def test_calc_bins(self):
#Initialization
dataRetriever = DataRetriever("../Datasets/metadata.json")
breastCancer = dataRetriever.retrieveData("breastCancer")
continousAttributes = [
"clumpThickness", "uniformityOfCellSize", "uniformityOfCellShape",
"marginalAdhesion", "singleEpithelialCellSize", "bareNuclei",
"blandChromatin", "normalNucleoli", "mitoses", "class"
]
bd = BinDiscretizer(breastCancer["clumpThickness"], bins=8)
fitted_data = bd.train_fit()
numpy_bins = np.histogram_bin_edges(breastCancer["clumpThickness"],
bins=8)
self.assertEqual(np.allclose(bd.bin_edges, numpy_bins[1:]), True,
"Should produce the same bins as np.histogram")
#Numpy digitize has this weird shenanigan where values of array.max() are considered the next highest bin:
#https://stackoverflow.com/questions/4355132/numpy-digitize-returns-values-out-of-range
#These lines convert the array returned to consider the max values as part of the right most bin
numpy_digitize = np.digitize(breastCancer["clumpThickness"],
numpy_bins)
max_vals = np.asarray(
np.where(numpy_digitize == numpy_digitize.max())).flatten()
numpy_digitize[max_vals] = numpy_digitize.max() - 1
print("Shape", breastCancer.shape)
self.assertEqual(np.allclose(numpy_digitize, fitted_data), True,
"Should produce the same results as np.digitize")
def test_data_retrieval(self):
dataRetriever = DataRetriever("../Datasets/metadata.json")
# This test is failing because the test itself isn't working
# self.assertEqual(dataRetriever.retrieveData("breastCancer"), pd.DataFrame() , "Should return a dataframe")
self.assertEqual(dataRetriever.retrieveData("dogDiseases"), None,
"Should return null since no data exist")
def test_train_test_sizes(self):
#Initialization
dataRetriever = DataRetriever("../Datasets/metadata.json")
dataRetriever.retrieveData("breastCancer")
breastCancer = dataRetriever.getDataSet()
continousAttributes = [
"clumpThickness", "uniformityOfCellSize", "uniformityOfCellShape",
"marginalAdhesion", "singleEpithelialCellSize", "bareNuclei",
"blandChromatin", "normalNucleoli", "mitoses", "class"
]
for test, train in KFolds(breastCancer, 10):
self.assertEqual(len(test) + len(train), len(breastCancer))
def test_train_test_independence(self):
#Initialization
dataRetriever = DataRetriever("../Datasets/metadata.json")
dataRetriever.retrieveData("breastCancer")
breastCancer = dataRetriever.getDataSet()
continousAttributes = [
"clumpThickness", "uniformityOfCellSize", "uniformityOfCellShape",
"marginalAdhesion", "singleEpithelialCellSize", "bareNuclei",
"blandChromatin", "normalNucleoli", "mitoses", "class"
]
for test, train in KFolds(breastCancer, 10):
#https://stackoverflow.com/questions/3170055/test-if-lists-share-any-items-in-python
self.assertFalse(bool(set(test.index) & set(train.index)))
def test_q_calculation(self):
dataRetriever = DataRetriever("../Datasets/metadata.json")
dataRetriever.retrieveData("breastCancer")
naiveBayes = NaiveBayes(dataRetriever.getDataSet(),
dataRetriever.getDataClass())
#seperatedByClass = naiveBayes.calculateQ()
#print(seperatedByClass)
self.assertEqual(
dataRetriever.getDataMenu(),
["breastCancer", "glass", "iris", "soybeanSmall", "vote"],
"should return list of data sets")
def testKMeans(self):
data = DataRetriever("../Datasets/metadata.json")
data.retrieveData("computerHardware")
kValue = 15
t = Timer()
t.start()
mediods = KMediods(data.getDataSet(), data.getDataClass(),
data.getDescreteAttributes(),
data.getContinuousAttributes(),
data.getPredictionType(), kValue, 100)
t.stop()
print(f"Time: {t}")
print(mediods)
mediods.to_csv('kmedoids.csv', index=False)
def test_proper_number_of_folds(self):
