def check(self, a_in, a_out, a_mask, act='Softmax'):
nn = MLPC(layers=[L(act)],
learning_rule='adam',
learning_rate=0.05,
n_iter=250,
n_stable=25)
nn.fit(a_in, a_out, a_mask)
return nn.predict_proba(a_in)
class TestClassifierFunctionality(unittest.TestCase):
def setUp(self):
self.nn = MLPC(layers=[L("Linear")], n_iter=1)
def test_FitAutoInitialize(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,))
self.nn.fit(a_in, a_out)
assert_true(self.nn.is_initialized)
def test_ExplicitValidSet(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,))
self.nn.valid_set = (a_in, a_out)
self.nn.fit(a_in, a_out)
assert_true(self.nn.is_initialized)
def test_PartialFit(self):
a_in, a_out = numpy.zeros((8,4)), numpy.random.randint(0, 5, (8,))
self.nn.partial_fit(a_in, a_out, classes=[0,1,2,3])
self.nn.partial_fit(a_in*2.0, a_out+1, classes=[0,1,2,3])
def test_PredictUninitializedNoUnitCount(self):
a_in = numpy.zeros((8,16))
assert_raises(AssertionError, self.nn.predict, a_in)
def test_PredictUninitializedNoLabels(self):
self.nn.layers[-1].units = 4
a_in = numpy.zeros((8,16))
assert_raises(AssertionError, self.nn.predict, a_in)
def test_PredictClasses(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,))
self.nn.fit(a_in, a_out)
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape[0], a_test.shape[0])
def test_PredictMultiClass(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 3, (8,2))
self.nn.fit(a_in, a_out)
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape, a_test.shape)
def test_EstimateProbalities(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,))
self.nn.fit(a_in, a_out)
a_test = self.nn.predict_proba(a_in)
assert_equal(type(a_out), type(a_test))
def test_CalculateScore(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,))
self.nn.fit(a_in, a_out)
f = self.nn.score(a_in, a_out)
assert_equal(type(f), numpy.float64)
class TestClassifierFunctionality(unittest.TestCase):
def setUp(self):
self.nn = MLPC(layers=[L("Linear")], n_iter=1)
def test_FitAutoInitialize(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, ))
self.nn.fit(a_in, a_out)
assert_true(self.nn.is_initialized)
def test_ExplicitValidSet(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, ))
self.nn.valid_set = (a_in, a_out)
self.nn.fit(a_in, a_out)
assert_true(self.nn.is_initialized)
def test_PartialFit(self):
a_in, a_out = numpy.zeros((8, 4)), numpy.random.randint(0, 5, (8, ))
self.nn.partial_fit(a_in, a_out, classes=[0, 1, 2, 3])
self.nn.partial_fit(a_in * 2.0, a_out + 1, classes=[0, 1, 2, 3])
def test_PredictUninitializedNoUnitCount(self):
a_in = numpy.zeros((8, 16))
assert_raises(AssertionError, self.nn.predict, a_in)
def test_PredictUninitializedNoLabels(self):
self.nn.layers[-1].units = 4
a_in = numpy.zeros((8, 16))
assert_raises(AssertionError, self.nn.predict, a_in)
def test_PredictClasses(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, ))
self.nn.fit(a_in, a_out)
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape[0], a_test.shape[0])
def test_PredictMultiClass(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 3, (8, 2))
self.nn.fit(a_in, a_out)
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape, a_test.shape)
def test_EstimateProbalities(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, ))
self.nn.fit(a_in, a_out)
a_test = self.nn.predict_proba(a_in)
assert_equal(type(a_out), type(a_test))
def test_CalculateScore(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, ))
self.nn.fit(a_in, a_out)
f = self.nn.score(a_in, a_out)
assert_equal(type(f), numpy.float64)
def load_train_data(path, modelNo=1):
X = []
with open(path + '/train_newFeat_sparse_mat.dat', 'rb') as infile:
X = pickle.load(infile)
random.seed(modelNo)
np.random.seed(modelNo)
r = random.sample(xrange(0, X.shape[1]), int(round(0.8 * X.shape[1])))
X = X[:, r]
y = pd.read_csv(path + '/labels.csv', index_col=False, header=None)
y = np.array(y).astype('int')
X_train, X_val, y_train, y_val = train_test_split(X,
y,
test_size=0.2,
random_state=modelNo,
stratify=y)
nn = Classifier(
layers=[
Layer("Rectifier", units=200, dropout=0.5),
Layer("Rectifier", units=200, dropout=0.5),
Layer("Rectifier", units=200, dropout=0.5),
Layer("Sigmoid")
],
learning_rate=0.02,
n_iter=40,
# valid_set=(X,y),
n_stable=15,
debug=True,
verbose=True)
print "Model No is", modelNo
if (modelNo == 1):
print "Model No is", modelNo
nn.valid_set = (X_val, y_val)
#rbm1 = SVC(C=100.0, gamma = 0.1, probability=True, verbose=1).fit(X[0:9999,:], y[0:9999])
#rbm2 = RandomForestClassifier(n_estimators=300, criterion='entropy', max_features='auto', bootstrap=False, oob_score=False, n_jobs=1, verbose=1).fit(X[0:9999,:], y[0:9999])
