def kd_tree_classification(k, lenData, pctTest, params, neightboards):
clear_csv()
samples = []
if (params[0] == "PAIS"):
samples = generar_muestra_pais(lenData)
else:
samples = generar_muestra_provincia(lenData, params[1])
quantity_for_testing = int(lenData * pctTest)
normalizer = Normalizer()
data = normalizer.prepare_data(samples, quantity_for_testing)
kdTree = Kd_Tree(neightboards)
firstRound = cross_validation(k, kdTree, data, lenData, "trainingFeatures",
"testingFeatures", "First")
secondRound = cross_validation(k, kdTree, data, lenData,
"trainingFeatures", "testingFeatures",
"Second")
secondWithFirst = cross_validation(k, kdTree, data, lenData,
"trainingFeaturesFirstInclude",
"testingFeaturesFirstInclude", "Second")
normalData = normalizer.get_normal_data()
predictions = [firstRound, secondRound, secondWithFirst]
show_accuracy("KD-TREE", predictions)
make_csv(k, normalData, lenData, pctTest, predictions)
def verify_address(adress, results, borough):
zips = Normalizer.select_zipcode_class(Normalizer.get_neighborhood(borough))
for r in results:
zip3dig = int(r[2]) / 100
if zip3dig in zips:
return r[0], r[1], adress+", "+r[2]
return None
def desicion_tree(k, lenData, pctTest, params, threshold):
clear_csv()
samples = []
if (params[0] == "PAIS"):
samples = generar_muestra_pais(lenData)
else:
samples = generar_muestra_provincia(lenData, params[1])
quantity_for_testing = int(lenData * pctTest)
normalizer = Normalizer()
data = normalizer.separate_data_2(samples, quantity_for_testing)
decisionTree = DecisionTree(threshold)
firstRound = cross_validation(k, decisionTree, data, lenData,
"trainingFeaturesFirst",
"testingFeaturesFirst", "First")
secondRound = cross_validation(k, decisionTree, data, lenData,
"trainingFeaturesSecond",
"testingFeaturesSecond", "Second")
secondWithFirst = cross_validation(k, decisionTree, data, lenData,
"trainingFeaturesFirstInclude",
"testingFeaturesFirstInclude", "Second")
normalData = normalizer.get_normal_data()
predictions = [firstRound, secondRound, secondWithFirst]
show_accuracy("DT", predictions)
make_csv(k, normalData, lenData, pctTest, predictions)
def lr_classification(k, lenData, pctTest, l_regulizer=1):
clear_csv()
samples = generar_muestra_pais(lenData)
quantity_for_testing = int(lenData * pctTest)
normalizer = Normalizer()
data = normalizer.prepare_data_tensor(samples, quantity_for_testing)
lrClassifier = LogisticRegression(1, l_regulizer)
firstRound = cross_validation(k, lrClassifier, data, lenData,
"trainingFeatures", "testingFeatures",
"First")
lrClassifier = LogisticRegression(2, l_regulizer)
print("Paso primero")
secondRound = cross_validation(k, lrClassifier, data, lenData,
"trainingFeatures", "testingFeatures",
"Second")
print("Paso segundo")
secondWithFirst = cross_validation(k, lrClassifier, data, lenData,
"trainingFeaturesFirstInclude",
"testingFeaturesFirstInclude", "Second")
print("Paso tercero")
normalData = normalizer.get_normal_data()
# predictions = [firstRound, secondRound, secondWithFirst]
predictions = [secondRound]
show_accuracy("LR", predictions)
make_csv(k, normalData, lenData, pctTest, predictions)
class Predicter(object):
def __init__(self, thetafile='theta.csv', datafile='data.csv'):
try:
df = pd.read_csv(thetafile)
self.theta = [
float(df['theta0'].astype(float)),
float(df['theta1'].astype(float))
]
except:
print("the theta file was not found")
self.theta = None
self.norm = Normalizer(datafile)
def predict(self, km):
"""
Model function:
f(x) = ax + b
"""
return self.theta[0] + (self.theta[1] * km)
def run(self, km):
if self.theta == None:
return None
norm_km = self.norm.normalize_km(km)
result = self.predict(norm_km)
return self.norm.denormalize_price(result)
def train_data_set():
normalizer = Normalizer()
data_set = []
labels = []
for i in range(0, 256, 8):
n = normalizer.norm(int(random.uniform(0, 256)))
data_set.append(n)
labels.append(n)
return labels, data_set
def svm_classification(
k, lenData, pctTest, params, C=1, gamma=1, kernel="rbf"):
clear_csv()
samples = []
print(params)
if (params[0] == "PAIS"):
samples = generar_muestra_pais(lenData)
else:
samples = generar_muestra_provincia(lenData, params[1])
quantity_for_testing = int(lenData * pctTest)
normalizer = Normalizer()
data = normalizer.prepare_data(samples, quantity_for_testing)
svmClassifier = SVMClassifier(kernel, C, gamma)
firstRound = cross_validation(
k,
svmClassifier,
data,
lenData,
"trainingFeatures",
"testingFeatures",
"First"
)
secondRound = cross_validation(
k,
svmClassifier,
data,
lenData,
"trainingFeatures",
"testingFeatures",
"Second"
)
secondWithFirst = cross_validation(
k,
svmClassifier,
data,
lenData,
"trainingFeaturesFirstInclude",
"testingFeaturesFirstInclude",
"Second"
)
normalData = normalizer.get_normal_data()
predictions = [firstRound, secondRound, secondWithFirst]
show_accuracy("SVM", predictions)
make_csv(k, normalData, lenData, pctTest, predictions)
def preprocess(self):
tuples = CsvManager.read(self.input)
num = CsvManager.read_progress()
print num
if num == 0:
CsvManager.write_geo_codes([], self.output)
CsvManager.write_progress('0')
self.progress.set_size(len(tuples))
self.progress.update_progress(num)
Normalizer.set_tuple(num, tuples)
return tuples
def __init__(self, thetafile='theta.csv', datafile='data.csv'):
try:
df = pd.read_csv(thetafile)
self.theta = [
float(df['theta0'].astype(float)),
float(df['theta1'].astype(float))
]
except:
print("the theta file was not found")
self.theta = None
self.norm = Normalizer(datafile)
def __init__(self, datafile, outfile, theta=[0, 0], learning_rate=0.1, train_range=1000):
self.t_history = []
self.theta = theta
self.datafile = datafile
self.output = outfile
self.data = pd.read_csv(datafile)
self.learning_rate = learning_rate
self.range = train_range
norm = Normalizer(self.datafile)
self.km = norm.normalize_km_list(self.data['km'])
self.price = norm.normalize_price_list(self.data['price'])
def t(network, data):
'''
test the model with input data
:param network:
:param data:
:return:
'''
normalizer = Normalizer()
norm_data = normalizer.norm(data)
predict_data = network.predict(norm_data)
print '\ttestdata(%u)\tpredict(%u)' % (data,
normalizer.denorm(predict_data))
def lr_classification(k, lenData, pctTest, l_regulizer=1):
clear_csv()
samples = generar_muestra_pais(lenData)
quantity_for_testing = int(lenData * pctTest)
normalizer = Normalizer()
data = normalizer.prepare_data_tensor(samples, quantity_for_testing)
lrClassifier = LogisticRegression(1, l_regulizer)
firstRound = cross_validation(
k,
lrClassifier,
data,
lenData,
"trainingFeatures",
"testingFeatures",
"First"
)
lrClassifier = LogisticRegression(2, l_regulizer)
print("Paso primero")
secondRound = cross_validation(
k,
lrClassifier,
data,
lenData,
"trainingFeatures",
"testingFeatures",
"Second"
)
print("Paso segundo")
secondWithFirst = cross_validation(
k,
lrClassifier,
data,
lenData,
"trainingFeaturesFirstInclude",
"testingFeaturesFirstInclude",
"Second"
)
