def ls_sklearn_sgd(x, y):
# Parameter estimation by sklearn SGD
sgd = SGDRegressor(fit_intercept=True)
sgd.fit(x.reshape((N, 1)), y)
beta_0_sk = sgd.intercept_
beta_1_sk = sgd.coef_[0]
return beta_0_sk, beta_1_sk
def initialize_regressors(sgd_grid):
l = SGDRegressor()
regressions_to_fit = []
for params in ParameterGrid(sgd_grid):
l.set_params(**params)
regressions_to_fit.append(clone(l))
return regressions_to_fit
def sgd(X, y, weight, X_test=False):
from sklearn.linear_model import SGDRegressor
from sklearn import cross_validation
from sklearn.metrics import confusion_matrix
from sklearn.preprocessing import StandardScaler
#X_train, X_test, y_train, y_test, weight_train, weight_test = cross_validation.train_test_split(
# X, y, weight, test_size=0.2, random_state=0)
clf = SGDRegressor(loss="huber", n_iter=100, penalty="l1")
#clf = LogisticRegression( max_iter=100)
X_train = X
y_train = y
scaler = StandardScaler(with_mean=False)
scaler.fit(X_train) # Don't cheat - fit only on training data
X_train = scaler.transform(X_train)
X_test = scaler.transform(X_test) # apply same transformation to test data
clf.fit(X_train, y_train, sample_weight=weight)
print(clf.score(X_train,y_train,weight))
y_pred = clf.predict(X_test)
from sklearn.externals import joblib
import scipy.io as sio
joblib.dump(clf, 'models/sgd_.pkl')
sio.savemat('predict_y_forward.mat', {'y':y_pred})
def linear_regression(features, values):
"""
Perform linear regression given a data set with an arbitrary number of features.
"""
model = SGDRegressor()
fit = model.fit(features, values)
return intercept, params
def predict_age():
mask = ~np.isnan(train["Age"])
age_train = train[mask]
age_test = train[~mask]
features = []
features.append(embarked_enc.transform(age_train["Embarked"]))
features.append(sex_enc.transform(age_train["Sex"]))
features.append(title_enc.transform(age_train["Title"]))
features.append(pclass_enc.transform(age_train["Pclass"]))
age_clf = SGDRegressor()
X = np.hstack(features)
y = np.array(train["Age"][mask]).T
age_clf.fit(X, y)
features = []
features.append(embarked_enc.transform(age_test["Embarked"]))
features.append(sex_enc.transform(age_test["Sex"]))
features.append(title_enc.transform(age_test["Title"]))
features.append(pclass_enc.transform(age_test["Pclass"]))
ages = age_clf.predict(np.hstack(features))
j = 0
for i in range(len(train)):
if ~mask[i]:
train.loc[i, "Age"] = ages[j]
j += 1
def build_sgd_regressor(X_test, X_train_full, y_train_full):
#print "Building SGD regressor..."
rf = SGDRegressor(loss="modified_huber", penalty="elasticnet", n_iter=20000, alpha=0.1, epsilon=0.01)
probas_rf = rf.fit(X_train_full, y_train_full).predict(X_test)
return probas_rf
def __init__(self, name, actions):
self.debug = True
self.actions = actions
self.initial_data = np.array([[3, 3.80, 0.40, 1.50], [3.0783744000000004, 1.7999999999999998, 0.04, 2.5],
[3.6603792000000004, 5.8500000000000005, 0.08, -1.5], [2.8383936000000003, 5.8500000000000005, 0.04, -3.0],
[4.5679104000000015, 5.8500000000000005, 0.04, -2.0], [2.885976, 4.05, 0.04, 1.0]])
self.initial_labels = np.array([[-1.0], [-0.2], [-0.5], [-.25], [0.0], [-2.1]])
self.model_name = str(name)
self.observation_samples = np.array([self.sample_states() for obv in range(10000)])
self.featurizer = sklearn.pipeline.FeatureUnion([("rbf1", RBFSampler(gamma=5.0, n_components=100)),
("rbf2", RBFSampler(gamma=2.0, n_components=100)),
("rbf3", RBFSampler(gamma=1.0, n_components=100)),
("rbf4", RBFSampler(gamma=0.5, n_components=100))])
self.featurizer.fit(self.observation_samples)
self.feature_scaler = StandardScaler()
self.feature_scaler.fit(self.observation_samples)
if self.model_name == 'svr':
self.model = SVR(kernel='rbf')
#self.model.fit(self.initial_data, self.initial_labels)
elif self.model_name == 'extra_trees':
self.model = ExtraTreesRegressor().fit(self.initial_data, self.initial_labels)
elif self.model_name == 'sgd':
self.models = {}
for a in range(len(self.actions)):
model = SGDRegressor(learning_rate="constant")
model.partial_fit([ self.featurize_state([3.6603792000000004, 2.8500000000000005, 0.08]) ], [0])
self.models[a] = model
#self.model = SGDRegressor(penalty='none')
#self.model.fit(self.feature_scaler.transform(self.initial_data), self.initial_labels)
else:
self.model = None
def predict(self, df):
# get time frame
time_frame = settings.time_frame
# copy of data
df_copy = df.copy()
from sklearn.linear_model import SGDRegressor
from sklearn.metrics import mean_absolute_error, mean_squared_error
# partition data
X_train, y_train, X_val, y_val, X_test, y_test = self.partition(df_copy)
# normalize features
X_train_std, X_val_std, X_test_std = self.feature_scale(X_train, X_val, X_test)
# instance of Linear Regression classifier
lr = SGDRegressor()
# fit model
lr.fit(X_train_std, y_train)
# predictions on validation set
predictions = lr.predict(X_val_std)
# R^2 score
score = lr.score(X_val_std, y_val)
# error
test_error = (mean_squared_error(y_val, predictions)**.5)
print test_error
def linear_regression_GD(features, values):
means, std_devs, features = normalized_features(features)
model = SGDRegressor(eta0=0.001)
results = model.fit(features, values)
intercept = results.intercept_
params = results.coef_
return intercept, params
class EdenRegressor(BaseEstimator, RegressorMixin):
"""Build a regressor for graphs."""
def __init__(self, r=3, d=8, nbits=16, discrete=True,
normalization=True, inner_normalization=True,
penalty='elasticnet', loss='squared_loss'):
"""construct."""
self.set_params(r, d, nbits, discrete,
normalization, inner_normalization,
penalty, loss)
def set_params(self, r=3, d=8, nbits=16, discrete=True,
normalization=True, inner_normalization=True,
penalty='elasticnet', loss='squared_loss'):
"""setter."""
self.r = r
self.d = d
self.nbits = nbits
self.normalization = normalization
self.inner_normalization = inner_normalization
self.discrete = discrete
self.model = SGDRegressor(
loss=loss, penalty=penalty,
average=True, shuffle=True,
max_iter=5, tol=None)
self.vectorizer = Vectorizer(
r=self.r, d=self.d,
normalization=self.normalization,
inner_normalization=self.inner_normalization,
discrete=self.discrete,
nbits=self.nbits)
return self
def transform(self, graphs):
"""transform."""
x = self.vectorizer.transform(graphs)
return x
@timeit
def kernel_matrix(self, graphs):
"""kernel_matrix."""
x = self.transform(graphs)
return metrics.pairwise.pairwise_kernels(x, metric='linear')
def fit(self, graphs, targets, randomize=True):
"""fit."""
x = self.transform(graphs)
self.model = self.model.fit(x, targets)
return self
def predict(self, graphs):
"""predict."""
x = self.transform(graphs)
preds = self.model.predict(x)
return preds
def decision_function(self, graphs):
"""decision_function."""
return self.predict(graphs)
def linear_regression(features, values):
sgd = SGDRegressor()
results = sgd.fit(values, features)
intercept = sgd.intercept_
params = results.get_params()
return intercept, params
def __init__(self, env, feature_transformer, learning_rate):
self.env = env
self.models = []
self.feature_transformer = feature_transformer
for i in range(env.action_space.n):
model = SGDRegressor(learning_rate=learning_rate)
model.partial_fit(feature_transformer.transform( [env.reset()] ), [0])
self.models.append(model)
def gradiantDescent(trainData,testData,trainOuts,testOuts): clf = SGDRegressor(loss="squared_loss") print(clf.fit(trainData,trainOuts)) print(clf.coef_) predictions = clf.predict(testData) print(predictions) misses,error = sup.crunchTestResults(predictions,testOuts,.5) print(1-error)
def linear_regression(features, values):
"""
Perform linear regression given a data set with an arbitrary number of features.
"""
clf = SGDRegressor(n_iter = 20)
clf.fit(features, values)
intercept,params = clf.intercept_,clf.coef_
return intercept, params
def linear_regression(features, values):
model = SGDRegressor(n_iter=1000)
results = model.fit(features, values)
intercept = results.intercept_
params = results.coef_
return intercept, params
def slim_train(A, l1_reg=0.001, l2_reg=0.0001):
"""
Computes W matrix of SLIM
This link is useful to understand the parameters used:
http://web.stanford.edu/~hastie/glmnet_matlab/intro.html
Basically, we are using this:
Sum( yi - B0 - xTB) + ...
