def forwardProp(self, allKids, words_embedded, updateWlab, label, theta,
freq):
(W1, W2, W3, W4, Wlab, b1, b2, b3, blab, WL) = self.getParams(theta)
sl = np.size(words_embedded, 1)
sentree = rnntree.rnntree(self.d, sl, words_embedded)
collapsed_sentence = range(sl)
if updateWlab:
temp_label = np.zeros(self.cat)
temp_label[label - 1] = 1.0
nodeUnder = np.ones([2 * sl - 1, 1])
for i in range(
sl, 2 * sl - 1
): # calculate n1, n2 and n1+n2 for each node in the sensentree and store in nodeUnder
kids = allKids[i]
n1 = nodeUnder[kids[0]]
n2 = nodeUnder[kids[1]]
nodeUnder[i] = n1 + n2
cat_size = self.cat
sentree.catDelta = np.zeros([cat_size, 2 * sl - 1])
sentree.catDelta_out = np.zeros([self.d, 2 * sl - 1])
# classifier on single words
for i in range(sl):
sm = softmax(np.dot(Wlab, words_embedded[:, i]) + blab)
lbl_sm = (1 - self.alpha) * (temp_label - sm)
sentree.nodeScores[i] = 1.0 / 2.0 * (np.dot(
lbl_sm, (temp_label - sm)))
sentree.catDelta[:, i] = -np.dot(lbl_sm, softmax_prime(sm))
# sm = sigmoid(self.Wlab*words_embedded + self.blab)
#lbl_sm = (1-self.alpha)*(label[:,np.ones(sl,1)] - sm)
#sentree.nodeScores[:sl] = 1/2*(lbl_sm.*(label(:,ones(sl,1)) - sm))
#sentree.catDelta[:, :sl] = -(lbl_sm).*sigmoid_prime(sm)
for i in range(sl, 2 * sl - 1):
kids = allKids[i]
c1 = sentree.nodeFeatures[:, kids[0]]
c2 = sentree.nodeFeatures[:, kids[1]]
# Eq. [2] in the paper: p = f(W[1][c1 c2] + b[1])
p = tanh(np.dot(W1, c1) + np.dot(W2, c2) + b1)
# See last paragraph in Section 2.3
p_norm1 = p / norm(p)
# Eq. (7) in the paper (for special case of 1d label)
#sm = sigmoid(np.dot(Wlab,p_norm1) + blab)
sm = softmax(np.dot(Wlab, p_norm1) + blab)
beta = 0.5
#lbl_sm = beta * (1.0-self.alpha)*(label - sm)
lbl_sm = beta * (1.0 - self.alpha) * (temp_label - sm)
#lbl_sm = beta * (1.0-self.alpha) * (temp_label-sm)
#sentree.catDelta[:, i] = -softmax_prime(sm)[:,label-1]
#J=-(1.0-self.alpha)*np.log(sm[label-1])
#sentree.catDelta[:, i] = -np.dot(lbl_sm,sigmoid_prime(sm))
sentree.catDelta[:, i] = -np.dot(lbl_sm, softmax_prime(sm))
#J = 1.0/2.0*(np.dot(lbl_sm,(label - sm)))
J = 1.0 / 2.0 * (np.dot(lbl_sm, (temp_label - sm)))
sentree.nodeFeatures[:, i] = p_norm1
sentree.nodeFeatures_unnormalized[:, i] = p
sentree.nodeScores[i] = J
sentree.numkids = nodeUnder
sentree.kids = allKids
else:
# Reconstruction Error
for j in range(sl - 1):
size2 = np.size(words_embedded, 1)
c1 = words_embedded[:, 0:-1]
c2 = words_embedded[:, 1:]
freq1 = freq[0:-1]
freq2 = freq[1:]
p = tanh(
np.dot(W1, c1) + np.dot(W2, c2) +
np.reshape(b1, [self.d, 1]) * ([1] * (size2 - 1)))
p_norm1 = p / np.sqrt(sum(p**2))
y1_unnormalized = tanh(
np.dot(W3, p_norm1) + np.reshape(b2, [self.d, 1]) *
([1] * (size2 - 1)))
y2_unnormalized = tanh(
np.dot(W4, p_norm1) + np.reshape(b3, [self.d, 1]) *
([1] * (size2 - 1)))
y1 = y1_unnormalized / np.sqrt(sum(y1_unnormalized**2))
y2 = y2_unnormalized / np.sqrt(sum(y2_unnormalized**2))
y1c1 = self.alpha * (y1 - c1)
y2c2 = self.alpha * (y2 - c2)
# Eq. (4) in the paper: reconstruction error
J = 1.0 / 2.0 * sum((y1c1) * (y1 - c1) + (y2c2) * (y2 - c2))
