def setUpClass(self):
self.DEBUG = False
self.METRICS = False
self.data_api_impl = DataApi('../../../data/')
self.cross_validator_impl = CrossValidator()
self.preprocessor_impl = Preprocessor()
def setUpClass(self):
self.DEBUG = False
self.METRICS = False
# construct DataApi instance with path prefix to data directory (relative from here)
self.data_api_impl = DataApi('../../../data/')
self.distance_functions_impl = DistanceFunctions()
def __init__(self):
# logger instance - VERBOSE level is highest (most verbose) level for logging
self.logger = Logger('DEMO') # configure log level here
# datalayer instance - read csv data files and convert into raw data frames
self.datalayer = DataApi('../../data/')
# preprocessor instance - everything for prerocessing data frames
self.preprocessor = Preprocessor()
# cross_validator instance - setup cross validation partitions
self.cross_validator = CrossValidator()
# utils instance - random things
self.utils = Utils()
def __init__(self):
KNN.__init__(self)
self.DEBUG = True
self.VERBOSE = False
self.data_api_impl = DataApi('../../data/')
self.utilities_impl = Utilities()
self.distance_functions_impl = DistanceFunctions()
# threshold for clustering convergence
# stop iterating when differences between consecutive centroids is smaller than this
self.CONVERGENCE_THRESHOLD = 0.25
# maximum clustering iterations allowed before returning answer
self.MAX_ITERATIONS = 5
class ViewData():
def __init__(self):
self.DEBUG = False
# construct DataApi instance with path prefix to data directory (relative from here)
self.data_api_impl = DataApi('../../data/')
def view_iris_data(self):
print('\nIRIS DATA:\n')
print(self.data_api_impl.get_iris_data())
def __init__(self):
self.DEBUG = True
self.VERBOSE = False
self.data_api_impl = DataApi('../../../data/')
self.data_set = None
self.CLASSIFICATION = True
self.REGRESSION = False
self.algorithm_name = None
def __init__(self):
self.DEBUG = False
# get instances of all the classes needed to run an experiment
self.data_api_impl = DataApi('../../data/')
self.preprocessor_impl = Preprocessor()
self.cross_validator_impl = CrossValidator()
self.parameter_tuner_impl = ParameterTuner()
# algorithm implementations
self.knn_impl = KNN()
self.enn_impl = EditedKNN()
self.cnn_impl = CondensedKNN()
self.kmeans_knn_impl = KMeansClustering()
self.k_medoids_clustering_impl = KMedoidsClustering()
self.results_processor_impl = Results()
self.CLASSIFICATION = False
self.REGRESSION = False
class ViewData():
def __init__(self):
self.DEBUG = False
# construct DataApi instance with path prefix to data directory (relative from here)
self.data_api_impl = DataApi('../../data/')
def view_abalone_data(self):
print('\nABALONE DATA:\n')
print(self.data_api_impl.get_raw_data_frame('abalone'))
def view_car_data(self):
print('\nCAR DATA:\n')
print(self.data_api_impl.get_raw_data_frame('car'))
def view_forestfires_data(self):
print('\nFORESTFIRES DATA:\n')
print(self.data_api_impl.get_raw_data_frame('forestfires'))
def view_machine_data(self):
print('\nMACHINE DATA:\n')
print(self.data_api_impl.get_raw_data_frame('machine'))
def view_segmentation_data(self):
print('\nSEGMENTATION DATA:\n')
print(self.data_api_impl.get_raw_data_frame('segmentation'))
def view_wine_data(self):
print('\nWINE DATA:\n')
print(self.data_api_impl.get_raw_data_frame('wine'))
class DistanceFunctionsTests(unittest.TestCase):
# SETUP
@classmethod
def setUpClass(self):
self.DEBUG = False
self.METRICS = False
# construct DataApi instance with path prefix to data directory (relative from here)
self.data_api_impl = DataApi('../../../data/')
self.distance_functions_impl = DistanceFunctions()
@classmethod
def tearDownClass(self):
pass
# TESTS
# test get manhattan distance
def test_get_manhattan_distance(self):
pass
# test get euclidean distance
def test_get_euclidean_distance(self):
abalone_data = self.data_api_impl.get_raw_data_frame('abalone')
self.assertTrue(abalone_data is not None)
distance_1_2 = self.distance_functions_impl.get_euclidean_distance(abalone_data[0,:], abalone_data[1,:])
print('distance_1_2: ' + str(distance_1_2))
self.weights = [
np.random.randn(self.layer_sizes[2], self.layer_sizes[1])
]
# the first node in every list correpsonds to the the 1 output from top to bottom. Print-wise, this means each row
#,as it is printed, correpsonds to the to all weights cominging in to output node i.
