def main():
if len(sys.argv) == 4:
dynamic_path, static_path, out_path = sys.argv[1:]
begin_train(dynamic_path, static_path, out_path)
if len(sys.argv) == 3:
base, n_jobs = sys.argv[1:]
n_jobs = int(n_jobs)
static_path = os.path.join(os.path.join(base, "static"),
"adjective_phase_set")
dynamic_path = os.path.join(os.path.join(base, "dynamic"),
"adjective_phase_set")
static_features = load_adjective_phase(static_path)
dynamic_features = load_adjective_phase(dynamic_path)
p = Parallel(n_jobs=n_jobs, verbose=0)
p(
delayed(begin_train_phase)(base, adjective, phase, static_features,
dynamic_features)
for adjective, phase in itertools.product(adjectives, phases))
else:
print "Usage:"
print "%s path adjective phase n_jobs" % sys.argv[0]
print "%s path adjective n_jobs" % sys.argv[0]
print "%s path n_jobs" % sys.argv[0]
def __init__(self):
rospy.loginfo("Loading the mkl adjective classifiers")
# Create paths to files
DIR = roslib.packages.get_pkg_dir("hadjective_mkl_classifier",
required=True) + "/classifiers/"
all_classifiers_file = os.path.join(DIR, "all_trained_classifiers.pkl")
hmm_chains_file = os.path.join(DIR, "loaded_chains.pkl")
dynamic_train_file = os.path.join(DIR, "adjective_phase_set_dynamic")
static_train_file = os.path.join(DIR, "adjective_phase_set_static")
# Load pickle files containing classifiers and chains
self.all_classifiers = cPickle.load(open(all_classifiers_file))
self.hmm_chains = cPickle.load(open(hmm_chains_file))
# Load all training features for later use in kernel building
self.training_dynamic = utilities.load_adjective_phase(
dynamic_train_file)
self.training_static = utilities.load_adjective_phase(
static_train_file)
# Setup subscriber and publisher nodes
rospy.Subscriber("new_hadjective_motion_pickle", String, self.callback)
self.adjectives_pub = rospy.Publisher("/hadjective_speech", AdjList)
rospy.loginfo("All MKL classifiers loaded")
def compute_probability_density_kernel(base_directory, classifiers, adjective,
motion):
"""
Compute the probability density of positive and
negative adjective per motion
"""
# Load static and dynamic
dynamic_features = utilities.load_adjective_phase(
'/media/data_storage/vchu/all_classifiers/icra2014/dynamic/adjective_phase_set/'
)
static_features = utilities.load_adjective_phase(
'/media/data_storage/vchu/all_classifiers/icra2014/static/adjective_phase_set/'
)
dynamic_train = dynamic_features[adjective][motion]['train']['features']
dynamic_train_scaler = preprocessing.StandardScaler().fit(dynamic_train)
dynamic_train_scaled_X = dynamic_train_scaler.transform(dynamic_train)
dynamic_train_kernel = linear_kernel(dynamic_train_scaled_X, -2)
#dynamic_train_kernel = standardize(dynamic_train_kernel)
#Static Train
static_train = static_features[adjective][motion]['train']['features']
static_train_scaler = preprocessing.StandardScaler().fit(static_train)
static_train_scaled_X = static_train_scaler.transform(static_train)
static_train_kernel = linear_kernel(static_train_scaled_X, -2)
# Load alphas
classifier = []
for mkl in classifiers:
if (mkl['adjective'] == adjective) and (mkl['phase'] == motion):
classifier = mkl
alpha = classifier['alpha']
# Pull out features for adjective/motion
#all_features = load_adjective_phase(base_directory)
#adj_motion_feat = all_features[adjective][motion]['train']
#features = adj_motion_feat['features']
#selected_components = compute_kernel(features)
train_X = (alpha) * static_train_kernel + (1 -
alpha) * dynamic_train_kernel
labels = dynamic_features[adjective][motion]['train']['labels']
selected_components = compute_kernel(train_X)
# Now split them according to positive and negative
#labels = adj_motion_feat['labels']
positive_components = selected_components[labels == 1]
negative_components = selected_components[labels == 0]
return (positive_components, negative_components)
def compute_probability_density_kernel(base_directory, classifiers, adjective, motion):
"""
Compute the probability density of positive and
negative adjective per motion
"""
