互信息

import numpy as np import matplotlib.pyplot as plt import seaborn as sns import torch import numexpr as en sns.set() X = np.array([1,0,1,1,0]) Y = np.array([1,1,1,0,0]) def checkType(X): if isinstance(X,(np.ndarray,list,tuple)): return torch.FloatTensor(X) elif isinstance(X,(torch.TensorType,torch.FloatType)): return X else: print("Type Error") def NumpyProb(X,symbol,x): n = X.size Lambda = "{}{}{}".format("X",symbol,"x") expr = en.evaluate(Lambda) return (expr).dot(np.ones(n))/n def NumpyEntropy(X): NE = 0 for x in np.unique(X): PX = NumpyProb(X,'==',x) NE += (- PX * np.log2(PX)) return NE NumpyEntropy(X) def TorchEntropy(X): NE = 0 init_X = checkType(X) m = torch.tensor(init_X.size())[0].float() for x in torch.unique(init_X): PX = (init_X == x).sum().float()/m NE += (- PX * np.log2(PX)) return NE TorchEntropy(X) def NumpyJointEntropy(X,Y): Xn,Yn = X.size,Y.size init_XY = np.vstack((X,Y)).T m,n = init_XY.shape enumerate_ = np.array([(x,y) for x in np.unique(X) for y in np.unique(Y)]) temp = np.array([((init_XY==e).dot(np.ones(n))==n).dot(np.ones(m))/m for e in enumerate_]) return (-temp*np.log2(temp)).sum() NumpyJointEntropy(X,Y) def TorchJointEntropy(X,Y): m = X.size comb = lambda SET : np.array([[i,j] for i in SET[::-1] for j in SET]) element = checkType(comb(np.union1d(np.unique(X),np.unique(Y)))) XY = torch.cat([checkType(X)[:,None],checkType(Y)[:,None]],1) PXY = torch.tensor([(((XY == e).float()).mm(torch.ones((2,1)))==2).sum().float()/m for e in element]) return (-PXY*torch.log2(PXY)).sum() TorchJointEntropy(X,Y) def MutualInfo(X,Y,keyelement="X<->Y",keyType='numpy'): if keyType == 'numpy': HX,HY,HXY = NumpyEntropy(X),NumpyEntropy(Y),NumpyJointEntropy(X,Y) elif keyType == "torch": HX,HY,HXY = TorchEntropy(X), TorchEntropy(Y),TorchJointEntropy(X,Y) return {"X<-Y":HX-HXY,"X->Y":HY-HXY,"X<->Y":HX+HY-HXY}[keyelement]

MutualInfo(X,Y,keyelement='X<->Y',keyType='torch')

MutualInfo(X,Y,keyelement='X->Y',keyType='torch')

MutualInfo(X,Y,keyelement='X<-Y',keyType='torch')