#ensure that we only have layers 1 and 2 b/c RBF will always be of size 3
self.biases = [np.random.randn(self.layer_sizes[2], 1)]
# EXECUTE SCRIPT
if __name__ == '__main__':
print('\nrunning RBFNetwork...\n')
data_set_name = 'segmentation'
data_api_impl = DataApi('../../data/')
data = data_api_impl.get_raw_data_frame(data_set_name)
rbf_network_impl = RBFNetwork(data_set_name, [19, 2, 7])
#initializing neuron values
rbf_network_impl.init_RBF_Neurons(data[0:2])
#print(rbf_network_impl.rbf_neurons)
rbf_network_impl.init_weights_biases()
print('weights: %s' % str(rbf_network_impl.weights))
print('biases: %s' % str(rbf_network_impl.biases))
#test_vector = [121.0,60.0,9,0.0,0.0,2.277778,2.329629,2.888889,2.8740742,26.74074,24.666666,35.22222,20.333334,-6.2222223,25.444445,-19.222221,35.22222,0.4223002,-1.776113]
#rbf_network_impl.get_forward_output(test_vector)
#print(rbf_network_impl.rbf_neurons)
#print(data.loc[0,:])
#print(rbf_network_impl.calculate_std(data))
class ExperimentRunner:
'''
CONSTRUCTOR
'''
def __init__(self):
# logger instance - VERBOSE level is highest (most verbose) level for logging
self.logger = Logger('DEMO') # configure log level here
# datalayer instance - read csv data files and convert into raw data frames
self.datalayer = DataApi('../../data/')
# preprocessor instance - everything for prerocessing data frames
self.preprocessor = Preprocessor()
# cross_validator instance - setup cross validation partitions
self.cross_validator = CrossValidator()
# utils instance - random things
self.utils = Utils()
# get average result given cross validation results dictionary
def get_avg_result(self, cv_results):
result_vals = []
# for each cross validation partition, append result value to corresponding list
for test_data_key in cv_results:
test_result = cv_results[test_data_key]
result_vals.append(test_result)
# should always equal the value of the 'folds' variable in cross validator
test_data_count = len(cv_results)
# calculate average values
avg_result = sum(result_vals) / test_data_count
# return average result
return avg_result
'''
get preprocessed data ready for consumption by experiment running logic
INPUT:
- data_set_name: name of data set to fetch data for
OUTPUT:
- preprocessed data frame - fully ready for experiment consumption
'''
def get_experiment_data(self, data_set_name):
data = self.datalayer.get_raw_data_frame(data_set_name)
self.logger.log('DEMO', 'data_set_name: \t%s\n' % str(data_set_name))
self.logger.log(
'DEMO',
'raw data: \n\n%s, shape: %s\n' % (str(data), str(data.shape)))
self.logger.log('DEMO', '----------------------------------------------------' \
+ '-----------------------------------------------\n')
data = self.preprocessor.preprocess_raw_data_frame(data, data_set_name)
self.logger.log(
'DEMO', 'preprocessed data: \n\n%s, shape: %s\n' %
(str(data), str(data.shape)))
self.logger.log('DEMO', '----------------------------------------------------' \
+ '-----------------------------------------------\n')
return data
'''
run experiment
INPUT:
- data_set_name: name of data set to run experiment on
- neural_network: instance of neural network to train/test with data
- hyperparams: hyperparameters and corresponding values to use in experiment
OUTPUT:
- <void> - logs all the important stuff at DEMO level
'''
def run_experiment(self, data_set_name, neural_network, hyperparams):
# LAYER ACTIVATION FUNCTION SPECIFICATION
self.logger.log(
'DEMO', 'layer_activation_funcs: %s\n' %
str(hyperparams["layer_activation_funcs"]))
# DATA RETRIEVAL AND PREPROCESSING
data = self.get_experiment_data(data_set_name)
self.logger.log('DEMO', 'data_set_name: %s\n' % str(data_set_name))
# CROSS VALIDATION PARTITIONING
# get cross validation partitions for data
cv_partitions = self.cross_validator.get_cv_partitions(data)
# dictionary for storing accuracy results
cv_results = {}
# list of sizes of test sets used for getting average test set size
test_data_sizes = []
# NEURAL NETWORK TRAINING AND TESTING
for partition in cv_partitions:
# initialize key and corresponding nested dictionary in results dictionary
test_data_key = 'test_data_' + str(partition)
cv_results[test_data_key] = {}
# get training set and test set for given cross validation partition
train_data, test_data = cv_partitions[partition]
test_data_sizes.append(
test_data.shape[0]
) # add number of rows in test set to test_set_sizes list
# HANDLE RBF NETWORK P2 RESULTS
if neural_network.network_name == 'RBF':
# configure RBF network shape based on training data
neural_network.configure_rbf_network(train_data, data,
data_set_name,
hyperparams["k"])
# GRADIENT DESCENT
# run gradient descent for given neural network instance
test_result_vals = neural_network.train_gradient_descent(
train_data, hyperparams, partition, test_data)
self.logger.log('DEMO', ('accuracy_vals' if neural_network.CLASSIFICATION else 'error_vals') \
+ ' for partition %s: %s\n' % (str(partition+1), str(test_result_vals)), True)
# append accuracy/error result of final gradient descent iteration to results dictionary
cv_results[test_data_key] = test_result_vals[-1]
# FINAL RESULTS (THE MODEL)
self.logger.log('DEMO', '------------------------------------------------------------' \
+ ' TRAINING DONE ------------------------------------------------------------')
self.logger.log('DEMO', 'trained network: weights --> \n\n%s, shapes: %s\n' \
% (str(neural_network.weights), str(self.utils.get_shapes(neural_network.weights))), True)
self.logger.log('DEMO', 'trained network: biases --> \n\n%s, shapes: %s\n' \
% (str(neural_network.biases), str(self.utils.get_shapes(neural_network.biases))), True)
self.logger.log('DEMO', 'data_set_name: %s\n' % str(data_set_name),
True)
self.logger.log('DEMO', 'trained network: AVERAGE ' \
+ ('ACCURACY' if neural_network.CLASSIFICATION else 'ERROR') + ' --> %s\n' \
% str(self.get_avg_result(cv_results)), True)
def __init__(self):
self.DEBUG = False
self.data_api_impl = DataApi('../../data/')
info = {
"PATTERN": "looper",
"LOOPER": {
"initial_capital": 100000,
"margin_ratio": {
"rb2010.CTP": 0.00003,
},
"commission_ratio": {
"rb2010.CTP": {
"close": 0.00001
},
},
"size_map": {
"rb2010.CTP": 10
}
}
}
app.config.from_mapping(info)
strategy = DoubleMaStrategy("ma")
data_api = DataApi()
data = data_api.get_tick("rb2010",
start_date="2020-04-10",
end_date="2020-07-21",
today=False)
# data = data_support.get_future_min("rb2010.SHFE", frq="1min", start="2019-10-01", end="2020-07-15")
app.add_data(data)
app.add_extension(strategy)
app.start()
result = app.get_result(report=True, auto_open=True)
class ExperimentRunner():
def __init__(self):
self.DEBUG = False
# get instances of all the classes needed to run an experiment
self.data_api_impl = DataApi('../../data/')