As:
Sum( aj - 0 - ATwj) + ...
Remember that we are wanting to learn wj. If you don't undestand this
mathematical notation, I suggest you to read section III of:
http://glaros.dtc.umn.edu/gkhome/slim/overview
"""
alpha = l1_reg + l2_reg
l1_ratio = l1_reg / alpha
model = SGDRegressor(
penalty='elasticnet',
fit_intercept=False,
alpha=alpha,
l1_ratio=l1_ratio,
)
# TODO: get dimensions in the right way
m, n = A.shape
# Fit each column of W separately
W = lil_matrix((n, n))
for j in range(n):
if j % 50 == 0:
print('-> %2.2f%%' % ((j / float(n)) * 100))
aj = A[:, j].copy()
# We need to remove the column j before training
A[:, j] = 0
model.fit(A, aj.toarray().ravel())
# We need to reinstate the matrix
A[:, j] = aj
w = model.coef_
# Removing negative values because it makes no sense in our approach
w[w < 0] = 0
for el in w.nonzero()[0]:
W[(el, j)] = w[el]
return W
def SGD_Regression(kf,data,label,k):
val=0
for train, test in kf:
X_train, X_test, y_train, y_test = data[train,:], data[test,:], label[train], label[test]
log = SGDRegressor(loss='squared_loss', penalty='l2', alpha=0.0001, l1_ratio=0.15,n_iter=5)
logit = log.fit(X_train,y_train)
y_pred = logit.predict(X_test)
val += metrics.mean_squared_error(y_test, y_pred)
return val/3
# print "SGD_Regression, Mean Squared Error ", "{0:.4f}".format(val/3)
def linear_regression(features, values):
"""
Perform linear regression given a data set with an arbitrary number of features.
"""
regressor = SGDRegressor()
result = regressor.fit(features, values)
intercept = result.intercept_
params = result.coef_
return intercept, params
def perform_sgd_regression(features, values):
clf = SGDRegressor(n_iter=20)
clf = clf.fit(features, values)
intercept = clf.intercept_
params = clf.coef_
print "intercept:"
print intercept
print "params:"
for i in range(len(params)):
print "%s: %f" %(features.columns.values[i], params[i])
def linear_regression(features, values):
"""
Perform linear regression given a data set with an arbitrary number of features.
"""
X = features
y = values
clf = SGDRegressor(n_iter=20)
results = clf.fit(X, y)
intercept = results.intercept_[0]
params = results.coef_
return intercept, params
def linear_regression(features, values):
"""
Perform linear regression given a data set with an arbitrary number of features.
"""
model = SGDRegressor(n_iter=100)
model.fit(features,values)
print 'SCORE: ',model.score(features,values)
intercept = model.intercept_
params = model.coef_
return intercept, params
def __init__(self):
# We create a separate model for each action in the environment's
# action space. Alternatively we could somehow encode the action
# into the features, but this way it's easier to code up.
self.models = []
for _ in range(env.action_space.n):
model = SGDRegressor(learning_rate="constant")
# We need to call partial_fit once to initialize the model
# or we get a NotFittedError when trying to make a prediction
# This is quite hacky.
model.partial_fit([self.featurize_state(env.reset())], [0])
self.models.append(model)
def sgd(pd, pl, qd, ql):
params = {'loss':['squared_loss', 'huber', 'epsilon_insensitive',
'squared_epsilon_insensitive'],
'alpha':expon(scale=1),
'epsilon':expon(scale=1),
'l1_ratio':uniform(),
'penalty':[ 'l2', 'l1', 'elasticnet']}
clf = SGDRegressor()
#clf = RandomizedSearchCV(clf, params, n_jobs=2, n_iter=10, verbose=10)
print("Training Linear SVM Randomly")
clf.fit(pd, pl)
print("Score: " + str(clf.score(qd, ql)))
return clf
def linear_regression(features, values):
"""
Perform linear regression given a data set with an arbitrary number of features.
"""
###########################
### YOUR CODE GOES HERE ###
###########################
clf = SGDRegressor()
model = clf.fit(features,values)
params = model.coef_
intercept = model.intercept_
return intercept, params
def linear_regression(features, values):
"""
Perform linear regression given a data set with an arbitrary number of features.
"""
###########################
### YOUR CODE GOES HERE ###
###########################
classifier = SGDRegressor(n_iter = 20)
classifier.fit(features, values)
intercept = classifier.intercept_
params = classifier.coef_
return intercept, params
def predictScores(trainFeatures,trainTargets,testFeatures,testItemIds,isRegression = False):
logging.info("Feature preparation done, fitting model...")
predicted_scores = []
if isRegression:
clf = SGDRegressor( penalty="l2",
alpha=1e-4)
print("trainFeatures rows::"+str(trainFeatures.shape[0]))
print("trainTargets rows::"+str(len(trainTargets)))
clf.fit(trainFeatures,trainTargets)
logging.info("Predicting...")
predicted_scores = clf.predict(testFeatures)
else:
clf = SGDClassifier( loss="log",
penalty="l2",
alpha=1e-4,
class_weight="auto")
print("trainFeatures rows::"+str(trainFeatures.shape[0]))
print("trainTargets rows::"+str(len(trainTargets)))
clf.fit(trainFeatures,trainTargets)
logging.info("Predicting...")
predicted_scores = clf.predict_proba(testFeatures).T[1]
logging.info("Write results...")
output_file = "avito_starter_solution.csv"
logging.info("Writing submission to %s" % output_file)
f = open(os.path.join(dataFolder,output_file), "w")
f.write("id\n")
for pred_score, item_id in sorted(zip(predicted_scores, testItemIds), reverse = True):
f.write("%d\n" % (item_id))
f.close()
def main(train_file, model_file):
#train_x, train_y = load_sparse_trainingData_memory(train_file, 2 * get_len_vector())
train_x, train_y = load_long_training_data_memory()
#train_x, train_y = load_trainingData(train_file)
logging('len of y: %d' % train_y.shape)
logging(train_x.shape)
#LR = LinearRegression(copy_X = False, normalize = True)
LR = SGDRegressor(verbose=1)
logging("training model...")
starttime = datetime.now()
LR.fit(train_x, train_y)
logging("training model, eplased time:%s" % str(datetime.now() - starttime))
logging("saving model")
joblib.dump(LR, model_file)
def linear_regression(features, values):
"""
Perform linear regression given a data set with an arbitrary number of features.
"""
y = values
X = features
clf = SGDRegressor()
clf.fit(X, y)
intercept = clf.intercept_
params = clf.coef_
return intercept, params
def linear_regression_gradient_descent(features, values):
"""
Perform linear regression given a data set with an arbitrary number of features.
