def __file_parser(self) -> None:
"""
Iterates through the file object and creates mapping
in the data.
"""
try:
file_obj = open(file=self.file_name, mode='r')
for i in file_obj:
field = i.split('|')
if "Applicant\n" in field:
d = {field[1]: None}
self._user.update(d)
if "Mentor\n" in field:
d = {field[1]: field[2]}
self._user.update(d)
if 'ChatGroup' in field:
field[3] = clean(field[3])
d = {int(field[1]): (field[2], field[3])}
self._chat_groups.update(d)
if 'Message' in field:
field[4] = clean(field[4])
field[4] = datetime.datetime.fromtimestamp(float(field[4]))
self._messages.append(field)
except IOError:
print("File not accessible")
logger.error("File not Found")
finally:
file_obj.close()
def read_data():
df_train = pd.read_csv(Config.data_path + 'train.csv')
df_test = pd.read_csv(Config.data_path + 'test.csv')
df_train['comment_text_clean'] = df_train['comment_text'].apply(
lambda x: clean(x))
df_test['comment_text_clean'] = df_test['comment_text'].apply(
lambda x: clean(x))
df_test.fillna(' ', inplace=True)
return (df_train, df_test)
def costFunction(self, params):
"""Returns Error as a function of Lambda Params.
Args:
params: List of doubles, represents different lambda values,
lambda_phi is the first param; lambda_pi's are the next
6 parameters
Returns:
Cost - Integer representing number of misclassified results in the
training set.
"""
# Set Lambda Phi & Pi based upon params
# 1st Param - Lambda Phi
# 2nd Param - Lambda Pi
self.lambda_phi = params[0]
self.lambda_pi = map(lambda x: x / sum(params[1:]), params[1:])
# Train the classifier
self.train()
# Calculate Error as
# Number of Training Errors
error = 0
for (ques, ans) in self.training_set:
res = self.get_classification(ques)
pred_ans = ut.clean(ut.key_max_val_dict(res))
if pred_ans != ans:
error += 1
return error
def get_classification(self, text): text = ut.clean(text) uni = nltk.tokenize.word_tokenize(text) bi = nltk.bigrams (uni) tri = nltk.trigrams (uni) temp_lambda = self.lambda_pi # Map to store answer to its divergence pairs list_of_ans = dict() for (ques, ans) in self.training_set: fin_val = 0.0 for t in uni: fin_val += temp_lambda[5] * (float(self.unigram_tot_dict.get(t,0))/self.len) fin_val += temp_lambda[4] * (float(self.unigram_dict.get((ques,t),0))/len(ques)) for t in bi: fin_val += temp_lambda[3] * (float(self.bigram_tot_dict.get(t,0))/self.unigram_tot_dict.get(t[:1],1)) fin_val += temp_lambda[2] * (float(self.bigram_dict.get((ques,t),0))/self.unigram_dict.get((ques,t[:1]),1)) for t in tri: fin_val += temp_lambda[1] * (float(self.trigram_tot_dict.get(t,0))/self.bigram_tot_dict.get(t[:2],1)) fin_val += temp_lambda[0] * (float(self.trigram_dict.get((ques,t),0))/self.bigram_dict.get((ques,t[:2]),1)) list_of_ans[self.training_orig.get(ans, ans)] = fin_val # Return Weighted list of responses return list_of_ans
def classify(self,Xtest): Prediction = list() for x in Xtest: Prediction.append(self.classifier.classify(self.feature_extractor(ut.clean(x)))) return Prediction
def __init__(self, lambda_phi, lambda_pi, training_set):
self.lambda_phi = lambda_phi
self.lambda_pi = lambda_pi
# Clean (Lower Case; Remove Punctuation) Training Set
# Keep backup of formatted data as well
self.training_orig = dict()
self.training_set = []
for (ques, ans) in training_set:
self.training_set.append((ut.clean(ques), ut.clean(ans)))
self.training_orig[ut.clean(ans)] = ans
print self.training_set
# Initialize Maps for Phi and Pi, to learn their values
self.pi_dict = dict()
self.phi_dict = dict()
def __init__ (self, lambda_phi, lambda_pi, training_set): self.lambda_phi = lambda_phi self.lambda_pi = lambda_pi # Clean (Lower Case; Remove Punctuation) Training Set # Keep backup of formatted data as well self.training_orig = dict() self.training_set = [] for (ques, ans) in training_set: self.training_set.append((ut.clean(ques),ut.clean(ans))) self.training_orig[ut.clean(ans)] = ans print self.training_set # Initialize Maps for Phi and Pi, to learn their values self.pi_dict = dict() self.phi_dict = dict()
def classify(self, Xtest):
Prediction = list()
for x in Xtest:
Prediction.append(
self.classifier.classify(self.feature_extractor(ut.clean(x))))
return Prediction
def __init__ (self, path = "Data/Dataset1.csv"):
cr = csv.reader(open(path,"rb"))
temp = [(row[1], row[0]) for row in cr]
self.tag_list = nltk.data.load('taggers/maxent_treebank_pos_tagger/english.pickle')
self.data_set_total = []
self.data_set_total = [(self.feature_extractor(ut.clean(n)), g) for (n,g) in temp]
self.train_set = self.data_set_total
self.train()
def __init__(self, lambda_pi, training_set):
self.lambda_pi = lambda_pi
