer/Bachelorarbeit/weights/first_it.csv
 
 #check the size of a csv file given a filter for the second object
 # assumes csv file has a header      
-def get_size(path="/home/programmer/Bachelorarbeit/split/splits.csv",filter=["train","validation"]) -> int:
+def get_size(path="/home/programmer/Bachelorarbeit/split/splits.csv",filter=["train"]) -> int:
     with open(path,'r') as file:
         reader = csv.reader(file)
         size = -1
@@ -57,18 +57,27 @@ def find_weight(path:str,ident:int)-> float:
             if row[0] == str(ident):
                 return float(row[2])
 
-                
-    print(f"{ident} is not in file ")
+    label = find_label(id=ident)
+    print(f"{ident} is not in file with {label} ")
+
+def find_label(id:int,path="/home/programmer/Bachelorarbeit/split/splits.csv")-> str:
+    with open(path,'r') as file:
+        reader = csv.reader(file)
+        for row in reader:
+            if row[0] == str(id):
+                return row[1]
+
 
 #to do 
 # return should be a tuple of weigths matching the sequenece of the target and label tensor
 def create_data_weights(batchsize:int,dim:int,weights:dict[str,float],idents:tuple[int,...])-> torch.tensor:
-    weight = torch.empty(batchsize,dim)
+    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    weight = torch.empty(batchsize,dim,device=device)
     index = 0
     for i in idents:
         w = weights[str(i)]
         for j in range(0,dim):
-            weight[index][j] = w
+            weight[index][j] = float(w)
         index = index + 1
     return weight
 
@@ -90,6 +99,6 @@ def create_weight_tensor(weight:float)-> torch.tensor:
 def create_class_weights()-> torch.tensor:
     pass
 
-# mock_init_weights()
+mock_init_weights()
 # print(get_weights((233713,51990)))