#rbm3 = GradientBoostingClassifier(n_estimators=50,max_depth=11,subsample=0.8,min_samples_leaf=5,verbose=1).fit(X[0:9999,:], y[0:9999])
nn.fit(X_train, y_train)
Y = []
with open(path + '/test_newFeat_sparse_mat.dat', 'rb') as infile:
Y = pickle.load(infile)
Y = Y[:, r]
preds2 = np.zeros((Y.shape[0], 38))
for i in xrange(0, 10):
s = i * 10000
e = min(preds2.shape[0], s + 10000)
preds2[s:e, :] = nn.predict_proba(Y[s:e, :])
p2 = pd.DataFrame(preds2)
p2.to_csv("p2_" + str(modelNo) + ".csv", index=None, header=None)
return p2
class TestClassifierFunctionality(unittest.TestCase):
def setUp(self):
self.nn = MLPC(layers=[L("Linear")], n_iter=1)
def test_FitAutoInitialize(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.zeros((8, ),
dtype=numpy.int32)
self.nn.fit(a_in, a_out)
assert_true(self.nn.is_initialized)
def test_PartialFit(self):
a_in, a_out = numpy.zeros((8, 4)), numpy.zeros((8, ),
dtype=numpy.int32)
self.nn.partial_fit(a_in, a_out, classes=[0, 1, 2, 3])
self.nn.partial_fit(a_in * 2.0, a_out + 1, classes=[0, 1, 2, 3])
def test_PredictUninitialized(self):
a_in = numpy.zeros((8, 16))
assert_raises(ValueError, self.nn.predict, a_in)
def test_PredictClasses(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.zeros((8, ),
dtype=numpy.int32)
self.nn.fit(a_in, a_out)
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape, a_test.shape)
def test_EstimateProbalities(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.zeros((8, ),
dtype=numpy.int32)
self.nn.fit(a_in, a_out)
a_test = self.nn.predict_proba(a_in)
assert_equal(type(a_out), type(a_test))
def test_CalculateScore(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.zeros((8, ),
dtype=numpy.int32)
self.nn.fit(a_in, a_out)
f = self.nn.score(a_in, a_out)
assert_equal(type(f), numpy.float64)
class TestClassifierFunctionality(unittest.TestCase):
def setUp(self):
self.nn = MLPC(layers=[L("Linear")], n_iter=1)
def test_FitAutoInitialize(self):
a_in, a_out = numpy.zeros((8,16)), numpy.zeros((8,), dtype=numpy.int32)
self.nn.fit(a_in, a_out)
assert_true(self.nn.is_initialized)
def test_PartialFit(self):
a_in, a_out = numpy.zeros((8,4)), numpy.zeros((8,), dtype=numpy.int32)
self.nn.partial_fit(a_in, a_out, classes=[0,1,2,3])
self.nn.partial_fit(a_in*2.0, a_out+1, classes=[0,1,2,3])
def test_PredictUninitialized(self):
a_in = numpy.zeros((8,16))
assert_raises(ValueError, self.nn.predict, a_in)
def test_PredictClasses(self):
a_in, a_out = numpy.zeros((8,16)), numpy.zeros((8,), dtype=numpy.int32)
self.nn.fit(a_in, a_out)
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape, a_test.shape)
def test_EstimateProbalities(self):
a_in, a_out = numpy.zeros((8,16)), numpy.zeros((8,), dtype=numpy.int32)
self.nn.fit(a_in, a_out)
a_test = self.nn.predict_proba(a_in)
assert_equal(type(a_out), type(a_test))
def test_CalculateScore(self):
a_in, a_out = numpy.zeros((8,16)), numpy.zeros((8,), dtype=numpy.int32)
self.nn.fit(a_in, a_out)
f = self.nn.score(a_in, a_out)
assert_equal(type(f), numpy.float64)
# Saving the NN
pickle.dump(nn, open("Convoluted.pk1", "wb"))
# ----------------------------------------------------------------------------------------------------------- #
# Estimating the generalisation error with CV: all classes indivudually and multiclass log-loss
print("CV for class-wise generalisation errors")
num_folds = 2
kf = KFold(y, n_folds=num_folds)
y_pred = y * 0
l_loss = np.zeros((num_folds,1), dtype= float)
p = 0
for train, test in kf:
X_train, X_test, y_train, y_test = X[train,:], X[test,:], y[train], y[test]
nn_cv = Classifier(layers=lay, learning_rate=0.001, n_iter=2)
nn_cv.fit(X=X_train, y=y_train)
y_pred[test] = nn_cv.predict(X_test)
y_pred2 = nn_cv.predict_proba(X_test)
l_loss[p, 0] = log_loss(y_test, y_pred2)
p += 1
print(classification_report(y, y_pred, target_names=namesClasses))
log_loss_CV = np.average(l_loss, axis=0)
# Calculating the multiclass log-loss
print("Multiclass Log-loss by CV: ", log_loss_CV)
print("Finished program")
print("--- %s seconds ---" % (round(time.time() - start_time, 4))) # Calculates machine time for the program
learning_rate=0.01,
batch_size=100,
n_iter=2000,
valid_size=0.25,
n_stable=200)
# Training
nn.fit(X_train, Y)
pickle.dump(nn, open('nn_susy_classification.pkl', 'wb'))
if not runTraining:
nn = pickle.load(open('nn_susy_classification.pkl', 'rb'))
# Testing
pred_train = nn.predict(X_train)
pred_test = nn.predict(X_test)
probabilities_train = nn.predict_proba(X_train)
probabilities_test = nn.predict_proba(X_test)
print "Training sample...."