self.preprocessor_impl = Preprocessor()
self.cross_validator_impl = CrossValidator()
self.parameter_tuner_impl = ParameterTuner()
# algorithm implementations
self.knn_impl = KNN()
self.enn_impl = EditedKNN()
self.cnn_impl = CondensedKNN()
self.kmeans_knn_impl = KMeansClustering()
self.k_medoids_clustering_impl = KMedoidsClustering()
self.results_processor_impl = Results()
self.CLASSIFICATION = False
self.REGRESSION = False
# run algorithm on data set with various parameters
def run_experiment(self, data_frame_name, algorithm):
self.set_experiment_type(data_frame_name)
# get raw data frame to run experiment against
raw_data_frame = self.data_api_impl.get_raw_data_frame(data_frame_name)
print(raw_data_frame)
# preprocess data
preprocessed_data_frame = self.preprocessor_impl.preprocess_raw_data_frame(raw_data_frame, data_frame_name)
print(preprocessed_data_frame)
# get indexes list for data frame cross validation - a list of row numbers used to partition the data
data_frame_indexes_list = self.cross_validator_impl.get_indexes_list(preprocessed_data_frame)
if self.DEBUG:
print('\ndata_frame_name --> ' + data_frame_name)
print('\nraw_data_frame:\n')
print(raw_data_frame)
print('\npreprocessed_data_frame:\n')
print(preprocessed_data_frame)
print('\ndata_frame_indexes_list for cross validation:\n')
print(data_frame_indexes_list)
# nested dictionary to hold algorithm performance results for each combination of training/test sets
# key pattern --> key = test_set_1 , where the number at the end of the key is the test set index
# each value is another dictionary with keys = { 'zero_one_loss', 'mean_squared_error' }
# the nested dictionary values are the corresponding loss function metrics for predictions using the test set
cross_validation_results = {}
# list of sizes of test sets used for getting average test set size
test_set_sizes = []
algorithm_parameters = self.parameter_tuner_impl.get_params(data_frame_name, algorithm)
# dictionary where key is parameter and value is tuple of average loss function results
results_by_parameter = {}
# get all cross validation partitions for given data frame
cv_partitions = self.cross_validator_impl.get_cv_partitions(preprocessed_data_frame)
# for each parameter value in the list of algorithm parameter values (see ParameterTuner)
for parameter in algorithm_parameters:
if self.DEBUG:
print('\n' + str(self.parameter_tuner_impl.get_parameter_key(algorithm)) + ': ' + str(parameter) + '\n')
# for each test set used in cross validation (number of folds)
for partition in cv_partitions:
# initialize key and corresponding nested dictionary in results dictionary
test_set_key = 'test_set_' + str(partition)
cross_validation_results[test_set_key] = {}
# get training set and test set for given cross validation partition
training_set, test_set = cv_partitions[partition]
test_set_sizes.append(test_set.shape[0]) # add number of rows in test set to test_set_sizes list
if self.DEBUG:
print('preprocessed dataframe before running algorithm:')
print(preprocessed_data_frame)
# run algorithms on training set / test set combination
# returns dictionary where key is the row index (as string) and value is the predicted class for that row
prediction_results = self.run_algorithm(data_frame_name, algorithm, training_set, test_set, \
preprocessed_data_frame, parameter)
# calculate loss function results given prediction results - measure prediction accuracy
accuracy, mean_squared_error = self.results_processor_impl.loss_function_analysis(test_set, prediction_results)
cross_validation_results[test_set_key]['accuracy'] = accuracy
cross_validation_results[test_set_key]['mean_squared_error'] = mean_squared_error
# calculate average loss function results over all cross validation folds
avg_accuracy, avg_mean_squared_error = self.results_processor_impl.get_avg_loss_vals(cross_validation_results)
avg_test_set_size = sum(test_set_sizes) / len(test_set_sizes) # get average test set size for reference
results_by_parameter[str(parameter)] = (avg_accuracy, avg_mean_squared_error)
print('\n\nRESULTS: average test set size: ' + str(avg_test_set_size) + \
((' --> accuracy: ' + str(avg_accuracy)) if self.CLASSIFICATION \
else (' --> mean_squared_error: ' + str(avg_mean_squared_error))))
print('\n---------------------------------------------------------------------------------------------------------------------')
# return dictionary of results by parameter
return results_by_parameter
def set_experiment_type(self, data_frame_name):
if data_frame_name in ['abalone', 'car', 'segmentation']:
self.CLASSIFICATION = True
self.REGRESSION = False
elif data_frame_name in ['machine', 'forestfires', 'wine']:
self.REGRESSION = True
self.CLASSIFICATION = False
else:
raise Exception('ERROR: unknown data_set_name --> ' + str(data_frame_name))
'''
run algorithm execution handler given algorithm name
INPUT:
- algorithm_name: name of algorithm to run handler for
OUTPUT:
- prediction results dictionary, maps instance index to tuple: (prediction, actual)
'''
def run_algorithm(self, data_set_name, algorithm_name, training_set, \
test_set, preprocessed_data_frame, parameter):
if algorithm == 'knn':
self.knn_impl.set_data_set(data_set_name)
self.knn_impl.set_algorithm_name(algorithm_name)
return self.knn_impl.do_knn(training_set, test_set, preprocessed_data_frame, parameter)
elif algorithm == 'enn':
self.enn_impl.set_data_set(data_set_name)
self.enn_impl.set_algorithm_name(algorithm_name)
return self.enn_impl.do_enn(training_set, test_set, preprocessed_data_frame, parameter)
elif algorithm == 'cnn':
self.cnn_impl.set_data_set(data_set_name)
self.cnn_impl.set_algorithm_name(algorithm_name)
return self.cnn_impl.do_cnn(training_set, test_set, preprocessed_data_frame, parameter)
elif algorithm == 'kmeans_knn':
self.kmeans_knn_impl.set_data_set(data_set_name)
self.kmeans_knn_impl.set_algorithm_name(algorithm_name)
return self.kmeans_knn_impl.cluster_do_knn(training_set, test_set, preprocessed_data_frame, data_set_name, parameter)
elif algorithm == 'kmedoids_knn':
self.k_medoids_clustering_impl.set_data_set(data_set_name)
self.k_medoids_clustering_impl.set_algorithm_name(algorithm_name)
return self.k_medoids_clustering_impl.cluster(training_set, test_set, preprocessed_data_frame, data_set_name, parameter)
def __init__(self):
self.DEBUG = False
self.data_api_impl = DataApi('../../data/')
self.utilities_impl = Utilities()
print("Number of Previous Edits: ")
print(number_of_edits_previous)
loopcounter += 1
print("Number of While Loops: ")
return edited_train_set.reset_index(drop=True)
# EXECUTE SCRIPT
if __name__ == '__main__':
print('running edited knn...')