#Initialization
dataRetriever = DataRetriever("../Datasets/metadata.json")
dataRetriever.retrieveData("breastCancer")
breastCancer = dataRetriever.getDataSet()
continousAttributes = [
"clumpThickness", "uniformityOfCellSize", "uniformityOfCellShape",
"marginalAdhesion", "singleEpithelialCellSize", "bareNuclei",
"blandChromatin", "normalNucleoli", "mitoses", "class"
]
iterations = 0
for test, train in KFolds(breastCancer, 10):
iterations += 1
self.assertEqual(iterations, 10)
def test_creation(self):
#Initialization
dataRetriever = DataRetriever("../Datasets/metadata.json")
breastCancer = dataRetriever.retrieveData("breastCancer")
continousAttributes = [
"clumpThickness", "uniformityOfCellSize", "uniformityOfCellShape",
"marginalAdhesion", "singleEpithelialCellSize", "bareNuclei",
"blandChromatin", "normalNucleoli", "mitoses", "class"
]
with self.assertRaises(Exception) as context:
BinDiscretizer()
self.assertTrue('positional' in str(context.exception),
"Should Raise Error if no data is passed in")
with self.assertRaises(TypeError):
BinDiscretizer(breastCancer["clumpThickness"], bins=None)
def test_stratisfied(self):
#Initialization
dataRetriever = DataRetriever("../Datasets/metadata.json")
dataRetriever.retrieveData("breastCancer")
breastCancer = dataRetriever.getDataSet()
continousAttributes = [
"clumpThickness", "uniformityOfCellSize", "uniformityOfCellShape",
"marginalAdhesion", "singleEpithelialCellSize", "bareNuclei",
"blandChromatin", "normalNucleoli", "mitoses", "class"
]
iterations = 0
for test, train in KFolds(breastCancer, 10, stratisfied=True):
print("TestLen 2", len(test[test['class'] == 2]), "TestLen 4",
len(test[test['class'] == 4]))
print("TrainLen 2", len(train[train['class'] == 2]), "TrainLen 4",
len(train[train['class'] == 4]))
iterations += 1
def test_range_normalizer_bounds(self):
#Initialization
dataRetriever = DataRetriever("../Datasets/metadata.json")
breastCancer = dataRetriever.retrieveData("breastCancer")
continousAttributes = [
"clumpThickness", "uniformityOfCellSize", "uniformityOfCellShape",
"marginalAdhesion", "singleEpithelialCellSize", "bareNuclei",
"blandChromatin", "normalNucleoli", "mitoses", "class"
]
rn = RangeNormalizer(breastCancer[continousAttributes])
fitted = rn.train_fit()
#Check if the mins/maxes of all the fitted columns are 0/1, respectively
self.assertEqual(np.allclose(np.ones(fitted.shape[1]), fitted.max()),
True)
self.assertEqual(np.allclose(np.zeros(fitted.shape[1]), fitted.min()),
True)
def test_untrained(self):
#Initialization
dataRetriever = DataRetriever("../Datasets/metadata.json")
breastCancer = dataRetriever.retrieveData("breastCancer")
continousAttributes = [
"clumpThickness", "uniformityOfCellSize", "uniformityOfCellShape",
"marginalAdhesion", "singleEpithelialCellSize", "bareNuclei",
"blandChromatin", "normalNucleoli", "mitoses", "class"
]
rn = RangeNormalizer(breastCancer[continousAttributes])
sn = StandardNormalizer(breastCancer[continousAttributes])
#Test range normalizer
with self.assertRaises(UntrainedUtilityError):
rn.fit(breastCancer[continousAttributes])
#Test standard normalizer
with self.assertRaises(UntrainedUtilityError):
sn.fit(breastCancer[continousAttributes])
def test_test_set_coverage(self):
#Initialization
dataRetriever = DataRetriever("../Datasets/metadata.json")
dataRetriever.retrieveData("breastCancer")
breastCancer = dataRetriever.getDataSet()
continousAttributes = [
"clumpThickness", "uniformityOfCellSize", "uniformityOfCellShape",
"marginalAdhesion", "singleEpithelialCellSize", "bareNuclei",
"blandChromatin", "normalNucleoli", "mitoses", "class"
]
tested_vals = []