print " Signal identified as signal (%) : ", 100.0 * np.sum(
pred_train[nBackgroundEvents:nBackgroundEvents +
nSignalEvents] == 1.0) / nSignalEvents
print " Signal identified as background (%) : ", 100.0 * np.sum(
pred_train[nBackgroundEvents:nBackgroundEvents +
nSignalEvents] == 0.0) / nSignalEvents
print " Background identified as signal (%) : ", 100.0 * np.sum(
pred_train[0:nBackgroundEvents] == 1.0) / nBackgroundEvents
print " Background identified as background (%): ", 100.0 * np.sum(
pred_train[0:nBackgroundEvents] == 0.0) / nBackgroundEvents
print ""
print "Testing sample...."
def two_layers_nnet(X_train,
Y_train,
X_test,
Y_test,
method1="Tanh",
neurons1=5,
method2="",
neurons2=0,
decay=0.0001,
learning_rate=0.001,
n_iter=25,
random_state=1):
"""
Parameters
----------
X_train : pandas data frame
data frame of features for the training set
Y_train : pandas data frame
data frame of labels for the training set
X_test : pandas data frame
data frame of features for the test set
Y_test : pandas data frame
data frame of labels for the test set
method1 : str
method used for the first layer
neurons1 : int
number of neurons of the first layer
method2 : None
method used for the first layer
neurons2 : int
number of neurons of the first layer
decay : float
weight decay
learning_rate : float
learning rate
n_iter : int
number of iterations
random_state : int
seed for weight initialization
Result:
-------
numpy array
logloss : averaged logarithmic loss
miss_err : missclassification error rate
prec : precision
recall : recall
f1 : f1 score
parameters : previous parameters in the order previously specified
"""
labels = np.unique(Y_train)
## # Scale Data
scaler = MinMaxScaler()
X_test = scaler.fit_transform(X_test)
X_train = scaler.fit_transform(X_train)
# Layers
if neurons2 == 0 :
layers=[Layer(method1, weight_decay = decay, units = neurons1),
Layer("Softmax")]
else:
layers=[Layer(method1, weight_decay = decay, units = neurons1),
Layer(method2, weight_decay = decay, units = neurons2),
Layer("Softmax")]
## # Run nnet
# Define classifier
nn = Classifier(layers,
learning_rate=learning_rate,
random_state=random_state,
n_iter=n_iter)
# Fit
nn.fit(X_train, Y_train)
# Predict
Y_hat = nn.predict(X_test)
Y_probs = nn.predict_proba(X_test)
## # Misclassification error rate
miss_err = 1-accuracy_score(Y_test, Y_hat)
## # Log Loss
logloss = log_loss(Y_test, Y_probs)
## # Precision
prec = precision_score(y_true=Y_test, y_pred=Y_hat, labels=labels, average='micro')
## # Recal
recall = recall_score(y_true=Y_test, y_pred=Y_hat, labels=labels, average='micro')
## # F1
f1 = f1_score(y_true=Y_test, y_pred=Y_hat, labels=labels, average='micro')
# Summarized results
result = np.array([logloss,
miss_err,
prec,
recall,
f1,
method1,
neurons1,
method2,
neurons2,
decay,
learning_rate,
n_iter,
random_state])
return result
def two_layers_nnet_predict(X_train,
Y_train,
X_test,
method1="Tanh",
neurons1=5,
method2="",
neurons2=0,
decay=0.0001,
learning_rate=0.001,
n_iter=25,
random_state=1):
"""
Parameters
----------
X_train : pandas data frame
data frame of features for the training set
Y_train : pandas data frame
data frame of labels for the training set
X_test : pandas data frame
data frame of features for the test set
method1 : str
method used for the first layer
neurons1 : int
number of neurons of the first layer
method2 : None
method used for the first layer
neurons2 : int
number of neurons of the first layer
decay : float
weight decay
learning_rate : float
learning rate
n_iter : int
number of iterations
random_state : int
seed for weight initialization
Result:
-------
tuple of numpy arrays
(predicted classes, predicted probabilities)
"""
labels = np.unique(Y_train)
## # Scale Data
scaler = MinMaxScaler()
X_test = scaler.fit_transform(X_test)
X_train = scaler.fit_transform(X_train)
## # Split data set into train/test
# Layers
if neurons2 == 0 :
layers=[Layer(method1, weight_decay = decay, units = neurons1),
Layer("Softmax")]
else:
layers=[Layer(method1, weight_decay = decay, units = neurons1),
Layer(method2, weight_decay = decay, units = neurons2),