print("Paso tercero")
normalData = normalizer.get_normal_data()
# predictions = [firstRound, secondRound, secondWithFirst]
predictions = [secondRound]
show_accuracy("LR", predictions)
make_csv(k, normalData, lenData, pctTest, predictions)
def kd_tree_classification(k, lenData, pctTest, params, neightboards):
clear_csv()
samples = []
if (params[0] == "PAIS"):
samples = generar_muestra_pais(lenData)
else:
samples = generar_muestra_provincia(lenData, params[1])
quantity_for_testing = int(lenData * pctTest)
normalizer = Normalizer()
data = normalizer.prepare_data(samples, quantity_for_testing)
kdTree = Kd_Tree(neightboards)
firstRound = cross_validation(
k,
kdTree,
data,
lenData,
"trainingFeatures",
"testingFeatures",
"First"
)
secondRound = cross_validation(
k,
kdTree,
data,
lenData,
"trainingFeatures",
"testingFeatures",
"Second"
)
secondWithFirst = cross_validation(
k,
kdTree,
data,
lenData,
"trainingFeaturesFirstInclude",
"testingFeaturesFirstInclude",
"Second"
)
normalData = normalizer.get_normal_data()
predictions = [firstRound, secondRound, secondWithFirst]
show_accuracy("KD-TREE", predictions)
make_csv(k, normalData, lenData, pctTest, predictions)
def desicion_tree(k, lenData, pctTest, params, threshold):
clear_csv()
samples = []
if (params[0] == "PAIS"):
samples = generar_muestra_pais(lenData)
else:
samples = generar_muestra_provincia(lenData, params[1])
quantity_for_testing = int(lenData * pctTest)
normalizer = Normalizer()
data = normalizer.separate_data_2(samples, quantity_for_testing)
decisionTree = DecisionTree(threshold)
firstRound = cross_validation(
k,
decisionTree,
data,
lenData,
"trainingFeaturesFirst",
"testingFeaturesFirst",
"First"
)
secondRound = cross_validation(
k,
decisionTree,
data,
lenData,
"trainingFeaturesSecond",
"testingFeaturesSecond",
"Second"
)
secondWithFirst = cross_validation(
k,
decisionTree,
data,
lenData,
"trainingFeaturesFirstInclude",
"testingFeaturesFirstInclude",
"Second"
)
normalData = normalizer.get_normal_data()
predictions = [firstRound, secondRound, secondWithFirst]
show_accuracy("DT", predictions)
make_csv(k, normalData, lenData, pctTest, predictions)
def __init__(self, loss_function):
"""
Parameters
----------
loss_function : SquaredErrorLoss
loss function for the MLP
"""
self.layers = []
self.activations_functions = []
self.loss_function = loss_function
self.normalizer = Normalizer()
# All errors, returned in end for plotting
self.errors = []
self.predict_errors = 0
def writeEntryPointProfiles(outEntryPointHdl, entryPointName, agentTypes):
epNumber = entryPointName.split("_")[0]
outEntryPointHdl.write(indent1 + "\n<entrypoint ID=\"ep_" + epNumber + "\" geotype=\"point\" type=\"dynamic\">\n")
outEntryPointHdl.write(2 * indent2 + "<loader>entry_loader</loader>\n")
outEntryPointHdl.write(2 * indent2 + "<geometry_id>" + epNumber + "</geometry_id>\n")
outEntryPointHdl.write(2 * indent2 + "<profiles>\n")
for agent in agentTypes:
n = Normalizer()
agent = n.normalize(agent);
outEntryPointHdl.write(3 * indent2 + "<profile>\n")
outEntryPointHdl.write(4 * indent2 + "<agentDistribution pctofentries=\"100.0\" agent_type=\"agent_" + agent + "\"/>\n")
outEntryPointHdl.write(4 * indent2 + "<timetableReference ref=\"tt_" + agent + "\" scaleFactor=\"1.0\"/>\n")
outEntryPointHdl.write(3 * indent2 + "</profile>\n")
outEntryPointHdl.write(2 * indent2 + "</profiles>\n")
outEntryPointHdl.write(indent1 + "</entrypoint>\n")
def pruebas():
# svm_classification(1000, 0.2, C=10, gamma=0.00833333333, kernel="rbf")
lenData = 2500
print(lenData)
print("kernel: ", "sigmoid", " C: ", 1, " G: ", 0.000000001)
pctTest = 0.2
# samples = generar_muestra_provincia(lenData, "SAN JOSE")
# quantity_for_testing = int(lenData*pctTest)
# normalizer = Normalizer()
# data = normalizer.prepare_data(samples, quantity_for_testing)
# svm_classification(10, lenData, pctTest, C=1, gamma=1, kernel="rbf")
time1 = time.time()
for i in range(0, 30):
samples = generar_muestra_pais(lenData)
quantity_for_testing = int(lenData * pctTest)
normalizer = Normalizer()
data = normalizer.prepare_data(samples, quantity_for_testing)
svm_classification(10,
lenData,
pctTest,
C=1,
gamma=0.000000001,
kernel="sigmoid")
time2 = time.time()
print("ms: ", ((time2 - time1) * 1000.0))
totalacc = 0.0
for i in range(0, len(accList), 3):
totalacc += accList[i][1]
print("ER: ", 1 - (totalacc / 30.0))
totalacc = 0.0
for i in range(1, len(accList), 3):
totalacc += accList[i][1]
print("ER: ", 1 - (totalacc / 30.0))
totalacc = 0.0
for i in range(2, len(accList), 3):
totalacc += accList[i][1]
print("ER: ", 1 - (totalacc / 30.0))
def main():
test = [ 61.19499969, 57.31000137, 56.09249878, 61.72000122,
61.38000107, 64.61000061, 61.93500137, 63.70249939,
63.57249832, 60.22750092, 61.23249817, 60.35250092,
65.61750031, 64.85749817, 66.51750183, 66.99749756,
68.3125 , 71.76249695, 71.10749817, 71.67250061,
70.69999695, 69.23249817, 67.09249878, 69.02500153,
68.75749969, 70.74250031, 70.79250336, 69.64499664,
71.93250275, 73.44999695, 72.26750183, 73.29000092,
74.38999939, 75.15750122, 75.93499756, 77.53250122,
78.75250244, 77.85250092, 76.91249847, 77.38500214,
76.92749786, 78.73999786, 78.28500366, 79.80750275,
79.21250153, 79.72250366, 79.18250275, 79.52749634,
79.5625 , 79.48500061, 80.46250153, 80.83499908,
81.27999878, 80.58000183, 82.875 , 83.36499786,
85.99749756, 88.20999908, 83.97499847, 84.69999695,
85.74749756, 88.01999664, 87.89749908, 87.93250275,
87.43000031, 89.71749878, 91.63249969, 90.01499939,
91.20999908, 88.40750122, 90.44499969, 91.19999695,
91.02749634, 91.02749634, 93.46250153, 93.17250061,
95.34249878, 95.75250244, 95.91999817, 95.47750092,
97.05750275, 97.72499847, 96.52249908, 96.32749939,
98.35749817, 97. , 97.27249908, 92.84500122,
92.61499786, 94.80999756, 93.25250244, 95.04000092,
96.19000244, 106.26000214, 108.9375 , 109.66500092,
110.0625 , 113.90249634, 111.11250305, 112.72750092]
test = np.array(test)
test = np.reshape(test, (-1, 1))
normalizer = Normalizer()
normalized = normalizer.FeatureScaler.transform(test)
print(normalized)
def _do_work(self,queue,bbnp,param,action_space,state_space,max_length,seed_id,num_workers,env,continuous_action,cma_param,index_start):
start = time.time()
np.random.seed(seed_id)
if self.noise_type == 'Gaussian':
noises = self.sigma * np.random.randn(num_workers,len(param))
noisy_param = param + noises
elif self.noise_type == 'Hadamard':
h_size = 1<<((max(num_workers,len(param))-1).bit_length())
h = hadamard(h_size)
noises = self.sigma*([email protected](np.random.choice([-1,1], h_size)))[:num_workers,:len(param)]
noisy_param = param + noises
elif self.noise_type == 'CMA':
noisy_param = cma_param
elif self.noise_type == 'CDPP':
noisy_param = self.buffer[index_start:index_start+num_workers]
fitness = []
anti_fitness = []
worker_summary = {}
for ind in noisy_param:
#do the roll out
if self.state_renormalize == True:
normal = Normalizer(state_space)
ind_fit = bbnp.roll_out(ind,env,self.env,normal,render = False,state_renormalize = True)