# Load static and dynamic
dynamic_features = utilities.load_adjective_phase('/media/data_storage/vchu/all_classifiers/icra2014/dynamic/adjective_phase_set/')
static_features = utilities.load_adjective_phase('/media/data_storage/vchu/all_classifiers/icra2014/static/adjective_phase_set/')
dynamic_train = dynamic_features[adjective][motion]['train']['features']
dynamic_train_scaler = preprocessing.StandardScaler().fit(dynamic_train)
dynamic_train_scaled_X = dynamic_train_scaler.transform(dynamic_train)
dynamic_train_kernel = linear_kernel(dynamic_train_scaled_X, -2)
#dynamic_train_kernel = standardize(dynamic_train_kernel)
#Static Train
static_train = static_features[adjective][motion]['train']['features']
static_train_scaler = preprocessing.StandardScaler().fit(static_train)
static_train_scaled_X = static_train_scaler.transform(static_train)
static_train_kernel = linear_kernel(static_train_scaled_X, -2)
# Load alphas
classifier = []
for mkl in classifiers:
if (mkl['adjective']== adjective) and (mkl['phase'] == motion):
classifier = mkl
alpha = classifier['alpha']
# Pull out features for adjective/motion
#all_features = load_adjective_phase(base_directory)
#adj_motion_feat = all_features[adjective][motion]['train']
#features = adj_motion_feat['features']
#selected_components = compute_kernel(features)
train_X = (alpha)*static_train_kernel + (1-alpha)*dynamic_train_kernel
labels = dynamic_features[adjective][motion]['train']['labels']
selected_components = compute_kernel(train_X)
# Now split them according to positive and negative
#labels = adj_motion_feat['labels']
positive_components = selected_components[labels == 1]
negative_components = selected_components[labels == 0]
return (positive_components, negative_components)
def __init__(self):
rospy.loginfo("Loading the mkl adjective classifiers")
# Create paths to files
DIR = roslib.packages.get_pkg_dir("hadjective_mkl_classifier", required=True) + "/classifiers/"
all_classifiers_file = os.path.join(DIR, "all_trained_classifiers.pkl")
hmm_chains_file = os.path.join(DIR, "loaded_chains.pkl")
dynamic_train_file = os.path.join(DIR,"adjective_phase_set_dynamic")
static_train_file = os.path.join(DIR, "adjective_phase_set_static")
# Load pickle files containing classifiers and chains
self.all_classifiers = cPickle.load(open(all_classifiers_file))
self.hmm_chains = cPickle.load(open(hmm_chains_file))
# Load all training features for later use in kernel building
self.training_dynamic = utilities.load_adjective_phase(dynamic_train_file)
self.training_static = utilities.load_adjective_phase(static_train_file)
# Setup subscriber and publisher nodes
rospy.Subscriber("new_hadjective_motion_pickle", String, self.callback)
self.adjectives_pub = rospy.Publisher("/hadjective_speech", AdjList)
rospy.loginfo("All MKL classifiers loaded")
def main():
if len(sys.argv) == 4:
dynamic_path, static_path, out_path = sys.argv[1:]
begin_train(dynamic_path, static_path, out_path)
if len(sys.argv) == 3:
base, n_jobs = sys.argv[1:]
n_jobs = int(n_jobs)
static_path = os.path.join(os.path.join(base, "static"),"adjective_phase_set")
dynamic_path = os.path.join(os.path.join(base, "dynamic"), "adjective_phase_set")
static_features = load_adjective_phase(static_path)
dynamic_features = load_adjective_phase(dynamic_path)
p = Parallel(n_jobs=n_jobs,verbose=0)
p(delayed(begin_train_phase)(base, adjective, phase, static_features,dynamic_features)
for adjective, phase in itertools.product(adjectives,
phases))
else:
print "Usage:"
print "%s path adjective phase n_jobs" % sys.argv[0]
print "%s path adjective n_jobs" % sys.argv[0]
print "%s path n_jobs" % sys.argv[0]
def test_adjective(classifier, adjective_report, store_values):
adjective = classifier['adjective']
phase = classifier['phase']
dynamic_features = utilities.load_adjective_phase('/media/data_storage/vchu/all_classifiers/icra2014/dynamic/adjective_phase_set/')
static_features = utilities.load_adjective_phase('/media/data_storage/vchu/all_classifiers/icra2014/static/adjective_phase_set/')
true_positives = 0.0
true_negatives = 0.0