Layer("Softmax")]
## # Run nnet
# Define classifier
nn = Classifier(layers,
learning_rate=learning_rate,
random_state=random_state,
n_iter=n_iter)
# Fit
nn.fit(X_train, Y_train)
# Predict
Y_hat = nn.predict(X_test)
Y_probs = nn.predict_proba(X_test)
# Summarized results
result = (Y_hat,Y_probs)
return result
learning_rate=0.00018, #valid_set = ((X_valid, y_valid))
n_iter=1000)
print "Neural network specifications:"
print nn
nn.fit(trainingSet, trainingSetLabels)
score1 = nn.score(trainingSet, trainingSetLabels)
score3 = nn.score(testingSet, testingSetLabels)
print "Training accuracy = ", score1
print "Testing accuracy = ", score3
probNN = nn.predict_proba(testingSet)
fprNN, tprNN, threshNN = metrics.roc_curve(testingSetLabels, probNN[:, 0]) #true positive rate, false positive rate (ROC curve)
print "Time = ", time.time() - startTime, "seconds"
startTime = time.time()
print
print
#------------------------SVM----------------------------
print "Support Vector Machine Classifier"
clf = svm.SVC(C = 100, gamma = 0.1, probability=True)
clf_info = clf.fit(trainingSet, trainingSetLabels)
print clf_info
layers=[
Layer("Rectifier", units=49),
Layer("Softmax")],
learning_rate=0.01,
batch_size = 100,
n_iter=100)
# Training
nn.fit(X_train,Y)
pickle.dump(nn, open('nn_susy_classification.pkl', 'wb'))
if not runTraining:
nn = pickle.load(open('nn_susy_classification.pkl', 'rb'))
# Testing
pred_train = nn.predict(X_train)
pred_test = nn.predict(X_test)
probabilities_train = nn.predict_proba(X_train)
probabilities_test = nn.predict_proba(X_test)
print "Training sample...."
print " Signal identified as signal (%) : ",100.0*np.sum(pred_train[nBackgroundEvents:nBackgroundEvents+nSignalEvents]==1.0)/nSignalEvents
print " Signal identified as background (%) : ",100.0*np.sum(pred_train[nBackgroundEvents:nBackgroundEvents+nSignalEvents]==0.0)/nSignalEvents
print " Background identified as signal (%) : ",100.0*np.sum(pred_train[0:nBackgroundEvents]==1.0)/nBackgroundEvents
print " Background identified as background (%): ",100.0*np.sum(pred_train[0:nBackgroundEvents]==0.0)/nBackgroundEvents
print ""
print "Testing sample...."
print " Signal identified as signal (%) : ",100.0*np.sum(pred_test[nBackgroundEvents:nBackgroundEvents+nSignalEvents]==1.0)/nSignalEvents
print " Signal identified as background (%) : ",100.0*np.sum(pred_test[nBackgroundEvents:nBackgroundEvents+nSignalEvents]==0.0)/nSignalEvents
print " Background identified as signal (%) : ",100.0*np.sum(pred_test[0:nBackgroundEvents]==1.0)/nBackgroundEvents
print " Background identified as background (%): ",100.0*np.sum(pred_test[0:nBackgroundEvents]==0.0)/nBackgroundEvents
## Plotting - performance curves
layer1 = Layer("Rectifier", units=45, weight_decay=0.001)
layer2 = Layer("Rectifier", units=30, weight_decay=0.001)
layer3 = Layer("Softmax")
cls = Classifier(layers=[layer1, layer2, layer3],
learning_rule="adam",
learning_rate=0.003,
f_stable=0.01,
debug=True,
batch_size=200,
n_iter=100)
cls.fit(X_train, y_train)
# get the probability of prediction for cross data
y_predict = cls.predict_proba(X_cross, collapse=True)
print(y_predict[:, 1])
p_cross = y_predict[:, 1]
loss_cross = -np.multiply(y_cross, np.log(p_cross)) - np.multiply(
1 - y_cross, np.log(1 - p_cross))
print(loss_cross.sum() / loss_cross.size)
# get the probability of prediction for test data
y_predict_test = cls.predict_proba(X_test, collapse=True)
p_test = y_predict_test[:, 1]
loss_test = -np.multiply(y_test, np.log(p_test)) - np.multiply(
1 - y_test, np.log(1 - p_test))
new_y_train = np.zeros((len(X_train), len(Classes)))
for i in range(0, len(TrainData)):
new_y_train[i, int(y_train[i])] = 1
nn.fit(X_train, y_train)
#y_valid = nn.predict(TestData[:,1:])
X_test = TestData[:, 1:]
y_test = TestData[:, 0]
#score = nn.score(X_test, y_test)
X_test_pred = nn.predict(X_test)
conf = nn.predict_proba(X_test)
print conf
filename = "confidence_nn" + ".txt"
#print np.array(conf)
#conf=[np.ravel(i) for i in conf]
#print conf
#conf = map()
np.savetxt(filename, np.array(conf), '%1.5f', delimiter=' ')
counter = 0