# finding the pair with smallest reconstruction error for constructing sentree
J_min = min(J)
J_minpos = np.argmin(J)
sentree.node_y1c1[:, sl + j] = y1c1[:, J_minpos]
sentree.node_y2c2[:, sl + j] = y2c2[:, J_minpos]
sentree.nodeDelta_out1[:, sl + j] = np.dot(
norm1tanh_prime(y1_unnormalized[:, J_minpos]),
y1c1[:, J_minpos])
sentree.nodeDelta_out2[:, sl + j] = np.dot(
norm1tanh_prime(y2_unnormalized[:, J_minpos]),
y2c2[:, J_minpos])
words_embedded = np.delete(words_embedded, J_minpos + 1, 1)
words_embedded[:, J_minpos] = p_norm1[:, J_minpos]
sentree.nodeFeatures[:, sl + j] = p_norm1[:, J_minpos]
sentree.nodeFeatures_unnormalized[:, sl + j] = p[:, J_minpos]
sentree.nodeScores[sl + j] = J_min
sentree.pp[collapsed_sentence[J_minpos]] = sl + j
sentree.pp[collapsed_sentence[J_minpos + 1]] = sl + j
sentree.kids[sl + j, :] = [
collapsed_sentence[J_minpos],
collapsed_sentence[J_minpos + 1]
]
sentree.numkids[sl + j] = sentree.numkids[sentree.kids[
sl + j, 0]] + sentree.numkids[sentree.kids[sl + j, 1]]
freq = np.delete(freq, J_minpos + 1)
freq[J_minpos] = (
sentree.numkids[sentree.kids[sl + j, 0]] * freq1[J_minpos]
+
sentree.numkids[sentree.kids[sl + j, 1]] * freq2[J_minpos]
) / (sentree.numkids[sentree.kids[sl + j, 0]] +
sentree.numkids[sentree.kids[sl + j, 1]])
collapsed_sentence = np.delete(collapsed_sentence,
J_minpos + 1)
collapsed_sentence[J_minpos] = sl + j
return sentree
def backProp(self, sentree, updateWcat, words_embedded, gradW1, gradW2,
gradW3, gradW4, gradWlab, gradb1, gradb2, gradb3, gradblab,
gradL, theta):
(W1, W2, W3, W4, Wlab, b1, b2, b3, blab, WL) = self.getParams(theta)
sl = np.size(words_embedded, 1)
toPopulate = np.array([[2 * sl - 2], [0], [0]])
nodeFeatures = sentree.nodeFeatures
nodeFeatures_unnormalized = sentree.nodeFeatures_unnormalized
W0 = np.zeros([self.d, self.d])
W = np.zeros([self.d, self.d, 3])
W[:, :, 0] = W0
W[:, :, 1] = W1
W[:, :, 2] = W2
DEL = [np.zeros([self.d, 1]), sentree.node_y1c1, sentree.node_y2c2]
while np.size(toPopulate, 1) != 0:
parentNode = toPopulate[:, 0].copy()
mat = W[:, :, parentNode[1]]
delt = DEL[parentNode[1]][:, parentNode[2]]
if parentNode[0] > sl - 1: # Non-leaf?
kids = sentree.kids[parentNode[0], :]
kid1 = [kids[0], 1, parentNode[0]]
kid2 = [kids[1], 2, parentNode[0]]
#toPopulate = np.array([kid1, kid2, toPopulate[:, 1:]])
toPopulate[:, 0] = kid2
toPopulate = np.insert(toPopulate, 0, kid1, 1)
a1_unnormalized = nodeFeatures_unnormalized[:, parentNode[
0]] # unnormalized feature of pp
a1 = nodeFeatures[:, parentNode[0]] # normalized feature of pp
nd1 = sentree.nodeDelta_out1[:, parentNode[0]] # grad c1
nd2 = sentree.nodeDelta_out2[:, parentNode[0]] # grad c2
pd = sentree.parentDelta[:, parentNode[0]]
if updateWcat:
smd = sentree.catDelta[:, parentNode[0]]
gradblab = gradblab + smd
parent_d = np.dot(
norm1tanh_prime(a1_unnormalized),
(np.dot(W3, nd1) + np.dot(W4, nd2) + np.dot(mat, pd) +
np.dot(np.transpose(Wlab), smd) - delt))
gradWlab = gradWlab + np.outer(smd, a1)
else:
parent_d = np.dot(norm1tanh_prime(a1_unnormalized),
(np.dot(W3, nd1) + np.dot(W4, nd2) +
np.dot(mat, pd) - delt))
gradb1 = gradb1 + parent_d
gradb2 = gradb2 + nd1
gradb3 = gradb3 + nd2
sentree.parentDelta[:, toPopulate[0][0]] = parent_d
sentree.parentDelta[:, toPopulate[0][1]] = parent_d
gradW1 = gradW1 + np.outer(parent_d,
nodeFeatures[:, toPopulate[0][0]])
gradW2 = gradW2 + np.outer(parent_d,
nodeFeatures[:, toPopulate[0][1]])
gradW3 = gradW3 + np.outer(nd1, a1)
gradW4 = gradW4 + np.outer(nd2, a1)
else: # leaf
if updateWcat:
gradWlab = gradWlab + np.outer(
sentree.catDelta[:, parentNode[0]],
nodeFeatures[:, parentNode[0]])
gradblab = gradblab + sentree.catDelta[:, parentNode[0]]
gradL[:, toPopulate[0][0]] = gradL[:, toPopulate[0][0]] + (
np.dot(mat, sentree.parentDelta[:, toPopulate[0][0]]) +
np.dot(np.transpose(Wlab),
sentree.catDelta[:, toPopulate[0][0]]) - delt)
else:
gradL[:, toPopulate[0][0]] = gradL[:, toPopulate[0][0]] + (
np.dot(mat, sentree.parentDelta[:, toPopulate[0][0]]) -
delt)
toPopulate = np.delete(toPopulate, 0, 1)
return (gradW1, gradW2, gradW3, gradW4, gradWlab, gradb1, gradb2,
gradb3, gradblab, gradL)
def forwardProp(self,allKids,words_embedded,updateWlab,label,theta,freq):
(W1,W2,W3,W4,Wlab,b1,b2,b3,blab,WL)=self.getParams(theta)
sl=np.size(words_embedded,1)
sentree=rnntree.rnntree(self.d,sl,words_embedded)
collapsed_sentence = range(sl)
if updateWlab:
temp_label=np.zeros(self.cat)
temp_label[label-1]=1.0
nodeUnder = np.ones([2*sl-1,1])
for i in range(sl,2*sl-1): # calculate n1, n2 and n1+n2 for each node in the sensentree and store in nodeUnder
kids = allKids[i]
n1 = nodeUnder[kids[0]]
n2 = nodeUnder[kids[1]]
nodeUnder[i] = n1+n2
cat_size=self.cat
sentree.catDelta = np.zeros([cat_size, 2*sl-1])
sentree.catDelta_out = np.zeros([self.d,2*sl-1])
# classifier on single words
for i in range(sl):
sm = softmax(np.dot(Wlab,words_embedded[:,i]) + blab)
lbl_sm = (1-self.alpha)*(temp_label - sm)
sentree.nodeScores[i] = 1.0/2.0*(np.dot(lbl_sm,(temp_label- sm)))
sentree.catDelta[:, i] = -np.dot(lbl_sm,softmax_prime(sm))
# sm = sigmoid(self.Wlab*words_embedded + self.blab)
#lbl_sm = (1-self.alpha)*(label[:,np.ones(sl,1)] - sm)
#sentree.nodeScores[:sl] = 1/2*(lbl_sm.*(label(:,ones(sl,1)) - sm))
#sentree.catDelta[:, :sl] = -(lbl_sm).*sigmoid_prime(sm)
for i in range(sl,2*sl-1):
kids = allKids[i]
c1 = sentree.nodeFeatures[:,kids[0]]
c2 = sentree.nodeFeatures[:,kids[1]]
# Eq. [2] in the paper: p = f(W[1][c1 c2] + b[1])
p = tanh(np.dot(W1,c1) + np.dot(W2,c2) + b1)
# See last paragraph in Section 2.3
p_norm1 = p/norm(p)
# Eq. (7) in the paper (for special case of 1d label)
#sm = sigmoid(np.dot(Wlab,p_norm1) + blab)
sm=softmax(np.dot(Wlab,p_norm1) + blab)
beta=0.5
#lbl_sm = beta * (1.0-self.alpha)*(label - sm)
lbl_sm = beta * (1.0-self.alpha)*(temp_label - sm)
#lbl_sm = beta * (1.0-self.alpha) * (temp_label-sm)
#sentree.catDelta[:, i] = -softmax_prime(sm)[:,label-1]
#J=-(1.0-self.alpha)*np.log(sm[label-1])
#sentree.catDelta[:, i] = -np.dot(lbl_sm,sigmoid_prime(sm))