false_positives = 0.0
false_negatives = 0.0
false_positive_list = []
false_negative_list = []
true_positive_list = []
true_negative_list = []
print '\n \nTesting Adjective: %s and phase %s' % (adjective, phase)
#Pull out test features/labels
dynamic_train = dynamic_features[adjective][phase]['train']['features']
dynamic_train_scaler = preprocessing.StandardScaler().fit(dynamic_train)
dynamic_train_scaled_X = dynamic_train_scaler.transform(dynamic_train)
dynamic_train_kernel = linear_kernel(dynamic_train_scaled_X, -2)
#dynamic_train_kernel = standardize(dynamic_train_kernel)
#Static Train
static_train = static_features[adjective][phase]['train']['features']
static_train_scaler = preprocessing.StandardScaler().fit(static_train)
static_train_scaled_X = static_train_scaler.transform(static_train)
static_train_kernel = linear_kernel(static_train_scaled_X, -2)
dynamic_test = dynamic_features[adjective][phase]['test']['features']
dynamic_test_scaled_X = classifier['dynamic_scaler'].transform(dynamic_test)
dynamic_kernel = linear_kernel_test(dynamic_test_scaled_X, dynamic_train_scaled_X, -2)
#dynamic_kernel = (dynamic_kernel - classifier['dynamic_kernel_mean']) / classifier['dynamic_kernel_std']
static_test = static_features[adjective][phase]['test']['features']
static_test_scaled_X = classifier['static_scaler'].transform(static_test)
static_kernel = linear_kernel_test(static_test_scaled_X, static_train_scaled_X, -2)
alpha = classifier['alpha']
test_X = (alpha)*static_kernel + (1-alpha)*dynamic_kernel
test_Y = dynamic_features[adjective][phase]['test']['labels']
object_ids = dynamic_features[adjective][phase]['test']['object_ids']
object_names = dynamic_features[adjective][phase]['test']['object_names']
# Pull out the classifier and merge features
clf = classifier['classifier']
# Predict the labels!
output = clf.predict(test_X)
# Determine if the true label and classifier prediction match
for val in xrange(len(test_Y)):
true_label = test_Y[val]
predict_label = output[val]
if true_label == 1:
if predict_label == 1:
true_positives += 1.0
true_positive_list.append(object_names[val])
else:
false_negatives += 1.0
false_negative_list.append(object_names[val])
else: # label is 0
if predict_label == 1:
false_positives += 1.0
false_positive_list.append(object_names[val])
else:
true_negatives += 1.0
true_negative_list.append(object_names[val])
# Compute statistics for the adjective
try:
precision = true_positives / (true_positives + false_positives)
recall = true_positives / (true_positives + false_negatives)
except ZeroDivisionError: # The case when none are found
precision = 0
recall = 0
try:
f1 = 2.0 * precision*recall / (precision + recall)
except ZeroDivisionError:
f1 = 0
store_values[adjective+'_'+phase] = (f1,precision, recall)
print "%d False Positive Objects are: %s \n" % (false_positives, sorted(false_positive_list))
print "%d False Negative Objects are: %s \n" % (false_negatives, sorted(false_negative_list))
print "%d True Positive Objects are: %s\n" % (true_positives, sorted(true_positive_list))
print "%d True Negative Objects are: %s\n" % (true_negatives, sorted(true_negative_list))
return (precision, recall, f1)
def begin_train(dynamic_path, static_path, out_path):
dynamic_features = utilities.load_adjective_phase(dynamic_path)
static_features = utilities.load_adjective_phase(static_path)
adjective_list = utilities.adjectives
for adjective in adjective_list:
# File name
dataset_file_name = "_".join(("trained", adjective))+".pkl"
newpath = os.path.join(out_path, "trained_adjectives")
path_name = os.path.join(newpath, dataset_file_name)
if os.path.exists(path_name):
print "File %s already exists, skipping it." % path_name
continue
overall_best_score = 0.0
dataset = defaultdict()
dataset['classifier'] = None
dataset['training_score'] = overall_best_score
dynamic_train = utilities.get_all_train_test_features(adjective, dynamic_features)
y_labels = dynamic_train[1]
object_ids = dynamic_train[2]
dynamic_scaler = preprocessing.StandardScaler().fit(dynamic_train[0])