for i in range(0, len(X_test_pred)):
if TestData[i, 0] == X_test_pred[i]:
counter = counter + 1
with open("predictions_mlp.txt", "w") as resFile:
for i in range(0, len(X_test_pred)):
resFile.write(str(int(X_test_pred[i])) + "\n")
learning_rate=0.00018, #valid_set = ((X_valid, y_valid))
n_iter=1000)
print "Neural network specifications:"
print nn
nn.fit(trainingSet, trainingSetLabels)
score1 = nn.score(trainingSet, trainingSetLabels)
score3 = nn.score(testingSet, testingSetLabels)
print "Training accuracy = ", score1
print "Testing accuracy = ", score3
probNN = nn.predict_proba(testingSet)
fprNN, tprNN, threshNN = metrics.roc_curve(
testingSetLabels,
probNN[:, 0]) #true positive rate, false positive rate (ROC curve)
print "Time = ", time.time() - startTime, "seconds"
startTime = time.time()
print
print
#------------------------SVM----------------------------
print "Support Vector Machine Classifier"
clf = svm.SVC(C=100, gamma=0.1, probability=True)
class TestClassifierFunctionality(unittest.TestCase):
def setUp(self):
self.nn = MLPC(layers=[L("Softmax")], n_iter=1)
def test_IsClassifier(self):
assert_true(self.nn.is_classifier)
def test_FitAutoInitialize(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, ))
self.nn.fit(a_in, a_out)
assert_true(self.nn.is_initialized)
def test_ExplicitValidSet(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, ))
self.nn.valid_set = (a_in, a_out)
self.nn.fit(a_in, a_out)
assert_true(self.nn.is_initialized)
def test_PartialFit(self):
a_in, a_out = numpy.zeros((8, 4)), numpy.random.randint(0, 5, (8, ))
self.nn.partial_fit(a_in, a_out, classes=[0, 1, 2, 3])
self.nn.partial_fit(a_in * 2.0, a_out + 1, classes=[0, 1, 2, 3])
def test_PredictUninitializedNoUnitCount(self):
a_in = numpy.zeros((8, 16))
assert_raises(AssertionError, self.nn.predict, a_in)
def test_PredictUninitializedNoLabels(self):
self.nn.layers[-1].units = 4
a_in = numpy.zeros((8, 16))
assert_raises(AssertionError, self.nn.predict, a_in)
def test_PredictBinaryProbability(self):
a_in = numpy.random.uniform(-1.0, 1.0, size=(8, 16))
a_out = numpy.array((a_in.sum(axis=1) >= 0.0), dtype=numpy.int32)
a_out[0], a_out[-1] = 0, 1
self.nn.fit(a_in, a_out)
a_proba = self.nn.predict_proba(a_in)
a_test = self.nn.predict(a_in)
c_out = numpy.unique(a_out)
assert_equal(2, c_out.shape[0])
assert_equal((8, 2), a_proba.shape)
def test_PredictClasses(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, ))
self.nn.fit(a_in, a_out)
self.nn.batch_size = 4
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape[0], a_test.shape[0])
c_out = numpy.unique(a_out)
assert_equal(len(self.nn.classes_), 1)
assert_true((self.nn.classes_[0] == c_out).all())
def test_PredictLargerBatchSize(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, 1))
self.nn.batch_size = 32
self.nn.fit(a_in, a_out)
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape[0], a_test.shape[0])
def test_PredictMultiClass(self):
a_in, a_out = numpy.zeros(
(32, 16)), numpy.random.randint(0, 3, (32, 2))
self.nn.fit(a_in, a_out)
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape, a_test.shape)
assert_equal(len(self.nn.classes_), 2)
assert_equal(self.nn.classes_[0].shape[0], 3)
assert_equal(self.nn.classes_[1].shape[0], 3)
def test_EstimateProbalities(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, ))
self.nn.fit(a_in, a_out)
a_test = self.nn.predict_proba(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_in.shape[0], a_test.shape[0])
def test_CalculateScore(self):
a_in, a_out = numpy.zeros((8, 16)), numpy.random.randint(0, 5, (8, ))
self.nn.fit(a_in, a_out)
f = self.nn.score(a_in, a_out)
assert_equal(type(f), numpy.float64)
#!/usr/bin/python
from sknn.mlp import Classifier, Layer
import pandas as pd
train = pd.read_csv('../data/modeltrain.csv',index_col=0)
test = pd.read_csv('../data/modeltest.csv',index_col=0)
label = train['Response'].values
feat = train.columns.drop('Response',1)
nn = Classifier(
layers=[
Layer("Rectifier",units=10),
Layer("Softmax")],
learning_rate=0.001,
n_iter=25)
nn.fit(train[feat].values, label)
from sklearn.metrics import log_loss
