|
| 1 | +importtorch |
| 2 | + |
| 3 | +classMeanSquareLoss: |
| 4 | +def__init__(self):pass |
| 5 | + |
| 6 | +defloss(self,y,y_pred): |
| 7 | +returntorch.sum(torch.power((y-y_pred),2),dim=1)/y.shape[0] |
| 8 | + |
| 9 | +defgradient(self,y,y_pred): |
| 10 | +return-(y-y_pred) |
| 11 | + |
| 12 | +classCrossEntropy: |
| 13 | +def__init__(self):pass |
| 14 | + |
| 15 | +defloss(self,y,p): |
| 16 | +# Avoid division by zero |
| 17 | +p=np.clip(p,1e-15,1-1e-15) |
| 18 | +return-y*torch.log(p)- (1-y)*torch.log(1-p) |
| 19 | + |
| 20 | +defgradient(self,y,p): |
| 21 | +# Avoid division by zero |
| 22 | +p=torch.clip(p,1e-15,1-1e-15) |
| 23 | +return- (y/p)+ (1-y)/ (1-p) |
| 24 | + |
| 25 | +classMeanAbsoluteLoss: |
| 26 | +def__init__(self):pass |
| 27 | + |
| 28 | +defloss(self,y,y_pred): |
| 29 | +returntorch.sum(torch.abs(y-y_pred),dim=1)/y.shape[0] |
| 30 | + |
| 31 | +defgradient(self,y,y_pred): |
| 32 | +return-(y-y_pred) |
| 33 | + |
| 34 | +classHuberLoss: |
| 35 | +def__init__(self):pass |
| 36 | + |
| 37 | +defloss(self,y,y_pred,delta): |
| 38 | +iftorch.abs(y-y_pred)<=delta: |
| 39 | +return0.5*torch.pow(y-y_pred,2) |
| 40 | +else: |
| 41 | +return (delta*torch.abs(y-y_pred))- (0.5*torch.pow(delta,2)) |
| 42 | + |
| 43 | +classHingeLoss: |
| 44 | +def__init__(self): |
| 45 | +pass |
| 46 | + |
| 47 | +defloss(self,y,y_pred): |
| 48 | +returntorch.max(0, (1-y)*y_pred).values |
| 49 | + |
| 50 | +classKLDivergence: |
| 51 | +def__init__(self): |
| 52 | +pass |
| 53 | + |
| 54 | +defloss(self,y,y_pred): |
| 55 | +returntorch.sum(y_pred*torch.log((y_pred/y))) |