def _setInternalFeatureStats(self, data):
# features [path length, path depth1, path depth2, sentence length, length between elements]
numFeatures = 5;
self.features_mean = 0;
self.features_std = np.ones(numFeatures);
totaltrees = 0
for s in range(len(data.allSStr)):
nverbs = len(data.verbIndices[s].flatten())
sentLength = len(data.allSStr[s].flatten())
totaltrees += nverbs*sentLength
allFeatures = np.zeros((totaltrees,numFeatures));
thisTree = Tree()
for s in range(len(data.allSStr)):
thisTree.pp = data.allSTree[s].flatten()
verbIndexesList = data.verbIndices[s].flatten()
for vid in verbIndexesList:
for wid in range(len(data.allSStr[s])):
indices = [vid, wid]
fullPath = findPath(thisTree, indices)
allFeatures[s,:] = getInternalFeatures(fullPath, data.allSNum[s], indices, self.features_mean, self.features_std);
# Find the mean
self.features_mean = np.mean(allFeatures,axis=0); # check this
allFeatures = allFeatures - self.features_mean
# Find the standard deviation
self.features_std = np.std(allFeatures, axis=0)
def writeVectors():
vecFileName = config.results_path + "vectors.out"
vecFile = open(vecFileName, 'w')
mats = sio.loadmat(config.corpus_path + 'vars.normalized.100.mat')
We_orig = mats.get('We')
params = sio.loadmat(config.corpus_path + 'params_rae.mat')
W1 = params.get('W1')
W2 = params.get('W2')
b1 = params.get('b1')
We = params.get('We')
b = params.get('b')
W = params.get('W')
hiddenSize = 100
nExamples = 5
print "loading data.."
rnnData_train = RNNDataCorpus()
rnnData_train.load_data_srl(load_file=config.train_data_srl,
nExamples=nExamples)
print 'writing vectors to: ', vecFileName
for ii in range(len(rnnData_train.allSNum)):
sNum = rnnData_train.allSNum[ii]
sStr = rnnData_train.allSStr[ii]
sTree = rnnData_train.allSTree[ii]
sKids = rnnData_train.allSKids[ii]
words_indexed = np.where(sNum >= 0)[0]
#L is only the part of the embedding matrix that is relevant for this sentence
#L is deltaWe
if We.shape[1] != 0:
L = We[:, words_indexed]
words_embedded = We_orig[:, words_indexed] + L
else:
words_embedded = We_orig[:, words_indexed]
# sl = words_embedded.shape[1]
tree = Tree()
tree.pp = all #np.zeros(((2*sl-1),1))
tree.nodeScores = np.zeros(len(sNum))
# tree.nodeNames = np.arange(1,(2*sl-1))
tree.kids = np.zeros((len(sNum), 2))
tree.nodeFeatures = np.zeros((hiddenSize, len(sNum)))
tree.nodeFeatures[:, :len(words_indexed)] = words_embedded
toMerge = np.zeros(shape=(words_indexed.shape), dtype='int32')
toMerge[:] = words_indexed[:]
while len(toMerge) > 1:
# find unpaired bottom leaf pairs (initially words) that share parent
i = -1
foundGoodPair = False
while (not foundGoodPair):
i += 1
if sTree[toMerge[i]] == sTree[toMerge[i + 1]]:
foundGoodPair = True
newParent = sTree[toMerge[i]]
kid1 = toMerge[i]
kid2 = toMerge[i + 1]
tree.kids[newParent, :] = [kid1, kid2]
# set new parent to be possible merge candidate
toMerge[i] = newParent
# delete other kid
toMerge = np.delete(toMerge, i + 1)
c1 = tree.nodeFeatures[:, kid1]
c2 = tree.nodeFeatures[:, kid2]
p = np.tanh(np.dot(W1, c1) + np.dot(W2, c2) + b1.flatten())
tree.nodeFeatures[:, newParent] = p
vec = tree.nodeFeatures[-1]
vecFile.write(" ".join([str(x) for x in vec]) + '\n')
vecFile.close()
print "finished! "
def writeVectors():
vecFileName = config.results_path+"vectors.out"
vecFile = open(vecFileName, 'w')
mats = sio.loadmat(config.corpus_path+'vars.normalized.100.mat')
We_orig = mats.get('We')
params = sio.loadmat(config.corpus_path+'params_rae.mat')
W1 = params.get('W1')
W2 = params.get('W2')
b1 = params.get('b1')
We = params.get('We')
b = params.get('b')
W = params.get('W')
hiddenSize = 100
nExamples = 5
print "loading data.."