sentree.catDelta[:, i] = -np.dot(lbl_sm,softmax_prime(sm))
#J = 1.0/2.0*(np.dot(lbl_sm,(label - sm)))
J = 1.0/2.0*(np.dot(lbl_sm,(temp_label - sm)))
sentree.nodeFeatures[:,i] = p_norm1
sentree.nodeFeatures_unnormalized[:,i] = p
sentree.nodeScores[i] = J
sentree.numkids = nodeUnder
sentree.kids = allKids
else:
# Reconstruction Error
for j in range(sl-1):
size2=np.size(words_embedded,1)
c1 = words_embedded[:,0:-1]
c2 = words_embedded[:,1:]
freq1 = freq[0:-1]
freq2 = freq[1:]
p = tanh(np.dot(W1,c1) + np.dot(W2,c2) + np.reshape(b1,[self.d,1])*([1]*(size2-1)))
p_norm1 =p/np.sqrt(sum(p**2))
y1_unnormalized = tanh(np.dot(W3,p_norm1) + np.reshape(b2,[self.d,1])*([1]*(size2-1)))
y2_unnormalized = tanh(np.dot(W4,p_norm1) + np.reshape(b3,[self.d,1])*([1]*(size2-1)))
y1 = y1_unnormalized/np.sqrt(sum(y1_unnormalized**2))
y2 = y2_unnormalized/np.sqrt(sum(y2_unnormalized**2))
y1c1 = self.alpha*(y1-c1)
y2c2 = self.alpha*(y2-c2)
# Eq. (4) in the paper: reconstruction error
J = 1.0/2.0*sum((y1c1)*(y1-c1) + (y2c2)*(y2-c2))
# finding the pair with smallest reconstruction error for constructing sentree
J_min= min(J)
J_minpos=np.argmin(J)
sentree.node_y1c1[:,sl+j] = y1c1[:,J_minpos]
sentree.node_y2c2[:,sl+j] = y2c2[:,J_minpos]
sentree.nodeDelta_out1[:,sl+j] = np.dot(norm1tanh_prime(y1_unnormalized[:,J_minpos]) , y1c1[:,J_minpos])
sentree.nodeDelta_out2[:,sl+j] = np.dot(norm1tanh_prime(y2_unnormalized[:,J_minpos]) , y2c2[:,J_minpos])
words_embedded=np.delete(words_embedded,J_minpos+1,1)
words_embedded[:,J_minpos]=p_norm1[:,J_minpos]
sentree.nodeFeatures[:, sl+j] = p_norm1[:,J_minpos]
sentree.nodeFeatures_unnormalized[:, sl+j]= p[:,J_minpos]
sentree.nodeScores[sl+j] = J_min
sentree.pp[collapsed_sentence[J_minpos]] = sl+j
sentree.pp[collapsed_sentence[J_minpos+1]] = sl+j
sentree.kids[sl+j,:] = [collapsed_sentence[J_minpos], collapsed_sentence[J_minpos+1]]
sentree.numkids[sl+j] = sentree.numkids[sentree.kids[sl+j,0]] + sentree.numkids[sentree.kids[sl+j,1]]
freq=np.delete(freq,J_minpos+1)
freq[J_minpos] = (sentree.numkids[sentree.kids[sl+j,0]]*freq1[J_minpos] + sentree.numkids[sentree.kids[sl+j,1]]*freq2[J_minpos])/(sentree.numkids[sentree.kids[sl+j,0]]+sentree.numkids[sentree.kids[sl+j,1]])
collapsed_sentence=np.delete(collapsed_sentence,J_minpos+1)
collapsed_sentence[J_minpos]=sl+j
return sentree
def backProp(self,sentree,updateWcat,words_embedded,gradW1,gradW2,gradW3,gradW4,gradWlab,gradb1,gradb2,gradb3,gradblab,gradL,theta):
(W1,W2,W3,W4,Wlab,b1,b2,b3,blab,WL)=self.getParams(theta)
sl=np.size(words_embedded,1)
toPopulate = np.array([[2*sl-2],[0],[0]])
nodeFeatures = sentree.nodeFeatures
nodeFeatures_unnormalized = sentree.nodeFeatures_unnormalized
W0 = np.zeros([self.d,self.d])
W = np.zeros([self.d,self.d,3])
W[:,:,0] = W0
W[:,:,1] = W1
W[:,:,2] = W2
DEL = [np.zeros([self.d,1]), sentree.node_y1c1, sentree.node_y2c2]
while np.size(toPopulate,1)!=0:
parentNode = toPopulate[:,0].copy()
mat = W[:,:,parentNode[1]]
delt = DEL[parentNode[1]][:,parentNode[2]]
if parentNode[0]>sl-1: # Non-leaf?