#Add a dummy index to the dataset so we can see which rows are selected each test
breastCancer["dummyIndex"] = np.arange(len(breastCancer)) + 1
for test, train in KFolds(breastCancer, 10):
tested_vals.extend(test["dummyIndex"])
self.assertTrue(set(tested_vals) == set(breastCancer["dummyIndex"]))
def testOneHotEncoder(self):
dataRetriver = DataRetriever("../Datasets/metadata.json")
glassData = dataRetriver.retrieveData("breastCancer")
data = glassData.getDataSet()
unknown = glassData.getDataClass()
train = data.sample(n=6, random_state=69)
test = data.sample(n=6, random_state=420)
ohe = OneHotEncoder()
encodedDataFrame = ohe.train_fit(train,
glassData.getDescreteAttributes())
encodedDict = ohe.encodedDict
encodedTest = ohe.fit(test)
# print(encodedDataFrame)
# print(encodedDict)
print("=============Train============")
print(encodedDataFrame[unknown])
print(train[unknown])
print("=============Test=============")
print(encodedTest[unknown])
print(test[unknown])
from MLAlgorithms.Utils.DataRetriever import DataRetriever
from MLAlgorithms.KNN.KMediods import KMediods
from MLAlgorithms.Utils.StandardNormalizer import StandardNormalizer
"""
This is the driver we used to create the medoids CSVs. We went with the archaic route and manually
changed the data set for each generation. This was the method chosen since some data sets took longer to calculate
There are a couple of lines that are data set specific
"""
data = DataRetriever("../Datasets/metadata.json")
dataSetName = "computerHardware"
print(f"Creating CSV for {dataSetName}")
data.retrieveData(dataSetName)
maxItter = 100
kValue = 78
# These are only used for image segmentation and abalone
# frac = .25
# random_state = 69
# kValue = m.floor(frac * kValue)
dataSetUnNormalized = data.getDataSet()
# dataSetUnNormalized[data.getDataClass()] = np.log(dataSetUnNormalized[data.getDataClass()] + 0.001) // This is for Forest Fires
sn = StandardNormalizer(dataSetUnNormalized[data.getContinuousAttributes()])
dataSetUnNormalized[data.getContinuousAttributes()] = sn.train_fit()
def network_tuner(*nodes_per_hidden_layer):
"""
This function is used to calcuate the optimal network architecture
The user should input the dataset they would like to operate with and change the performance metric in accordance to the data set type IE regression or classification
"""
MSEs = []
bestNetwork = {}
learning_rate = 0.0001
maxItter = 500
batch_size = .5
dataRetriever = DataRetriever("../Datasets/metadata.json")
dataRetriever.retrieveData("glass")
dataset = dataRetriever.getDataSet().dropna()
dataset = dataset.reset_index(drop=True)
# This line is used to normalize the data for Forest Fires
# dataset[dataRetriever.getDataClass()] = np.log(dataset[dataRetriever.getDataClass()]+0.1)
dataset[dataRetriever.getContinuousAttributes()] = (dataset[dataRetriever.getContinuousAttributes()]-dataset[dataRetriever.getContinuousAttributes()].mean())/dataset[dataRetriever.getContinuousAttributes()].std()
test_set = dataset.sample(frac=0.1, random_state=69)
train_set = dataset.drop(test_set.index)
test_set = test_set.reset_index(drop=True)
train_set = train_set.reset_index(drop=True)
ohe = OneHotEncoder()
discrete_attr = dataRetriever.getDescreteAttributes()
if dataRetriever.getDataClass() in discrete_attr:
discrete_attr.remove(dataRetriever.getDataClass())
datasetEncoded = ohe.train_fit(train_set, dataRetriever.getDescreteAttributes())
testEncoded = ohe.fit(test_set)
output = None
nn = NeuralNetwork(datasetEncoded, 0, [], dataRetriever.getPredictionType(), dataRetriever.getDataClass())
for i in range(maxItter):