rnnData_train = RNNDataCorpus()
rnnData_train.load_data_srl(load_file=config.train_data_srl, nExamples=nExamples)
print 'writing vectors to: ', vecFileName
for ii in range(len(rnnData_train.allSNum)):
sNum = rnnData_train.allSNum[ii]
sStr = rnnData_train.allSStr[ii]
sTree = rnnData_train.allSTree[ii]
sKids = rnnData_train.allSKids[ii]
words_indexed = np.where(sNum >= 0)[0]
#L is only the part of the embedding matrix that is relevant for this sentence
#L is deltaWe
if We.shape[1] != 0:
L = We[:, words_indexed]
words_embedded = We_orig[:, words_indexed] + L;
else :
words_embedded = We_orig[:, words_indexed]
# sl = words_embedded.shape[1]
tree = Tree()
tree.pp = all#np.zeros(((2*sl-1),1))
tree.nodeScores = np.zeros(len(sNum))
# tree.nodeNames = np.arange(1,(2*sl-1))
tree.kids = np.zeros((len(sNum),2))
tree.nodeFeatures = np.zeros((hiddenSize, len(sNum)))
tree.nodeFeatures[:,:len(words_indexed)] = words_embedded;
toMerge = np.zeros(shape=(words_indexed.shape), dtype='int32')
toMerge[:] = words_indexed[:]
while len(toMerge)>1 :
# find unpaired bottom leaf pairs (initially words) that share parent
i=-1;
foundGoodPair = False
while (not foundGoodPair ) :
i += 1
if sTree[toMerge[i]]==sTree[toMerge[i+1]]:
foundGoodPair = True
newParent = sTree[toMerge[i]]
kid1 = toMerge[i]
kid2 = toMerge[i+1]
tree.kids[newParent,:] = [kid1, kid2];
# set new parent to be possible merge candidate
toMerge[i] = newParent;
# delete other kid
toMerge = np.delete(toMerge,i+1)
c1 = tree.nodeFeatures[:,kid1]
c2 = tree.nodeFeatures[:,kid2]
p = np.tanh(np.dot(W1,c1) + np.dot(W2,c2) + b1.flatten())
tree.nodeFeatures[:,newParent] = p;
vec = tree.nodeFeatures[-1]
vecFile.write(" ".join([str(x) for x in vec])+'\n')
vecFile.close()
print "finished! "
def forwardPropTree(W, WO, Wcat, Wv, Wo, sNum,sTree, sStr=None, sNN=None, indicies=None, params=None):
wsz = params.wordSize
r = params.rankWo
words = np.where(sNum>=0)[0]
numTotalNodes = len(sNum)
allV = Wv[:,sNum[words]]
allO = Wo[:,sNum[words]]
thisTree = Tree()
# set tree structure of tree
thisTree.pp = sTree #to check
# set which nodes are leaf nodes
thisTree.isLeafVec = np.zeros(numTotalNodes);
thisTree.isLeafVec[words] = 1;
thisTree.nodeNames = np.arange(len(sTree))
thisTree.nodeLabels = sNum;
# the inputs to the parent
thisTree.ParIn_z = np.zeros((wsz,numTotalNodes)) # empty for leaf nodes
thisTree.ParIn_a = np.zeros((wsz,numTotalNodes))
#node vectors
thisTree.nodeAct_a = np.zeros((wsz, numTotalNodes))
# the new operators
thisTree.nodeOp_A = np.zeros((wsz**2,numTotalNodes))
# the scores for each decision
thisTree.score = np.zeros(numTotalNodes);
# the children of each node (for speed)
thisTree.kids = np.zeros((numTotalNodes,2), dtype='int32');
# initialize the vectors and operators of the words (leaf nodes)
thisTree.nodeAct_a[:,words] = allV;
for thisWordNum in range(len(words)):
diag_a = np.diag(allO[:wsz,thisWordNum])
U = allO[wsz:wsz*(1+r),thisWordNum].reshape(wsz,r)
V = allO[wsz*(1+r):,thisWordNum].reshape(wsz, r)
A = diag_a + np.dot(U, np.transpose(V))
A = A.reshape(wsz**2)
thisTree.nodeOp_A[:, thisWordNum] = A
toMerge = np.zeros(shape=(words.shape), dtype='int32')
toMerge[:] = words[:]
while len(toMerge)>1 :
# find unpaired bottom leaf pairs (initially words) that share parent
i=-1;