kids = sentree.kids[parentNode[0],:]
kid1 = [kids[0], 1, parentNode[0]]
kid2 = [kids[1], 2, parentNode[0]]
#toPopulate = np.array([kid1, kid2, toPopulate[:, 1:]])
toPopulate[:,0]=kid2
toPopulate=np.insert(toPopulate,0,kid1,1)
a1_unnormalized = nodeFeatures_unnormalized[:,parentNode[0]] # unnormalized feature of pp
a1 = nodeFeatures[:,parentNode[0]] # normalized feature of pp
nd1 = sentree.nodeDelta_out1[:,parentNode[0]] # grad c1
nd2 = sentree.nodeDelta_out2[:,parentNode[0]] # grad c2
pd = sentree.parentDelta[:,parentNode[0]]
if updateWcat:
smd = sentree.catDelta[:,parentNode[0]];
gradblab =gradblab + smd
parent_d = np.dot(norm1tanh_prime(a1_unnormalized) , (np.dot(W3,nd1) +
np.dot(W4,nd2) +
np.dot(mat,pd) + np.dot(np.transpose(Wlab),smd) - delt))
gradWlab = gradWlab + np.outer(smd,a1)
else:
parent_d = np.dot(norm1tanh_prime(a1_unnormalized) , (np.dot(W3,nd1) +
np.dot(W4,nd2) +
np.dot(mat,pd) - delt))
gradb1 = gradb1 + parent_d
gradb2 = gradb2 + nd1
gradb3 = gradb3 + nd2
sentree.parentDelta[:,toPopulate[0][0]] = parent_d
sentree.parentDelta[:,toPopulate[0][1]] = parent_d
gradW1 = gradW1 + np.outer(parent_d,nodeFeatures[:,toPopulate[0][0]])
gradW2 = gradW2 + np.outer(parent_d,nodeFeatures[:,toPopulate[0][1]])
gradW3 = gradW3 + np.outer(nd1,a1)
gradW4 = gradW4 + np.outer(nd2,a1)
else: # leaf
if updateWcat:
gradWlab = gradWlab + np.outer(sentree.catDelta[:, parentNode[0]], nodeFeatures[:,parentNode[0]])
gradblab = gradblab + sentree.catDelta[:,parentNode[0]]
gradL[:,toPopulate[0][0]] = gradL[:,toPopulate[0][0]] +(np.dot(mat,sentree.parentDelta[:,toPopulate[0][0]]) + np.dot(np.transpose(Wlab),sentree.catDelta[:,toPopulate[0][0]]) - delt)
else:
gradL[:,toPopulate[0][0]] = gradL[:,toPopulate[0][0]] +(np.dot(mat,sentree.parentDelta[:,toPopulate[0][0]]) - delt)
toPopulate=np.delete(toPopulate,0,1)
return (gradW1,gradW2,gradW3,gradW4,gradWlab,gradb1,gradb2,gradb3,gradblab,gradL)
def forwardProp(self,allKids,words_embedded,updateWlab,label,theta,freq):
(W1,W2,W3,W4,Wlab,b1,b2,b3,blab,WL)=self.getParams(theta)
#sl是words_embedded的个数，一句话单词的个数
# allKids一开始没有值，是因为训练之前，语法树本来就没有构建完，树结构是训练完了以后才出现的。但是，allkids内容应该会随着算法的进行而变化
sl=np.size(words_embedded,1)
sentree=rnntree.rnntree(self.d,sl,words_embedded)
collapsed_sentence = range(sl)
# updateWlab主要是获得情感误差，修正情感的权值
# 情感误差也是需要p作为输入的，因此也需要计算出p
if updateWlab:
temp_label=np.zeros(self.cat)
#假设cat ＝ 4， temp_label就是(0,0,0,0)。下面这句话的意思是label对应的位置为1
temp_label[label-1]=1.0
nodeUnder = np.ones([2*sl-1,1])
# 这个for循环是计算出，某个节点底下一共有多少个子节点
# kids存了两个值，分别代表左右孩子。
# 可以推测出，allkids存的东西，allkids[i]代表第i个非叶子节点，allkids[i][0]是左孩子，allkids[i][1]是右孩子
for i in range(sl,2*sl-1): # calculate n1, n2 and n1+n2 for each node in the sensentree and store in nodeUnder
kids = allKids[i]
n1 = nodeUnder[kids[0]]
n2 = nodeUnder[kids[1]]
nodeUnder[i] = n1+n2
cat_size=self.cat