# We don't call an inital feedforward because backpropagate starts with a feedforward call
# batch_size represents the number of data points per batch
output = nn._back_propagate(learning_rate=learning_rate, batch_size=batch_size)
final = nn.test(testEncoded.drop(dataRetriever.getDataClass(), axis=1))
output = nn._feed_forward(testEncoded.drop(dataRetriever.getDataClass(), axis=1), testing=True)
actual = testEncoded[dataRetriever.getDataClass()]
## ===================== Classification =================
correct = 0
acc = 0
for i, row in enumerate(final):
if row == actual.iloc[i]: correct += 1
# final = final.reshape(final.shape[0])
# MSE = ((actual-final)**2).mean()
# MSEs.append(MSE)
bestNetwork['network'] = nn
bestNetwork['acc'] = acc
bestNetwork['arc'] = [0]
# # ============================================
# # ============ Compare Acc to Most Common Class
values = test_set[dataRetriever.getDataClass()].value_counts()
# USED FOR CLASSIFICATION
# print(f'Accuracy: {acc}')
# print(f'Max Class Prior: {values.max()/values.sum()}')
# print(f"Class Distribution:\n{values}")
# print("Final: ", final)
# print("Actual: ", list(actual))
# print()
numOfLayer = len(nodes_per_hidden_layer)
print("Number of Hidden Layers: ", numOfLayer)
for layer in range(numOfLayer):
print(f"Layer Number: {layer + 1}")
combinations = list(itertools.product(*nodes_per_hidden_layer[:layer+1]))
for combo in combinations:
output = None
print("Node Combination: ",list(combo))
print(combo)
nn = NeuralNetwork(datasetEncoded, layer, list(combo), dataRetriever.getPredictionType(), dataRetriever.getDataClass())
for i in range(maxItter):
# We don't call an inital feedforward because backpropagate starts with a feedforward call
# batch_size represents the number of data points per batch
output = nn._back_propagate(learning_rate=learning_rate, batch_size=batch_size)
final = nn.test(testEncoded.drop(dataRetriever.getDataClass(), axis=1))
output = nn._feed_forward(testEncoded.drop(dataRetriever.getDataClass(), axis=1), testing=True)
actual = testEncoded[dataRetriever.getDataClass()]
## ===================== Classification =================
correct = 0
acc = 0
for i, row in enumerate(final):
if row == actual.iloc[i]: correct += 1
acc = correct/len(test_set)
# # # ============================================
# # # ============ Compare Acc to Most Common Class
values = test_set[dataRetriever.getDataClass()].value_counts()
# USED FOR CLASSIFICATION
# print(f'Accuracy: {acc}')
# print(f'Max Class Prior: {values.max()/values.sum()}')
# # print(f"Class Distribution:\n{values}")
# print("Final: ", final)
# print("Actual: ", list(actual))
# print()
if acc > bestNetwork['acc']:
bestNetwork['network'] = nn
bestNetwork['acc'] = acc
bestNetwork['arc'] = combo
# final = final.reshape(final.shape[0])
# MSE = ((actual-final)**2).mean()
# MSEs.append(MSE)
# if MSE < bestNetwork['acc']:
# bestNetwork['network'] = nn
# bestNetwork['acc'] = MSE
# bestNetwork['arc'] = combo
return bestNetwork#, MSEs
def run_driver(current_data_set,
mutation_rate=0.5,
maxIter=1000,
batch_size=0.6,
population_size=110,
network_architecture=[15],
pb_actor=None):
cost_func = {
"breastCancer": "bin_cross",
"glass": "log_cosh",
"soybeanSmall": "log_cosh",
"abalone": "log_cosh",
"forestFires": "log_cosh",
"computerHardware": "log_cosh"
}
title_text = r"""
______ __ _ ___ __ _ __ __
/ ____/___ ____ ___ / /_ (_)_____ / / / /____ _ ____ _____ (_)/ /_ / /_ ____ ___ _____
/ / __ / _ \ / __ \ / _ \ / __// // ___/ / /| / / // __ `// __ \ / ___// // __// __ \ / __ `__ \ / ___/
/ /_/ // __// / / // __// /_ / // /__ / ___ / / // /_/ // /_/ // / / // /_ / / / // / / / / /(__ )
\____/ \___//_/ /_/ \___/ \__//_/ \___/ /_/ |_//_/ \__, / \____//_/ /_/ \__//_/ /_//_/ /_/ /_//____/