normal = Normalizer(state_space)
ind_fit_anti = bbnp.roll_out(-ind,env,self.env,normal,render = False,state_renormalize = True)
else:
normal = Normalizer(state_space)
ind_fit = bbnp.roll_out(ind,env,self.env,normal,render = False,init = False)
ind_fit_anti = bbnp.roll_out(-ind,env,self.env,normal,render = False,init = False)
fitness.append(ind_fit)
anti_fitness.append(ind_fit_anti)
end = time.time()
worker_summary['fit'] = fitness
worker_summary['anti_fit'] = anti_fitness
worker_summary['seed_id'] = seed_id
queue.put(worker_summary)
def pruebas():
# svm_classification(1000, 0.2, C=10, gamma=0.00833333333, kernel="rbf")
lenData = 2500
print(lenData)
print("kernel: ", "sigmoid", " C: ", 1, " G: ", 0.000000001)
pctTest = 0.2
# samples = generar_muestra_provincia(lenData, "SAN JOSE")
# quantity_for_testing = int(lenData*pctTest)
# normalizer = Normalizer()
# data = normalizer.prepare_data(samples, quantity_for_testing)
# svm_classification(10, lenData, pctTest, C=1, gamma=1, kernel="rbf")
time1 = time.time()
for i in range(0, 30):
samples = generar_muestra_pais(lenData)
quantity_for_testing = int(lenData*pctTest)
normalizer = Normalizer()
data = normalizer.prepare_data(samples, quantity_for_testing)
svm_classification(
10, lenData, pctTest, C=1, gamma=0.000000001, kernel="sigmoid")
time2 = time.time()
print("ms: ", ((time2-time1)*1000.0))
totalacc = 0.0
for i in range(0, len(accList), 3):
totalacc += accList[i][1]
print("ER: ", 1-(totalacc/30.0))
totalacc = 0.0
for i in range(1, len(accList), 3):
totalacc += accList[i][1]
print("ER: ", 1-(totalacc/30.0))
totalacc = 0.0
for i in range(2, len(accList), 3):
totalacc += accList[i][1]
print("ER: ", 1-(totalacc/30.0))
def svm_classification(k,
lenData,
pctTest,
params,
C=1,
gamma=1,
kernel="rbf"):
clear_csv()
samples = []
print(params)
if (params[0] == "PAIS"):
samples = generar_muestra_pais(lenData)
else:
samples = generar_muestra_provincia(lenData, params[1])
quantity_for_testing = int(lenData * pctTest)
normalizer = Normalizer()
data = normalizer.prepare_data(samples, quantity_for_testing)
svmClassifier = SVMClassifier(kernel, C, gamma)
firstRound = cross_validation(k, svmClassifier, data, lenData,
"trainingFeatures", "testingFeatures",
"First")
secondRound = cross_validation(k, svmClassifier, data, lenData,
"trainingFeatures", "testingFeatures",
"Second")
secondWithFirst = cross_validation(k, svmClassifier, data, lenData,
"trainingFeaturesFirstInclude",
"testingFeaturesFirstInclude", "Second")
normalData = normalizer.get_normal_data()
predictions = [firstRound, secondRound, secondWithFirst]
show_accuracy("SVM", predictions)
make_csv(k, normalData, lenData, pctTest, predictions)
def fix_acris(self):
tuples = self.preprocess()
real_estates = []
while tuples:
try:
t = tuples.pop(0)
bbl = Normalizer.set_bbl(t[0], t[1], t[2])
address = t[3]+" "+t[4]
address = Normalizer.set_address(address, bbl)
date = Normalizer.set_str_to_epoch(t[5])
price = t[6]
real_estates.append((bbl, address, date, price))
except ValueError:
self.error_log.open()
self.error_log.write(t[1]+", "+str(t[0]))
self.error_log.close()
except KeyboardInterrupt:
print ""
print "Stopped"
CsvManager.append_geo_codes(real_estates, self.output)
CsvManager.append_geo_codes(real_estates, self.output)
def testLogProbabilities(self): probs = numpy.asarray([ [0.01, 0.04], [0.1, 0.4], [0.4, 0.1], [0.00004, 0.00001]]) expectedProbs = numpy.asarray([ [0.2, 0.8], [0.2, 0.8], [0.8, 0.2], [0.8, 0.2]]) logProbs = numpy.log(probs) logExpectedProbs = numpy.log(expectedProbs) target = Normalizer() normalizedLogProbs = target.normalizeLogProbabilities(logProbs) normalizedProbs = numpy.exp(normalizedLogProbs) numpy.testing.assert_array_almost_equal(normalizedProbs, expectedProbs) numpy.testing.assert_array_almost_equal(normalizedLogProbs, logExpectedProbs)
def normalizeTokens(self):
'''Normalizes each token in the instance and stores the normalized tokens in the token dictionary'''
normalizedTokenDictionary = {}
normalizedTokenId = 1
for tokenObject in self.tokenDictionary.values():
normalizedTokenObjects = Normalizer.getInstance().normalizeToken(
tokenObject)
for normalizedTokenObject in normalizedTokenObjects:
normalizedTokenDictionary[
normalizedTokenId] = normalizedTokenObject
normalizedTokenId += 1
return normalizedTokenDictionary
def _get_block_dict(self):
self._block_dict = {}
for section in self._config.sections():
self._block_dict[section] = []
for option in self._config[section]:
if option not in [
'url_base', 'query_interval', 'tmp', 'output',
'iterations'
]:
normalized_option = Normalizer.normalize_name(option)
output_xml_file = self._output + '/' + section + '/' + normalized_option + ".xml"
self._block_dict[section].append((option, output_xml_file))
def _process_article(self, doc_key, article):
normalizer = Normalizer()
all_terms = []
for i in article.findall('*'):
if 'encoded' in i.tag or 'description' in i.tag:
if i.text:
i = BeautifulSoup(i.text, "lxml")
if i.text is not None:
for y in i.text.split():
all_terms.append(y)
cleaned_terms = []
for i in all_terms:
normalized = normalizer.normalize_name(i).strip()
if not len(normalized) < 3 and not normalizer.is_stop_word(
normalized) and not normalizer.is_link(normalized):
cleaned_terms.append(normalized)
for term in cleaned_terms:
stemmed_term = self._stemmer.stemWord(term)
term_id = self._get_or_create_term_id(stemmed_term)
self._ii_dict.setdefault(term_id, set())
self._ii_dict[term_id].add(self._document_dict.get(doc_key))
def __init__(self):
self.NaiveBayesClassifier = NaiveBayesClassifier()
# Sentence Splitters
self.RuleBasedSentenceSplitter = RuleBasedSentenceSplitter()
self.MLBasedSentenceSplitter = MLBasedSentenceSplitter()
# Tokenizers
self.RuleBasedTokenizer = RuleBasedTokenizer()
self.MLBasedTokenizer = MLBasedTokenizer()
# Normalizer
self.Normalizer = Normalizer()
# Stemmer
self.Stemmer = Stemmer()
# Stopword Eliminators
self.StaticStopWordEliminator = StaticStopwordRemover()
self.DynamicStopWordEliminator = DynamicStopWordEliminator()
def search_lat_long(self):
tuples = self.preprocess()
count = 1
nominatim, google, opencage, bing, tiger = self.build_geocodings()
while tuples:
t = tuples.pop(0)
status, found = self.geocode_process(t, nominatim)
if not found:
if status == -1:
status, found = self.geocode_process(t, bing)
if not found and status == -1:
self.geocode_process(t, tiger)
elif status == -2:
i = 1
while i < 3:
print "Waiting 45' for the "+Normalizer.set_order(str(i))+" time"
time.sleep(2700)
status, found = self.geocode_process(t, nominatim)
if found:
continue
elif status == -2:
i += 1
elif status == -3:
return
if count % 100 == 0:
for i in range(3):
t = tuples.pop(0)
status, found = self.geocode_process(t, google)
time.sleep(3)
if not found:
self.geocode_process(t, opencage)
time.sleep(3)
else:
t = tuples.pop(0)
self.geocode_process(t, opencage)
time.sleep(3)
count += 1
def _apply(self):
'''
Apply transformations set ...