dynamic_X = dynamic_scaler.transform(dynamic_train[0])
dynamic_kernel = linear_kernel(dynamic_X, -2)
#dynamic_kernel = standardize(dynamic_kernel)
static_train = utilities.get_all_train_test_features(adjective, static_features)
static_scaler = preprocessing.StandardScaler().fit(static_train[0])
static_X = static_scaler.transform(static_train[0])
static_kernel = linear_kernel(static_X, -2)
#static_kernel = standardize(static_kernel)
alpha_range = [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]
#alpha_range = [0.5]
for alpha in alpha_range:
print "Beginning %s, alpha %1.1f at %s" % (adjective,alpha, time.asctime())
combined_kernel = (alpha)*static_kernel + (1-alpha)*dynamic_kernel;
trained_clf, best_score = gram_grid_search(combined_kernel, labels=y_labels, object_ids=object_ids, refined_range = adjective)
print "F1: %1.5f at %s" % (best_score, time.asctime())
if best_score > overall_best_score:
overall_best_score = best_score
dataset['classifier'] = trained_clf
dataset['training_score'] = best_score
dataset['alpha'] = alpha
dataset['dynamic_features'] = dynamic_features[adjective]
dataset['static_features'] = static_features[adjective]
dataset['adjective'] = adjective
dataset['dynamic_scaler'] = dynamic_scaler
dataset['dynamic_train_scaled'] = dynamic_X;
dataset['dynamic_kernel_mean'] = np.mean(dynamic_kernel)
dataset['dynamic_kernel_std'] = np.std(dynamic_kernel)
dataset['static_scaler'] = static_scaler
dataset['static_train_scaled'] = static_X;
dataset['static_kernel_mean'] = np.mean(static_kernel)
dataset['static_kernel_std'] = np.std(static_kernel)
print "Saving trained_classifier"
# Save the results in the folder
with open(path_name, "w") as f:
print "Saving file: ", path_name
cPickle.dump(dataset, f, protocol=cPickle.HIGHEST_PROTOCOL)
def test_adjective(classifier, adjective_report, store_values):
adjective = classifier['adjective']
phase = classifier['phase']
dynamic_features = utilities.load_adjective_phase(
'/media/data_storage/vchu/all_classifiers/icra2014/dynamic/adjective_phase_set/'
)
static_features = utilities.load_adjective_phase(
'/media/data_storage/vchu/all_classifiers/icra2014/static/adjective_phase_set/'
)
true_positives = 0.0
true_negatives = 0.0
false_positives = 0.0
false_negatives = 0.0
false_positive_list = []
false_negative_list = []
true_positive_list = []
true_negative_list = []
print '\n \nTesting Adjective: %s and phase %s' % (adjective, phase)
#Pull out test features/labels
dynamic_train = dynamic_features[adjective][phase]['train']['features']
dynamic_train_scaler = preprocessing.StandardScaler().fit(dynamic_train)
dynamic_train_scaled_X = dynamic_train_scaler.transform(dynamic_train)
dynamic_train_kernel = linear_kernel(dynamic_train_scaled_X, -2)
#dynamic_train_kernel = standardize(dynamic_train_kernel)
#Static Train
static_train = static_features[adjective][phase]['train']['features']
static_train_scaler = preprocessing.StandardScaler().fit(static_train)
static_train_scaled_X = static_train_scaler.transform(static_train)
static_train_kernel = linear_kernel(static_train_scaled_X, -2)
dynamic_test = dynamic_features[adjective][phase]['test']['features']
dynamic_test_scaled_X = classifier['dynamic_scaler'].transform(
dynamic_test)
dynamic_kernel = linear_kernel_test(dynamic_test_scaled_X,
dynamic_train_scaled_X, -2)
#dynamic_kernel = (dynamic_kernel - classifier['dynamic_kernel_mean']) / classifier['dynamic_kernel_std']
static_test = static_features[adjective][phase]['test']['features']
static_test_scaled_X = classifier['static_scaler'].transform(static_test)
static_kernel = linear_kernel_test(static_test_scaled_X,
static_train_scaled_X, -2)
alpha = classifier['alpha']
test_X = (alpha) * static_kernel + (1 - alpha) * dynamic_kernel
test_Y = dynamic_features[adjective][phase]['test']['labels']
object_ids = dynamic_features[adjective][phase]['test']['object_ids']
object_names = dynamic_features[adjective][phase]['test']['object_names']
# Pull out the classifier and merge features
clf = classifier['classifier']