sentree.catDelta = np.zeros([cat_size, 2*sl-1])
sentree.catDelta_out = np.zeros([self.d,2*sl-1])
# classifier on single words
# 处理所有单词，即叶子节点
# 这里有个问题就是，为什么叶子节点也要计算情感误差
for i in range(sl):
sm = softmax(np.dot(Wlab,words_embedded[:,i]) + blab)
#这里不管情感误差是如何计算的，sentree.nodeScores存的是情感误差没错了。
#sentree.catDelta存的什么不清楚，但是和情感误差有关
lbl_sm = (1-self.alpha)*(temp_label - sm)
sentree.nodeScores[i] = 1.0/2.0*(np.dot(lbl_sm,(temp_label- sm)))
sentree.catDelta[:, i] = -np.dot(lbl_sm,softmax_prime(sm))
# sm = sigmoid(self.Wlab*words_embedded + self.blab)
#lbl_sm = (1-self.alpha)*(label[:,np.ones(sl,1)] - sm)
#sentree.nodeScores[:sl] = 1/2*(lbl_sm.*(label(:,ones(sl,1)) - sm))
#sentree.catDelta[:, :sl] = -(lbl_sm).*sigmoid_prime(sm)
#超过sl的部分是单词的父亲节点
for i in range(sl,2*sl-1):
kids = allKids[i]
#c1,c2,是左右孩子的向量
c1 = sentree.nodeFeatures[:,kids[0]]
c2 = sentree.nodeFeatures[:,kids[1]]
# Eq. [2] in the paper: p = f(W[1][c1 c2] + b[1])
#计算p，显然p是个数值，即得分，用于判断哪两个节点合并
p = tanh(np.dot(W1,c1) + np.dot(W2,c2) + b1)
# See last paragraph in Section 2.3
p_norm1 = p/norm(p)
# Eq. (7) in the paper (for special case of 1d label)
#sm = sigmoid(np.dot(Wlab,p_norm1) + blab)
#这里是计算节点的情感标签，sm
sm = softmax(np.dot(Wlab,p_norm1) + blab)
beta=0.5
#lbl_sm = beta * (1.0-self.alpha)*(label - sm)
lbl_sm = beta * (1.0-self.alpha)*(temp_label - sm)
#lbl_sm = beta * (1.0-self.alpha) * (temp_label-sm)
#sentree.catDelta[:, i] = -softmax_prime(sm)[:,label-1]
#J=-(1.0-self.alpha)*np.log(sm[label-1])
#sentree.catDelta[:, i] = -np.dot(lbl_sm,sigmoid_prime(sm))
sentree.catDelta[:, i] = -np.dot(lbl_sm,softmax_prime(sm))
#J = 1.0/2.0*(np.dot(lbl_sm,(label - sm)))
J = 1.0/2.0*(np.dot(lbl_sm,(temp_label - sm)))
sentree.nodeFeatures[:,i] = p_norm1
sentree.nodeFeatures_unnormalized[:,i] = p
sentree.nodeScores[i] = J
sentree.numkids = nodeUnder
sentree.kids = allKids
else:
# 这里主要是计算重构误差
# Reconstruction Error
for j in range(sl-1):
size2=np.size(words_embedded,1)
"""
经过测试，p有多个值
也就不难怪这里c1，c2里面分别存了多个单词的向量
因此，这个算法并不是一个个依次算p的，而是一次性一起算出来p
也因此J的值应该也是有多个值。代表两两单词计算的不同结果。
"""
c1 = words_embedded[:,0:-1] # 去掉最后一个单词
c2 = words_embedded[:,1:] # 去掉第一个单词
freq1 = freq[0:-1]
freq2 = freq[1:]
p = tanh(np.dot(W1,c1) + np.dot(W2,c2) + np.reshape(b1,[self.d,1])*([1]*(size2-1)))
p_norm1 =p/np.sqrt(sum(p**2))
y1_unnormalized = tanh(np.dot(W3,p_norm1) + np.reshape(b2,[self.d,1])*([1]*(size2-1)))
y2_unnormalized = tanh(np.dot(W4,p_norm1) + np.reshape(b3,[self.d,1])*([1]*(size2-1)))
y1 = y1_unnormalized/np.sqrt(sum(y1_unnormalized**2))
y2 = y2_unnormalized/np.sqrt(sum(y2_unnormalized**2))
y1c1 = self.alpha*(y1-c1)
y2c2 = self.alpha*(y2-c2)
# Eq. (4) in the paper: reconstruction error
J = 1.0/2.0*sum((y1c1)*(y1-c1) + (y2c2)*(y2-c2))
# finding the pair with smallest reconstruction error for constructing sentree
J_min= min(J)
J_minpos=np.argmin(J)
"""
只有非叶子节点才会有重构节点，因此，sentree.node_y1c1需要从sl+j开始存y1c1.