edited_knn = EditedKNN()
data_api_impl = DataApi('../../data/')
cross_validator_impl = CrossValidator()
preprocessor_impl = Preprocessor()
wine_data = data_api_impl.get_raw_data_frame('segmentation')
prep_wine_data = preprocessor_impl.preprocess_raw_data_frame(
wine_data, 'segmentation')
wine_data_train_set = cross_validator_impl.get_training_set(
prep_wine_data, test_set_number=3)
print('wine_data_train_set.shape: ' + str(wine_data_train_set.shape))
wine_data_test_set = cross_validator_impl.get_test_set(
prep_wine_data, test_set_number, indexes_list)
edited_knn.enn(wine_data_train_set, wine_data_test_set, prep_wine_data, k)
def __init__(self):
KNN.__init__(self)
self.DEBUG = False
self.data_api_impl = DataApi('../../data/')
self.enn_impl = EditedKNN()
self.cnn_impl = CondensedKNN()
import base64
from data_api import DataApi
from flask import Flask, url_for
from flask_cors import CORS
from flask_restful import Api, Resource
from model_api import ExplanationModel
from PIL import Image
from flask_restful import reqparse
from attribute_chunker import CounterFactualGenerator
app = Flask(__name__)
api = Api(app)
CORS(app)
data_api = DataApi()
explanation_model = ExplanationModel()
cf_gen = CounterFactualGenerator()
def any_response(data):
ALLOWED = ["http://localhost:8888"]
response = make_response(data)
origin = request.headers["Origin"]
if origin in ALLOWED:
response.headers["Access-Control-Allow-Origin"] = origin
return response
def npimg2base64(img):
pil_img = Image.fromarray(img)
class DataApiTests(unittest.TestCase):
# SETUP
@classmethod
def setUpClass(self):
self.DEBUG = False
self.METRICS = False
# construct DataApi instance with path prefix to data directory (relative from here)
self.data_api_impl = DataApi('../../../data/')
@classmethod
def tearDownClass(self):
pass
# TESTS
# test abalone data retrieval, number of rows/columns
def test_get_abalone_data(self):
abalone_data = self.data_api_impl.get_raw_data_frame('abalone')
self.assertTrue(abalone_data is not None)
self.assertTrue(
abalone_data.shape[0] == 4177) # 4177 rows in abalone data matrix
self.assertTrue(abalone_data.shape[1] ==
9) # 9 attribute columns in abalone data matrix
# test car data retrieval, number of rows/columns
def test_get_car_data(self):
car_data = self.data_api_impl.get_raw_data_frame('car')
self.assertTrue(car_data is not None)
self.assertTrue(
car_data.shape[0] == 1728) # 1728 rows in car data matrix
self.assertTrue(
car_data.shape[1] == 7) # 7 attribute columns in car data matrix
# test forestfires data retrieval, number of rows/columns
def test_get_forestfires_data(self):
forestfires_data = self.data_api_impl.get_raw_data_frame('forestfires')
self.assertTrue(forestfires_data is not None)
self.assertTrue(forestfires_data.shape[0] ==
518) # 518 rows in forestfires data matrix
self.assertTrue(forestfires_data.shape[1] ==
13) # 13 attribute columns in forestfires data matrix
# test machine data retrieval, number of rows/columns
def test_get_machine_data(self):
machine_data = self.data_api_impl.get_raw_data_frame('machine')
self.assertTrue(machine_data is not None)
self.assertTrue(
machine_data.shape[0] == 209) # 209 rows in machine data matrix
self.assertTrue(machine_data.shape[1] ==
10) # 10 attribute columns in machine data matrix
# test segmentation data retrieval, number of rows/columns
def test_get_segmentation_data(self):
segmentation_data = self.data_api_impl.get_raw_data_frame(
'segmentation')
self.assertTrue(segmentation_data is not None)
self.assertTrue(segmentation_data.shape[0] ==
213) # 213 rows in segmentation data matrix
self.assertTrue(segmentation_data.shape[1] ==
20) # 20 attribute columns in segmentation data matrix
# test wine data retrieval, number of rows/columns
def test_get_wine_data(self):
wine_data = self.data_api_impl.get_raw_data_frame('wine')
self.assertTrue(wine_data is not None)
self.assertTrue(
wine_data.shape[0] == 6497) # 6497 rows in wine data matrix
self.assertTrue(wine_data.shape[1] ==
12) # 12 attribute columns in wine data matrix
class CrossValidatorTests(unittest.TestCase):
# SETUP
@classmethod
def setUpClass(self):
self.DEBUG = False
self.METRICS = False
self.data_api_impl = DataApi('../../../data/')
self.cross_validator_impl = CrossValidator()
self.preprocessor_impl = Preprocessor()
@classmethod
def tearDownClass(self):
pass
# TESTS
'''
# test get indexes list for abalone data
def test_get_indexes_list_abalone_data(self):
abalone_data = self.data_api_impl.get_raw_data_frame('abalone')
self.assertTrue(abalone_data is not None)
abalone_indexes = self.cross_validator_impl.get_indexes_list(abalone_data)
self.assertTrue(len(abalone_indexes) == 4177) # 4177 rows in abalone data frame
for i in range(1, 10):
self.assertTrue(abalone_indexes.count(i) == 417) # each subset has 417 rows
self.assertTrue(abalone_indexes.count(10) == 424) # last subset has 417 + remaining...
# test get indexes list for car data
def test_get_indexes_list_car_data(self):
car_data = self.data_api_impl.get_raw_data_frame('car')
self.assertTrue(car_data is not None)
car_indexes = self.cross_validator_impl.get_indexes_list(car_data)
self.assertTrue(len(car_indexes) == 1728) # 1728 rows in car data frame
for i in range(1, 10):
self.assertTrue(car_indexes.count(i) == 172) # each subset has 172 rows
self.assertTrue(car_indexes.count(10) == 180) # last subset has 172 + remaining...
# test get indexes list for forest fires data
def test_get_indexes_list_ff_data(self):
ff_data = self.data_api_impl.get_raw_data_frame('forestfires')
self.assertTrue(ff_data is not None)
ff_indexes = self.cross_validator_impl.get_indexes_list(ff_data)
self.assertTrue(len(ff_indexes) == 518) # 518 rows in forest fires data frame
for i in range(1, 10):
self.assertTrue(ff_indexes.count(i) == 51) # each subset has 51 rows
self.assertTrue(ff_indexes.count(10) == 59) # last subset has 51 + remaining...