# Predict the labels!
output = clf.predict(test_X)
# Determine if the true label and classifier prediction match
for val in xrange(len(test_Y)):
true_label = test_Y[val]
predict_label = output[val]
if true_label == 1:
if predict_label == 1:
true_positives += 1.0
true_positive_list.append(object_names[val])
else:
false_negatives += 1.0
false_negative_list.append(object_names[val])
else: # label is 0
if predict_label == 1:
false_positives += 1.0
false_positive_list.append(object_names[val])
else:
true_negatives += 1.0
true_negative_list.append(object_names[val])
# Compute statistics for the adjective
try:
precision = true_positives / (true_positives + false_positives)
recall = true_positives / (true_positives + false_negatives)
except ZeroDivisionError: # The case when none are found
precision = 0
recall = 0
try:
f1 = 2.0 * precision * recall / (precision + recall)
except ZeroDivisionError:
f1 = 0
store_values[adjective + '_' + phase] = (f1, precision, recall)
print "%d False Positive Objects are: %s \n" % (
false_positives, sorted(false_positive_list))
print "%d False Negative Objects are: %s \n" % (
false_negatives, sorted(false_negative_list))
print "%d True Positive Objects are: %s\n" % (true_positives,
sorted(true_positive_list))
print "%d True Negative Objects are: %s\n" % (true_negatives,
sorted(true_negative_list))
return (precision, recall, f1)
def begin_train(dynamic_path, static_path, out_path):
dynamic_features = utilities.load_adjective_phase(dynamic_path)
static_features = utilities.load_adjective_phase(static_path)
adjective_list = utilities.adjectives
for adjective in adjective_list:
# File name
dataset_file_name = "_".join(("trained", adjective)) + ".pkl"
newpath = os.path.join(out_path, "trained_adjectives")
path_name = os.path.join(newpath, dataset_file_name)
if os.path.exists(path_name):
print "File %s already exists, skipping it." % path_name
continue
overall_best_score = 0.0
dataset = defaultdict()
dataset['classifier'] = None
dataset['training_score'] = overall_best_score
dynamic_train = utilities.get_all_train_test_features(
adjective, dynamic_features)
y_labels = dynamic_train[1]
object_ids = dynamic_train[2]
dynamic_scaler = preprocessing.StandardScaler().fit(dynamic_train[0])
dynamic_X = dynamic_scaler.transform(dynamic_train[0])
dynamic_kernel = linear_kernel(dynamic_X, -2)
#dynamic_kernel = standardize(dynamic_kernel)
static_train = utilities.get_all_train_test_features(
adjective, static_features)
static_scaler = preprocessing.StandardScaler().fit(static_train[0])
static_X = static_scaler.transform(static_train[0])
static_kernel = linear_kernel(static_X, -2)
#static_kernel = standardize(static_kernel)
alpha_range = [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]
#alpha_range = [0.5]
for alpha in alpha_range:
print "Beginning %s, alpha %1.1f at %s" % (adjective, alpha,
time.asctime())
combined_kernel = (alpha) * static_kernel + (
1 - alpha) * dynamic_kernel
trained_clf, best_score = gram_grid_search(combined_kernel,
labels=y_labels,
object_ids=object_ids,
refined_range=adjective)
print "F1: %1.5f at %s" % (best_score, time.asctime())
if best_score > overall_best_score:
overall_best_score = best_score
dataset['classifier'] = trained_clf
dataset['training_score'] = best_score
dataset['alpha'] = alpha
dataset['dynamic_features'] = dynamic_features[adjective]
dataset['static_features'] = static_features[adjective]
dataset['adjective'] = adjective
dataset['dynamic_scaler'] = dynamic_scaler
dataset['dynamic_train_scaled'] = dynamic_X