log_loss(label,nn.predict_proba(train[feat].values)[:,1])
class TestClassifierFunctionality(unittest.TestCase):
def setUp(self):
self.nn = MLPC(layers=[L("Softmax")], n_iter=1)
def test_IsClassifier(self):
assert_true(self.nn.is_classifier)
def test_FitAutoInitialize(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,))
self.nn.fit(a_in, a_out)
assert_true(self.nn.is_initialized)
def test_ExplicitValidSet(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,))
self.nn.valid_set = (a_in, a_out)
self.nn.fit(a_in, a_out)
assert_true(self.nn.is_initialized)
def test_PartialFit(self):
a_in, a_out = numpy.zeros((8,4)), numpy.random.randint(0, 5, (8,))
self.nn.partial_fit(a_in, a_out, classes=[0,1,2,3])
self.nn.partial_fit(a_in*2.0, a_out+1, classes=[0,1,2,3])
def test_PredictUninitializedNoUnitCount(self):
a_in = numpy.zeros((8,16))
assert_raises(AssertionError, self.nn.predict, a_in)
def test_PredictUninitializedNoLabels(self):
self.nn.layers[-1].units = 4
a_in = numpy.zeros((8,16))
assert_raises(AssertionError, self.nn.predict, a_in)
def test_PredictBinaryProbability(self):
a_in = numpy.random.uniform(-1.0, 1.0, size=(8,16))
a_out = numpy.array((a_in.sum(axis=1) >= 0.0), dtype=numpy.int32)
a_out[0], a_out[-1] = 0, 1
self.nn.fit(a_in, a_out)
a_proba = self.nn.predict_proba(a_in)
a_test = self.nn.predict(a_in)
c_out = numpy.unique(a_out)
assert_equal(2, c_out.shape[0])
assert_equal((8, 2), a_proba.shape)
assert_true((a_proba >= 0.0).all())
assert_true((a_proba <= 1.0).all())
assert_true((abs(a_proba.sum(axis=1) - 1.0) < 1E-9).all())
def test_PredictClasses(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,))
self.nn.fit(a_in, a_out)
self.nn.batch_size = 4
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape[0], a_test.shape[0])
c_out = numpy.unique(a_out)
assert_equal(len(self.nn.classes_), 1)
assert_true((self.nn.classes_[0] == c_out).all())
def test_PredictLargerBatchSize(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,1))
self.nn.batch_size = 32
self.nn.fit(a_in, a_out)
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape[0], a_test.shape[0])
def test_PredictMultiClass(self):
a_in, a_out = numpy.zeros((32,16)), numpy.random.randint(0, 3, (32,2))
self.nn.fit(a_in, a_out)
a_test = self.nn.predict(a_in)
assert_equal(type(a_out), type(a_test))
assert_equal(a_out.shape, a_test.shape)
assert_equal(len(self.nn.classes_), 2)
assert_equal(self.nn.classes_[0].shape[0], 3)
assert_equal(self.nn.classes_[1].shape[0], 3)
def test_EstimateProbalities(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,))
self.nn.fit(a_in, a_out)
a_proba = self.nn.predict_proba(a_in)
assert_equal(type(a_out), type(a_proba))
assert_equal(a_in.shape[0], a_proba.shape[0])
assert_true((a_proba >= 0.0).all())
assert_true((a_proba <= 1.0).all())
assert_true((abs(a_proba.sum(axis=1) - 1.0) < 1E-9).all())
def test_MultipleProbalitiesAsList(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,4))
self.nn.fit(a_in, a_out)
a_proba = self.nn.predict_proba(a_in)
assert_equal(list, type(a_proba))
assert_equal(4, len(a_proba))
for p in a_proba:
assert_equal(a_in.shape[0], p.shape[0])
assert_less_equal(p.shape[1], 5)
assert_true((p >= 0.0).all())
assert_true((p <= 1.0).all())
assert_true((abs(p.sum(axis=1) - 1.0) < 1E-9).all())
def test_CalculateScore(self):
a_in, a_out = numpy.zeros((8,16)), numpy.random.randint(0, 5, (8,))
self.nn.fit(a_in, a_out)
f = self.nn.score(a_in, a_out)
assert_equal(type(f), numpy.float64)
def check(self, a_in, a_out, a_mask, act='Softmax', n_iter=100):
nn = MLPC(layers=[L(act)], learning_rule='rmsprop', n_iter=n_iter)
nn.fit(a_in, a_out, a_mask)
return nn.predict_proba(a_in)
class ClassifierScikitNN():
packageName = 'com.brodagroup.machinelearning.ClassifierScikitNN'
logger = None
hidden_units = None
classifier = None
# Initializer
def __init__(self,
num_classes=None,
num_features=None,
learning_rate=0.01,
learning_rule='sgd',
learning_momentum=0.9,
dropout_rate=None,
weight_decay=None,
random_state=0,
n_iter=10):
self.logger = Logger(self.packageName).getLogger()
self.logger.debug('Starting...')