# Clean (Lower Case; Remove Punctuation) Training Set
# Keep backup of formatted data as well
self.training_orig = dict()
self.training_set = []
for (ques, ans) in training_set:
self.training_set.append((ut.clean(ques), ut.clean(ans)))
self.training_orig[ut.clean(ans)] = ans
self.unigram_dict = dict()
self.bigram_dict = dict()
self.trigram_dict = dict()
self.unigram_tot_dict = dict()
self.bigram_tot_dict = dict()
self.trigram_tot_dict = dict()
self.len = 0
self.train()
def get_classification(self, text):
text = ut.clean(text)
# Map to store answer to its divergence pairs
list_of_ans = dict()
# Calculate Denominator of Cond. Prob. P(v|W); v - Single word in the answer; W - User utterance
# Calculate once, since its independent of indv. responses/ answers
pv_den = 0.0
# Store PI_Ws(w_i) [i = 1...n] for each Ws for use in the numerator as well
pi_prod = dict()
# Iterate over all questions
for (ques, ans) in self.training_set:
for word in text.split():
pi_prod[ques] = pi_prod.get(ques, 1) * self.pi_dict.get(
(ques, word), 1
) # Smoothing, if not available (i.e count = 0), make it 1 so that the rest doesnt become 0
if (pi_prod[ques] == 0):
break
pv_den += pi_prod[ques]
# Calculate metric for answer that can be given
# Iterate over all answers in training_Set
for (question, answer) in self.training_set:
# Iterate over individual words in every answer
for word in answer.split():
# Calculate numerator of Cond. Prob. P(v|W)
pv_num = 0.0
for ques, ans in self.training_set:
pv_num += self.phi_dict.get((ans, word), 0) * pi_prod[
ques] # Smoothing, if not available (i.e count = 0), make it 1 so that the rest doesnt become 0
# Calculate pv for each word 'v'
pv = pv_num / pv_den
# In KL Divergence; log0 = 0
log_term = pv / self.phi_dict.get(
(answer, word), 0
) # Smoothing, if not available (i.e count = 0), make it 1 so that the rest doesnt become 0
if (log_term != 0):
log_term = np.log10(log_term)
list_of_ans[self.training_orig.get(
answer, answer)] = list_of_ans.get(
self.training_orig.get(answer, answer),
0) + (pv * log_term)
# Return Weighted list of responses
return list_of_ans
def __init__(self, path="Data/Dataset1.csv"):
cr = csv.reader(open(path, "rb"))
temp = [(row[1], row[0]) for row in cr]
self.tag_list = nltk.data.load(
'taggers/maxent_treebank_pos_tagger/english.pickle')
self.data_set_total = []
self.data_set_total = [(self.feature_extractor(ut.clean(n)), g)
for (n, g) in temp]
self.train_set = self.data_set_total
self.train()
def __init__ (self,lambda_pi, training_set): self.lambda_pi = lambda_pi # Clean (Lower Case; Remove Punctuation) Training Set # Keep backup of formatted data as well self.training_orig = dict() self.training_set = [] for (ques, ans) in training_set: self.training_set.append((ut.clean(ques),ut.clean(ans))) self.training_orig[ut.clean(ans)] = ans self.unigram_dict = dict() self.bigram_dict = dict() self.trigram_dict = dict() self.unigram_tot_dict = dict() self.bigram_tot_dict = dict() self.trigram_tot_dict = dict() self.len = 0 self.train()
def data_prep(df_train, df_test, fast_text_embeddings):
def extract_coefs(word, *arr):
return word, np.asarray(arr, dtype='float32')
# cleaning the comment text
df_train['comment_text'] = df_train['comment_text'].apply(
lambda x: clean(x))
df_test['comment_text'] = df_test['comment_text'].apply(lambda x: clean(x))
# Handling the nan values
train_features = df_train["comment_text"].fillna("fillna").values
train_labels = df_train[Config.labels].values
test_features = df_test["comment_text"].fillna("fillna").values
# data preprocessing
tokenizer = text.Tokenizer(num_words=Config.max_features)
tokenizer.fit_on_texts(list(train_features) + list(test_features))
train_features = tokenizer.texts_to_sequences(train_features)
test_features = tokenizer.texts_to_sequences(test_features)
train_features = sequence.pad_sequences(train_features,
maxlen=Config.max_sentence_len)
test_features = sequence.pad_sequences(test_features,
maxlen=Config.max_sentence_len)
# creating the embedding matrix
embeddings_index = dict(
extract_coefs(*emb.rstrip().rsplit(' '))
for emb in open(fast_text_embeddings))
word_index = tokenizer.word_index
nb_words = min(Config.max_features, len(word_index))
embedding_matrix = np.zeros((nb_words, Config.embedding_size))
for word, idx in word_index.items():
if idx >= Config.max_features: continue
embedding_vector = embeddings_index.get(word)
if embedding_vector is not None:
embedding_matrix[idx] = embedding_vector
return train_features, train_labels, test_features, embedding_matrix
def __init__(self, lambda_phi, lambda_pi, training_set):
"""Constructor; Initialize all class attributes.