# test get indexes list for machine data
def test_get_indexes_list_machine_data(self):
machine_data = self.data_api_impl.get_raw_data_frame('machine')
self.assertTrue(machine_data is not None)
machine_indexes = self.cross_validator_impl.get_indexes_list(machine_data)
self.assertTrue(len(machine_indexes) == 209) # 209 rows in machine data frame
for i in range(1, 10):
self.assertTrue(machine_indexes.count(i) == 20) # each subset has 20 rows
self.assertTrue(machine_indexes.count(10) == 29) # last subset has 20 + remaining...
# test get indexes list for segmentation data
def test_get_indexes_list_segmentation_data(self):
segmentation_data = self.data_api_impl.get_raw_data_frame('segmentation')
self.assertTrue(segmentation_data is not None)
segmentation_indexes = self.cross_validator_impl.get_indexes_list(segmentation_data)
self.assertTrue(len(segmentation_indexes) == 213) # 213 rows in segmentation data frame
for i in range(1, 10):
self.assertTrue(segmentation_indexes.count(i) == 21) # each subset has 21 rows
self.assertTrue(segmentation_indexes.count(10) == 24) # last subset has 21 + remaining...
# test get indexes list for wine data
def test_get_indexes_list_wine_data(self):
wine_data = self.data_api_impl.get_raw_data_frame('wine')
self.assertTrue(wine_data is not None)
wine_indexes = self.cross_validator_impl.get_indexes_list(wine_data)
self.assertTrue(len(wine_indexes) == 6497) # 6497 rows in wine data frame
for i in range(1, 10):
self.assertTrue(wine_indexes.count(i) == 649) # each subset has 649 rows
self.assertTrue(wine_indexes.count(10) == 656) # last subset has 649 + remaining...
# TRAINING SET
# test get training set 2 with wine data
def test_get_training_set(self):
wine_data = self.data_api_impl.get_raw_data_frame('wine')
wine_data_training_set = self.cross_validator_impl.get_training_set(wine_data, 2)
self.assertTrue(wine_data_training_set.shape[0] == 5848) # 6497 - 649 rows in test set 2 means 5484 rows in training set
self.assertTrue(wine_data_training_set.shape[1] == 12) # number of columns does not change
# TEST SET
# test get test set (-2) with wine data
def test_get_test_set(self):
wine_data = self.data_api_impl.get_raw_data_frame('wine')
wine_data_test_set = self.cross_validator_impl.get_test_set(wine_data, 2)
self.assertTrue(wine_data_test_set.shape[0] == 649) # 649 rows in test set 2
self.assertTrue(wine_data_test_set.shape[1] == 12) # number of columns does not change
'''
def test_cv_partitions(self):
abalone_data = self.data_api_impl.get_raw_data_frame('abalone')
prep_abalone_data = self.preprocessor_impl.preprocess_raw_data_frame(abalone_data, 'abalone')
cv_partitions = self.cross_validator_impl.get_cv_partitions(prep_abalone_data)
self.assertTrue(cv_partitions is not None)
for partition in cv_partitions:
train_data_indexes = list(cv_partitions[partition][0].index.values)
test_data_indexes = list(cv_partitions[partition][1].index.values)
for test_index in test_data_indexes:
self.assertTrue(test_index not in train_data_indexes)
centroids_data = self.get_cluster_centroids(cluster_assignments, centroids_data, train_data)
# return a dictionary where key is the instance index and value is a tuple: (prediction, actual)
print('K MEANS CLUSTERING CONVERGED. iterations: ' + str(iteration_count))
return centroids_data
# EXECUTE SCRIPT
if __name__ == '__main__':
print('k means clustering...')
k_means_clustering_impl = KMeansClustering()
data_api_impl = DataApi('../../data/')
preprocessor_impl = Preprocessor()
cross_validator_impl = CrossValidator()
'''
wine_data = data_api_impl.get_raw_data_frame('wine')
prep_wine_data = preprocessor_impl.preprocess_raw_data_frame(wine_data, 'wine')
'''
abalone_data = data_api_impl.get_raw_data_frame('abalone')
prep_abalone_data = preprocessor_impl.preprocess_raw_data_frame(abalone_data, 'abalone')
print('\npossible classes: ' + str(list(set(abalone_data.loc[:, 'CLASS'].values))) + '\n')
training_set, test_set = cross_validator_impl.get_cv_partitions(prep_abalone_data)[0]
trading_record: 格式为 [(1,datetime1)]
1: 开多
2: 开空
-1: 平多
-2： 平空
"""
self.data.setdefault(local_symbol, {})["record"] = trading_record
self.data.setdefault(local_symbol, {})["kline"] = [[
str(kline.datetime), kline.open_price, kline.high_price,
kline.low_price, kline.close_price, kline.volume
] for kline in klines]
def render(self, path):
for local_symbol, obj in self.data.items():
with open(path, "w") as f:
print(obj)
kline_string = kline_template.render(draw_klines=obj["kline"],
bs=obj["record"])
f.write(kline_string)
if __name__ == '__main__':
plot = Plot("some")
from data_api import DataApi
code = "rb2105.SHFE"
data_api = DataApi(uri="http://192.168.1.239:8124")
kline = data_api.get_n_min_bar(code, 1, "2021-04-15", "2021-04-16")
plot.add_kline(code, klines=kline, trading_record=[])
plot.render("x.html")
class KMeansClustering(KNN):
def __init__(self):
KNN.__init__(self)
self.DEBUG = True
self.VERBOSE = False
self.data_api_impl = DataApi('../../data/')
self.utilities_impl = Utilities()
self.distance_functions_impl = DistanceFunctions()
# threshold for clustering convergence
# stop iterating when differences between consecutive centroids is smaller than this
self.CONVERGENCE_THRESHOLD = 0.25
# maximum clustering iterations allowed before returning answer
self.MAX_ITERATIONS = 5
#self.MAX_CLUSTER_TIME = 5 # minutes
'''
perform k-means-clustering against full_data_frame using k value as parameter
INPUT:
- data_set_name: name of data set to cluster
- train_data: training data set to cluster
- k: value for parameter k, i.e. the number of clusters to partition the data set into
OUTPUT:
- tuple:
- index 1: list of cluster keys representing the cluster each data point belongs to
- index 2: centroids dataframe for k centroids (without class values)
'''
def cluster(self, data_set_name, train_data, k):
print('\nk means clustering with k: ' + str(k))
# get list of column labels for data set, not including the CLASS column label
data_column_labels = self.data_api_impl.get_column_labels(data_set_name, include_class=False)
# get training data without class column
train_data = train_data.loc[:, train_data.columns != 'CLASS']
# randomly generate k initial centroids from training data
centroids = self.generate_initial_centroids(train_data, k)
#print('generated initial centroids')
if self.DEBUG:
print('centroids:')
print(centroids)
if not isinstance(centroids, pd.DataFrame):
# convert list of centroids to data frame using same column labels
centroids_df = pd.DataFrame.from_records(centroids, columns=data_column_labels)
#print('created centroids_df from records from centroids')
else:
centroids_df = centroids
# combine the centroids_df and train_data dataframes into one frame with centroids first
centroids_and_data_df = centroids_df.append(train_data, ignore_index=True)
#print('created centroids_and_data_df')
# get distance matrix - distance from every training point to every centroid
distance_from_centroids = self.get_distance_matrix(centroids_and_data_df)
#print('calculated distance matrix for centroids_and_data_df')
if self.DEBUG and self.VERBOSE:
print('cluster: train_data.shape: ' + str(train_data.shape))
print('cluster: k: ' + str(k))
print('cluster: number of centroids: ' + str(len(centroids)))
print('centroids_and_data_df:')
print(centroids_and_data_df)
print('cluster: centroids_and_data_df.shape: ' + str(centroids_and_data_df.shape))
print('cluster: distance_from_centroids.shape: ' + str(distance_from_centroids.shape))
cluster_assignments = []
iteration_count = 1
# initial centroids
previous_centroids = centroids_and_data_df.iloc[:k, :]
# set initial centroids diffs to maximum number sizes
centroids_diff = [sys.maxsize for i in range(k)]
cluster_start_time = time.time()
#print('right before clustering...')