"""
###########################
### YOUR CODE GOES HERE ###
###########################
clf = SGDRegressor(n_iter=15)
results = clf.fit(features, values)
intercept= results.intercept_
params = results.coef_
return intercept, params
def SGDRegressor_pred(X_train, X_test, y_train_normalized, y_train_mean, y_test):
# The learning rate:
# ---constant: eta = eta0 [assign to the initial one, eta0]
# ---optimal: eta = 1.0/(t+t0)
# ---invscaling: eta = eta0 / pow(t, power_t) [default]
clf = SGDRegressor(alpha=0.0001, eta0=0.001, n_iter=150, fit_intercept=False, shuffle=True, verbose=0)
clf = clf.fit(X_train, y_train_normalized)
# Conveting to back, (could be used sklearn standardization function for both decoding and encoding)
predictions_train = clf.predict(X_train) + y_train_mean
predictions = clf.predict(X_test) + y_train_mean
score_test = clf.score(X_test, y_test)
return predictions, predictions_train, score_test
plt.show()
'''
Spot check Algorithms
'''
Rmodels = []
Rmodels.append(('LR',
LinearRegression(copy_X=True,
fit_intercept=True,
n_jobs=None,
normalize=True)))
Rmodels.append(('LGBM', LGBMRegressor()))
Rmodels.append(('XGB', XGBRegressor()))
#Rmodels.append(('CatBoost', CatBoostRegressor()))
Rmodels.append(('SGD', SGDRegressor()))
Rmodels.append(('KernelRidge', KernelRidge()))
Rmodels.append(('ElasticNet', ElasticNet()))
Rmodels.append(('BayesianRidge', BayesianRidge()))
Rmodels.append(('GradientBoosting', GradientBoostingRegressor()))
Rmodels.append(('SVR', SVR(gamma='auto'))) # kernel = linear, svr
Rmodels.append(('NN', MLPRegressor(solver='lbfgs'))) #neural network
Rmodels.append(('KNN', KNeighborsRegressor())) # kneighbor
Rmodels.append(('RF', RandomForestRegressor(
n_estimators=10))) # Ensemble method - collection of many decision trees
# decision tree
# Gradient boosting
Cmodels = []
Cmodels.append(('Logistic', LogisticRegression()))
Cmodels.append(('SVM', SVC(gamma='auto')))
def run_sgd_reg():
sgd_reg = SGDRegressor(max_iter=1000, tol=1e-3, penalty=None, eta0=0.1)
return sgd_reg.fit(X_train_scaled, y_train)
# Model configuration
base = make_pipeline(
StackingEstimator(estimator=LassoLarsCV(normalize=True)),
StackingEstimator(estimator=LinearSVR(C=0.01, dual=True, epsilon=0.001, loss="epsilon_insensitive", tol=0.1)),
MaxAbsScaler(),
StackingEstimator(estimator=RidgeCV()),
Normalizer(norm="l2"),
StackingEstimator(estimator=LinearSVR(C=0.5, dual=False, epsilon=0.1, loss="squared_epsilon_insensitive", tol=0.1)),
StackingEstimator(estimator=ExtraTreesRegressor(bootstrap=False, max_features=0.4, min_samples_leaf=2, min_samples_split=4, n_estimators=100)),
MinMaxScaler(),
StackingEstimator(estimator=RidgeCV()),
StackingEstimator(estimator=LinearSVR(C=5.0, dual=True, epsilon=0.1, loss="epsilon_insensitive", tol=0.0001)),
StackingEstimator(estimator=RidgeCV()),
StackingEstimator(estimator=SGDRegressor()),
RobustScaler(),
StackingEstimator(estimator=LinearSVR(C=15.0, dual=True, epsilon=0.01, loss="epsilon_insensitive", tol=0.1)),
StackingEstimator(estimator=ElasticNetCV(l1_ratio=0.75, tol=0.001)),
StackingEstimator(estimator=XGBRegressor(learning_rate=0.1, max_depth=1, min_child_weight=6, n_estimators=100, nthread=1, objective="reg:squarederror", subsample=0.6500000000000001)),
MinMaxScaler(),
StackingEstimator(estimator=ExtraTreesRegressor(bootstrap=False, max_features=0.2, min_samples_leaf=2, min_samples_split=4, n_estimators=100)),
StackingEstimator(estimator=LinearSVR(C=5.0, dual=True, epsilon=0.1, loss="epsilon_insensitive", tol=0.0001)),
MaxAbsScaler(),
RandomForestRegressor(bootstrap=False, max_features=0.05, min_samples_leaf=1, min_samples_split=4, n_estimators=100)
)
parameters = {'base_estimator': base,
'n_estimators': 100, #default = 50
'learning_rate': 0.3, #default = 1.0
'loss': 'linear',
# scikit-learn 中的 SGD 随机梯度下降法
# 只能解决线性模型
from time import time
from sklearn.linear_model import SGDRegressor
from GradientDescent.own_SGD import X_train_standard, X_test_standard, y_train, y_test
# 初始化
sgd_reg = SGDRegressor()
# 进行训练
start_time1 = time()
sgd_reg.fit(X_train_standard, y_train)
end_time1 = time()
sgd_reg.score(X_test_standard, y_test)
if __name__ == "__main__":
print(end_time1 - start_time1)
class SGDPolyCartPoleSolver:
def __init__(self, n_episodes=1000, max_env_steps=None, gamma=0.9, epsilon=1.0, epsilon_min=0.01,
epsilon_decay=0.005, alpha=0.0001, batch_size=32, monitor=False):
self.memory = deque(maxlen=100000)
self.env = gym.make('CartPole-v0')
if monitor: # whether or not to display video
self.env = gym.wrappers.Monitor(self.env, '../data/cartpole-1', force=True)
# hyper-parameter setting
self.gamma = gamma
self.epsilon = epsilon
self.epsilon_min = epsilon_min
self.epsilon_decay = epsilon_decay
self.alpha = alpha
self.n_episodes = n_episodes
self.batch_size = batch_size
self.feature_tuning = PolynomialFeatures(interaction_only=True)
if max_env_steps is not None:
self.env._max_episode_steps = max_env_steps
# Init model
self.model = SGDRegressor(
alpha=self.alpha,
learning_rate='optimal',
shuffle=False,
warm_start=True)
# Initialize feature tunning
self.feature_tuning.fit(np.reshape(np.hstack((self.env.reset(), 0)), [1, 5]))
# Initialize model
self.model.partial_fit(self.preprocess_state(self.env.reset(), 0), [0])
def remember(self, state, action, reward, next_state, done):
"""In this method, the (s, a, r, s') tuple is stored in the memory"""
self.memory.append((state, action, reward, next_state, done))
def choose_action(self, state, epsilon):
"""Chooses the next action according to the model trained and the policy"""
qsa = np.asarray([self.model.predict(self.preprocess_state(state, a))
for a in range(self.env.action_space.n)]).flatten()
return self.env.action_space.sample() if (np.random.random() <= epsilon) \
else np.argmax(qsa) # exploits the current knowledge if the random number > epsilon, otherwise explores
def get_epsilon(self, episode):
"""Returns an epsilon that decays over time until a minimum epsilon value is reached; in this case the minimum
value is returned"""
return max(self.epsilon_min, self.epsilon * math.exp(-self.epsilon_decay * episode))
def preprocess_state(self, state, action):
"""State and action are stacked horizontally and its features are combined as a polynomial to be passed as an
input of the approximator"""
# poly_state converts the horizontal stack into a combination of its parameters i.e.
# [1, s_1, s_2, s_3, s_4, a_1, s_1 s_2, s_1 s_3, ...]
poly_state = self.feature_tuning.transform(np.reshape(np.hstack((state, action)), [1, 5]))
return poly_state
def replay(self, batch_size):
"""Previously stored (s, a, r, s') tuples are replayed (that is, are added into the model). The size of the
tuples added is determined by the batch_size parameter"""
x_batch, y_batch = [], []
minibatch = random.sample(self.memory, min(len(self.memory), batch_size))
for state, action, reward, next_state, done in minibatch:
qsa_s_prime = np.asarray([self.model.predict(self.preprocess_state(next_state, a))
for a in range(self.env.action_space.n)])
qsa_s = reward if done \
else reward + self.gamma * np.max(qsa_s_prime)
x_batch.append(self.preprocess_state(state, action)[0])
y_batch.append(qsa_s)
self.model.partial_fit(np.array(x_batch), np.array(y_batch))
def run(self):
"""Main loop that controls the execution of the agent"""
scores100 = deque(maxlen=100)
scores = []
for e in range(self.n_episodes):
state = self.env.reset()
done = False
t = 0 # t counts the number of time-steps the pole has been kept up
while not done:
action = self.choose_action(state, self.get_epsilon(e))
next_state, reward, done, _ = self.env.step(action)
self.remember(state, action, reward, next_state, done)
self.replay(self.batch_size)
state = next_state
t += 1
scores100.append(t)
scores.append(t)
mean_score = np.mean(scores100)
if e % 100 == 0:
print('[Episode {}] - Mean survival time over last 100 episodes was {} ticks.'.format(e, mean_score))
# noinspection PyUnboundLocalVariable
print('[Episode {}] - Mean survival time over last 100 episodes was {} ticks.'.format(e, mean_score))
return scores
exported_pipeline = make_pipeline(
SelectPercentile(score_func=f_regression, percentile=89),
StackingEstimator(
estimator=KNeighborsRegressor(n_neighbors=48, p=1, weights="uniform")),
StackingEstimator(estimator=XGBRegressor(learning_rate=0.001,
max_depth=1,
min_child_weight=3,
n_estimators=50,
n_jobs=1,
objective="reg:squarederror",
subsample=0.9500000000000001,
verbosity=0)), MinMaxScaler(),
StackingEstimator(estimator=SGDRegressor(alpha=0.01,
eta0=0.01,
fit_intercept=False,
l1_ratio=0.0,
learning_rate="constant",
loss="huber",
penalty="elasticnet",
power_t=0.0)),
StackingEstimator(estimator=LinearSVR(
C=25.0, dual=True, epsilon=0.1, loss="epsilon_insensitive",
tol=0.0001)), FeatureAgglomeration(affinity="l2", linkage="average"),
SelectPercentile(score_func=f_regression, percentile=6),
StackingEstimator(estimator=ExtraTreesRegressor(bootstrap=False,
max_features=0.8,
min_samples_leaf=19,
min_samples_split=10,
n_estimators=400)),
ZeroCount(), FeatureAgglomeration(affinity="l2", linkage="complete"),
StackingEstimator(estimator=RidgeCV()), RidgeCV())
def run(config, train=True):
"""
Trains our pipeline according to the configuration provided.
