def calculateError_validation(path_Folder, mode, WV, dim, W1 , W2, OutputFile, WSat, WNu, iteration, activationFunc):
posFileList_test = os.listdir(path_Folder + "pos/")
negFileList_test = os.listdir(path_Folder + "neg/")
numberofSamples_test = min(len(posFileList_test), len(negFileList_test))
sumErr=0.0
for k in range(0, numberofSamples_test):
path_File_test = path_Folder + "pos/" + posFileList_test[k]
EDUs = readTree_att_NSWeight(path_File_test, W1, WV, dim, WSat, WNu, activationFunc)
y = [1.0, 0]
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
y_in = feedforward(input2, W2)
output = feedforward_act(input2, W2, activationFunc)
sumErr += MSE(y, output)
path_File_test = path_Folder + "neg/" + negFileList_test[k]
#print negFileList_test[k]
EDUs = readTree_att_NSWeight(path_File_test, W1, WV, dim, WSat, WNu, activationFunc)
y = [0, 1.0]
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
y_in = feedforward(input2, W2)
output = feedforward_act(input2, W2, activationFunc)
sumErr += MSE(y, output)
totall_Err = sumErr/(2*numberofSamples_test)
return totall_Err
print iteration, " ", mode , " ", totall_Err
OutputFile.write("%s,%s,%s\n" % (iteration, mode, totall_Err))
def calculateError_validation_EDUs(allEDUs, mode, WV, dim, W1 , W2, OutputFile, WSat, WNu, iteration, activationFunc):
target_list = np.zeros([0, 2])
output_list = np.zeros([0, 2])
EDU_key = allEDUs.keys()
for EDUid in EDU_key:
EDUs = allEDUs [EDUid]
EDUs = update_EDU(EDUs, W1, WSat, WNu, dim, activationFunc)
if (len(EDUs) > 0 and EDUid>0):
y = [1.0, 0]
#eduKeys = sortEduKey(EDUs.keys(), reverse=True)
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
output = feedforward_act(input2, W2, global_outputActivation)
target_list = np.concatenate((target_list, [y]), 0)
output_list = np.concatenate((output_list, [output]), 0)
if (len(EDUs) > 0 and EDUid < 0):
y = [0, 1.0]
# eduKeys = sortEduKey(EDUs.keys(), reverse=True)
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
#y_in = feedforward(input2, W2)
output = feedforward_act(input2, W2, global_outputActivation)
target_list = np.concatenate((target_list, [y]), 0)
output_list = np.concatenate((output_list, [output]), 0)
totall_Err = MSE(target_list, output_list)
print(iteration, " ", mode , " ", totall_Err)
OutputFile.write("%s,%s,%s\n" % (iteration, mode, totall_Err))
return totall_Err
def test_AttWeight(path_Folder, mode, WV, dim, W1 , W2, OutputFile, WSat, WNu, iteration, activationFunc):
posFileList_test = os.listdir(path_Folder + "pos/")
negFileList_test = os.listdir(path_Folder + "neg/")
numberofSamples_test = min(len(posFileList_test), len(negFileList_test))
#numberofSamples_test=100
tp = 0
fp = 0
tn = 0
fn = 0
for k in range(0, numberofSamples_test):
path_File_test = path_Folder + "pos/" + posFileList_test[k]
#print posFileList_test[k]
EDUs = readTree_att_NSWeight(path_File_test, W1, WV, dim, WSat, WNu, activationFunc)
y = [1.0, 0]
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
# print "pos"
# print input2
y_in = feedforward(input2, W2)
output = feedforward_act(input2, W2, activationFunc)
#print output
if output[0] > output[1]:
tp += 1
else:
fn += 1
path_File_test = path_Folder + "neg/" + negFileList_test[k]
#print negFileList_test[k]
EDUs = readTree_att_NSWeight(path_File_test, W1, WV, dim, WSat, WNu, activationFunc)
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
# print "neg"
# print input2
y_in = feedforward(input2, W2)
output = feedforward_act(input2, W2, activationFunc)
if output[0] < output[1]:
tn += 1
else:
fp += 1
accuracy = float(tp + tn) / (tp + tn + fp + fn)
if (tp+fp) == 0:
precision = 0
else:
precision = float(tp) / (tp + fp)
recall = float(tp) / (tp + fn)
if (precision + recall) == 0:
F1 =0
else:
F1 = 2 * (float(precision * recall)) / (precision + recall)
print iteration, " ", mode , " ", tp, " ", tn, " ", fp, " ", fn, " ", accuracy, " ", precision, " ", recall, " ", F1
OutputFile.write("%s,%s,%s,%s,%s,%s,%s,%s,%s,%s\n" % (iteration, mode, tp, tn, fp, fn, accuracy, precision, recall, F1))
def train_for_each_Sample_AttWeight (EDUs, EDUs_test, y, W1, W21, W22, eta, dim, activationFunc, dropOutPercent):
W1_copy = W1.copy()
W21_copy = W21.copy()
W22_copy = W22.copy()
indexNode = dropOut(len(W1[0]), dropOutPercent)
#W1_doc = dropcolrow(W1, indexNode, False)
#indexNode = dropOut(len(W1_query[0]), dropOutPercent)
W1 = dropcolrow(W1, indexNode, False)
indexNode2 = []
indexNode2.extend(indexNode)
indexNode2.extend(indexNode*2)
W21 = dropcolrow(W21, indexNode2, True)
eduKeys_test = sortEduKey(EDUs_test.keys(), reverse=True)
input2_test = EDUs_test[str(eduKeys_test[0])].vector
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
input = (np.concatenate([input2, input2_test], 0))
#y_in1 = feedforward(input, W21)
output1 = feedforward_act(input, W21, activationFunc)
y_in = feedforward(output1, W22)
output = feedforward_act(output1, W22, global_outputActivation)
error_soft = softmax_error(y, output, y_in, global_outputActivation)
delta_W22 = calculate_deltaW(error_soft, output1)
error_hidden = non_softmax_error(error_soft, W22, input, W21, activationFunc)
delta_W21 = calculate_deltaW(error_hidden, input)
delta_W1_doc = BpthroughTree(EDUs, error_hidden, W1, W21, dim, activationFunc, True)
delta_W1_query = BpthroughTree(EDUs_test, error_hidden, W1, W21, dim, activationFunc, False)
#print ("=============== : ", np.sum(delta_W1_doc[:, indexNode]))
delta_W1_doc = dropcolrow(delta_W1_doc, indexNode, False)
delta_W1_query = dropcolrow(delta_W1_query, indexNode, False)
delta_W21 = dropcolrow(delta_W21, indexNode2, True)
delta_W = np.divide(np.add(delta_W1_doc, delta_W1_query), 2)
W21 = update_weight(eta, W21_copy, delta_W21)
W22 = update_weight(eta, W22_copy, delta_W22)
W1 = update_weight(eta, W1_copy, delta_W)
return W1, W21, W22
def calculateError_validation_pair(allEDUs, mode, WV, dim, W1, W21, W22, OutputFile, iteration, activationFunc, pairs):
#FileList = os.listdir(path_Folder)
numberofSamples = len(pairs)
target_list=[]
output_list=[]
for pair in pairs: #j in range(0, numberofSamples):
filenames = pair.split(' ')
#path_File_test = path_Folder + filenames[0]
EDUs_test = allEDUs[filenames[0]]#readTree_att_NSWeight(path_File_test, W1_query, WV, dim, activationFunc)
EDUs_test = update_EDU(EDUs_test, W1, dim, activationFunc)
eduKeys_test = sortEduKey(EDUs_test.keys(), reverse=True)
input2_test = EDUs_test[str(eduKeys_test[0])].vector
#path_File_test = path_Folder + filenames[1]
EDUs = allEDUs[filenames[1]]#readTree_att_NSWeight(path_File_test, W1_doc, WV, dim, activationFunc)
EDUs = update_EDU(EDUs, W1, dim, activationFunc)
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
input = np.concatenate([input2, input2_test], 0)
output1 = feedforward_act(input, W21, activationFunc)
output = feedforward_act(output1, W22, global_outputActivation)
if (filenames[0].split("-")[1] == filenames[1].split("-")[1]):
if len(W22[0]) == 1:
y = [1.0]
else:
y = [0.8, -0.8]
target_list.append(y)
output_list.append(output)
else:
if len(W22[0]) == 1:
y = [-1.0]
else:
y = [-0.8, 0.8]
target_list.append(y)
output_list.append(output)
totall_Err = MSE(output_list, target_list)
print(iteration, " ", mode , " ", totall_Err)
OutputFile.write("%s,%s,%s\n" % (iteration, mode, totall_Err))