'''
if 'filter' not in self.config:
raise ValueError('Config map should include "filter."')
opt = DataParserOpt(**self.config['filter'])
dr = DataReader(self.map_file, self.data, opt=opt)
A, y = dr.extract_features()
if 'sample' in self.config:
try:
method = self.config['sample']['method']
method_args = self.config['sample']['method_args']
ds = DataSampler(A, y, method, **method_args)
except Exception as e:
ds = DataSampler(A, y)
print('{} or\nSample structure has no method_args'.format(e))
A, y = ds.sample()
if 'normalize' in self.config:
nml = Normalizer(map_file=self.map_file)
for method_str in self.config['normalize']['methods']:
method = getattr(nml, method_str)
A = method(A)
return A, y
def test_distance(self):
normalizer = Normalizer()
self.assertEqual(distance([0, 1, 2], [3, 4, 5]), 5.196152422706632)
print("Accuracy Training:",
accuracy.eval({
X: train_x,
y: train_y
}))
## sess = tf.Session()
## with sess.as_default():
## return self.toparty(self.y.eval({self.X: test_x, self.y: test_y}).tolist())
##
samples = generar_muestra_pais(100)
quantity_for_testing = int(100 * 0.2)
normalizer = Normalizer()
data = normalizer.prepare_data(samples, quantity_for_testing)
classes = np.append(data["trainingClassesFirst"],
data["testingClassesFirst"],
axis=0)
sample = {
"trainingFeatures": data["trainingFeatures"],
"trainingClasses": data["trainingClassesFirst"],
"testingFeatures": data["testingFeatures"],
"testingClasses": data["testingClassesFirst"]
}
sample2 = {
"testingFeatures": data["testingFeatures"],
"testingClasses": data["testingClassesFirst"]
}
print(sample2["testingClasses"])
# empieza el entrenamiento
_, c = sess.run([optimizer, cost], feed_dict={self.X: train_x,
self.y: train_y})
cost_in_each_epoch += c
## # you can uncomment next two lines of code for printing cost when training
## if (epoch+1) % display_step == 0:
## print("Epoch: {}".format(epoch + 1), "cost={}".format(cost_in_each_epoch))
print("Accuracy Training:", accuracy.eval({X: train_x, y: train_y}))
## sess = tf.Session()
## with sess.as_default():
## return self.toparty(self.y.eval({self.X: test_x, self.y: test_y}).tolist())
##
samples = generar_muestra_pais(100)
quantity_for_testing = int(100*0.2)
normalizer = Normalizer()
data = normalizer.prepare_data(samples, quantity_for_testing)
classes = np.append(
data["trainingClassesFirst"],
data["testingClassesFirst"],
axis=0
)
sample = { "trainingFeatures": data["trainingFeatures"], "trainingClasses": data["trainingClassesFirst"],"testingFeatures": data["testingFeatures"], "testingClasses": data["testingClassesFirst"]}
sample2 = { "testingFeatures": data["testingFeatures"], "testingClasses": data["testingClassesFirst"]}
print(sample2["testingClasses"])
prueba = logistic_regression_classifier(1,classes)
prueba.train(sample)
print(prueba.classify(sample2))
def train(self):
#need to put into function
if not os.path.exists("./" + param_dir):
os.mkdir(param_dir)
for test in range(self.num_test):
fit_list = []
iteration_list = []
plt.ion()
plt.show()
#set up environment
env = gym.make(self.env)
if self.continuous_action == True:
action_space = env.action_space.shape[0]
else:
action_space = env.action_space.n
state_space = env.observation_space.low.size
param_dict = {}
#get param for the len
best_reward = 0
bbnp = bb_numpy(action_space,state_space,self.max_length,continuous_action = self.continuous_action,state_renormalize = self.state_renormalize)
state = np.zeros(state_space)[None,...]
bbnp.forward_propagate(state,init = True)
#initialize param
param = np.array(bbnp.get_flat_param())
SGD_ = SGD(param, self.lr)
cma_es = cma.CMAEvolutionStrategy(param,self.sigma,{'popsize': self.num_perturbations,})
for iteration in range(self.iterations):
ts = time.time()
cma_param = np.array(cma_es.ask())
if self.method_type == 'CDPP':
if iteration == 0:
X = np.random.randn(self.num_perturbations*3,len(param))
cond_indices = cond_kdpp([], X, k = self.num_perturbations)
self.buffer = X[cond_indices]
else:
dists = np.linalg.norm(self.buffer-param,axis=1)
num_closest= int(self.perc_reuse*self.num_perturbations)
closest_indices = dists.argsort()[:num_closest]
reused_samples = self.buffer[closest_indices]
X = np.random.randn(self.num_perturbations*3,len(param)) + param
cond_indices = cond_kdpp(self.buffer[closest_indices],X,k=(self.num_perturbations-num_closest))
self.buffer = np.vstack((self.buffer[closest_indices],X[cond_indices]))
all_indices = []
queue = Queue()
num_workers = []
num_perts = self.num_perturbations
shared_amt = self.num_perturbations//self.num_cpu
if self.noise_type == 'CDPP' and iteration!=0:
num_perts = self.num_perturbations - int(self.perc_reuse*self.num_perturbations)
shared_amt = (self.num_perturbations - int(self.perc_reuse*self.num_perturbations))//self.num_cpu
while num_perts > shared_amt:
num_perts -= shared_amt
num_workers.append(shared_amt)
num_workers.append(num_perts)
start_indices = [0]
# print('num_workers=',num_workers)
for i in range(1,len(num_workers)):
start_indices.append(start_indices[i-1]+num_workers[i-1])
cma_param_slicer = [0]
cma_param_slicer.extend(num_workers)
cma_param_slicer = np.cumsum(cma_param_slicer)
seed_id = np.random.randint(np.iinfo(np.int32(10)).max, size=len(num_workers))
workers = [Process(target = self._do_work,args = (queue,bbnp,param,action_space,state_space,self.max_length,seed_id[i],num_workers[i],env,self.continuous_action,cma_param[cma_param_slicer[i]:cma_param_slicer[i+1],:],start_indices[i])) for i in range(len(seed_id))]
for worker in workers:
worker.start()
results = [queue.get() for p in workers]
# Swapping this with the above line so deadlock is avoided
for worker in workers:
worker.join()
if self.noise_type == 'CDPP':
if iteration != 0:
old_pert_fitness = pert_fitness[closest_indices]
old_anti_pert_fitness = anti_pert_fitness[closest_indices]
pert_fitness,anti_pert_fitness,seed_id = get_info_summary(results)
pert_fitness = np.array(pert_fitness)[...,None]
anti_pert_fitness = np.array(anti_pert_fitness)[...,None]
if self.noise_type == 'CDPP':
if iteration != 0:
pert_fitness = np.vstack((old_pert_fitness,pert_fitness))
anti_pert_fitness = np.vstack((old_anti_pert_fitness,anti_pert_fitness))
if self.noise_type in ['Gaussian','Hadamard']:
noises = get_noise_matrices(seed_id,num_workers,len(param),self.sigma,self.noise_type)
#record average_fit
average_fit = np.sum(pert_fitness)/self.num_perturbations
fit_list.append(average_fit)
iteration_list.append(iteration)
#dynamic plot the graph
plt.plot(iteration_list,fit_list,'r')
plt.draw()
plt.pause(0.3)
#Ranking best
if self.method_type == 'Rank':
top_ind = np.sort(np.argsort(pert_fitness,axis =0)[-self.best:][::-1],axis = 0).flatten()
pert_fitness = pert_fitness[top_ind]
gradient = (1 / len(top_ind) / self.sigma * (noises[top_ind,:].T@pert_fitness)).flatten()
SGD_gradient = SGD_.get_gradients(gradient)
param = param + SGD_gradient
print('param',param)
#Vanilla
elif self.method_type == 'Vanilla':
gradient = (1 / self.num_perturbations / self.sigma * (noises.T@pert_fitness)).flatten()
SGD_gradient = SGD_.get_gradients(gradient)
param = param + SGD_gradient
#CMA
elif self.method_type == 'CMA':
cma_es.tell(cma_param,-pert_fitness[:,0] - compute_weight_decay(0.01,cma_param))
param = cma_es.result[5] # mean of all perturbations - for render and save - not used to update new space
#ARS
elif self.method_type == 'ARS':
fb_fitness = np.hstack((pert_fitness,anti_pert_fitness))
top_ind = (np.argsort(np.max(fb_fitness,axis = 1,keepdims = True),axis = 0)[-self.best:][::-1]).flatten()
fit_diff = pert_fitness - anti_pert_fitness
reward_noise = np.std(np.vstack((pert_fitness,anti_pert_fitness)))
fit_diff = fit_diff[top_ind]
gradient = (1 / len(top_ind) / self.sigma/reward_noise * (noises[top_ind,:].T@fit_diff)).flatten()
SGD_gradient = SGD_.get_gradients(gradient)
param = param + SGD_gradient
#CDPP
elif self.method_type == 'CDPP':
cond_noise = self.buffer - param
if iteration != 0:
noises = np.vstack((reused_samples,cond_noise))
else:
noises = cond_noise
top_ind = np.sort(np.argsort(pert_fitness,axis =0)[-self.best:][::-1],axis = 0).flatten()
best_pert_fitness = pert_fitness[top_ind]
gradient = (1 / self.num_perturbations / self.sigma * ((self.buffer - param)[top_ind,:].T@best_pert_fitness)).flatten()
param = param + self.lr * gradient
if iteration % self.video_save_interval == 0 and iteration !=0:
normal = Normalizer(state_space)
video_env = gym.wrappers.Monitor(env, './videos/' + str(self.env) + '/'+ str(self.method_type) + '_perturbations_' + str(self.num_perturbations) + '_' +'state_renormalize_' + str(self.state_renormalize)+ '_Simga_' + str(self.sigma) + "_best_" + str(self.best) + "_iter_" +str(iteration) + "_test_" + str(self.test))
bbnp.roll_out(param,video_env,self.env,normal,render = True)
print("-" * 100)
te = time.time()
#print the results
print('iteration: {} | average_fit: {} | # params: {} | time: {:2.2f}s'.format(iteration,average_fit,len(param),te-ts))
self.save_param(path,name,param,average_fit,iteration)
class MLP:
"""
This is the class used for processing the data and creating a Multi
layered perceptron.