foundGoodPair = False
while (not foundGoodPair ) :
i += 1
if sTree[toMerge[i]]==sTree[toMerge[i+1]]:
foundGoodPair = True
newParent = sTree[toMerge[i]]
kid1 = toMerge[i]
kid2 = toMerge[i+1]
thisTree.kids[newParent,:] = [kid1, kid2];
# set new parent to be possible merge candidate
toMerge[i] = newParent;
# delete other kid
toMerge = np.delete(toMerge,i+1)
a = thisTree.nodeAct_a[:,kid1];
A = thisTree.nodeOp_A[:,kid1].reshape(wsz,wsz)
b = thisTree.nodeAct_a[:,kid2];
B = thisTree.nodeOp_A[:,kid2].reshape(wsz,wsz)
l_a = np.dot(B,a)
r_a = np.dot(A,b)
C = np.concatenate((l_a,r_a, np.ndarray([1])))
thisTree.nodeAct_a[:,newParent] = np.tanh(np.dot(W,C))
P_A = (np.dot(WO,np.vstack((A,B)))).reshape(wsz**2)
# save all this for backprop:
thisTree.ParIn_a[:,kid1] = l_a
thisTree.ParIn_a[:,kid2] = r_a
thisTree.nodeOp_A[:,newParent] = P_A
return thisTree
def forwardPropTree(W,
WO,
Wcat,
Wv,
Wo,
sNum,
sTree,
sStr=None,
sNN=None,
indicies=None,
params=None):
wsz = params.wordSize
r = params.rankWo
words = np.where(sNum >= 0)[0]
numTotalNodes = len(sNum)
allV = Wv[:, sNum[words]]
allO = Wo[:, sNum[words]]
thisTree = Tree()
# set tree structure of tree
thisTree.pp = sTree #to check
# set which nodes are leaf nodes
thisTree.isLeafVec = np.zeros(numTotalNodes)
thisTree.isLeafVec[words] = 1
thisTree.nodeNames = np.arange(len(sTree))
thisTree.nodeLabels = sNum
# the inputs to the parent
thisTree.ParIn_z = np.zeros((wsz, numTotalNodes)) # empty for leaf nodes
thisTree.ParIn_a = np.zeros((wsz, numTotalNodes))
#node vectors
thisTree.nodeAct_a = np.zeros((wsz, numTotalNodes))
# the new operators
thisTree.nodeOp_A = np.zeros((wsz**2, numTotalNodes))
# the scores for each decision
thisTree.score = np.zeros(numTotalNodes)
# the children of each node (for speed)
thisTree.kids = np.zeros((numTotalNodes, 2), dtype='int32')
# initialize the vectors and operators of the words (leaf nodes)
thisTree.nodeAct_a[:, words] = allV
for thisWordNum in range(len(words)):
diag_a = np.diag(allO[:wsz, thisWordNum])
U = allO[wsz:wsz * (1 + r), thisWordNum].reshape(wsz, r)
V = allO[wsz * (1 + r):, thisWordNum].reshape(wsz, r)
A = diag_a + np.dot(U, np.transpose(V))
A = A.reshape(wsz**2)
thisTree.nodeOp_A[:, thisWordNum] = A
toMerge = np.zeros(shape=(words.shape), dtype='int32')
toMerge[:] = words[:]
while len(toMerge) > 1:
# find unpaired bottom leaf pairs (initially words) that share parent
i = -1
foundGoodPair = False
while (not foundGoodPair):
i += 1
if sTree[toMerge[i]] == sTree[toMerge[i + 1]]:
foundGoodPair = True
newParent = sTree[toMerge[i]]
kid1 = toMerge[i]
kid2 = toMerge[i + 1]
thisTree.kids[newParent, :] = [kid1, kid2]
# set new parent to be possible merge candidate
toMerge[i] = newParent
# delete other kid
toMerge = np.delete(toMerge, i + 1)
a = thisTree.nodeAct_a[:, kid1]
A = thisTree.nodeOp_A[:, kid1].reshape(wsz, wsz)
b = thisTree.nodeAct_a[:, kid2]
B = thisTree.nodeOp_A[:, kid2].reshape(wsz, wsz)
l_a = np.dot(B, a)
r_a = np.dot(A, b)
C = np.concatenate((l_a, r_a, np.ndarray([1])))
thisTree.nodeAct_a[:, newParent] = np.tanh(np.dot(W, C))
P_A = (np.dot(WO, np.vstack((A, B)))).reshape(wsz**2)
# save all this for backprop:
thisTree.ParIn_a[:, kid1] = l_a
thisTree.ParIn_a[:, kid2] = r_a
thisTree.nodeOp_A[:, newParent] = P_A
return thisTree