"""
train_dir = config["train_dir"]
val_dir = config["val_dir"]
print("Reading in data...")
train_data = NucleiDataset(train_dir).load()
val_data = NucleiDataset(val_dir).load()
x_train = train_data.images_
y_train = train_data.masks_ # value in 0, 1, ..., n
y_train_bin = (y_train > 0).astype(y_train.dtype) # value in {0, 1}
x_val = val_data.images_
y_val = val_data.masks_
y_val_bin = (y_val > 0).astype(y_val.dtype)
print("Preprocessing data...")
preprocesser = Preprocesser()
x_train_pre = preprocesser.fit_transform(x_train)
x_val_pre = preprocesser.transform(x_val)
bilateral_d = 2
bilateral_sigma_color = 75
bilateral_sigma_space = 75
equalize_hist_clip_limit = 0.03
dialation_kernel = disk(radius=3)
dialation_iters = 1
print("Transforming data...")
print(x_train_pre.min())
print(x_train_pre.max())
print(x_val_pre.min())
print(x_val_pre.max())
transformer = BasisTransformer(
bilateral_d=bilateral_d,
bilateral_sigma_color=bilateral_sigma_color,
bilateral_sigma_space=bilateral_sigma_space,
equalize_hist_clip_limit=equalize_hist_clip_limit,
dialation_kernel=dialation_kernel,
dialation_iters=dialation_iters)
x_train_feat = transformer.fit_transform(x_train_pre)
x_val_feat = transformer.fit_transform(x_val_pre)
sgd_params = {
"regressor":
SGDRegressor(penalty='elasticnet', l1_ratio=0.11, max_iter=5,
tol=None),
"batch_size":
1000,
"num_iters":
25000,
}
pa_params = {
"regressor": PassiveAggressiveRegressor(C=.2, max_iter=5, tol=None),
"batch_size": 1000,
"num_iters": 25000,
}
sgd = MiniBatchRegressor(**sgd_params)
pa = MiniBatchRegressor(**pa_params)
print("Fitting linear models...")
sgd.fit(x_train_feat, y_train_bin)
pa.fit(x_train_feat, y_train_bin)
x_train_extended = extend_features(x_train_feat, sgd, pa)
x_val_extended = extend_features(x_val_feat, sgd, pa)
# Now we train UNet
numchannels = x_train_extended.shape[-1]
unet_config = {
"numchannels": numchannels,
"epochs": 50,
"callbacks": [],
"weights": none
}
unet = UNet(**unet_config)
if unet_config["weights"] is not None:
unet.load_weights(unet_config["weights"])
print("Fitting UNet...")
unet.fit(x_train_extended, y_train_bin, x_val_extended, y_val_bin)
# begin inference and print out test scores
x_train_pred = unet.predict(x_train_extended)
x_val_pred = unet.predict(x_val_extended)
segmenter_params = {"nms_min_distance": 3, "watershed_line": True}
segmenter = NucleiSegmenter(**segmenter_params)
print("Segmenting nuclei...")
train_components = segmenter.fit_transform(x_train_pred, x_train_pre)
val_components = segmenter.fit_transform(x_val_pred, x_val_pre)
def dict_method_reg():
dict_method = {}
# 1st part
"""4KNR"""
me4 = neighbors.KNeighborsRegressor(n_neighbors=5,
weights='uniform',
algorithm='auto',
leaf_size=30,
p=2,
metric='minkowski')
cv4 = 5
scoring4 = 'r2'
param_grid4 = [{
'n_neighbors': [3, 4, 5, 6, 7],
"weights": ['uniform', "distance"],
"leaf_size": [10, 20, 30]
}]
dict_method.update({"KNR-set": [me4, cv4, scoring4, param_grid4]})
"""1SVR"""
me1 = SVR(kernel='rbf',
gamma='auto',
degree=3,
tol=1e-3,
epsilon=0.1,
shrinking=False,
max_iter=2000)
cv1 = 5
scoring1 = 'r2'
param_grid1 = [{
'C': [10000, 100, 50, 10, 5, 2.5, 1, 0.5, 0.1, 0.01],
'kernel': ker
}]
dict_method.update({"SVR-set": [me1, cv1, scoring1, param_grid1]})
"""5kernelridge"""
me5 = kernel_ridge.KernelRidge(alpha=1,
gamma="scale",
degree=3,
coef0=1,
kernel_params=None)
cv5 = 5
scoring5 = 'r2'
param_grid5 = [{
'alpha': [100, 50, 10, 5, 2.5, 1, 0.5, 0.1, 0.01, 0.001, 1e-4, 1e-5],
'kernel':
ker
}]
dict_method.update({'KRR-set': [me5, cv5, scoring5, param_grid5]})
"""6GPR"""
me6 = gaussian_process.GaussianProcessRegressor(kernel=kernel,
alpha=1e-10,
optimizer='fmin_l_bfgs_b',
n_restarts_optimizer=0,
normalize_y=False,
copy_X_train=True,
random_state=0)
cv6 = 5
scoring6 = 'r2'
param_grid6 = [{'alpha': [1e-3, 1e-2], 'kernel': ker}]
dict_method.update({"GPR-set": [me6, cv6, scoring6, param_grid6]})
# 2nd part
"""6RFR"""
me7 = RandomForestRegressor(n_estimators=100,
max_depth=None,
min_samples_split=2,
min_samples_leaf=1,
min_weight_fraction_leaf=0.0,
max_leaf_nodes=None,
min_impurity_decrease=0.0,
min_impurity_split=None,
bootstrap=True,
oob_score=False,
random_state=None,
verbose=0,
warm_start=False)
cv7 = 5
scoring7 = 'r2'
param_grid7 = [{'max_depth': [3, 4, 5, 6], 'min_samples_split': [2, 3]}]
dict_method.update({"RFR-em": [me7, cv7, scoring7, param_grid7]})
"""7GBR"""
me8 = GradientBoostingRegressor(loss='ls',
learning_rate=0.1,
n_estimators=100,
subsample=1.0,
criterion='friedman_mse',
min_samples_split=2,
min_samples_leaf=1,
min_weight_fraction_leaf=0.,
max_depth=3,
min_impurity_decrease=0.,
min_impurity_split=None,
init=None,
random_state=None,
max_features=None,
alpha=0.9,
verbose=0,
max_leaf_nodes=None,
warm_start=False,
presort='auto')
cv8 = 5
scoring8 = 'r2'
param_grid8 = [{
'max_depth': [3, 4, 5, 6],
'min_samples_split': [2, 3],
'learning_rate': [0.1, 0.05]
}]
dict_method.update({'GBR-em': [me8, cv8, scoring8, param_grid8]})
"AdaBR"
dt = DecisionTreeRegressor(criterion="mse",
splitter="best",
max_features=None,
max_depth=5,
min_samples_split=4)
me9 = AdaBoostRegressor(dt,
n_estimators=200,
learning_rate=0.05,
loss='linear',
random_state=0)
cv9 = 5
scoring9 = 'explained_variance'
param_grid9 = [{'n_estimators': [100, 200], 'learning_rate': [0.1, 0.05]}]
dict_method.update({"AdaBR-em": [me9, cv9, scoring9, param_grid9]})
'''DTR'''
me10 = DecisionTreeRegressor(criterion="mse",
splitter="best",
max_depth=None,
min_samples_split=2,
min_samples_leaf=1,
min_weight_fraction_leaf=0.,
max_features=None,
random_state=0,
max_leaf_nodes=None,
min_impurity_decrease=0.,
min_impurity_split=None,
presort=False)
cv10 = 5
scoring10 = 'r2'
param_grid10 = [{
'max_depth': [2, 3, 4, 5, 6, 7, 8],
"min_samples_split": [2, 3, 4],
"min_samples_leaf": [1, 2]
}]
dict_method.update({'DTR-em': [me10, cv10, scoring10, param_grid10]})
'ElasticNet'
me11 = ElasticNet(alpha=1.0,
l1_ratio=0.7,
fit_intercept=True,
normalize=False,
precompute=False,
max_iter=1000,
copy_X=True,
tol=0.0001,
warm_start=False,
positive=False,
random_state=None)
cv11 = 5
scoring11 = 'r2'
param_grid11 = [{
'alpha': [0.0001, 0.001, 0.01, 0.1, 1, 10],
'l1_ratio': [0.3, 0.5, 0.8]
}]
dict_method.update(
{"ElasticNet-L1": [me11, cv11, scoring11, param_grid11]})
'Lasso'
me12 = Lasso(
alpha=1.0,
fit_intercept=True,
normalize=False,
precompute=False,
copy_X=True,
max_iter=1000,
tol=0.001,
warm_start=False,
positive=False,
random_state=None,
)
cv12 = 5
scoring12 = 'r2'
param_grid12 = [
{
'alpha': [
0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1, 10, 100,
1000
]
},
]
dict_method.update({"LASSO-set": [me12, cv12, scoring12, param_grid12]})
"""2BayesianRidge"""
me2 = BayesianRidge(alpha_1=1e-06,
alpha_2=1e-06,
compute_score=False,
copy_X=True,
fit_intercept=True,
lambda_1=1e-06,
lambda_2=1e-06,
n_iter=300,
normalize=False,
tol=0.01,
verbose=False)
cv2 = 5
scoring2 = 'r2'
param_grid2 = [{
'alpha_1': [1e-07, 1e-06, 1e-05],
'alpha_2': [1e-07, 1e-06, 1e-05]
}]
dict_method.update({'BRR-set': [me2, cv2, scoring2, param_grid2]})
"""3SGDRL2"""
me3 = SGDRegressor(alpha=0.0001,
average=False,
epsilon=0.1,
eta0=0.01,
fit_intercept=True,
l1_ratio=0.15,
learning_rate='invscaling',
loss='squared_loss',
max_iter=1000,
penalty='l2',
power_t=0.25,
random_state=0,
shuffle=True,
tol=0.01,
verbose=0,
warm_start=False)
cv3 = 5
scoring3 = 'r2'
param_grid3 = [{
'alpha': [100, 10, 1, 0.1, 0.01, 0.001, 0.0001, 1e-05],
'loss': ['squared_loss', "huber"],
"penalty": ["l1", "l2"]
}]
dict_method.update({'SGDR-set': [me3, cv3, scoring3, param_grid3]})
"""PassiveAggressiveRegressor"""
me14 = PassiveAggressiveRegressor(C=1.0,
fit_intercept=True,
max_iter=1000,
tol=0.001,
early_stopping=False,
validation_fraction=0.1,
n_iter_no_change=5,
shuffle=True,
verbose=0,