...
Attributes
----------
layers : list of the class Layer
a list of the layers included in the MLP
activations_functions : list of the class Activation
a list of the activations included in the MLP
loss_function : Loss
the loss function defined for the MLP
normalizer : Normalizer
a normalizer class to normalize the data used in the MLP
errors : float []
list for the average error for each training epoch in the MLP
predict_errors : float
a float for the average error of all data rows in predict
Methods
-------
add_layer(self, size, activation_function)
adds a layer to the MLP
_backprop(self, d_loss, learning_rate)
backpropagates through the MLP
train(self, x, y, learning_rate=0.01, n_epochs=10)
method for training the MLP
predict(self, x, y)
method for predicting given data
plot(self, dataset_name, nr_epochs, y_test, y_pred, train_time, pred_time)
method for plotting result from MLP
"""
def __init__(self, loss_function):
"""
Parameters
----------
loss_function : SquaredErrorLoss
loss function for the MLP
"""
self.layers = []
self.activations_functions = []
self.loss_function = loss_function
self.normalizer = Normalizer()
# All errors, returned in end for plotting
self.errors = []
self.predict_errors = 0
def add_layer(self, size, activation_function):
"""Adds a layer to the MLP.
Parameters
----------
size : int
The size of the layer to be added
activation_function : Activation
The activation function for the layer to be added
"""
n_inputs = size
if self.layers:
n_inputs = self.layers[-1].n_nodes
self.layers.append(Layer(size, n_inputs))
self.activations_functions.append(activation_function)
def _backprop(self, d_loss, learning_rate):
"""Backpropagates through the MLP.
Parameters
----------
d_loss : numpy array
The derivated loss
learning_rate : float
The rate at which the MLP learns
"""
for i in range(len(self.layers) - 1, 1, -1):
input_data = self.layers[i].get_output()
loss = self.activations_functions[i].backward(input_data).reshape(
len(input_data), -1) * d_loss
d_loss = self.layers[i].backprop(loss, learning_rate)
def train(self, x, y, learning_rate=0.01, n_epochs=10):
"""Backpropagates through the MLP.
Parameters
----------
n_epochs : int
The number of epochs the MLP will run
x : floats [][]
The input data for training the MLP
y : floats [][]
The target data for each training data
learning_rate : float
The rate at which the MLP learns
"""
# Normalize data
self.normalizer.fit(x, y)
x, y = self.normalizer.normalize(x, y)
# Main loop, handles forward and backprop
for _ in range(n_epochs):
output = x
for j, layer in enumerate(self.layers):
# Activation forward, sets the input and output for each layers
output = self.activations_functions[j].forward(
layer.forward(output))
layer.set_output(output)
# Current error for the epoch is saved
error = (np.average(self.loss_function.forward(output, y)))
self.errors.append(error)
# Get the derivative loss from the output node
d_loss = self.loss_function.backward(output, y)
# Backprop
self._backprop(d_loss, learning_rate)
def predict(self, x, y):
"""Predicts a given input with the MLP.
Parameters
----------
x : floats [][]
The input data for training the MLP
y : floats [][]
The target data for each training data
"""
# Normalize data
x, y = self.normalizer.normalize(x, y)
output = x
for j, layer in enumerate(self.layers):
output = self.activations_functions[j].forward(
layer.forward(output))
self.predict_errors = (np.average(self.loss_function.forward(
output, y)))
return self.normalizer.renormalize(output)
def plot(self, dataset_name, nr_epochs, y_test, y_pred, train_time,
pred_time):
"""Plots the results from a dataset trained and predicted with the MLP.