"""
sentree.node_y1c1[:,sl+j] = y1c1[:,J_minpos]
sentree.node_y2c2[:,sl+j] = y2c2[:,J_minpos]
sentree.nodeDelta_out1[:,sl+j] = np.dot(norm1tanh_prime(y1_unnormalized[:,J_minpos]) , y1c1[:,J_minpos])
sentree.nodeDelta_out2[:,sl+j] = np.dot(norm1tanh_prime(y2_unnormalized[:,J_minpos]) , y2c2[:,J_minpos])
#一对节点被选中以后，需要删除words_embedded对应的向量
#还要把合成的节点加入words_embedded
words_embedded=np.delete(words_embedded,J_minpos+1,1)
words_embedded[:,J_minpos]=p_norm1[:,J_minpos]
sentree.nodeFeatures[:, sl+j] = p_norm1[:,J_minpos]
sentree.nodeFeatures_unnormalized[:, sl+j]= p[:,J_minpos]
sentree.nodeScores[sl+j] = J_min
# pp存的可能是父节点信息，因为两个孩子拥有同一个父亲
sentree.pp[collapsed_sentence[J_minpos]] = sl+j
sentree.pp[collapsed_sentence[J_minpos+1]] = sl+j
sentree.kids[sl+j,:] = [collapsed_sentence[J_minpos], collapsed_sentence[J_minpos+1]]
sentree.numkids[sl+j] = sentree.numkids[sentree.kids[sl+j,0]] + sentree.numkids[sentree.kids[sl+j,1]]
freq=np.delete(freq,J_minpos+1)
freq[J_minpos] = (sentree.numkids[sentree.kids[sl+j,0]]*freq1[J_minpos] + sentree.numkids[sentree.kids[sl+j,1]]*freq2[J_minpos])/(sentree.numkids[sentree.kids[sl+j,0]]+sentree.numkids[sentree.kids[sl+j,1]])
collapsed_sentence=np.delete(collapsed_sentence,J_minpos+1)
collapsed_sentence[J_minpos]=sl+j
print("@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@")
print(sentree.pp)
print("^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^")
print(sentree.kids)
return sentree
def forwardProp(self,allKids,words_embedded,updateWlab,label,theta,freq):
#allkids存的是所有节点，第i行存第i个节点，列表示第i行节点所包含的子节点
(W1,W2,W3,W4,Wlab,b1,b2,b3,blab,WL)=self.getParams(theta)
#s1可能是词汇表的大小
sl=np.size(words_embedded,1)
sentree=rnntree.rnntree(self.d,sl,words_embedded)
collapsed_sentence = range(sl)
#计算情感误差
if updateWlab:
temp_label=np.zeros(self.cat)
#label表示当前标签，label-1主要是因为list从0开始，即当前标签的位置为1
temp_label[label-1]=1.0
nodeUnder = np.ones([2*sl-1,1])
#n1,n2是kids的子节点数
for i in range(sl,2*sl-1): # calculate n1, n2 and n1+n2 for each node in the sensentree and store in nodeUnder
kids = allKids[i]
n1 = nodeUnder[kids[0]] #左节点
n2 = nodeUnder[kids[1]] #右节点
nodeUnder[i] = n1+n2 #第i个节点的子节点数目
cat_size=self.cat
sentree.catDelta = np.zeros([cat_size, 2*sl-1])
sentree.catDelta_out = np.zeros([self.d,2*sl-1])
# classifier on single words
for i in range(sl):
sm = softmax(np.dot(Wlab,words_embedded[:,i]) + blab)
#这里代码部分计算情感误差和论文不太一样，这里直接用yi-h(x)来表示情感误差
lbl_sm = (1-self.alpha)*(temp_label - sm)
#这里貌似是在计算J
sentree.nodeScores[i] = 1.0/2.0*(np.dot(lbl_sm,(temp_label- sm))) #sentree.nodeScores分为2个部分，这里计算0-s1，下面计算2*s1-1
sentree.catDelta[:, i] = -np.dot(lbl_sm,softmax_prime(sm))
# sm = sigmoid(self.Wlab*words_embedded + self.blab)
#lbl_sm = (1-self.alpha)*(label[:,np.ones(sl,1)] - sm)
#sentree.nodeScores[:sl] = 1/2*(lbl_sm.*(label(:,ones(sl,1)) - sm))
#sentree.catDelta[:, :sl] = -(lbl_sm).*sigmoid_prime(sm)
for i in range(sl,2*sl-1):
#kids,c1,c2 是什么
kids = allKids[i]
c1 = sentree.nodeFeatures[:,kids[0]] #左孩子的词向量
c2 = sentree.nodeFeatures[:,kids[1]] #右孩子的词向量