Args:
lambda_phi: Lambda Values to indicate the relative importance of individual Answers vs The entire Set.
lambda_pi: Lambda Values to indicate relative importance of indv. words vs bigram words vs trigram words in
both individual questions and the entire set.
training_set : Training data for Classification [Tuple Format].
"""
self.lambda_phi = lambda_phi
self.lambda_pi = lambda_pi
# Clean (Lower Case; Remove Punctuation) Training Set
# Keep backup of formatted data as well
self.training_orig = dict()
self.training_set = []
for (ques, ans) in training_set:
self.training_set.append((ut.clean(ques), ut.clean(ans)))
self.training_orig[ut.clean(ans)] = ans
# Initialize Maps for Phi and Pi, to learn their values
self.pi_dict = dict()
self.phi_dict = dict()
def get_classification(self, text): text = ut.clean(text) # Map to store answer to its divergence pairs list_of_ans = dict() # Calculate Denominator of Cond. Prob. P(v|W); v - Single word in the answer; W - User utterance # Calculate once, since its independent of indv. responses/ answers pv_den = 0.0 # Store PI_Ws(w_i) [i = 1...n] for each Ws for use in the numerator as well pi_prod = dict() # Iterate over all questions for (ques, ans) in self.training_set: for word in text.split(): pi_prod[ques] = pi_prod.get(ques , 1) * self.pi_dict.get((ques,word), 1) # Smoothing, if not available (i.e count = 0), make it 1 so that the rest doesnt become 0 if(pi_prod[ques] == 0): break pv_den += pi_prod[ques] # Calculate metric for answer that can be given # Iterate over all answers in training_Set for (question, answer) in self.training_set: # Iterate over individual words in every answer for word in answer.split(): # Calculate numerator of Cond. Prob. P(v|W) pv_num = 0.0 for ques,ans in self.training_set: pv_num += self.phi_dict.get((ans, word), 0) * pi_prod[ques] # Smoothing, if not available (i.e count = 0), make it 1 so that the rest doesnt become 0 # Calculate pv for each word 'v' pv = pv_num / pv_den # In KL Divergence; log0 = 0 log_term = pv / self.phi_dict.get((answer,word), 0) # Smoothing, if not available (i.e count = 0), make it 1 so that the rest doesnt become 0 if(log_term != 0): log_term = np.log10(log_term) list_of_ans[self.training_orig.get(answer, answer)] = list_of_ans.get(self.training_orig.get(answer, answer),0) + (pv * log_term) # Return Weighted list of responses return list_of_ans
def get_classification(self, text):
text = ut.clean(text)
uni = nltk.tokenize.word_tokenize(text)
bi = nltk.bigrams(uni)
tri = nltk.trigrams(uni)
temp_lambda = self.lambda_pi
# Map to store answer to its divergence pairs
list_of_ans = dict()
for (ques, ans) in self.training_set:
fin_val = 0.0
for t in uni:
fin_val += temp_lambda[5] * (
float(self.unigram_tot_dict.get(t, 0)) / self.len)
fin_val += temp_lambda[4] * (
float(self.unigram_dict.get((ques, t), 0)) / len(ques))
for t in bi:
fin_val += temp_lambda[3] * (
float(self.bigram_tot_dict.get(t, 0)) /
self.unigram_tot_dict.get(t[:1], 1))
fin_val += temp_lambda[2] * (
float(self.bigram_dict.get(
(ques, t), 0)) / self.unigram_dict.get(
(ques, t[:1]), 1))
for t in tri:
fin_val += temp_lambda[1] * (
float(self.trigram_tot_dict.get(t, 0)) /
self.bigram_tot_dict.get(t[:2], 1))
fin_val += temp_lambda[0] * (
float(self.trigram_dict.get(
(ques, t), 0)) / self.bigram_dict.get(
(ques, t[:2]), 1))
list_of_ans[self.training_orig.get(ans, ans)] = fin_val
# Return Weighted list of responses
return list_of_ans
def costFunction (self, params): self.lambda_phi = params[0] self.lambda_pi = map(lambda x: x/sum(params[1:]), params[1:] ) #print self.lambda_pi # Train the classifier self.train() # Calculate Error error = 0.0 for (ques,ans) in self.training_set: res = self.get_classification(ques) pred_ans = ut.clean(ut.key_max_val_dict (res)) # Add one for every misclassified result if(pred_ans != ans): # print pred_ans, ans # raw_input() error += 1 return error