Parameters
----------
dataset_name : str
The name of the dataset being plot
nr_epochs : int
The number of epochs the MLP has been trained
y_test : floats [][]
The target data for each training data
y_pred : floats [][]
The predicted result from the MLP
train_time : float
The elapsed time of the MLP training on the dataset
pred_time : float
The elapsed time of the MLP prediction on the dataset
"""
xy_max = max(max(y_pred), max(y_test))
xy_min = min(min(y_pred), min(y_test))
plt.figure(figsize=(10, 6))
plt.suptitle("Data: {}".format(dataset_name))
plt.subplot(121)
plt.scatter(np.arange(1, nr_epochs + 1), self.errors, label='loss')
plt.title("Average Loss by epoch")
plt.xlabel('Epochs')
plt.ylabel('Loss')
pred_info = "Total pred time: {:.6f}\n".format(pred_time)
train_info = "Total train time: {:.2f}\n".format(train_time)
average_loss = "Pred MSE: {:.2f}\n".format(
np.average(self.predict_errors))
epochs = "Number of epochs: {}".format(nr_epochs)
text = pred_info + train_info + average_loss + epochs
plt.annotate(text,
xy=(1, 1),
xytext=(-15, -15),
fontsize=10,
xycoords='axes fraction',
textcoords='offset points',
bbox=dict(facecolor='white', alpha=0.8),
horizontalalignment='right',
verticalalignment='top')
plt.subplot(122)
plt.scatter(y_test, y_pred)
plt.xlim(xy_min, xy_max)
plt.ylim(xy_min, xy_max)
plt.xlabel("Target")
plt.ylabel("Predicted")
plt.title("Actual Y vs Predicted ")
plt.show()
while line:
strLine = line.split(delim)
try:
entryPointNumber = int(strLine[0])
epName = genericAgentTypeName(strLine[2])
activityEntryPointName = epName + "_" + strLine[0].replace(" ", "_") + "_" + strLine[1].replace(" ", "_")
entryPointName = strLine[0].replace(" ", "_") + "_" + strLine[1].replace(" ", "_")
totalEntries = int(strLine[16])
if totalEntries > 0:
## Generating list of agenttypes over entry points
if not entryPointName in agentTypesAtEntryPoints.keys():
agentTypesAtEntryPoints[entryPointName] = []
agentTypesAtEntryPoints[entryPointName].append(activityEntryPointName)
## Writing Time Tables
n = Normalizer()
timetableHeader = "\n<timetableData ID=\"tt_" + epName + "_" + n.normalize(entryPointName) + "\">"
timetableNote = "<!-- Time Table for Entrypoint ID: " + str(entryPointNumber) + " -->"
timetableFooter = "</timetableData>"
outTTHdl.write(indent1 + timetableHeader + "\n")
outTTHdl.write(indent2 + timetableNote + "\n")
time1 = 0
timeIncrement = 2
for i in range(3, 15):
try:
pctCount = int(strLine[i])
except:
pctCount = 0
time2 = time1 + timeIncrement
count = int(round(totalEntries * (pctCount / 100.0)))
class LDA(object):
def __init__(self, vectorizer="tfidf", stopwords_path=None):
self.vectorizer = vectorizer
self.stopwords_path = stopwords_path
self.normalizer = Normalizer(self.stopwords_path)
self.normalizer.load_stopwords()
@print_run_time
def tranform_corpora(self, corpora_dir, corpora_path, id2word_path):
"""转化语料
1. 从{corpora_dir}文件夹下提取所有.txt文件作为语料
2. 文件总每一行经过预处理后作为一行,存入{corpora_path}文件
3. 保存id2word到{id2word_path}文件
"""
self.corpora_dir = corpora_dir
self.corpora_path = corpora_path
self.id2word_path = id2word_path
self._transform_corpora(self.normalizer, self.corpora_dir, self.corpora_path, self.id2word_path)
@print_run_time
def train_lda(self, model_dir, model_fname, num_topics):
self.model_dir = model_dir
self.model_fname = model_fname
self.num_topics = num_topics
self.model = self._train_lda(self.vectorizer, self.corpora_path, self.id2word_path,
self.model_dir, self.model_fname, self.num_topics)
self.model_path = os.path.join(self.model_dir, self.vectorizer, self.model_fname)
@staticmethod
def _transform_corpora(normalizer, corpora_dir, corpora_path, id2word_path):
"""转化语料
1. 从{corpora_dir}文件夹下提取所有.txt文件作为语料
2. 文件总每一行经过预处理后作为一行,存入{corpora_path}文件
3. 保存id2word到{id2word_path}文件
Args:
corpora_dir(path) :- 语料文件所在的文件夹
corpora_path(path) :- 汇总所有语料的.txt文件
id2word_path(path) :- gensim的字典文件
"""
corpora = []
if not os.path.isdir(corpora_dir):
raise OSError(corpora_dir, "doesn't exist")
if not os.path.isdir(os.path.dirname(corpora_path)):
raise OSError(os.path.dirname(corpora_path), " doesn't exist")
if not os.path.isdir(os.path.dirname(os.path.dirname(id2word_path))):
raise OSError("the grandparent directory of ", id2word_path, " doesnt't exist")
output_tfidf = open(corpora_path, 'a', encoding="utf8")
for file in os.listdir(corpora_dir):
if file.endswith('txt'):
file = os.path.join(corpora_dir, file)
print(file+' read')
with open(file, encoding="utf8") as f:
lines = f.readlines()
for line in lines:
words = normalizer.tokenize(line)
if len(words) > 0:
corpora.append(words)
output_tfidf.write('{}\n'.format(" ".join(words)))
f.close()
output_tfidf.close()
id2word = gensim.corpora.Dictionary(corpora)
parent_dir = os.path.dirname(id2word_path)
make_dir(parent_dir)
if not os.path.isfile(id2word_path):
id2word.save(id2word_path)
print('id2word saved')
else:
print(id2word_path, ' already exists')
@staticmethod
def _train_lda(vectorizer, corpora_path, id2word_path, model_dir, model_fname=model_fname, num_topics=10):
"""训练和保存基于tfidf的lda模型
基于{corpora_path}文件保存的语料和{id2word_path}保存的gensim字典来训练lda_tfidf模型,
保存该模型到{model_dir}文件夹下
Args:
vectorizer(str) :- 向量化方法, choices=["bow", "tfidf"]
corpora_path(path) :- 保存语料的.txt文件
id2word_path(path) :- 保存gensim字典的文件
model_dir(path) :- 保存gensim LDA模型的文件夹
model_fname(path) :- 模型文件名
num_topics(int) :- lda的超参,主题数
"""
try:
assert vectorizer in ["bow", "tfidf"]
except AssertionError:
raise AssertionError("vectorizer must be bow or tfidf")
if not os.path.isdir(model_dir):
raise OSError(model_dir, "doesn't exist")
corpora = []
with open(corpora_path, 'r', encoding="utf8") as fp:
lines = fp.readlines()
for line in lines:
corpora.append(line.strip())
id2word = gensim.corpora.Dictionary.load(id2word_path)
corpus = [id2word.doc2bow(corpus.split(" ")) for corpus in corpora]
# tfidf的话需要计算idf
if vectorizer == "tfidf":
MmCorpus.serialize(corpus_tfidf_mm, corpus)
corpus = MmCorpus(corpus_tfidf_mm)
model = gensim.models.LdaModel(corpus=corpus, id2word=id2word, num_topics=num_topics)
model_path = os.path.join(model_dir, vectorizer)
make_dir(model_path)
model_path = os.path.join(model_path, model_fname)
if not os.path.isfile(model_path):
model.save(model_path)
print('model saved')
else:
print(f"{model_path} already exists")
return model
@staticmethod
def analysis_topics(fname):
"""将各个主题的关键字打印出来
把self.model.print_topics(10)保存到fname后打印出来
"""
f = open(fname, 'r')
lines = f.readlines()
for line in lines:
print(re.findall(r'\"([^\"]*)\"', line))
@staticmethod
def _short_long_similarity(model_path, normalizer, id2word_path, short, long):
"""计算长短文本的相似度
Args:
model_path(path) :- gensim.models.ldamodel的保存路径
id2word_path(path) :- gensim.corpora.Dictionary的保存路径
short(str) :- 短文本
long(str) :- 长文本
Returns:
prob(float) :- 长短文本的匹配度
theta(iterables) :- 长文本在lda模型下的主题分布概率,
每个元素为(主题的序号, 对应主题的概率)
"""
lda = gensim.models.LdaModel.load(model_path)
id2word = gensim.corpora.Dictionary.load(id2word_path)
theta = lda[id2word.doc2bow(normalizer.tokenize(long))]
short = normalizer.tokenize(short)
short = set(short)
short = id2word.doc2idx(short)
prob = 0
for word in short:
prob_w = sum([lda.expElogbeta[k][word]*1000 * p_zk for (k, p_zk) in theta])
prob += math.log(prob_w)