return totall_Err
def train_for_each_Sample (EDUs, y, W1, W2, eta, activationFunc):
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
y_in = feedforward(input2, W2)
output = feedforward_act(input2, W2, activationFunc)
#print output
error_soft = softmax_error(y, output, y_in, activationFunc)
delta_W2 = calculate_deltaW(error_soft, input2)
delta_W1 = BpthroughTree(EDUs, error_soft, W1, W2, activationFunc)
W2 = update_weight(eta, W2, delta_W2)
W1 = update_weight(eta, W1, delta_W1)
return W1, W2
def calculateError_validation(path_Folder, mode, WV, dim, W1_doc, W1_query, W2, OutputFile, iteration, activationFunc):
FileList = os.listdir(path_Folder)
numberofSamples = len(FileList)
target_list=[]
output_list=[]
for j in range(0, numberofSamples):
path_File_test = path_Folder + FileList[j]
EDUs_test = readTree_att_NSWeight(path_File_test, W1_query, WV, dim, activationFunc)
eduKeys_test = sortEduKey(EDUs_test.keys(), reverse=True)
input2_test = EDUs_test[str(eduKeys_test[0])].vector
for k in range(0, numberofSamples):
path_File_test = path_Folder + FileList[k]
EDUs = readTree_att_NSWeight(path_File_test, W1_doc, WV, dim, activationFunc)
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
input = np.concatenate([input2, input2_test], 0)
output = feedforward_act(input, W2, global_outputActivation)
#print output
if (FileList[j].split("-")[1] == FileList[k].split("-")[1]):
y = [1.0]
target_list.append(y)
output_list.append(output)
else:
if (k % 5 == 0):
y = [0.0]
target_list.append(y)
output_list.append(output)
totall_Err = cross_entropy(target_list, output_list)
print(iteration, " ", mode , " ", totall_Err)
OutputFile.write("%s,%s,%s\n" % (iteration, mode, totall_Err))
return totall_Err
def test_AttWeight_DrHarati_pair(allEDUs, mode, WV, dim, W1, W21, W22, OutputFile, iteration, activationFunc, pairs):
numberofSamples = len(pairs)
#numberofSamples_test = 50
tp = 0
fp = 0
tn = 0
fn = 0
sim = 0
notsim = 0
for pair in pairs: #j in range(0, numberofSamples):
filenames = pair.split(' ')
EDUs_test = allEDUs[filenames[0]]
EDUs_test = update_EDU(EDUs_test, W1, dim, activationFunc)
eduKeys_test = sortEduKey(EDUs_test.keys(), reverse=True)
input2_test = EDUs_test[str(eduKeys_test[0])].vector
EDUs = allEDUs[filenames[1]]
EDUs = update_EDU(EDUs, W1, dim, activationFunc)
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
input = np.concatenate([input2, input2_test], 0)
output1 = feedforward_act(input, W21, activationFunc)
output = feedforward_act(output1, W22, global_outputActivation)
if (filenames[0].split("-")[1] == filenames[1].split("-")[1]):
sim += 1
if (sim % 500 == 0):
print("Similar ", filenames[0].split("-")[1], filenames[1].split("-")[1], output)
#if output[0] > output[1]:
if (len(W22[0]) == 1):
if output[0] > 0:
# print("Similar ", filenames[0].split("-")[1], filenames[1].split("-")[1], output)
tp += 1
else:
# print("Similar ", filenames[0].split("-")[1], filenames[1].split("-")[1], output)
fn += 1
else:
if output[0] > output[1]:
# print("Similar ", filenames[0].split("-")[1], filenames[1].split("-")[1], output)
tp += 1
else:
# print("Similar ", filenames[0].split("-")[1], filenames[1].split("-")[1], output)
fn += 1
else:
notsim += 1
if (notsim % 500 == 0):
print("Not-Similar ", filenames[0].split("-")[1], filenames[1].split("-")[1], output)
#if output[0] < output[1]:
if (len(W22[0]) == 1):
if output[0] < 0:
# print("Not-Similar ", filenames[0].split("-")[1], filenames[1].split("-")[1], output)
tn += 1
else:
# print("Not-Similar ", filenames[0].split("-")[1], filenames[1].split("-")[1], output)
fp += 1
else:
if output[0] < output[1]:
# print("Not-Similar ", filenames[0].split("-")[1], filenames[1].split("-")[1], output)
tn += 1
else:
# print("Not-Similar ", filenames[0].split("-")[1], filenames[1].split("-")[1], output)
fp += 1
print (sim, notsim)
accuracy = float(tp + tn) / (tp + tn + fp + fn)
precision, recall, F1 = calculate_eval_metrics(tp, tn, fp, fn)
print(iteration, " ", mode , " ", tp, " ", tn, " ", fp, " ", fn, " ", accuracy, " ", precision, " ", recall, " ", F1)
OutputFile.write("%s,%s,%s,%s,%s,%s,%s,%s,%s,%s\n" % (iteration, mode, tp, tn, fp, fn, accuracy, precision, recall, F1))
def readTree_att_Scalar(filePath, W1, WV, dim, hierarchyType, attScaler, activationFunc):
EDUs = {}
EDUs_main = {}
if os.path.exists(filePath):
if os.stat(filePath).st_size > 0 :
file_1 = open(filePath, "r")
for line in file_1:
line = re.sub(r"[\n(\[\])]", "", line)
arr = re.split("\s*,\s*",line)
arr2 = re.split("\'", line)
relation = arr2[len(arr2)-2]
hierarchy = arr2[len(arr2)-4]
if arr[0] == arr[1]:
if int(arr[0]) / 10 != 0:
arr[0] = "9" + arr[0]
vector = WordAveraging(preprocessor1(re.sub(r"[\n,.:'(\[\])]", "", arr2[1])), WV, dim)
#vector = tanh(vector)
if hierarchy == hierarchyType:
vector = np.multiply(attScaler, vector)
EDUs_main [arr[0]] = ClassNode.Node(True, False, 0, 0, "", hierarchy, relation, vector, "")
EDUs[arr[0]] = "hi"
else:
numconcat = ""
for num in range(int(arr[0]), int(arr[1])+1):
if num/10 !=0 :
num="9"+str(num)
numconcat += `int(num)`
childs={}
i=1
numconcat2 = numconcat
EDUitem = sortEduKey(EDUs_main.keys(), reverse=True)
for key in EDUitem:
key = str(key)
if numconcat!='' and key in numconcat:
childs[i] = EDUs_main[key].vector
if i==1:
rightChild = key
EDUs_main[key].child = "right"
elif i==2:
leftChild = key
EDUs_main[key].child = "left"
#print len(childs[i]), key
i = 2
numconcat = numconcat.replace(key,"")
del EDUs[key]
EDUs[numconcat2]="hi"
vector = feedforward_act(np.concatenate([childs[2], childs[1]], 0), W1, activationFunc)
if hierarchy == hierarchyType:
vector = np.multiply(attScaler, vector)
EDUs_main[numconcat2]=ClassNode.Node(False, False, leftChild, rightChild, "" , hierarchy, relation, vector, "")
eduKey=EDUs.keys()
eduKey.sort()
if len(eduKey)>1:
vector = feedforward_act(np.concatenate([EDUs_main[eduKey[0]].vector, EDUs_main[eduKey[1]].vector], 0), W1, activationFunc)
EDUs[eduKey[0]+eduKey[1]] = "hi"
EDUs_main[eduKey[0]].child = "left"
EDUs_main[eduKey[1]].child = "right"
if hierarchy == hierarchyType:
vector = np.multiply(attScaler, vector)
EDUs_main[eduKey[0]+eduKey[1]] = ClassNode.Node(False, True, eduKey[0], eduKey[1],"", "", "", vector, "")
del EDUs[eduKey[0]]
del EDUs[eduKey[1]]
return EDUs_main
else:
return EDUs_main
def test_AttWeight_DrHarati(path_Folder, mode, WV, dim, W1 , W2, OutputFile, WSat, WNu, iteration, activationFunc):
posFileList_test = os.listdir(path_Folder + "pos/")
negFileList_test = os.listdir(path_Folder + "neg/")
numberofSamples_test = min(len(posFileList_test), len(negFileList_test))
#numberofSamples_test=100
tp = 0
fp = 0
tn = 0
fn = 0
for k in range(0, numberofSamples_test):
path_File_test = path_Folder + "pos/" + posFileList_test[k]
EDUs = readTree_att_NSWeight(path_File_test, W1, WV, dim, WSat, WNu, activationFunc)
y = [1.0, 0]
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
y_in = feedforward(input2, W2)
output = feedforward_act(input2, W2, activationFunc)
if output[0] == output[1]:
print "input", input2
# print "W2", W2[0, :]
# print "W1", W1[0, :]
# print "WN", WNu[0, :]
# print "WS", WSat[0, :]
# print "pos", output
if output[0] > output[1]:
#print "pos ", output
tp += 1
else:
fn += 1