while not self.threshold_reached(centroids_diff) and iteration_count < self.MAX_ITERATIONS:
print('clustering... iteration: ' + str(iteration_count))
new_cluster_assignments = []
# for each training point in the training data (start indexing at k)
for instance_idx in range(k, centroids_and_data_df.shape[0]):
data_point = centroids_and_data_df.iloc[instance_idx, :]
# get list of distances from instance to each centroid
idx_distances = np.array(distance_from_centroids[instance_idx][:k])
# get index of centroid with least distance from instance
closest_centroid_idx = np.argmin(idx_distances)
# map instance index to centroid index (0 - k)
new_cluster_assignments.append(closest_centroid_idx)
#print('calculated new cluster assignments')
cluster_assignments = new_cluster_assignments
updated_centroids = []
# update each centroid to mean of all points assigned to that centroid
for centroid_idx in range(k):
#print('updating mean for centroid: ' + str(centroid_idx))
cluster_points = pd.DataFrame(columns=data_column_labels)
np_cluster_assignments = np.array(cluster_assignments)
# get list of indexes for all instances assigned to given centroid index
idxs_for_cluster_val = np.where(np_cluster_assignments == centroid_idx)[0]
for idx in idxs_for_cluster_val:
# add point to cluster points dataframe, add k to row index to skip centroid points
# NOT EFFICIENT - CHANGE THIS SO IT IS FASTER
cluster_points = cluster_points.append(centroids_and_data_df.iloc[idx+k, :])
#print('built cluster points list')
if self.DEBUG and self.VERBOSE:
print('centroid_idx: ' + str(centroid_idx))
print('idxs_for_cluster_val:' + str(idxs_for_cluster_val))
print('points in cluster ' + str(centroid_idx) + ': ' + str(cluster_points.shape))
print('cluster_points:')
print(cluster_points)
avg_centroid = None
# if there are points that were assigned to the cluster (centroid index)
if cluster_points.shape[0] > 0:
avg_centroid = self.get_avg_centroid(cluster_points)
#print('got average centroid for idx: ' + str(centroid_idx))
if self.DEBUG and self.VERBOSE:
print('avg_centroid:')
print(avg_centroid)
# do not update the centroid if any of the values in the centroid are nan
if avg_centroid is not None and not np.isnan(np.array(avg_centroid)).any():
updated_centroids.append(avg_centroid)
#print('appended avg_centroid to updated_centroids list')
else:
if self.DEBUG and self.VERBOSE:
print('ERROR: avg_centroid is none!')
print('bad cluster points:')
print(cluster_points)
print('avg_centroid:')
print(avg_centroid)
# keep previous centroid if the average centroid is still null (no points in cluster)
updated_centroids.append(list(previous_centroids.iloc[centroid_idx, :].values))
#print('appended previous centroid to updated_centroids list')
# update centroids dataframe using list of updated centroids representing new average centroids
updated_centroids_df = pd.DataFrame.from_records(updated_centroids, columns=data_column_labels)
#print('created updated_centroids_df from records of updated_centroids')
# update centroids in reference data frame (the one that contains the points too)
for row_num in range(updated_centroids_df.shape[0]):
centroids_and_data_df.iloc[row_num, :] = updated_centroids_df.iloc[row_num, :]
#print('updated centroids in reference data frame centroids_and_data_df')
# update distance matrix using new centroids for distance calculations
distance_from_centroids = self.get_distance_matrix(centroids_and_data_df)
#print('updated distance matrix using new centroids')
# calculate distance between pairs of previous/new centroids to see if we've satisfied the threshold
centroids_diff = self.get_centroids_diff(previous_centroids, updated_centroids_df)
#print('calculated centroids diff')
# BUG: fix the issue where the first centroid diff is always zero
if centroids_diff[0] == 0:
#iteration_count = iteration_count - 1 # remove this line?