prob = prob/len(short)
prob -= math.log(1000)
return prob, theta
def short_long_sim(self, short, long):
"""用self.model计算长短文本相似度
"""
return self._short_long_similarity(self.model_path, self.normalizer, self.id2word_path, short, long)
def GetJointLogProbability(logProbabilities, labels, ids): modelAggregator = ModelAggregator() normalizer = Normalizer() unormalizedLogProbs, labels, ids = modelAggregator.Aggregate(logProbabilities, labels, ids) return (normalizer.normalizeLogProbabilities(unormalizedLogProbs), labels,ids)
def __init__(self, vectorizer="tfidf", stopWordsFile=None):
self.vectorizer = vectorizer
self.stopWordsFile = stopWordsFile
self.normalizer = Normalizer(self.stopWordsFile)
self.normalizer.load_stopWords()
class LDA():
def __init__(self, vectorizer="tfidf", stopWordsFile=None):
self.vectorizer = vectorizer
self.stopWordsFile = stopWordsFile
self.normalizer = Normalizer(self.stopWordsFile)
self.normalizer.load_stopWords()
@print_run_time
def create_corporaListAndCorporaText(self, corpora_source, corpora_txt,
id2word_fname):
self.corpora_source = corpora_source
self.corpora_txt = corpora_txt
self.id2word_fname = id2word_fname
self._create_corporaListAndCorporaText(self.normalizer,
self.corpora_source,
self.corpora_txt,
self.id2word_fname)
@print_run_time
def createAndSave_lda(self, ldaModel_save_repo, num_topics):
self.ldaModel_save_repo = ldaModel_save_repo
self.num_topics = num_topics
if self.vectorizer == "tfidf":
self.model = self._createAndSave_lda_tfidf(self.corpora_txt,
self.id2word_fname,
self.ldaModel_save_repo,
self.num_topics)
self.model_fname = self.ldaModel_save_repo + '/gensim_tfidf/crawl_news.model'
else:
self.model = self._createAndSave_lda_bow(self.corpora_txt,
self.id2word_fname,
self.ldaModel_save_repo,
self.num_topics)
self.model_fname = self.ldaModel_save_repo + '/gensim_bow/crawl_news.model'
@staticmethod
def _create_corporaListAndCorporaText(normalizer, corpora_source,
corpora_txt, id2word_fname):
''' 从{corpora_source}文件夹下提取所有.txt文件作为语料
文件总每一行经过预处理后作为一行,存入{corpora_txt}文件
并保存id2word到{id2word_fname}文件
Args:
corpora_source(path) :- 语料文件所在的文件夹
corpora_txt(path) :- 汇总所有语料的.txt文件
id2word_fname(path) :- gensim的字典文件
'''
corpora = []
if not os.path.isdir(corpora_source):
raise OSError(corpora_source, "doesn't exist")
if not os.path.isdir(os.path.dirname(corpora_txt)):
raise OSError(os.path.dirname(corpora_txt), " doesn't exist")
if not os.path.isdir(os.path.dirname(os.path.dirname(id2word_fname))):
raise OSError("the grandparent directory of ", id2word_fname,
" doesnt't exist")
output_tfidf = open(corpora_txt, 'a', encoding="utf8")
for file in os.listdir(corpora_source):
if file.endswith('txt'):
file = os.path.join(corpora_source, file)
print(file + ' read')
with open(file, encoding="utf8") as f:
lines = f.readlines()
for line in lines:
words = normalizer.tokenize(line)
if len(words) > 0:
corpora.append(words)
output_tfidf.write('{}\n'.format(" ".join(words)))
f.close()
output_tfidf.close()
id2word = gensim.corpora.Dictionary(corpora)
parent_dir = os.path.dirname(id2word_fname)
make_dir(parent_dir)
if not os.path.isfile(id2word_fname):
id2word.save(id2word_fname)
print('id2word saved')
else:
print(id2word_fname, ' already exists')
@staticmethod
def _createAndSave_lda_bow(corpora_txt,
id2word_fname,
ldaModel_save_repo,
num_topics=10):
''' 训练和保存基于bow的lda模型
基于{corpora_txt}文件保存的语料和{id2word_fname}保存的gensim字典来训练lda_bow模型,
主题数为{num_topics}
保存该模型到{ldaModel_save_repo}文件夹下
Args:
corpora_txt(path) :- 保存语料的.txt文件
id2word_fname(path) :- 保存gensim字典的文件
ldaModel_save_repo(path) ：- 保存gensim LDA模型的文件夹
num_topics(int) :- lda的超参,主题数
'''
if not os.path.isdir(ldaModel_save_repo):
raise OSError(ldaModel_save_repo, "doesn't exist")
corpora = []
with open(corpora_txt, 'r', encoding="utf8") as fp:
lines = fp.readlines()
for line in lines:
corpora.append(line.strip())
id2word = gensim.corpora.Dictionary.load(id2word_fname)
corpus = [id2word.doc2bow(corpus.split(" ")) for corpus in corpora]
lda_bow = gensim.models.LdaModel(corpus=corpus,
id2word=id2word,
num_topics=num_topics)
make_dir(ldaModel_save_repo + '/gensim_bow')
if not os.path.isfile(ldaModel_save_repo +
'/gensim_bow/crawl_news.model'):
lda_bow.save(ldaModel_save_repo + '/gensim_bow/crawl_news.model')
print('lda_bow saved')
else:
print(ldaModel_save_repo,
'/gensim_bow/crawl_news.model already exists')
return lda_bow
@staticmethod
def _createAndSave_lda_tfidf(corpora_txt,
id2word_fname,
ldaModel_save_repo,
num_topics=10):
''' 训练和保存基于tfidf的lda模型
基于{corpora_txt}文件保存的语料和{id2word_fname}保存的gensim字典来训练lda_tfidf模型,
主题数为{num_topics}
保存该模型到{ldaModel_save_repo}文件夹下
Args:
corpora_txt(path) :- 保存语料的.txt文件
id2word_fname(path) :- 保存gensim字典的文件
ldaModel_save_repo(path) ：- 保存gensim LDA模型的文件夹
num_topics(int) :- lda的超参,主题数
'''
if not os.path.isdir(ldaModel_save_repo):
raise OSError(ldaModel_save_repo, "doesn't exist")
corpora = []
with open(corpora_txt, 'r', encoding="utf8") as fp:
lines = fp.readlines()
for line in lines:
corpora.append(line.strip())
id2word = gensim.corpora.Dictionary.load(id2word_fname)
MmCorpus.serialize(
'corpus_tfidf.mm',
[id2word.doc2bow(corpus.split(" ")) for corpus in corpora])
mm = MmCorpus('corpus_tfidf.mm')
lda_tfidf = gensim.models.LdaModel(corpus=mm,
id2word=id2word,
num_topics=num_topics)
make_dir(ldaModel_save_repo + '/gensim_tfidf')
if not os.path.isfile(ldaModel_save_repo +
'/gensim_tfidf/crawl_news.model'):
lda_tfidf.save(ldaModel_save_repo +
'/gensim_tfidf/crawl_news.model')
print('lda_tfidf saved')
else:
print(ldaModel_save_repo,
'/gensim_tfidf/crawl_news.model already exists')
return lda_tfidf
@staticmethod
def analysis_topics(fname):
'''将各个主题的关键字打印出来
把self.model.print_topics(10)保存到fname后打印出来
'''
f = open(fname, 'r')
lines = f.readlines()
for line in lines:
print(re.findall(r'\"([^\"]*)\"', line))
@staticmethod
def _short_long_similarity(lda_fname, normalizer, id2word_fname, short,
long):
'''计算长短文本的相似度
Args:
lda_fname(path) :- gensim.models.ldamodel的保存路径
id2word_fnmae(path) :- gensim.corpora.Dictionary的保存路径
short(str) :- 短文本
long(str) :- 长文本
Returns:
prob(float) :- 长短文本的匹配度
Theta(iterables) :- 长文本在lda模型下的主题分布概率,
每个元素为(主题的序号, 对应主题的概率)
'''
lda = gensim.models.LdaModel.load(lda_fname)
id2word = gensim.corpora.Dictionary.load(id2word_fname)
Theta = lda[id2word.doc2bow(normalizer.tokenize(long))]
short = normalizer.tokenize(short)
short = set(short)
short = id2word.doc2idx(short)
prob = 0
for word in short:
prob_w = sum([
lda.expElogbeta[k][word] * 1000 * p_zk for (k, p_zk) in Theta
])
prob += math.log(prob_w)
prob = prob / len(short)
prob -= math.log(1000)
return prob, Theta
def short_long_sim(self, short, long):
return self._short_long_similarity(self.model_fname, self.normalizer,
self.id2word_fname, short, long)
def writeStates(self, outStateHdl, entryPointID, entryPointName, wayPointProbabilities, countFunctions):
## Entry-/waypoint probability states
if entryPointID == 10101:
pass
# import pdb;pdb.set_trace()
i = 2
propSum = 0
for wayPoint in wayPointProbabilities:
if wayPoint[i] > 0:
propSum += 1
if 0 == 0:
#if propSum > 0:
entryPointName = entryPointName.replace(" ", "_")
outStateHdl.write("<stateType ID=\"" + "entryState_" + self.genericAgentId + "_" + str(entryPointID) + "_" + entryPointName + "\">\n")