self.num_classes = num_classes
self.num_features = num_features
self.learning_rule = learning_rule
self.learning_momentum = learning_momentum
self.dropout_rate = dropout_rate
self.learning_rate = learning_rate
self.weight_decay = weight_decay
self.random_state = random_state
self.n_iter = n_iter
self.hidden_units = round( (self.num_features + self.num_classes)/3, 0)
# Layer('Tanh', units=self.num_features),
# Layer('Maxout', units=self.num_features, pieces=2),
self.classifier = Classifier(
layers=[
Layer('Maxout', units=self.num_features, pieces=2),
Layer('Sigmoid', units=self.hidden_units),
Layer('Softmax', units=self.num_classes)
],
learning_rule=self.learning_rule,
learning_rate=self.learning_rate,
learning_momentum=self.learning_momentum,
dropout_rate=self.dropout_rate,
weight_decay=self.weight_decay,
random_state=self.random_state,
n_iter=self.n_iter)
return
def __str__(self):
x = self.packageName + '('
x = x + '\n\t num_classes={0}, num_features: {1}'.format(self.num_classes, self.num_features)
x = x + '\n\t learning_rule={0}, learning_rate: {1}'.format(self.learning_rule, self.learning_rate)
x = x + '\n\t learning_momentum={0}, dropout_rate: {1}'.format(self.learning_momentum, self.dropout_rate)
x = x + '\n\t hidden_units={0}, weight_decay: {1}'.format(self.hidden_units, self.weight_decay)
x = x + '\n\t random_state={0}, n_iter: {1}'.format(self.random_state, self.n_iter)
return(x)
def fit(self, X, y):
self.classifier.fit(X,y)
def predict(self, X):
y_pred = self.classifier.predict(X)
return(y_pred)
def predict_proba(self, X):
y_pred = self.classifier.predict_proba(X)
return(y_pred)
def get_params(self, deep=True):
return {
"num_classes": self.num_classes,
"num_features": self.num_features,
"learning_rule": self.learning_rule,
"learning_rate": self.learning_rate,
"learning_momentum": self.learning_momentum,
"dropout_rate": self.dropout_rate,
"weight_decay": self.weight_decay,
"random_state": self.random_state,
"n_iter": self.n_iter
}
def set_params(self, **parameters):
for parameter, value in parameters.items():
setattr(self, parameter, value)
return self
def check(self, a_in, a_out, a_mask, act='Softmax'):
nn = MLPC(layers=[L(act)], learning_rule='adam', learning_rate=0.05, n_iter=250, n_stable=25)
nn.fit(a_in, a_out, a_mask)
return nn.predict_proba(a_in)
# ====================================================
clf = Classifier(
layers=[Layer('Sigmoid', units=hu), Layer('Softmax', units=3)],
learning_rule=lr,
learning_rate=lrt,
n_iter=ni
)
startTime = datetime.datetime.now()
clf.fit(X_train, y_train)
endTime = datetime.datetime.now()
y_score = clf.predict_proba(X_test)
y_hat = clf.predict(X_test)
ys = [y_s[y_h-1] for y_s, y_h in zip(y_score, y_hat)]
tmp = np.append(X_test, np.reshape(y_test, (1,y_test.shape[0])).T, axis=1)
tmp = np.append(tmp, np.reshape(y_hat, (1,y_hat.shape[0])).T, axis=1)
tmp = np.append(tmp, y_score, axis=1)
tmp = np.append(tmp, np.asarray(ys), axis=1)
output['data'] = [['X_' + str(i) for i in range(1, X_test.shape[1] + 1)] +
['y_label', 'y_hat', 'y_score_1', 'y_score_2', 'y_score_3', 'ys']] + \
tmp.tolist()
acc = accuracy_score(y_hat, y_test)
confMatrix = confusion_matrix(y_test, y_hat).tolist()
def check(self, a_in, a_out, a_mask, act='Softmax', n_iter=100):
nn = MLPC(layers=[L(act)], learning_rule='rmsprop', n_iter=n_iter)
nn.fit(a_in, a_out, a_mask)
return nn.predict_proba(a_in)
#!/usr/bin/python
from sknn.mlp import Classifier, Layer
import pandas as pd
train = pd.read_csv('../data/modeltrain.csv', index_col=0)
test = pd.read_csv('../data/modeltest.csv', index_col=0)