def feature_extractor(self,sentence):
features = {}
words = word_tokenize(ut.clean(sentence))
features["length"] = len(words)
for q in ["how","where","when","what","why"]:
features["has(%s)" % q] = (q in words)
temp = [ (b) for (a,b) in pos_tag(words)]
tag_dict = dict ()
for tag in temp:
tag_dict [tag] = tag_dict.get(tag,0) + 1
for tags in self.tag_list.classifier().labels():
features["count(%s)" % tags] = tag_dict.get(tags,0)
#features["has(%s)" % tags] = (tag_dict.get(tags,0) != 0)
return features
def costFunction(self, params):
self.lambda_pi = map(lambda x: x / sum(params), params)
#print self.lambda_pi
# Train the classifier
self.train()
# Calculate Error
error = 0.0
for (ques, ans) in self.training_set:
res = self.get_classification(ques)
pred_ans = ut.clean(ut.key_max_val_dict(res))
# Add one for every misclassified result
if (pred_ans != ans):
# print pred_ans, ans
# raw_input()
error += 1
return error
def feature_extractor(self, sentence):
features = {}
words = word_tokenize(ut.clean(sentence))
features["length"] = len(words)
for q in ["how", "where", "when", "what", "why"]:
features["has(%s)" % q] = (q in words)
temp = [(b) for (a, b) in pos_tag(words)]
tag_dict = dict()
for tag in temp:
tag_dict[tag] = tag_dict.get(tag, 0) + 1
for tags in self.tag_list.classifier().labels():
features["count(%s)" % tags] = tag_dict.get(tags, 0)
#features["has(%s)" % tags] = (tag_dict.get(tags,0) != 0)
return features
def prep_dummy_csv():
from hashlib import sha256
# We need to hash customer emails as a semi-unique identifier
def hash(inp):
bytes=inp.encode('utf-8')
h_e = sha256()
h_e.update(bytes)
return h_e.hexdigest()[:6]
concat = u.concat('CSV_Files')
to_dummy_csv = u.clean(concat)
email_hashes = [hash(i) for i in to_dummy_csv['Email_Billing']]
to_dummy_csv['Email_Billing'] = email_hashes
to_dummy_csv[['Full_Name_Billing', 'Full_Name_Shipping', 'Address_1_Shipping']]='Restricted'
to_dummy_csv.drop('Unnamed:_0', axis=1, inplace=True)
to_dummy_csv['State_Name_Shipping'].loc[to_dummy_csv['State_Name_Shipping'].isna()] = 'Missing'
to_dummy_csv['Discount_Amount'].loc[to_dummy_csv['Discount_Amount'].isna()] = to_dummy_csv['Discount_Amount'].mean()
print(to_dummy_csv.isna().sum())
to_dummy_csv.dropna(inplace=True)
to_dummy_csv= to_dummy_csv.loc[to_dummy_csv['Order_Date']<='2020-12-31-23:59']
print(to_dummy_csv.head(5))
print(to_dummy_csv.columns)
print(to_dummy_csv.tail(5))
to_dummy_csv.to_csv('dummy.csv')
import trysearch as ts
import AlchemyAPI as AP
Dialog_Manager = DM.dialog_manager()
q_class = QC.q_classification()
var = 1
while var:
print '\n'
string = raw_input(" Enter the question: ")
if string in ["end","End","exit","Exit"]:
var = 0
else:
temp = [ (a,b) for (a,b) in pos_tag(nltk.tokenize.word_tokenize(ut.clean(string)))]
temp1 = dict()
temp2 = ''
for (a,b) in temp:
if a == 'i':
a = 'i'
elif b == 'RB' or b == 'VB':
temp2 += 'action '
temp1['action'] = a
elif b == 'VBP' or b == 'NN':
temp2 += 'object '
temp1['object'] = a
else:
temp2 += a + ' '
result_string = Dialog_Manager.get_reply(temp2)
res_class = q_class.classify([string])
elif sys.argv[1] == 'pcp':
frame = serve_frame('CSV_Files')
dfpa = ma.ProductAnalysis(frame)
dfpa.highest_positive_product_change_over_month_analysis()
elif sys.argv[1] == 'pcp':
frame = serve_frame('CSV_Files')
dfpa = ma.ProductAnalysis(frame)
dfpa.highest_negative_product_change_over_month_analysis()
elif sys.argv[1] == 'plg':
frame = serve_frame('CSV_Files')
dfpa = ma.ProductAnalysis(frame)
dfpa.product_line_change_over_month_graph()
file = u.concat('CSV_FIles')
# Test slice for functionality
file = u.clean(file)
cp1 = u.FieldLoader(file)
u.date_difference()
u.date_difference_email_list_compiler()
#profiles_master_init()
'''Work log:
2/25
functionality and indexing for time: reading, saving and analyzing time fields (completed?)