loss='epsilon_insensitive',
epsilon=0.1,
random_state=None,
warm_start=False,
average=False)
cv14 = 5
scoring14 = 'r2'
param_grid14 = [{
'C': [1.0e8, 1.0e6, 10000, 100, 50, 10, 5, 2.5, 1, 0.5, 0.1, 0.01]
}]
dict_method.update({'PAR-set': [me14, cv14, scoring14, param_grid14]})
return dict_method
blend_train = []
blend_test = []
train_time = timer(None)
kf = KFold(n_splits=folds, random_state=1001)
for i, (train_index, val_index) in enumerate(kf.split(X_train, y_train)):
start_time = timer(None)
Xtrain, Xval = X_train[train_index], X_train[val_index]
ytrain, yval = y_train[train_index], y_train[val_index]
model = SGDRegressor(penalty='l2',
loss='squared_epsilon_insensitive',
max_iter=200,
tol=0.00001,
epsilon=0.0001,
learning_rate='invscaling',
fit_intercept=False,
alpha=1e-10,
l1_ratio=0.09,
shuffle=True,
verbose=0,
random_state=1001)
model.fit(Xtrain, ytrain)
sgd_scores_val = model.predict(Xval)
sgd_RMSLE = np.sqrt(mean_squared_error(yval, sgd_scores_val))
print('\n Fold %02d SGD RMSLE: %.6f' % ((i + 1), sgd_RMSLE))
sgd_y_pred = model.predict(X_test)
model = Ridge(alpha=4.75,
solver='sag',
fit_intercept=False,
random_state=1001,
from nlpia.models import LinearRegressor
line = LinearRegressor()
line = line.fit(sms['topic4'], sms['vader'])
print('{:.4f}'.format(line.slope))
# 0.29
sms['line'] = line.predict(sms['topic4'])
##########################
from sklearn.linear_model import SGDRegressor
sgd = SGDRegressor(n_iter=20000)
sgd = sgd.fit(sms[['topic4']], sms['vader'])
print('{:.4f}'.format(sgd.coef_[0]))
# 0.2930
sms['sgd'] = sgd.predict(sms[['topic4']])
##########################
from nlpia.models import OneNeuronRegressor
nn = OneNeuronRegressor(alpha=100, n_iter=200)
nn = nn.fit(sms[['topic4']], sms['vader'])
print(nn.W[0, 1])
# 0.29386408
import json
import matplotlib.pyplot as plt
from sklearn.linear_model import SGDClassifier, SGDRegressor, PassiveAggressiveClassifier, PassiveAggressiveRegressor, Perceptron
from sklearn.preprocessing import MinMaxScaler, StandardScaler, MaxAbsScaler
penalty = ['l1', 'l2', 'elasticnet'][2]
sgd_classifiers = [
('SGD-h', SGDClassifier(loss='hinge', penalty=penalty)),
('SGD-l', SGDClassifier(loss='log', penalty=penalty)),
('SGD-mh', SGDClassifier(loss='modified_huber', penalty=penalty)),
('SGD-sh', SGDClassifier(loss='squared_hinge', penalty=penalty)),
('SGD-p', SGDClassifier(loss='perceptron', penalty=penalty))
]
sgd_regressors = [('SGD-sl', SGDRegressor(loss='squared_loss',
penalty=penalty)),
('SGD-h', SGDRegressor(loss='huber', penalty=penalty)),
('SGD-ei',
SGDRegressor(loss='epsilon_insensitive', penalty=penalty)),
('SGD-sei',
SGDRegressor(loss='squared_epsilon_insensitive',
penalty=penalty))]
selected_classifiers = [('SGD', SGDClassifier(loss='hinge', penalty='l1')),
('Perceptron-I', Perceptron(penalty='l1')),
('Perceptron-II', Perceptron(penalty='l2')),
('Perceptron-II', Perceptron(penalty='elasticnet')),
('PA-I', PassiveAggressiveClassifier(loss='hinge')),
('PA-II',
PassiveAggressiveClassifier(loss='squared_hinge'))]
def learning_schedule(t):
return t0 / (t + t1)
theta = np.random.randn(2, 1)
for epoch in range(n_epochs):
for i in range(m):
random_index = np.random.randint(m)
xi = X_b[random_index:random_index + 1]
yi = y[random_index:random_index + 1]
gradients = 2 * xi.T.dot(xi.dot(theta) - yi)
eta = learning_schedule(epoch * m + i)
theta = theta - eta * gradients
print("SGD: ", theta.flatten())
# =============================================================================
# SGD using sklearn's inbuilt SGDregressor class
# =============================================================================
from sklearn.linear_model import SGDRegressor
sgd_reg = SGDRegressor(max_iter=500,
penalty=None,
eta0=0.1,
early_stopping=True)
sgd_reg.fit(X, y.ravel())
print(sgd_reg.intercept_, sgd_reg.coef_)
def create(X, X_column_types, y, y_column_types, arm, **kwargs):
method = kwargs.get("method", None)
#method = kwargs.get("method", "Binary_operators")
#method = kwargs.get("method", "Binning")
#method = kwargs.pop("method", "Cluster")
categorical_cols = [c for c, t in zip(X.columns, X_column_types) if t in [DATATYPE_CATEGORY_INT, DATATYPE_CATEGORY_STRING]]
numerical_cols = [c for c, t in zip(X.columns, X_column_types) if t == DATATYPE_NUMBER]
categorical = X[categorical_cols]
numerical = X[numerical_cols]
# feature selection using Genetic Algorithm
if method == "fs_GA":
print("fs_GA")
enc = OneHotEncoder()
enc.fit(categorical)
Data_cat=pd.DataFrame(enc.transform(categorical).toarray())
X_data = pd.concat([numerical, Data_cat], axis=1)
if y_column_types[0] == DATATYPE_NUMBER:
y = y
else:
from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
le.fit(y)
y = le.transform(y)
X_train, X_test, y_train, y_test = train_test_split(X_data, y, train_size=0.8, random_state=42)
def get_fitness(individual):
if y_column_types[0] == DATATYPE_NUMBER:
rg = RandomForestRegressor(random_state=42)
else:
rg = RandomForestClassifier(random_state=42)
columns = [column for (column, binary_value) in zip(X_train.columns, individual) if binary_value]
training_set = X_train[columns]
test_set = X_test[columns]
rg.fit(training_set.values, y_train)
preds = rg.predict(test_set.values)
return 100 / np.sqrt(mean_squared_error(y_test, preds))
individual = [1] * 100
get_fitness(individual)
def get_population_fitness(population):
return sorted([(individual, get_fitness(individual)) for individual in population], key=lambda tup: tup[1], reverse=True)
def crossover(individual_a, individual_b):
crossing_point = random.randint(0, 99)
offspring_a = individual_a[0:crossing_point] + individual_b[crossing_point:100]
offspring_b = individual_b[0:crossing_point] + individual_a[crossing_point:100]
return offspring_a, offspring_b
def tournament(current_population):
index = sorted(random.sample(range(0, 20), 5))
tournament_members = [current_population[i] for i in index]
total_fitness = sum([individual[1] for individual in tournament_members])
probabilities = [individual[1] / total_fitness for individual in tournament_members]
index_a, index_b = np.random.choice(5, size=2, p=probabilities)
return crossover(tournament_members[index_a][0], tournament_members[index_b][0])
def mutation(individual):
mutation_point = random.randint(0, 99)
if(individual[mutation_point]):
individual[mutation_point] = 0
else:
individual[mutation_point] = 1
def build_next_generation(current_population, mutation_rate):
next_generation = []
next_generation.append(current_population[0][0]) # elitism
next_generation.append(current_population[random.randint(1,19)][0]) # randomness
for i in range(9): # tournaments
offspring_a, offspring_b = tournament(current_population)
next_generation.append(offspring_a)
next_generation.append(offspring_b)
for individual in next_generation: # mutation
if(random.randint(1,mutation_rate) == 1):
mutation(individual)
return next_generation
def run_ga(current_population, num_of_generations, mutation_rate=1000):
fittest_individuals = []
for i in range(num_of_generations):
current_population = get_population_fitness(current_population) # get pop fitness
fittest_individuals.append(current_population[0]) # record fittest individual (for graphing and analysis)
current_population = build_next_generation(current_population, mutation_rate) # make new population
return fittest_individuals
initial_population = [[random.randint(0, 1) for i in range(100)] for i in range(20)]
high_mutation_fittest = run_ga(initial_population, 100, mutation_rate=5)
high_mutation_fitness = [ind[1] for ind in high_mutation_fittest]
for item in high_mutation_fittest[:-1]:
if item[1] == max(high_mutation_fitness):
top_performer = item
break
#print("Total features included: " + str(top_performer[0].count(1)))