label = train['Response'].values
feat = train.columns.drop('Response', 1)
nn = Classifier(layers=[Layer("Rectifier", units=10),
Layer("Softmax")],
learning_rate=0.001,
n_iter=25)
nn.fit(train[feat].values, label)
from sklearn.metrics import log_loss
log_loss(label, nn.predict_proba(train[feat].values)[:, 1])
#Set random_state=0 for testing
X_train, X_test, y_train, y_test, i_train, i_test = train_test_split(
X_data, y_data, range(0, len(y_data)), test_size=0.20)
classifier = Classifier(
layers=[Layer("Sigmoid", units=int(sys.argv[4])),
Layer("Softmax")],
learning_rate=float(sys.argv[2]),
n_iter=int(sys.argv[3]))
classifier.fit(X_train, y_train)
old_stdout = sys.stdout
sys.stdout = open(os.devnull, 'w')
results = classifier.predict(X_test) #May produce junk output
results_proba = classifier.predict_proba(X_test) #May produce junk output
sys.stdout.close()
sys.stdout = old_stdout
results = np.reshape(results, (results.shape[0], 1))
results_proba = np.reshape(results_proba, (results.shape[0], 2))
y_test = np.reshape(y_test, results.shape)
Acc = 100 * (results == y_test).sum() / float(len(y_test))
Pre = 100 * (np.logical_and(results == 1, y_test == 1)).sum() / float(
(results == 1).sum())
Rec = 100 * (np.logical_and(results == 1, y_test == 1)).sum() / float(
(y_test == 1).sum())
F1S = 2 * (Pre * Rec) / float(Pre + Rec)
print 'Acc\tPre\tRec\tF1S'
n_iter=10000,
#n_stable=50,
#f_stable=0.01,
valid_set=(base['validation']['data'],base['validation']['target']),
callback={'on_epoch_finish': store_errors},
verbose = verbose
)
if opt_samp == Oversampling.DontUse:
nn.fit(base['training']['data'],base['training']['target'],w_train)
else:
nn.fit(base['training']['data'],base['training']['target'])
print('Testing')
predictions = np.squeeze(np.asarray(nn.predict(base['testing']['data'])))
prob_predictions = nn.predict_proba(base['testing']['data'])
target = base['testing']['target']
targetByClass = np.array([0,0])
errors_total = 0
vp = 0
fp = 0
vn = 0
fn = 0
test_mse = 0
for predicted, obj in zip(predictions,base['testing']['target']):
predicted
if predicted != obj:
# print(' error')
clf = Classifier(layers=[
Layer('Sigmoid', units=hu),
Layer('Softmax', units=2)
],
learning_rule=lr,
learning_rate=lrt,
n_iter=ni)
startTime = datetime.now()
clf.fit(X_train, y_train)
endTime = datetime.now()
y_score = clf.predict_proba(X_test)
y_hat = clf.predict(X_test)
ys = [y_s[y_h] for y_s, y_h in zip(y_score, y_hat)]
tmp = np.append(X_test,
np.reshape(y_test, (1, y_test.shape[0])).T,
axis=1)
tmp = np.append(tmp,
np.reshape(y_hat, (1, y_hat.shape[0])).T,
axis=1)
tmp = np.append(tmp, y_score, axis=1)
tmp = np.append(tmp, np.asarray(ys), axis=1)
output['data'] = [['X_' + str(i) for i in range(1, X_test.shape[1] + 1)] +
['y_label', 'y_hat', 'y_score', 'y_score', 'ys']] + \
tmp.tolist()
np.save("data_x_fire_a", data_x)
np.save("data_y_fire_a", data_y)
print 'saved'
data_x = np.load("data_x_fire_a.npy")[100:]
data_y = 1*np.load("data_y_fire_a.npy")[100:]
nn = Classifier(
layers=[
Layer("Sigmoid", units=512),
Layer("Softmax")],
learning_rate=0.01,
n_iter=40,
# callback=my_callback
)
print("Generating Fit")
nn.fit(data_x, data_y)
print("Fit generated")
fs = open('nn_fire_a.pkl', 'wb')
pickle.dump(nn, fs)
fs = open('nn_fire_a.pkl', 'rb')
nn = pickle.load(fs)
n = 259
# print(nn.predict_proba(data_x[n:n+20]))
out = np.column_stack((nn.predict_proba(data_x[n:n+20]), nn.predict(data_x[n:n+20]), data_y[n:n+20]))
print(out)
# print(data_y[n:n+20])
# nn.score(data_x, data_y)
# fs.close()
# print("NN Pickled")
# pickle.save()