Streamline initializations: maybe make a initialization function in Analysis so we can simply click on the py file
and run the program
**As of now customer_profile line 63 should protect from repeated inputs of sales information. This will need to be tested.
***Concat in Utility needs to be corrected! atm there is only one customer!! Best guess is recent addition--concat
def serve_frame(file):
frame = u.concat(file)
frame = u.clean(frame)
return frame
def get_classification(self, text):
"""Classify 'text' based upon learned values in 'train', and return weighted list of responses.
Args:
text: The text (or Question, in this case) to be classified.
Returns:
A list (as a dict) of the weighted responses for the given 'text',
The key being the response, and the value being the KL-Divergence.
For Example,
{
["This is the way"] => -0.0032,
["That is not true"] => -0.023,
.
.
.
}
.
"""
# Clean the input (Remove Punctuation, Lower Case)
text = ut.clean(text)
# Map to store answer to its divergence pairs
list_of_ans = dict()
# Calculate Denominator of Cond. Prob. P(v|W); v - Single word in the answer; W - User utterance
# Calculate once, since its independent of indv. responses/ answers
pv_den = 0.0
# Store PI_Ws(w_i) [i = 1...n] for each Ws for use in the numerator as well
pi_prod = dict()
# Iterate over all questions
for (ques, ans) in self.training_set:
for word in text.split():
pi_prod[ques] = pi_prod.get(ques, 1) * self.pi_dict.get(
(ques, word), 1
) # Smoothing, if not available (i.e count = 0), make it 1 so that the rest doesnt become 0
if pi_prod[ques] == 0:
break
pv_den += pi_prod[ques]
# Calculate metric for answer that can be given
# Iterate over all answers in training_Set
for (question, answer) in self.training_set:
# Iterate over individual words in every answer
for word in answer.split():
# Calculate numerator of Cond. Prob. P(v|W)
pv_num = 0.0
for ques, ans in self.training_set:
pv_num += (
self.phi_dict.get((ans, word), 0) * pi_prod[ques]
) # Smoothing, if not available (i.e count = 0), make it 1 so that the rest doesnt become 0
# Calculate pv for each word 'v'
pv = pv_num / pv_den
# By Definition of KL Divergence; log0 = 0
log_term = pv / self.phi_dict.get(
(answer, word), 0
) # Smoothing, if not available (i.e count = 0), make it 1 so that the rest doesnt become 0
if log_term != 0:
log_term = np.log10(log_term)
list_of_ans[self.training_orig.get(answer, answer)] = list_of_ans.get(
self.training_orig.get(answer, answer), 0
) + (pv * log_term)
# Return Weighted list of responses
return list_of_ans
Dialog_Manager = DM.dialog_manager()
q_class = QC.q_classification()
var = 1
while var:
print '\n'
string = raw_input(" Enter the question: ")
if string in ["end", "End", "exit", "Exit"]:
var = 0
else:
temp = [
(a, b)
for (a,
b) in pos_tag(nltk.tokenize.word_tokenize(ut.clean(string)))
]
temp1 = dict()
temp2 = ''
for (a, b) in temp:
if a == 'i':
a = 'i'
elif b == 'RB' or b == 'VB':
temp2 += 'action '
temp1['action'] = a
elif b == 'VBP' or b == 'NN':
temp2 += 'object '
temp1['object'] = a
else:
temp2 += a + ' '
result_string = Dialog_Manager.get_reply(temp2)