outStateHdl.write(indent2 + "<iconcolour>" + self.iconColor + "</iconcolour>\n")
outStateHdl.write(indent2 + "<speed>\n")
outStateHdl.write(indent2 + indent2 + "<alpha>" + str(self.speedAlpha) + "</alpha>\n")
outStateHdl.write(indent2 + indent2 + "<beta>" + str(self.speedBeta) + "</beta>\n")
outStateHdl.write(indent2 + "</speed>\n")
outStateHdl.write(indent2 + "<headOnFactor>1.0</headOnFactor>\n");
outStateHdl.write(indent2 + "<agentMode>network</agentMode>\n")
outStateHdl.write(indent2 + "<type>normal</type>\n")
outStateHdl.write(indent2 + "<instantiationFunctions>\n" + indent2 + indent2 + "<function name=\"selectWaypoint\">\n")
for wayPoint in wayPointProbabilities:
if wayPoint[i] > 0:
blanks = " "
for cc in range(3 - len(str(wayPoint[i]))):
blanks += " "
n = Normalizer()
outStateHdl.write(indent2 + indent2 + "<waypoint id=\"" + str(wayPoint[0]) + "\" probability=\"" + str(wayPoint[i]) + "\" />" + blanks + "<!-- " + n.normalize(wayPoint[1]) + "-->\n")
outStateHdl.write(indent2 + indent2 + "</function>\n" + indent2 + "</instantiationFunctions>\n")
countFunctions += 1
outStateHdl.write("</stateType>\n\n")
i += 1
## Pause at waypoint behaviour
entryPointName = entryPointName.replace(" ", "_")
outStateHdl.write("<stateType ID=\"" + "atWayPointState_" + self.genericAgentId + "_" + str(entryPointID) + "_" + entryPointName + "\">\n")
outStateHdl.write(indent2 + "<iconcolour>" + self.iconColorPause + "</iconcolour>\n")
outStateHdl.write(indent2 + "<speed>\n")
outStateHdl.write(indent2 + indent2 + "<alpha>" + str(self.speedAlpha) + "</alpha>\n")
outStateHdl.write(indent2 + indent2 + "<beta>" + str(self.speedBeta) + "</beta>\n")
outStateHdl.write(indent2 + "</speed>\n")
outStateHdl.write(indent2 + "<headOnFactor>1.0</headOnFactor>\n");
outStateHdl.write(indent2 + "<agentMode>network</agentMode>\n")
outStateHdl.write(indent2 + "<type>normal</type>\n")
outStateHdl.write(indent2 + "<categoricTransitionFunctions>\n")
outStateHdl.write(indent2 + indent2 + "<function name=\"waitTime\">\n")
outStateHdl.write(indent2 + indent2 + indent2 + "<parameter>" + str(self.wayPointWait) + "</parameter>\n")
outStateHdl.write(indent2 + indent2 + indent2 + "<toState>" + "Exit_" + self.genericAgentId + "_" + str(entryPointID) + "_" + entryPointName + "</toState>\n")
outStateHdl.write(indent2 + indent2 + "</function>\n")
outStateHdl.write(indent2 + "</categoricTransitionFunctions>\n")
outStateHdl.write("</stateType>\n\n")
## Exit behaviour
outStateHdl.write("<stateType ID=\"" + "exitState_" + self.genericAgentId + "_" + str(entryPointID) + "_" + entryPointName + "\">\n")
outStateHdl.write(indent2 + "<iconcolour>" + self.iconColorExit + "</iconcolour>\n")
outStateHdl.write(indent2 + "<speed>\n")
outStateHdl.write(indent2 + indent2 + "<alpha>" + str(self.speedAlpha) + "</alpha>\n")
outStateHdl.write(indent2 + indent2 + "<beta>" + str(self.speedBeta) + "</beta>\n")
outStateHdl.write(indent2 + "</speed>\n")
outStateHdl.write(indent2 + "<headOnFactor>1.0</headOnFactor>\n");
outStateHdl.write(indent2 + "<agentMode>network</agentMode>\n")
outStateHdl.write(indent2 + "<type>normal</type>\n")
outStateHdl.write(indent2 + "<instantiationFunctions>\n")
outStateHdl.write(indent2 + indent2 + "<function name=\"selectWaypoint\">\n")
outStateHdl.write(indent2 + indent2 + indent2 + "<waypoint id=\"" + str(entryPointID) + "\" probability=\"100\"/>\n")
outStateHdl.write(indent2 + indent2 + "</function>\n")
outStateHdl.write(indent2 + "</instantiationFunctions>\n")
outStateHdl.write("</stateType>\n\n")
from HyperParameter import HyperParameter
from Normalizer import Normalizer
from AiPolicy import AIPolicy
from Explorator import Explorator
from Trainer import Trainer
# Running the main code
def mkdir(base, name):
path = os.path.join(base, name)
if not os.path.exists(path):
os.makedirs(path)
return path
workDir = mkdir('exp', 'brs')
monitorDir = mkdir(workDir, 'monitor')
hyperParameter = HyperParameter()
np.random.seed(hyperParameter.seed)
environment = gym.make(hyperParameter.environmentName)
environment = wrappers.Monitor(environment, monitorDir, force=True)
inputsNumber = environment.observation_space.shape[0]
outputsNumber = environment.action_space.shape[0]
policy = AIPolicy(inputsNumber, outputsNumber, hyperParameter)
normalizer = Normalizer(inputsNumber)
explorator = Explorator(hyperParameter, normalizer, policy, environment)
trainer = Trainer(policy, normalizer, hyperParameter, explorator)
trainer.train()
from NN import NeuralNetwork from Normalizer import Normalizer import numpy as np from sklearn.datasets import load_breast_cancer from sklearn.model_selection import train_test_split cancer = load_breast_cancer() X = cancer['data'] y = cancer['target'] length = len(cancer['feature_names']) nn = NeuralNetwork([length, 30, 20, 10, 5, 1]) X_train, X_test, y_train, y_test = train_test_split(X, y) normalize = Normalizer() normalize.fit(X_train) X_train = normalize.transform(X_train) X_test = normalize.transform(X_test) nn.fit(X_train, y_train, epochs=1000, verbose=False) predictions = nn.predict(X_test) print(nn.cost(predictions, y_test))
def __init__(self, vectorizer="tfidf", stopwords_path=None):
self.vectorizer = vectorizer
self.stopwords_path = stopwords_path
self.normalizer = Normalizer(self.stopwords_path)
self.normalizer.load_stopwords()
def collect_news(self):
self.output = self.config["DEFAULT"]["output"]
query_interval = int(self.config["DEFAULT"]["query_interval"])
iterations = int(self.config["DEFAULT"]["iterations"])
iterations_counter = 0
while iterations_counter <= iterations:
for section in self.config.sections():
for option in self.config[section]:
url_base = self.config[section]['url_base']
if option not in [
'url_base', 'query_interval', 'tmp', 'output',
'iterations'
]:
path = url_base + self.config[section][option]
try:
xml_data = requests.get(path)
except requests.exceptions.ChunkedEncodingError:
if self._callback:
self._callback("DLERR", path)
continue
try:
tree = ET.ElementTree(
ET.fromstring(xml_data.content))
except ET.ParseError:
if self._callback:
self._callback("PARSEERR", path)
continue
root = tree.getroot()
news_list = root.findall('./channel/item')
normalized_option = Normalizer.normalize_name(option)
self.create_dir_if_not_exists(section)
output_xml_file = self.create_output_file_if_not_exists(
normalized_option, section)
tree_output = ET.parse(output_xml_file)
root = tree_output.getroot()
for article in news_list:
article_title = article.find('title')
article_date = article.find('pubDate')
try:
article_title.text = self.normalize_value(
article_title.text)
article_date.text = self.normalize_value(
article_date.text)
search_filter = './item[title=' + '"' + article_title.text + '"' + "]" \
+ '[pubDate=' + '"' + article_date.text + '"' + ']'
all_items = root.findall(search_filter)
if len(all_items) == 0:
if self._callback:
self._callback("NEWARTICLE", section,
option,
article_title.text)
root.append(article)
except AttributeError:
if self._callback:
self._callback("BADFORMAT")
tree_output.write(output_xml_file)
if self._callback:
self._callback("WAITING", query_interval)
self._callback("CANINTERR")
time.sleep(query_interval)
iterations_counter += 1
