def improveHyperParameters(train, targetId, tuned_parameter, controlEst):
expReg = ExploreRegressor.ExploreRegressor(train, targetId, svm.SVR(),
0.25)
best = expReg.gridSearch(tuned_parameters=tuned_parameter,
cv=10,
verbose=1)
print '%s control est' % (targetId)
control = ExploreRegressor.ExploreRegressor(train, targetId, controlEst,
0.25)
control.reportCrossValidationError(cv=20)
print '%s optimal est. Parmeters: %s' % (targetId, best.best_params_)
optimal = ExploreRegressor.ExploreRegressor(train, targetId,
best.best_estimator_, 0.25)
optimal.reportCrossValidationError(cv=10)
def improveHyperParameters(train, targetId, tuned_parameter, controlEst):
# Shuffle rows to not over fit on groups
train = train.reindex(np.random.permutation(train.index))
expReg = ExploreRegressor.ExploreRegressor(
train, targetId, ensemble.ExtraTreesRegressor(n_jobs=10), 0.25)
best = expReg.gridSearch(tuned_parameters=tuned_parameter, cv=5, verbose=2)
print '%s control est' % (targetId)
control = ExploreRegressor.ExploreRegressor(train, targetId, controlEst,
0.25)
control.reportCrossValidationError(cv=10)
print '%s optimal est. Parmeters: %s' % (targetId, best.best_params_)
optimal = ExploreRegressor.ExploreRegressor(train, targetId,
best.best_estimator_, 0.25)
optimal.reportCrossValidationError(cv=10)
def testRegressor(train, regressor, target, id):
expReg = ExploreRegressor.ExploreRegressor(train, target, regressor, 0.25)
print '%s model' % (id)
expReg.reportCrossValidationError(cv=20)
def reportError( trainFile, testFile, dataId ):
train = pd.read_csv( trainFile, header=0 )
test = pd.read_csv( testFile, header=0 )
# Replace nominal parameters by 0 and 1s
train.replace( to_replace='Topsoil', value=0, inplace=True )
train.replace( to_replace='Subsoil', value=1, inplace=True )
test.replace( to_replace='Topsoil', value=0, inplace=True )
test.replace( to_replace='Subsoil', value=1, inplace=True )
est = svm.SVR(C=10000.0, verbose = 0)
targets = ['Ca','P','pH','SOC','Sand']
# ests = [
# svm.SVR(kernel='poly',C=10000.0, degree=2, gamma=0.001 ), # Control SVR for Ca
# svm.SVR(kernel='rbf',C=15000.0, degree=1, gamma=0.001 ), # Control SVR for P
# svm.SVR(kernel='rbf',C=10000.0, degree=1, gamma=0.001 ), # Control SVR for pH
# svm.SVR(kernel='rbf',C=5000.0, degree=1, gamma=0 ), # Control SVR for SOC
# svm.SVR(kernel='rbf',C=15000.0, degree=1, gamma=0 ), # Control SVR for Sand
# ]
# ests = [
# svm.SVR(kernel='poly',C=9000.0, degree=2, gamma=0.0009 ), # Control SVR for Ca
# svm.SVR(kernel='rbf',C=17500.0, degree=1, gamma=0.00125 ), # Control SVR for P
# svm.SVR(kernel='rbf',C=7500.0, degree=1, gamma=0.00075 ), # Control SVR for pH
# svm.SVR(kernel='rbf',C=5000.0, degree=1, gamma=0 ), # Control SVR for SOC
# svm.SVR(kernel='rbf',C=12500.0, degree=1, gamma=0 ), # Control SVR for Sand
# ]
# ests = [
# svm.SVR(C=10000.0), # Control SVR for Ca
# svm.SVR(C=10000.0), # Control SVR for P
# svm.SVR(C=10000.0), # Control SVR for pH
# svm.SVR(C=10000.0), # Control SVR for SOC
# svm.SVR(C=10000.0), # Control SVR for Sand
# ]
ests = [
svm.SVR( kernel='poly', C=17000, gamma=0.0075, degree=1 ), # Control SVR for Ca
svm.SVR( kernel='rbf', C=11000, gamma=0.0025, degree=1 ), # Control SVR for P
svm.SVR( kernel='rbf',C=5750, gamma=0, degree=1 ), # Control SVR for pH
svm.SVR( kernel='rbf', C=8250, gamma=0, degree=1 ), # Control SVR for SOC
svm.SVR( kernel='rbf', C=20500, gamma=0, degree=1 ), # Control SVR for Sand
]
errors = []
for i in range( len(targets) ):
target = targets[ i ]
if( target == 'Ca' ):
continue
if( target == 'P' ):
continue
if( target == 'pH' ):
continue
# if( target == 'SOC' ):
# continue
# if( target == 'Sand' ):
# continue
svmReg = ExploreRegressor.ExploreRegressor( train, target, ests[ i ], 0.25 )
scores = svmReg.reportCrossValidationError( cv=20 )
errors.append( np.sqrt( -scores.mean() ) )
print '%s, mean error %0.4f of errors %s' % ( dataId, np.mean( errors ), errors )
}], # Parameter tuning for Sand
]
# Regular est from 20 fold cv
controlEsts = [
linear_model.Ridge(alpha=0.06), # Control SVR for Ca
linear_model.Ridge(alpha=0.05), # Control SVR for P
linear_model.Ridge(alpha=0.0015), # Control SVR for pH
linear_model.Ridge(alpha=0.05), # Control SVR for SOC
linear_model.Ridge(alpha=0.015), # Control SVR for Sand
]
targetId = 'SOC'
print '%s control est' % (targetId)
control = ExploreRegressor.ExploreRegressor(train, targetId,
linear_model.Ridge(alpha=0.1),
0.25)
control.reportCrossValidationError(cv=100)
print '%s optimal est. Parmeters: %s' % (targetId, 'optimal')
optimal = ExploreRegressor.ExploreRegressor(train, targetId,
linear_model.Ridge(alpha=0.05),
0.25)
optimal.reportCrossValidationError(cv=100)
for i in range(len(targets)):
target = targets[i]
if (target == 'Ca'):
continue
# Load data from CSV file
df = pd.read_csv(
'/Users/carrillo/workspace/Kaggle/resources/AfSIS/trainingTransformed.csv',
header=0)
# Specify SOC regressor
targetId = 'P'
est = RandomForestRegressor(n_estimators=800,
max_depth=16,
min_samples_leaf=1,
n_jobs=10,
oob_score=True,
verbose=0)
regAll = ExploreRegressor.ExploreRegressor(df, targetId, est, 0.25)
trainMSE, testMSE = regAll.getErrors(iterations=20)
print 'All features'
print 'meanTrainingMSE\tsdTrainingMSE\tmeanTestMSE\tsdTestMSE'
print '%f\t%f\t%f\t%f' % (np.mean(trainMSE), np.std(trainMSE),
np.mean(testMSE), np.std(testMSE))
clv = regAll.gridSearch(verbose=0)
regNew = ExploreRegressor.ExploreRegressor(df, targetId, clv.best_estimator_,
0.25)
regNew.learn()
trainMSE, testMSE = regNew.getErrors(iterations=20)
print '%f\t%f\t%f\t%f\t%s' % (np.mean(trainMSE), np.std(trainMSE),
np.mean(testMSE), np.std(testMSE),
clv.best_params_)