# Eq. [2] in the paper: p = f(W[1][c1 c2] + b[1])
p = tanh(np.dot(W1,c1) + np.dot(W2,c2) + b1)
# See last paragraph in Section 2.3
p_norm1 = p/norm(p)
# Eq. (7) in the paper (for special case of 1d label)
#sm = sigmoid(np.dot(Wlab,p_norm1) + blab)
sm=softmax(np.dot(Wlab,p_norm1) + blab)
beta=0.5 #论文里面本来是没有beta这个值的
#lbl_sm = beta * (1.0-self.alpha)*(label - sm)
lbl_sm = beta * (1.0-self.alpha)*(temp_label - sm)
#lbl_sm = beta * (1.0-self.alpha) * (temp_label-sm)
#sentree.catDelta[:, i] = -softmax_prime(sm)[:,label-1]
#J=-(1.0-self.alpha)*np.log(sm[label-1])
#sentree.catDelta[:, i] = -np.dot(lbl_sm,sigmoid_prime(sm))
sentree.catDelta[:, i] = -np.dot(lbl_sm,softmax_prime(sm))
#J = 1.0/2.0*(np.dot(lbl_sm,(label - sm)))
J = 1.0/2.0*(np.dot(lbl_sm,(temp_label - sm)))
sentree.nodeFeatures[:,i] = p_norm1
sentree.nodeFeatures_unnormalized[:,i] = p
sentree.nodeScores[i] = J
sentree.numkids = nodeUnder
sentree.kids = allKids
#计算重构误差
else:
# Reconstruction Error
for j in range(sl-1):
size2=np.size(words_embedded,1)
c1 = words_embedded[:,0:-1]
c2 = words_embedded[:,1:]
freq1 = freq[0:-1]
freq2 = freq[1:]
p = tanh(np.dot(W1,c1) + np.dot(W2,c2) + np.reshape(b1,[self.d,1])*([1]*(size2-1)))
p_norm1 =p/np.sqrt(sum(p**2))
#下方y1，y2实际上就是论文的c1，c2，由p分解而来。
y1_unnormalized = tanh(np.dot(W3,p_norm1) + np.reshape(b2,[self.d,1])*([1]*(size2-1)))
y2_unnormalized = tanh(np.dot(W4,p_norm1) + np.reshape(b3,[self.d,1])*([1]*(size2-1)))
y1 = y1_unnormalized/np.sqrt(sum(y1_unnormalized**2))
y2 = y2_unnormalized/np.sqrt(sum(y2_unnormalized**2))
y1c1 = self.alpha*(y1-c1)
y2c2 = self.alpha*(y2-c2)
# Eq. (4) in the paper: reconstruction error：重构误差
#(y1-c1)*(y1-c1)的结果是一个数值
J = 1.0/2.0*sum((y1c1)*(y1-c1) + (y2c2)*(y2-c2))
#这个for循环的下面部分没看懂
# finding the pair with smallest reconstruction error for constructing sentree
#min(J)是什么意思，J是一个值
J_min= min(J)
J_minpos=np.argmin(J)
#重构误差最小的重构向量存入树中（c1'，c2'）
sentree.node_y1c1[:,sl+j] = y1c1[:,J_minpos]
sentree.node_y2c2[:,sl+j] = y2c2[:,J_minpos]
#可能是更新值
sentree.nodeDelta_out1[:,sl+j] = np.dot(norm1tanh_prime(y1_unnormalized[:,J_minpos]) , y1c1[:,J_minpos])
sentree.nodeDelta_out2[:,sl+j] = np.dot(norm1tanh_prime(y2_unnormalized[:,J_minpos]) , y2c2[:,J_minpos])
words_embedded=np.delete(words_embedded,J_minpos+1,1)
words_embedded[:,J_minpos]=p_norm1[:,J_minpos]
sentree.nodeFeatures[:, sl+j] = p_norm1[:,J_minpos]
sentree.nodeFeatures_unnormalized[:, sl+j]= p[:,J_minpos]
sentree.nodeScores[sl+j] = J_min
sentree.pp[collapsed_sentence[J_minpos]] = sl+j
sentree.pp[collapsed_sentence[J_minpos+1]] = sl+j
sentree.kids[sl+j,:] = [collapsed_sentence[J_minpos], collapsed_sentence[J_minpos+1]]
sentree.numkids[sl+j] = sentree.numkids[sentree.kids[sl+j,0]] + sentree.numkids[sentree.kids[sl+j,1]]
freq=np.delete(freq,J_minpos+1)
freq[J_minpos] = (sentree.numkids[sentree.kids[sl+j,0]]*freq1[J_minpos] + sentree.numkids[sentree.kids[sl+j,1]]*freq2[J_minpos])/(sentree.numkids[sentree.kids[sl+j,0]]+sentree.numkids[sentree.kids[sl+j,1]])
collapsed_sentence=np.delete(collapsed_sentence,J_minpos+1)
collapsed_sentence[J_minpos]=sl+j
return sentree