/____/
"""
output_json = {}
# ====================== Adjustable Variables ==============================
# current_data_set = "abalone"
# mutation_rate = .5
# maxIter = 10
# batch_size = .6
# population_size = 110
# network_architecture = []
# ===========================================================================
output_json["parameters"] = {
"mutation_rate": mutation_rate,
"population_size": population_size,
"network_architecture": network_architecture,
"cost_func": cost_func[current_data_set],
"maxIter": maxIter,
"batch_size": batch_size
}
# ================ Data pre-processing =================================================
dataRetriever = DataRetriever("../../Datasets/metadata.json")
dataRetriever.retrieveData(current_data_set)
dataset = dataRetriever.getDataSet().dropna()
discrete_attr = dataRetriever.getDescreteAttributes()
cont_attributes = dataRetriever.getContinuousAttributes()
# This line is used to normalize the data for Forest Fires
if current_data_set == "forestFires":
discrete_attr.remove('month')
discrete_attr.remove('day')
dataset['month'] = (pd.to_datetime(dataset.month,
format='%b').dt.month) - 1
dataset["day"] = dataset['day'].apply(
lambda x: list(calendar.day_abbr).index(x.capitalize()))
dataset["month_sin"] = np.sin(dataset['month'])
dataset["month_cos"] = np.sin(dataset['month'])
dataset["day_sin"] = np.sin(dataset['day'])
dataset["day_cos"] = np.sin(dataset['day'])
dataset = dataset.drop('day', axis=1)
dataset = dataset.drop('month', axis=1)
cont_attributes.append('month_sin')
cont_attributes.append('month_cos')
cont_attributes.append('day_sin')
cont_attributes.append('day_cos')
dataset[dataRetriever.getDataClass()] = np.log(
dataset[dataRetriever.getDataClass()] + 0.000001)
elif current_data_set == "computerHardware":
discrete_attr.remove('venderName')
discrete_attr.remove('modelName')
dataset = dataset.drop('venderName', axis=1)
dataset = dataset.drop('modelName', axis=1)
dataset = dataset.reset_index(drop=True)
if dataRetriever.getDataClass() in discrete_attr:
discrete_attr.remove(dataRetriever.getDataClass())
# ======================= Train Neural Network ================
print(title_text)
fold = 0
metrics = []
for test_set, train_set in KFolds(dataset, 10):
fold += 1
fitness_file = f"../DataDump/GA/{current_data_set}_layer{len(network_architecture)}_fold{fold}_fitness.csv"
output_file = f"../DataDump/GA/{current_data_set}_layer{len(network_architecture)}_fold{fold}_output.csv"
metrics.append(
multiprocess_func.remote(test_set,
train_set,
fold,
fitness_file,
output_file,
dataRetriever,
cost_func[current_data_set],
current_data_set,
mutation_rate,
maxIter,
batch_size,
population_size,
network_architecture,
pb_actor=None))
metrics = ray.get(metrics)
print(metrics)
print("Average Performance: ", np.asarray(metrics).mean())
output_json["Metrics"] = metrics
output_json["Average"] = np.asarray(metrics, dtype=np.float64).mean()
output_json["Std"] = np.asarray(metrics, dtype=np.float64).std()
with open(
f"../DataDump/GA_{current_data_set}_layer{len(network_architecture)}.json",
'w') as f:
json.dump(output_json, f, indent=4)
def test_menu(self):
dataRetriever = DataRetriever("../Datasets/metadata.json")
self.assertEqual(
dataRetriever.getDataMenu(),
["breastCancer", "glass", "iris", "soybeanSmall", "vote"],
"should return list of data sets")
def test_existence(self):
dataRetriever = DataRetriever("../Datasets/metadata.json")
self.assertEqual(dataRetriever.hasData("breastCancer"), True,
"Should have breast cancer data")
self.assertEqual(dataRetriever.hasData("dogDiseases"), False,
"Shouldn't have dog disease data")