path_File_test = path_Folder + "neg/" + negFileList_test[k]
EDUs = readTree_att_NSWeight(path_File_test, W1, WV, dim, WSat, WNu, activationFunc)
eduKeys = sortEduKey(EDUs.keys(), reverse=True)
input2 = EDUs[str(eduKeys[0])].vector
y_in = feedforward(input2, W2)
output = feedforward_act(input2, W2, activationFunc)
if output[0] == output[1]:
print "input", input2
# print "W2", W2[0, :]
# print "W1", W1[0,:]
# print "WN", WNu[0,:]
# print "WS", WSat[0,:]
# print "neg",output
if output[0] < output[1]:
#print "neg ", output
tn += 1
else:
fp += 1
accuracy = float(tp + tn) / (tp + tn + fp + fn)
precision, recall, F1 = calculate_eval_metrics(tp, tn, fp, fn)
print iteration, " ", mode , " ", tp, " ", tn, " ", fp, " ", fn, " ", accuracy, " ", precision, " ", recall, " ", F1
OutputFile.write("%s,%s,%s,%s,%s,%s,%s,%s,%s,%s\n" % (iteration, mode, tp, tn, fp, fn, accuracy, precision, recall, F1))
def readTree_att_NSWeight(filePath, W1, WV, dim, WSat, WNu, activationFunc):
#print filePath
EDUs = {}
EDUs_main = {}
if os.path.exists(filePath):
if os.stat(filePath).st_size > 0 :
file_1 = open(filePath, "r")
for line in file_1:
line = re.sub(r"[\n(\[\])]", "", line)
arr = re.split("\s*,\s*",line)
arr2 = re.split("\'", line)
relation = arr2[len(arr2)-2]
hierarchy = arr2[len(arr2)-4]
if arr[0] == arr[1]:
if int(arr[0]) / 10 != 0:
arr[0] = "9" + arr[0]
vector = WordAveraging(preprocessor1(re.sub(r"[\n,.:'(\[\])]", "", arr2[1])), WV, dim)
#print vector
if np.absolute(max(vector))>2:
print "vayyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyy"
if np.sum(np.absolute(vector))/dim >=0.8:
for item in EDUs_main.keys():
print item, " ", EDUs_main[item].vector
print "khodaaaaaaaaaaaaaaaaaa"
#vector = tanh (vector)
EDUs_main [arr[0]] = ClassNode.Node(True, False, 0, 0, "", hierarchy, relation, vector, "")
EDUs[arr[0]] = "hi"
else:
numconcat = ""
for num in range(int(arr[0]), int(arr[1])+1):
if num/10 !=0 :
num="9"+str(num)
numconcat += `int(num)`
childs={}
i=1
numconcat2 = numconcat
EDUitem = sortEduKey(EDUs_main.keys(), reverse=True)
for key in EDUitem:
key = str(key)
if numconcat!='' and key in numconcat:
childs[i] = EDUs_main[key].vector
#print key, " ", childs[i]
childs[i] = apply_attention(childs[i], EDUs_main[key].nodeHierarchy, WNu, WSat, activationFunc)
if np.sum(np.absolute(childs[i])) / dim >= 0.8:
for item in EDUs_main.keys():
print item, " ", EDUs_main[item].vector
print key, " ", childs[i]
print "khodaaaaaaaaaaaaaaaaaa"
if i==1:
rightChild = key
EDUs_main[key].child = "right"
elif i==2:
leftChild = key
EDUs_main[key].child = "left"
i = 2
numconcat = numconcat.replace(key,"")
del EDUs[key]
EDUs[numconcat2]="hi"
vector = feedforward_act(np.concatenate([childs[2], childs[1]], 0), W1, activationFunc)
EDUs_main[numconcat2]=ClassNode.Node(False, False, leftChild, rightChild, "" , hierarchy, relation, vector, "")
eduKey=EDUs.keys()
eduKey.sort()
if len(eduKey)>1:
EDU_0 = apply_attention(EDUs_main[eduKey[0]].vector, EDUs_main[eduKey[0]].nodeHierarchy, WNu, WSat, activationFunc)
EDU_1 = apply_attention(EDUs_main[eduKey[1]].vector, EDUs_main[eduKey[1]].nodeHierarchy, WNu, WSat, activationFunc)
vector = feedforward_act(np.concatenate([EDU_0, EDU_1], 0), W1, activationFunc)
EDUs[eduKey[0]+eduKey[1]] = "hi"
EDUs_main[eduKey[0]].child = "left"
EDUs_main[eduKey[1]].child = "right"
# if hierarchy == "Nucleus":
# vector = np.matmul(vector, WNu)
# elif hierarchy == "Satellite":
# vector = np.matmul(vector, WSat)
EDUs_main[eduKey[0]+eduKey[1]] = ClassNode.Node(False, True, eduKey[0], eduKey[1],"", "", "", vector, "")
del EDUs[eduKey[0]]
del EDUs[eduKey[1]]
return EDUs_main
else:
return EDUs_main
