def svr_lin_objective(input_dict):
from sklearn.svm import SVR
from sklearn.metrics.pairwise import linear_kernel
from sklearn.externals import joblib
from multichannel import MultiChannelModel, multichannel_KFoldCV
import numpy as np
import csv, os, time
#load the dataset
dataset_root = '/home/luke/projects/THE_dataset' #directory where features/labels kept
train_path = os.path.join(dataset_root, 'train_set_wc3d.pkl')
test_path = os.path.join(dataset_root, 'test_set_wc3d.pkl')
X_train, y_train = joblib.load(train_path)
X_test, y_test = joblib.load(test_path)
#run the exp
C = input_dict['C']
kpt = {'kernel_func': linear_kernel, 'param_dict': {}}
model = SVR(kernel='precomputed', C=C)
mcm = MultiChannelModel(num_channels=6, model=model, kernel_param_tuple=kpt)
scores = multichannel_KFoldCV(mcm, X_train, y_train, n_folds=3, verbose=False)
loss = 1-np.mean(scores)
eval_time = time.time()
#logging
with open('svr_lin_log.csv','a') as f:
fc = csv.writer(f)
row = [loss, eval_time, C]
fc.writerow(row)
print loss, input_dict
return {'loss': loss, 'eval_time': eval_time}
def test_kfold(self):
#dummy data
X_channel = np.array([[0.1, 0.5], [0.5, 0.1], [0.1, 0.6],
[0.7, 0.1]]) #the folds are the same
X_channel = np.vstack((X_channel, X_channel))
X_train = (X_channel, X_channel + .01)
Y = np.array([1, 10, 1, 10])
Y = np.hstack((Y, Y))
#set up the model
model = SVR(kernel='precomputed', C=1)
mcm = MultiChannelModel(num_channels=2, model=model)
#run a CV
score_list = multichannel_KFoldCV(mcm,
X_train,
Y,
n_folds=2,
verbose=True)
#assertions
print score_list
self.assertTrue(len(score_list) == 2)
self.assertTrue(score_list[0] == score_list[1] > .5)
def test_mcm(self):
X_channel = np.array([[0.1, 0.5], [0.5, 0.1], [0.1, 0.6], [0.7, 0.1]])
X_train = (X_channel, X_channel)
X_test = (X_channel + .1, X_channel + .01)
Y = np.array([1, 10, 1, 10])
#set up the model
model = SVR(kernel='precomputed', C=1)
mcm = MultiChannelModel(num_channels=2, model=model)
#train
mcm.fit(X_train, Y)
#predict
Y_pred = mcm.predict(X_test)
#assertions
self.assertTrue(Y.shape == (4, ))
#scores
train_score = mcm.score(X_train, Y)
test_score = mcm.score(X_test, Y)
#assertions
print 'train_score = {0}, test_score = {1}'.format(
train_score, test_score)
self.assertTrue(isinstance(test_score, float))
self.assertTrue(isinstance(train_score, float))
self.assertTrue(test_score < train_score)
def objective(x):
from sklearn.svm import SVR
from sklearn.metrics.pairwise import linear_kernel
from sklearn.externals import joblib
from multichannel import MultiChannelModel, multichannel_KFoldCV, theano_rbf as rbf_kernel, theano_chi2 as chi2_kernel
import numpy as np
import csv, os, time
#output the time taken
t0 = time.time()
def create_kpt(num_channels, gammas):
kernel_param_list = []
for channel in xrange(num_channels):
if channel < 4:
kdict = {'kernel_func': chi2_kernel, 'param_dict': {'gamma': gammas[channel]}}
#elif channel == 5:
# kdict = {'kernel_func': rbf_kernel, 'param_dict': {'gamma': gammas[channel]}}
kernel_param_list.append(kdict)
kernel_param_tuple = tuple(kernel_param_list)
return kernel_param_tuple
#load the dataset
dataset_root = '/home/luke/projects/THE_dataset' #directory where features/labels kept
train_path = os.path.join(dataset_root, 'train_set_wc3d.pkl')
test_path = os.path.join(dataset_root, 'test_set_wc3d.pkl')
X_train, y_train = joblib.load(train_path)
X_test, y_test = joblib.load(test_path)
num_channels=5
X_train = (X_train[i] for i in xrange(num_channels-1))
X_test = (X_test[i] for i in xrange(num_channels-1))
#run the exp
gammas = [x['traj_gamma'], x['hog_gamma'], x['hof_gamma'],
x['mbhx_gamma'], x['mbhy_gamma']]
kpt = create_kpt(num_channels, gammas)
C = x['C']
model = SVR(kernel='precomputed', C=C)
mcm = MultiChannelModel(num_channels=num_channels, model=model, kernel_param_tuple=kpt)
scores = multichannel_KFoldCV(mcm, X_train, y_train, n_folds=3, verbose=False)