selected_features = [column for (column, binary_value) in zip(X.columns, top_performer[0]) if binary_value]
excluded_features = [column for (column, binary_value) in zip(X.columns, top_performer[0]) if not binary_value]
X = X[selected_features]
categorical_cols = [c for c, t in zip(X.columns, X_column_types) if t in [DATATYPE_CATEGORY_INT, DATATYPE_CATEGORY_STRING]]
numerical_cols = [c for c, t in zip(X.columns, X_column_types) if t == DATATYPE_NUMBER]
categorical = X[categorical_cols]
numerical = X[numerical_cols]
if method == "Binary_operators":
print("binaryoperators")
Binary_operator = pd.DataFrame()
#Apply binary operators
for i in range(numerical.shape[1]):
a = numerical.iloc[:,i]
for j in range(i+1, numerical.shape[1]):
b = numerical.iloc[:,j]
result = a*b
Binary_operator = pd.concat([Binary_operator, result], axis=1)
# apply addition operation
for i in range(numerical.shape[1]):
a = numerical.iloc[:,i]
for j in range(i+1, numerical.shape[1]):
b = numerical.iloc[:,j]
result = a+b
Binary_operator = pd.concat([Binary_operator, result], axis=1)
numerical = Binary_operator.copy()
if method == "Binning":
print("Binning")
num_discretizer=pd.DataFrame()
for i in range(numerical.shape[1]):
d_f1 = pd.DataFrame(pd.cut(numerical.iloc[:,i], 6, labels=False, duplicates='drop'))
d_f2 = pd.DataFrame(pd.cut(numerical.iloc[:,i], 4, labels=False, duplicates='drop'))
num_discretizer = pd.concat([num_discretizer, d_f1, d_f2], axis=1)
else:
# discritize the numerical features
num_discretizer=pd.DataFrame()
for i in range(numerical.shape[1]):
d_f = pd.DataFrame(pd.cut(numerical.iloc[:,i], 10, labels=False))
d_f2 = pd.DataFrame(pd.cut(numerical.iloc[:,i], 5, labels=False))
d_f3 = pd.DataFrame(pd.cut(numerical.iloc[:,i], 4, labels=False))
d_f4 = pd.DataFrame(pd.cut(numerical.iloc[:,i], 3, labels=False))
num_discretizer = pd.concat([num_discretizer, d_f, d_f2, d_f3, d_f4], axis=1)
# function to rename the duplicate columns
def df_column_uniquify(df):
df_columns = df.columns
new_columns = []
for item in df_columns:
counter = 0
newitem = item
while newitem in new_columns:
counter += 1
newitem = "{}_{}".format(item, counter)
new_columns.append(newitem)
df.columns = new_columns
return df
num_discretizer=df_column_uniquify(num_discretizer)
# Categorical features encoding
cat_list=[]
for i in range(categorical.shape[1]):
if (len(categorical.iloc[:, i].unique()) >= 2):
cat_list.append(categorical.keys()[i])
categorical = categorical[cat_list]
# We cluster the categorical features by most frequently repated and kmeans clustering
if method == "cluster":
print("clustering")
categorical_strg = [c for c, t in zip(categorical.columns, X_column_types) if t in [DATATYPE_CATEGORY_STRING]]
categorical_int = [c for c, t in zip(categorical.columns, X_column_types) if t in [DATATYPE_CATEGORY_INT]]
strg_cate = categorical[categorical_strg]
int_cate = categorical[categorical_int]
frequent = pd.DataFrame()
frequent_col = []
if len(strg_cate.columns)>=1:
# clustering the string categorical variables by top 10 frequently repeated values
for i in range(len(strg_cate.columns)):
if (strg_cate[strg_cate.columns[i]].nunique() > 10):
frequent_col.append(strg_cate.columns[i])
n=10
top=strg_cate[strg_cate.columns[i]].value_counts()[:n].index.tolist()
for label in top:
strg_cate[label]= np.where(strg_cate[strg_cate.columns[i]]==label, 1, 0)
df1 = strg_cate[[strg_cate.columns[i]]+top]
frequent = pd.concat([frequent, df1.drop([strg_cate.columns[i]], axis=1)], axis=1)
if len(frequent_col)>=1:
strg_cate=strg_cate.drop(frequent_col, axis=1)
else:
strg_cate = strg_cate.copy()
if len(int_cate.columns)>=1:
# clustering the interger categorical variables by using kmeans clustering
int_col=[]
for i in range(len(int_cate.columns)):
if (int_cate[int_cate.columns[i]].nunique() > 10):
x = int_cate.iloc[:,i:i+1]
kmeans = KMeans(10)
kmeans.fit(x)
cluster = kmeans.fit_predict(x)
int_cate[int_cate.columns[i] + '_cluster']=cluster
int_col.append(int_cate.columns[i])
if len(int_col)>=1:
int_cate = int_cate.drop(int_col, axis=1)
else:
int_cate = int_cate.copy()
if (len(strg_cate.columns)>0 or len(int_cate.columns)>0):
categorical = pd.concat([strg_cate, int_cate], axis=1)
enc = OneHotEncoder()
enc.fit(categorical)
Data_cat=pd.DataFrame(enc.transform(categorical).toarray())
if len(frequent_col)>=1:
original_feats = pd.concat([numerical, Data_cat, frequent], axis=1)
num_discret = pd.concat([numerical, Data_cat, frequent, num_discretizer], axis=1)
else:
original_feats = pd.concat([numerical, Data_cat], axis=1)
num_discret = pd.concat([numerical, Data_cat, num_discretizer], axis=1)
else:
# One hot encode the categorical_features
#Data_cat = pd.get_dummies(categorical)
enc = OneHotEncoder()
enc.fit(categorical)
Data_cat=pd.DataFrame(enc.transform(categorical).toarray())
original_feats = pd.concat([numerical, Data_cat], axis=1)
num_discret = pd.concat([numerical, Data_cat, num_discretizer], axis=1)
#Select the best half of discretized features by Mini batch gradient descent
#clf = SGDClassifier(loss="log", penalty="l1")
mini_batches = []
batch_size=32
data = np.hstack((num_discret, (y.values).reshape(-1,1)))
#data =pd.concat([num_discretizer, y], axis=1)
np.random.shuffle(data)
n_minibatches = data.shape[0] // batch_size
i = 0
for i in range(n_minibatches + 1):
mini_batch = data[i * batch_size:(i + 1)*batch_size, :]
X_mini = mini_batch[:, :-1]
Y_mini = mini_batch[:, -1].reshape((-1, 1))
mini_batches.append((X_mini, Y_mini))
if data.shape[0] % batch_size != 0:
mini_batch = data[i * batch_size:data.shape[0]]
X_mini = mini_batch[:, :-1]
Y_mini = mini_batch[:, -1].reshape((-1, 1))
mini_batches.append((X_mini, Y_mini))
if (y_column_types[0] == DATATYPE_NUMBER):
model = SGDRegressor(loss="squared_loss", penalty="l1")
for X_mini, Y_mini in mini_batches:
model.partial_fit(X_mini, Y_mini)
coefs=model.coef_
else:
model = SGDClassifier(loss="log", penalty="l1")
for X_mini, Y_mini in mini_batches:
model.partial_fit(X_mini, Y_mini, classes=np.unique(y))
coefs=model.coef_[0]
num = len(numerical.columns)+len(Data_cat.columns)
#coefs=model.coef_
h=np.argsort(coefs[num:])[::-1][:int(num_discretizer.shape[1]/2)]
best_half_sorted = [x+num for x in h]
best_dicretized = num_discret.iloc[:,best_half_sorted]
total = pd.concat([categorical, best_dicretized], axis=1)
# one hot encode the interger discretized features
enc = OneHotEncoder()
enc.fit(best_dicretized)
dicretized_ohe=pd.DataFrame(enc.transform(best_dicretized).toarray())
# combine cat_ohe and disretized_ohe features
Data = pd.concat([Data_cat, dicretized_ohe], axis=1)
# Rename the features which has duplicates
Data = df_column_uniquify(Data)
second_order = pd.DataFrame()
final_feats = pd.DataFrame()
cnt = 0
cnt_1 = 0
for i in range(len(total.columns)-1):
a= Data.iloc[:,[o for o in range(cnt, cnt+len(total.iloc[:, i].unique()))]]
cnt = cnt+len(total.iloc[:, i].unique())
cnt_1 = cnt
for j in range(i+1, len(total.columns)):
b= Data.iloc[:,[p for p in range(cnt_1, cnt_1+len(total.iloc[:, j].unique()))]]
cnt_1 = cnt_1+len(total.iloc[:, j].unique())
first = pd.DataFrame()
for k in range(a.shape[0]):
c = a.iloc[[k]].values
d = b.iloc[[k]].values
result = np.outer(c, d).ravel()
first=first.append(pd.Series(result), ignore_index=True)
second_order = pd.concat([second_order, first], axis =1)
second_order = df_column_uniquify(second_order)
firstorder_select = pd.concat([original_feats, second_order], axis=1)
# slect the second order features using Logistic regression
#clf = SGDClassifier(loss="log", penalty="l1")
mini_batches = []
batch_size=32
data = np.hstack((firstorder_select, (y.values).reshape(-1,1)))
#data = pd.concat([second_order, y], axis=1)
np.random.shuffle(data)
n_minibatches = data.shape[0] // batch_size
i = 0
for i in range(n_minibatches + 1):
mini_batch = data[i * batch_size:(i + 1)*batch_size, :]