#print('decremented iteration_count to: ' + str(iteration_count) + ', because centroids_diff[0] == 0')
#print('centroids_diff[0] == 0 !!!')
iteration_count = iteration_count + 1
continue # workaround for now
else:
print('\nclustering iteration: ' + str(iteration_count + 1))
print('centroids_diff: ' + str(centroids_diff))
# update previous centroids dataframe to updated centroids dataframe
previous_centroids = updated_centroids_df
#print('updated previous centroids using updated_centroids_df')
iteration_count = iteration_count + 1
print('cluster returning: %s, %s, %s' % (str(len(cluster_assignments)), \
str(updated_centroids_df.shape), str(iteration_count)))
# return a tuple containing the final list of cluster assignments and the final centroids
return (cluster_assignments, updated_centroids_df, iteration_count - 1)
'''
generate initial cluster centroids with random values in min/max range for each column
INPUT:
- data_frame: data to generate centroids for
- k: k param value, i.e. number of centroids to generate
OUTPUT:
- list of centroid points with same dimensionality as regular data points
'''
def generate_initial_centroids(self, data_frame, k):
'''
# RANDOM GENERATION APPROACH
centroids = []
# get min/max values for each column (the bounds of the values for each column)
column_bounds = self.utilities_impl.get_column_bounds(data_frame)
num_cols = len(column_bounds)
for centroid_index in range(k):
centroid = []
for col_index in range(num_cols):
min_max_bounds = column_bounds[col_index]
# randomly generate value in min/max range for each attribute
centroid.append(random.uniform(min_max_bounds[0], min_max_bounds[1]))
centroids.append(centroid)
# return list of centroid points
return centroids
'''
# RANDOM POINTS APPROACH
indexes = random.sample(range(data_frame.shape[0]), k)
# BUG: change this so it doesn't throw the pandas error in the log
return data_frame.reindex(indexes)
# return boolean indicating whether the centroid diff threshold has been reached
def threshold_reached(self, centroids_diff):
np_diffs = np.array(centroids_diff)
# workaround for bug causing centroids_diff to be all zeros
if np_diffs is None or np_diffs[0] == 0:
return False
# return boolean indicating whether any centroid diffs are greater than threshold
return not list(np_diffs[np_diffs > self.CONVERGENCE_THRESHOLD])
'''
get average centroid from all points assigned to given cluster
INPUT:
- cluster_points: dataframe consisting of all points assigned to cluster
OUTPUT:
- centroid where each column is average value of respective column in cluster_points
'''
def get_avg_centroid(self, cluster_points):
avg_col_vals = []
# for each column in dataframe representing all points assigned to cluster
for column_label, _ in cluster_points.items():
column_vals = cluster_points.loc[:, column_label].values
column_vals = [float(val) for val in column_vals]
# calculate average column value and append to list
avg_col_vals.append(stats.mean(column_vals))
if self.DEBUG and self.VERBOSE:
print('column_label: ' + str(column_label))
print('len(column_vals): ' + str(len(column_vals)))
print('column_vals: ')
print(column_vals)
print('avg_column_vals: ' + str(stats.mean(column_vals)))
# return average centroid as list of average values for each column
return avg_col_vals
'''
get diff between centroids from iteration n and iteration n+1
'''
def get_centroids_diff(self, previous_centroids, updated_centroids_df):
if self.DEBUG and self.VERBOSE:
print('previous_centroids:')
print(previous_centroids)
print('updated_centroids_df:')
print(updated_centroids_df)
centroid_diffs = []
# for each centroid (instance) in the previous centroids dataframe
for row_num in range(previous_centroids.shape[0]):
prev_row = previous_centroids.iloc[row_num, :]
updated_row = updated_centroids_df.iloc[row_num, :]
# calculate euclidean distance between previous and updated centroid instance
diff_dist = self.distance_functions_impl.get_euclidean_distance(prev_row, updated_row)
centroid_diffs.append(diff_dist)
# return list containing distances between each corresponding pair of centroids
return centroid_diffs
'''
evaluate clustering - show counts
'''
def evaluate_clustering(self, data, clustering_assignments, k):
print('\nCLUSTERING EVALUATION:')
np_cluster_assignments = np.array(cluster_assignments)
for centroid_idx in range(k):
print('centroid_idx: ' + str(centroid_idx))
freqs = {}
idxs_for_cluster_val = np.where(np_cluster_assignments == centroid_idx)[0]
for idx in idxs_for_cluster_val:
if idx in data.index:
actual_class = str(data.loc[idx, 'CLASS'])
if actual_class in freqs:
freqs[actual_class] = freqs[actual_class] + 1
else:
freqs[actual_class] = 1
else:
print('ERROR: ' + str(idx) + ' not in data.index!')
print('freqs: ' + str(freqs))
'''
get distance matrix using pdist and squareform methods from scipy.spatial.distance
INPUT:
- data_frame: data frame we're working with
OUTPUT:
- distance matrix containing distances between every pair of points in data frame
'''
def get_distance_matrix(self, data_frame):
# get distance matrix (upper triangle) using distance metric
distances = pdist(data_frame.values, metric='euclidean')
# fill in lower triangle maintaining symmetry
dist_matrix = squareform(distances)
# return full distance matrix
return dist_matrix
'''
method for getting k cluster centroids using clustering output from cluster() method above
the input 'centroids_data' is a dataframe that contains all the attribute value for the centroids
this method is responsible for appending the corresponding class values to each centroid instance
INPUT:
- cluster_assignments - list of clustering assignments
- centroids_data - centroids rows without class values
OUTPUT:
- data frame with k rows, representing k centroids
'''
def get_cluster_centroids(self, cluster_assignments, centroids_data, dataframe):
if self.DEBUG and self.VERBOSE:
print('get_cluster_centroids: unique cluster assignments: ' + str(set(cluster_assignments)))
print('get_cluster_centroids: centroids_data.shape - BEFORE: ' + str(centroids_data.shape))
print('get_cluster_centroids: dataframe.shape: ' + str(dataframe.shape))
print('get_cluster_centroids: dataframe:')
print(dataframe)
# convert list of cluster assignments to np array to utilize np methods
np_cluster_assignments = np.array(cluster_assignments)
centroid_class_vals = []
#print('len set cluster assignments: ' + str(len(set(cluster_assignments))))
# for each unique cluster value (centroid index) that data points were assigned to
for unique_cluster_val in set(cluster_assignments):
# get dataframe row indexes that were assigned to the cluster
val_idxs = np.where(np_cluster_assignments == unique_cluster_val)[0]
#print('get_cluster_centroids: val_idxs assigned to unique_cluster_val: ' + str(unique_cluster_val) + ' --> ' + str(val_idxs))