loss = 1-np.mean(scores)
eval_time = time.time()
#logging
with open('svr_nonlin_IDT_log.csv','a') as f:
fc = csv.writer(f)
row = [loss, eval_time,C] + gammas
fc.writerow(row)
print x, loss, 'time taken: {}'.format(time.time()-t0)
return {'loss': loss, 'eval_time': eval_time}
def objective(x):
from sklearn.linear_model import Ridge
from sklearn.metrics.pairwise import linear_kernel
from sklearn.externals import joblib
from multichannel import MultiChannelModel, multichannel_KFoldCV, theano_rbf as rbf_kernel, theano_chi2 as chi2_kernel
import numpy as np
import csv, os, time
#output the time taken
t0 = time.time()
def create_kpt(num_channels, gammas):
kernel_param_list = []
for channel in xrange(num_channels):
if channel < 4:
kdict = {
'kernel_func': chi2_kernel,
'param_dict': {
'gamma': gammas[channel]
}
}
elif channel == 5:
kdict = {
'kernel_func': rbf_kernel,
'param_dict': {
'gamma': gammas[channel]
}
}
kernel_param_list.append(kdict)
kernel_param_tuple = tuple(kernel_param_list)
return kernel_param_tuple
#load the dataset
dataset_root = '/home/luke/projects/THE_dataset' #directory where features/labels kept
train_path = os.path.join(dataset_root, 'train_set_wc3d.pkl')
test_path = os.path.join(dataset_root, 'test_set_wc3d.pkl')
X_train, y_train = joblib.load(train_path)
X_test, y_test = joblib.load(test_path)
#run the exp
#set the krnel gammas set from previous experiment
gammas = [0.06386, 0.117941, 0.060457, 0.180092, 3405.482, 0.79326]
kpt = create_kpt(6, gammas)
#set the alpha level
alpha = x['alpha']
#set the channel weights
cw = [
x['traj_cw'], x['hog_cw'], x['hof_cw'], x['mbhx_cw'], x['mbhy_cw'],
x['c3d_cw']
]
model = Ridge(alpha=alpha)
mcm = MultiChannelModel(num_channels=6,
model=model,
kernel_param_tuple=kpt,
channel_weights=cw)
scores = multichannel_KFoldCV(mcm,
X_train,
y_train,
n_folds=3,
verbose=False)
loss = 1 - np.mean(scores)
eval_time = time.time()
print 'params: {0}, loss: {1}, time taken: {2}'.format(
x, loss,
time.time() - t0)
return {'loss': loss, 'eval_time': eval_time}
def objective(x):
from sklearn.linear_model import Ridge
from sklearn.metrics.pairwise import linear_kernel
from sklearn.externals import joblib
from multichannel import MultiChannelModel, multichannel_KFoldCV, theano_rbf as rbf_kernel, theano_chi2 as chi2_kernel
import numpy as np
import csv, os, time
#output the time taken
t0 = time.time()
def create_kpt(num_channels, gammas):
kernel_param_list = []
for channel in xrange(num_channels):
if channel < 4:
kdict = {
'kernel_func': chi2_kernel,
'param_dict': {
'gamma': gammas[channel]
}
}
elif channel == 5:
kdict = {
'kernel_func': rbf_kernel,
'param_dict': {
'gamma': gammas[channel]
}
}
kernel_param_list.append(kdict)
kernel_param_tuple = tuple(kernel_param_list)
return kernel_param_tuple
#load the dataset
dataset_root = '/home/luke/projects/THE_dataset' #directory where features/labels kept
train_path = os.path.join(dataset_root, 'train_set_wc3d.pkl')
test_path = os.path.join(dataset_root, 'test_set_wc3d.pkl')
X_train, y_train = joblib.load(train_path)
X_test, y_test = joblib.load(test_path)
#run the exp
#set the krnel gammas
gammas = [
x['traj_gamma'], x['hog_gamma'], x['hof_gamma'], x['mbhx_gamma'],
x['mbhy_gamma'], x['c3d_gamma']
]
kpt = create_kpt(6, gammas)
#set the alpha level
alpha = x['alpha']
#set the channel weights
cw = [
x['traj_cw'], x['hog_cw'], x['hof_cw'], x['mbhx_cw'], x['mbhy_cw'],
x['c3d_cw']
]
model = Ridge(alpha=alpha)
mcm = MultiChannelModel(num_channels=6,
model=model,
kernel_param_tuple=kpt,
channel_weights=cw)
scores = multichannel_KFoldCV(mcm,
X_train,
y_train,
n_folds=3,
verbose=False)
loss = 1 - np.mean(scores)
eval_time = time.time()
#logging
with open('ridge_nonlin_cweights_log.csv', 'a') as f:
fc = csv.writer(f)
row = [loss, eval_time, alpha] + gammas + cw
fc.writerow(row)
print x, loss, 'time taken: {}'.format(time.time() - t0)
return {'loss': loss, 'eval_time': eval_time}