X_mini = mini_batch[:, :-1]
Y_mini = mini_batch[:, -1].reshape((-1, 1))
mini_batches.append((X_mini, Y_mini))
if data.shape[0] % batch_size != 0:
mini_batch = data[i * batch_size:data.shape[0]]
X_mini = mini_batch[:, :-1]
Y_mini = mini_batch[:, -1].reshape((-1, 1))
mini_batches.append((X_mini, Y_mini))
#create_mini_batches(gen_feats, y, 32)
if (y_column_types[0] == DATATYPE_NUMBER):
model = SGDRegressor(loss="squared_loss", penalty="l1")
for X_mini, Y_mini in mini_batches:
model.partial_fit(X_mini, Y_mini)
coefs=model.coef_
else:
model = SGDClassifier(loss="log", penalty="l1")
for X_mini, Y_mini in mini_batches:
model.partial_fit(X_mini, Y_mini, classes=np.unique(y))
coefs=model.coef_[0]
num1 = len(original_feats.columns)
#selected top 10 features
g=np.argsort(coefs[num1:])[::-1][:10]
selected_sorted=[x+num1 for x in g]
selected_best = firstorder_select.iloc[:, selected_sorted]
selected = selected_best.copy()
new_col_types = X_column_types+[DATATYPE_CATEGORY_INT]*len(selected_best.columns)
total_feats = pd.concat([original_feats, selected_best], axis=1)
final_feats = pd.concat([X, selected_best], axis=1)
# higher order features generation
if len(categorical.columns)>2:
for i in range(len(categorical.columns)-2):
cnt = 0
Higher_order = pd.DataFrame()
for i in range(len(total.columns)):
a= Data.iloc[:,[o for o in range(cnt, cnt+len(total.iloc[:, i].unique()))]]
cnt = cnt+len(total.iloc[:, i].unique())
for j in range(selected_best.shape[1]):
b= selected_best.iloc[:,j]
second = pd.DataFrame()
for k in range(a.shape[0]):
c = a.iloc[[k]].values
d = b.iloc[[k]].values
result_1 = np.outer(c, d).ravel()
second=second.append(pd.Series(result_1), ignore_index=True)
Higher_order = pd.concat([Higher_order, second], axis =1)
Higher_order=df_column_uniquify(Higher_order)
High_order_sel = pd.concat([total_feats, Higher_order], axis=1)
mini_batches = []
batch_size=32
data = np.hstack((High_order_sel, (y.values).reshape(-1,1)))
#data = pd.concat([Higher_order, y], axis=1)
np.random.shuffle(data)
n_minibatches = data.shape[0] // batch_size
i = 0
for i in range(n_minibatches + 1):
mini_batch = data[i * batch_size:(i + 1)*batch_size, :]
X_mini = mini_batch[:, :-1]
Y_mini = mini_batch[:, -1].reshape((-1, 1))
mini_batches.append((X_mini, Y_mini))
if data.shape[0] % batch_size != 0:
mini_batch = data[i * batch_size:data.shape[0]]
X_mini = mini_batch[:, :-1]
Y_mini = mini_batch[:, -1].reshape((-1, 1))
mini_batches.append((X_mini, Y_mini))
#create_mini_batches(gen_feats, y, 32)
if (y_column_types[0] == DATATYPE_NUMBER):
model = SGDRegressor(loss="squared_loss", penalty="l1")
for X_mini, Y_mini in mini_batches:
model.partial_fit(X_mini, Y_mini)
coefs=model.coef_
else:
model = SGDClassifier(loss="log", penalty="l1")
for X_mini, Y_mini in mini_batches:
model.partial_fit(X_mini, Y_mini, classes=np.unique(y))
coefs=model.coef_[0]
#coefs=model.coef_
num2 = len(total_feats.columns)
sort=np.argsort(coefs[num2:])[::-1][:5]
selected_sorted=[x+num2 for x in sort]
selected_best = High_order_sel.iloc[:, selected_sorted]
selected = pd.concat([selected, selected_best], axis=1)
total_feats = pd.concat([total_feats, selected_best], axis=1)
final_feats = pd.concat([final_feats, selected_best], axis=1)
transformed_X = final_feats
new_col_types = X_column_types+[DATATYPE_CATEGORY_INT]*len(selected.columns)
else:
transformed_X = final_feats
return None, transformed_X, new_col_types
def create_sgd(arguments):
return SGDRegressor(**arguments)
preds = []
from sklearn.linear_model import (LinearRegression, SGDRegressor)
import lightgbm as lgb
sgdr= SGDRegressor(
penalty = 'l2' ,
random_state = SEED )
lgb_params = {
'feature_fraction': 0.75,
'metric': 'rmse',
'nthread':1,
'min_data_in_leaf': 2**7,
'bagging_fraction': 0.75,
def train_SGDRegressor(X_train, y_train):
model = SGDRegressor()
model.fit(X_train, y_train)
return model
user_avg = np.zeros(nusers)
user_std = np.zeros(nusers)
for i in range(0, nusers, batch_size):
users_current = R_u[i:min(i + batch_size, nusers), :]
batch_avg = ((users_current.sum(axis=1).flatten()) /
users_current.getnnz(axis=1))
user_avg[i:min(i + batch_size, nusers)] = batch_avg
user_std[i:min(i + batch_size, nusers)] = np.sqrt(
abs(
users_current.power(2).sum(axis=1).flatten() /
users_current.getnnz(axis=1) - batch_avg))
print 'done avging', movie_avg, user_avg
# sgd fitter
lin_model = SGDRegressor()
rat_num = len(probe_ratings)
for i in range(0, rat_num, batch_size):
given = probe_ratings[i:min(i + batch_size, rat_num)]
u_mean = user_avg[probe_users[i:min(i + batch_size, probe_num)]]
m_mean = movie_avg[probe_movies[i:min(i + batch_size, probe_num)]]
u_mean = np.array([u_mean]).T
m_mean = np.array([m_mean]).T
preding = np.concatenate((u_mean, m_mean), axis=1)
lin_model.partial_fit(preding, given)
model_2 = BayesianRidge()
model_3 = LassoLars(alpha=0.3, fit_intercept=False, normalize=True)
model_4 = DecisionTreeRegressor(min_samples_leaf=20)
model_5 = RandomForestRegressor(n_estimators=30)
model_6 = KNeighborsRegressor(n_neighbors = 30)
model_7 = xgb.XGBRegressor(objective ='reg:squarederror', colsample_bytree = 0.3, learning_rate = 0.1, max_depth = 5, alpha = 10, n_estimators = 10)
model_8=SGDRegressor(max_iter=np.ceil(10**6 / max([len(x) for x in train_X_drug_catg])), tol=1e-3,loss="squared_loss", penalty=None)
model_list=[model_1,model_2,model_3,model_4,model_5,model_6,model_7,model_8]
model_name=['Linear','Bayesian','LassoLars','DecisionTree','RandomForest','KNeighbors','XGB','SGD']
model_dict=dict(zip(model_list,model_name))
for model_x in model_list:
for which_drg_catg in range(len(drug_catgrs)):
for z in store_list_drug_catg[which_drg_catg]:
def train(features, labels):
lr = SGDRegressor()
lr.fit(features, labels)
weights = lr.coef_
return weights
# print(appIndex, score, timerEnd-timerStart)
return prediction
# initialize vars -------------------------------------------------------------
# define scaler for normalization
scaler = StandardScaler()
# define PCA for dimentionality reduction
pca = IncrementalPCA()
# define regression model
regressor = SGDRegressor(random_state=42, max_iter=1, tol=1e-3)
regressors, predictions = None, None
# load training data one batch at a time
trainFiles = glob(DATA_SUBFOLDER + '/trainBatch*')
timerStartLocal = time.time()
for fileNum, file in enumerate(trainFiles):
# load batch
timerStartLocal2 = time.time()
data, labels = load(file)
#init scaler and pca
scaler.partial_fit(data)
# pca.partial_fit(data)
def train_model(features, targets, *params):
model = SGDRegressor()
model.fit(features, targets)
return model
('tfidf_transformer', TfidfTransformer())])
)
## prediction
def fillna_and_log(x):
x = x.copy()
x[np.isnan(x)] = 0
return np.log(1 + x)
from sklearn.linear_model import SGDRegressor, SGDClassifier
title_word_1_2gram_dtm_0_predict_log_price = PredictionFeature('title_word_1_2gram_dtm_0_predict_log_price',
title_word_1_2gram_dtm_0,
SGDRegressor(penalty='elasticnet', l1_ratio=0.7,
random_state=132, n_iter=20), price,
y_transformer=fillna_and_log, keep_true=True,
true_name='log_price')
title_word_1_2gram_dtm_0_predict_is_test = PredictionFeature('title_word_1_2gram_dtm_0_predict_is_test',
title_word_1_2gram_dtm_0, \
SGDClassifier(penalty='elasticnet', l1_ratio=0.7,
random_state=132, n_iter=20), is_test,
y_transformer=None, keep_true=False,
only_predict=True, predict_binary_probability=True,
true_name='')
title_description_dtm_0_predict_log_price = PredictionFeature('title_description_dtm_0_predict_log_price',
title_description_dtm_0,
SGDRegressor(penalty='elasticnet', l1_ratio=0.7,
random_state=133, n_iter=30), price,
y_transformer=fillna_and_log)
eta = learning_schedule(epoch * m + i)
theta = theta - eta * gradients
theta_path_sgd.append(theta)
plt.plot(X, y, "b.")