idx_class_vals = self.get_idx_class_vals(dataframe, val_idxs)
#print('get_cluster_centroids: idx_class_vals for unique_cluster_val: ' + str(unique_cluster_val) + ' --> ' + str(idx_class_vals))
highest_freq_class = self.utilities_impl.get_mode(idx_class_vals)
print('highest_freq_class: ' + str(highest_freq_class))
# append highest frequency class to list of centroid class values
centroid_class_vals.append(highest_freq_class)
#print('len centroid_class_vals: ' + str(len(centroid_class_vals)))
#print('get_cluster_centroids: centroid_class_vals: ' + str(centroid_class_vals))
# these values will not match if there were clusters that had no points assigned to them
if len(centroid_class_vals) != centroids_data.shape[0]:
# get list of all possible class values for given dataframe
poss_class_vals = list(set(dataframe.loc[:, 'CLASS'].values))
# randomly assign class values to missing clusters to make dimensions match
# BUG: this shouldn't be necessary, handle missing clusters in a better way
centroid_class_vals = self.handle_cluster_count_mismatch(\
centroid_class_vals, centroids_data.shape[0], poss_class_vals)
# append generated class values column to centroids dataframe (assigning class to each centroid)
centroids_data['CLASS'] = centroid_class_vals
if self.DEBUG and self.VERBOSE:
print('get_cluster_centroids: centroids_data.shape - AFTER: ' + str(centroids_data.shape))
print('get_cluster_centroids: centroids_data - AFTER:')
print(centroids_data)
# return complete centroids dataframe (now containing the corresponding class values for each centroid)
return centroids_data
# get class values from dataframe for all indexes in idxs arg
def get_idx_class_vals(self, dataframe, idxs):
class_vals = []
for idx in idxs:
row_data = dataframe.iloc[idx, :]
class_vals.append(row_data['CLASS'])
# return list of class values for idxs
return class_vals
# this shouldn't be necessary, but for now workaround by returning random class value
def handle_cluster_count_mismatch(self, centroid_class_vals, expected_centroid_count, poss_class_vals):
while len(centroid_class_vals) < expected_centroid_count:
# BUG: this shouldn't be necessary, for now return random class value for missing clusters
centroid_class_vals.append(random.choice(poss_class_vals))
assert len(centroid_class_vals) == expected_centroid_count
# return list of centroid class values with expected length
return centroid_class_vals
'''
do full knn run through using k means clustering output as reduced data set for knn
NOTE: this always sets k equal to the number of possible class values for the given data set
INPUT:
- train_data: training data that will be clustered
- test_data: test data
- dataframe: full dataframe
- data_name: name of data set we're using
- k: value for k parameter
OUTPUT:
- returns a dictionary where key is the instance index and value is a tuple: (prediction, actual)
'''
def cluster_do_knn(self, train_data, test_data, dataframe, data_name, k):
# get number of possible class values for given data frame
num_poss_class_vals = len(set(dataframe.loc[:, 'CLASS'].values))
# k means cluster the training data to get list of final cluster assignments and resulting centroids
cluster_assignments, centroids_data, iteration_count = self.cluster(data_name, train_data, k=num_poss_class_vals)
'''
would be cool if we did some cluster evaluation but this method isn't working for some reason...
k_means_clustering_impl.evaluate_clustering(training_set, cluster_assignments, k=5)
'''
# get resulting centroids using highest frequency class value for each set of points in clusters
centroids_data = self.get_cluster_centroids(cluster_assignments, centroids_data, train_data)
# return a dictionary where key is the instance index and value is a tuple: (prediction, actual)
print('K MEANS CLUSTERING CONVERGED. iterations: ' + str(iteration_count))
return self.do_knn(centroids_data, test_data, dataframe, k)
'''
get cluster centroids for use in RBF network setup
INPUT:
- train_data: training data that will be clustered
- dataframe: full dataframe
- data_name: name of data set we're using
- k: value of k parameter in base knn
OUTPUT:
- return cluster centroids for given training data and k value
'''
def get_centroids_for_rbf_network(self, train_data, dataframe, data_name, k):
# get number of possible class values for given data frame
num_poss_class_vals = len(set(dataframe.loc[:, 'CLASS'].values))
# k means cluster the training data to get list of final cluster assignments and resulting centroids
cluster_assignments, centroids_data, iteration_count = self.cluster(data_name, train_data, k=num_poss_class_vals)
# get resulting centroids using highest frequency class value for each set of points in clusters
centroids_data = self.get_cluster_centroids(cluster_assignments, centroids_data, train_data)
# return a dictionary where key is the instance index and value is a tuple: (prediction, actual)
print('K MEANS CLUSTERING CONVERGED. iterations: ' + str(iteration_count))
return centroids_data
accuracy_vals = self.filter_vals(accuracy_vals)
mean_squared_error_vals = self.filter_vals(mean_squared_error_vals)
# calculate average values
avg_accuracy = sum(accuracy_vals) / test_set_count
avg_mean_squared_error = sum(mean_squared_error_vals) / test_set_count
# return tuple with average values for zero_one_loss and mean_squared_error
return (avg_accuracy, avg_mean_squared_error)
def filter_vals(self, vals):
filtered_vals = []
for val in vals:
if val is not None:
filtered_vals.append(val)
return filtered_vals
# EXECUTE SCRIPT
if __name__ == "__main__":
print('running results...')
results_impl = Results()
data_api_impl = DataApi('../../data/')
wine_data = data_api_impl.get_raw_data_frame('wine')
def __init__(self):
self.DEBUG = False
# construct DataApi instance with path prefix to data directory (relative from here)
self.data_api_impl = DataApi('../../data/')
def __init__(self):
KNN.__init__(self)
self.DEBUG = True
self.data_api_impl = DataApi('../../data/')
#Then scales to a 0-1 value (For comparison)
for i, valuetype in enumerate(uniquevals):
uniquevals[i] = i / (len(uniquevals) - 1)
normalized = attribute.replace(attribute.unique(), uniquevals)
return normalized
# EXECUTE SCRIPT
if __name__ == '__main__':
print('running preprocessor...')
preprocessor_impl = Preprocessor()
data_api_impl = DataApi('../../data/')
'''
raw_abalone_data = data_api_impl.get_raw_data_frame('abalone')
print('raw_abalone_data:')
print(raw_abalone_data)
prep_abalone_data = preprocessor_impl.preprocess_raw_data_frame(raw_abalone_data, 'abalone')
print('prep_abalone_data:')
print(prep_abalone_data)
'''
'''
raw_car_data = data_api_impl.get_raw_data_frame('car')
print('raw_car_data:')
print(raw_car_data)
prep_car_data = preprocessor_impl.preprocess_raw_data_frame(raw_car_data, 'car')
print('prep_car_data:')
print(prep_car_data)