plt.xlabel("$x_1$", fontsize=18)
plt.ylabel("$y$", rotation=0, fontsize=18)
plt.title('SGD')
plt.axis([0, 2, 0, 15])
plt.show()
#%% SGD SCikit-Learn
params = []
from sklearn.linear_model import SGDRegressor
sgd_reg = SGDRegressor(n_iter=50, penalty=None, eta0=0.1)
sgd_reg.fit(X, y)
params.append(sgd_reg.intercept_)
params.append(sgd_reg.coef_)
#%% Mini Batch Gradient Descendent
theta_path_mgd = []
n_iterations = 50
minibatch_size = 20
theta = np.random.randn(2, 1) # random initialization
t0, t1 = 10, 1000
def learning_schedule(t):
X_train, X_test, y_train, y_test = split_data(X, y)
X_train, X_test, y_train, y_test = preprocess_data(X_train, X_test,
y_train, y_test)
#print(X_train)
#X_values = np.delete(raw_data, raw_data.shape[1]-1, 1)
#Y_values = raw_data[:,raw_data.shape[1]-1]
weights_sk = np.full(
(1, X_train.shape[1]), 1.0
) #do not reuse the weights since sk-learn does inplace work with the coef_init matrix!
intercept_sk = 1
weights_own = np.full((1, X_train.shape[1]), 1.0)
intercept_own = 1
sk_gdc = SGDRegressor()
sk_gdc.fit(
X_train, y_train, coef_init=weights_sk, intercept_init=intercept_sk
) #coef_init is the same as our weights for comparison reasons (sklear does not pass w_0!)
print("Weights and intercept found by sk:", weights_sk, intercept_sk)
own_gdc = OwnGradientDescentRegressor(debug_output=True)
print(weights_own, weights_own.shape)
weights_own, intercept_own = own_gdc.fit(X_train,
y_train,
coef_init=weights_own,
intercept_init=intercept_own)
print("Weights and intercept found by own:", weights_own, intercept_own)
print("Prediction with sk-learn:", sk_gdc.predict(X_test))
print("Prediction with own-imp:", own_gdc.predict(X_test))
#!/usr/bin/python # -*- coding: UTF-8 -*- # 文件名: elastic_net.py import numpy as np from sklearn.linear_model import ElasticNet from sklearn.linear_model import SGDRegressor __author__ = 'yasaka' X = 2 * np.random.rand(100, 1) y = 4 + 3 * X + np.random.randn(100, 1) elastic_net = ElasticNet(alpha=0.0001, l1_ratio=0.15) elastic_net.fit(X, y) print(elastic_net.predict(1.5)) sgd_reg = SGDRegressor(penalty='elasticnet', n_iter=1000) sgd_reg.fit(X, y.ravel()) print(sgd_reg.predict(1.5))
eta0 = 0.01
max_iter = 100
from sklearn.model_selection import train_test_split
X_train_dataset, X_test, y_train_dataset, y_test = train_test_split(
X_scaled,y, test_size=0.2, random_state=42)
sgd_regressor = SGDRegressor(
eta0=eta0, max_iter=max_iter, warm_start=True, learning_rate="constant")
rmse_val_score = []
rmse_train_score = []
model_list = []
X_train, X_val, y_train, y_val = train_test_split(
X_train_dataset,y_train_dataset, test_size=0.2, random_state=42)
sgd_regressor.fit(X_train,y_train)
# kf = KFold(n_splits=100, shuffle=True)
# for train_index, test_index in kf.split(X_train_dataset):
for i in range(300):
y_pred = sgd_regressor.predict(X_train)
train_mse = mean_squared_error(y_true=y_train, y_pred=y_train_pred)
train_rmse = np.sqrt(train_mse)
print('Linear Regression Train RMSE:', train_rmse)
train_r2 = r2_score(y_train, y_train_pred)
print('Linear Regression Train R^2:', train_r2)
# 테스트 세트의 예측값
y_test_pred = lin_reg.predict(X_test)
test_mse = mean_squared_error(y_test, y_test_pred)
test_rmse = np.sqrt(test_mse)
test_r2 = r2_score(y_test, y_test_pred)
print('Linear Regression Test RMSE:', test_rmse)
print('Linear Regression Test R^2:', test_r2)
# LinearRegression vs SGDRegressor
sgd_reg = SGDRegressor(random_state=1) # 모델 생성
sgd_reg.fit(X_train, y_train) # 모델 훈련
y_train_pred = sgd_reg.predict(X_train) # 학습 세트 예측값
# -> 학습 세트의 RMSE, R2-score
y_test_pred = sgd_reg.predict(X_test) # 테스트 세트 예측값
# -> 테스트 세트의 RMSE, R2-Score
# Scaler 사용 -> Pipeline
pipe1 = Pipeline([('scaler', StandardScaler()),
('regressor', LinearRegression())])
pipe1.fit(X_train, y_train) # 학습
y_train_pred = pipe1.predict(X_train) # Train 예측값
# -> Train RMSE, R2-score
y_test_pred = pipe1.predict(X_test) # Test 예측값
scaler = StandardScaler()
#Borramos las 5 primeras columnas pues no aportan información
x = np.delete(x, 1, axis=1)
x = np.delete(x, 1, axis=1)
x = np.delete(x, 1, axis=1)
x = np.delete(x, 1, axis=1)
x = np.delete(x, 1, axis=1)
x_train, x_test, y_train, y_test = train_test_split(x,
y,
test_size=0.33,
shuffle=True)
#Preprocesado de los datos
preproc = [("missing", SimpleImputer()), ("var", VarianceThreshold(0.01)),
("poly", PolynomialFeatures(1)), ("standardize", StandardScaler())]
pipe = Pipeline(preproc + [('model', SGDRegressor())])
params_grid = [{
"model": [SGDRegressor(max_iter=500)],
"model__loss": [
'huber', 'squared_loss', 'epsilon_insensitive',
'squared_epsilon_insensitive'
],
"model__penalty": ['l1', 'l2'],
"model__alpha":
np.logspace(-5, 5, 5),
"poly__degree": [1, 2]
}, {
"model": [LinearRegression()],
"poly__degree": [1, 2]
}, {
scaled = scaler.transform(X) scaled_df = pd.DataFrame(scaled, columns= X.columns) scaled_df[:5] X = scaled_df X[:5] from sklearn.linear_model import LinearRegression model = LinearRegression() from sklearn.model_selection import train_test_split X_train, X_test, y_train, y_test = train_test_split(X,y,test_size=0.3, random_state=42) model.fit(X_train, y_train) pred = model.predict(X_test) from sklearn import metrics metrics.r2_score(y_test, pred) from sklearn.linear_model import SGDRegressor mod = SGDRegressor() mod.fit(X_train, y_train) predict = mod.predict(X_test) metrics.r2_score(y_test, predict)
from sklearn.linear_model import SGDRegressor
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.tree import DecisionTreeRegressor
from script import features as f
train, test, feature_names = f.get_sets(1)
dict = train[feature_names].to_dict()
print(dict)
X_train, X_test, y_train, y_test = train_test_split(train[feature_names],
train.Sales.values,
test_size=0.10,
random_state=2)
clf = make_pipeline(StandardScaler(),
SGDRegressor(loss='squared_loss', penalty='l2'))
f.score("SGDRegressor", clf, X_train, y_train, X_test, y_test)
f.kfold_validation("SGDRegressor", clf, X_train, y_train)
f.submission(clf, train, test, feature_names)
# clf = DecisionTreeRegressor()
# f.score("DecisionTreeRegressor", clf, X_train, y_train, X_test, y_test)
# f.kfold_validation("DecisionTreeRegressor", clf, X_train, y_train)
#
# clf = RandomForestRegressor(n_jobs=-1, n_estimators=25)
# f.score("RandomForestRegressor", clf, X_train, y_train, X_test, y_test)
# f.kfold_validation("RandomForestRegressor", clf, X_train, y_train)
# clf = GradientBoostingRegressor()
# f.score("GradientBoostingRegressor", clf, X_train, y_train, X_test, y_test)
# f.kfold_validation("GradientBoostingRegressor", clf, X_train, y_train)