dataset['dynamic_kernel_mean'] = np.mean(dynamic_kernel)
dataset['dynamic_kernel_std'] = np.std(dynamic_kernel)
dataset['static_scaler'] = static_scaler
dataset['static_train_scaled'] = static_X
dataset['static_kernel_mean'] = np.mean(static_kernel)
dataset['static_kernel_std'] = np.std(static_kernel)
print "Saving trained_classifier"
# Save the results in the folder
with open(path_name, "w") as f:
print "Saving file: ", path_name
cPickle.dump(dataset, f, protocol=cPickle.HIGHEST_PROTOCOL)
def test_adjective(classifier, adjective_report ):
true_positives = 0.0
true_negatives = 0.0
false_positives = 0.0
false_negatives = 0.0
false_positive_list = []
false_negative_list = []
true_positive_list = []
true_negative_list = []
adjective = classifier['adjective']
dynamic_features = utilities.load_adjective_phase('/home/imcmahon/Desktop/mkl/dynamic/adjective_phase_set')
static_features = utilities.load_adjective_phase('/home/imcmahon/Desktop/mkl/static/adjective_phase_set')
#dynamic_features = utilities.load_adjective_phase(dynamic_path)
#static_features = utilities.load_adjective_phase(static_path)
#import pdb; pdb.set_trace()
#Dynamic Train
dynamic_train = utilities.get_all_train_test_features(adjective, dynamic_features, train=True)
dynamic_train_scaler = preprocessing.StandardScaler().fit(dynamic_train[0])
dynamic_train_scaled_X = dynamic_train_scaler.transform(dynamic_train[0])
dynamic_train_kernel = linear_kernel(dynamic_train_scaled_X, -2)
#dynamic_train_kernel = standardize(dynamic_train_kernel)
#Static Train
static_train = utilities.get_all_train_test_features(adjective, static_features, train=True)
static_train_scaler = preprocessing.StandardScaler().fit(static_train[0])
static_train_scaled_X = static_train_scaler.transform(static_train[0])
static_train_kernel = linear_kernel(static_train_scaled_X, -2)
#static_train_kernel = standardize(static_train_kernel)
#Recompute the GRAM matrix
#alpha = classifier['alpha'];
#train_X = (alpha)*static_train_kernel + (1-alpha)*dynamic_train_kernel
#import pdb; pdb.set_trace()
dynamic_test = utilities.get_all_train_test_features(adjective, dynamic_features, train=False)
dynamic_test_scaled_X = classifier['dynamic_scaler'].transform(dynamic_test[0])
dynamic_kernel = linear_kernel_test(dynamic_test_scaled_X, dynamic_train_scaled_X, -2)
#dynamic_kernel = (dynamic_kernel - classifier['dynamic_kernel_mean']) / classifier['dynamic_kernel_std']
static_test = utilities.get_all_train_test_features(adjective, static_features, train=False)
static_test_scaled_X = classifier['static_scaler'].transform(static_test[0])
static_kernel = linear_kernel_test(static_test_scaled_X, static_train_scaled_X, -2)
#static_kernel = (static_kernel - classifier['static_kernel_mean']) / classifier['static_kernel_std']
alpha = classifier['alpha'];
test_X = (alpha)*static_kernel + (1-alpha)*dynamic_kernel
print '\n \nTesting Adjective: %s' % classifier['adjective']
#Pull out test features/labels
#for phase in phases:
#import pdb; pdb.set_trace()
#test_set = classifier['test']
#test_X.append(test_set['features'])
#test_X.append(test_set['features'])
test_Y = dynamic_test[1]
object_ids = dynamic_test[2]
#object_names = test_set['object_names']
# Pull out the classifier and merge features
#test_X = np.concatenate(test_X, axis=1)
#import pdb; pdb.set_trace()
clf = classifier['classifier']
c_dict[classifier['adjective']] = clf.C
#import pdb; pdb.set_trace()
print clf
# Predict the labels!
#if 'scaler' in classifier:
# if type(classifier['scaler']) == preprocessing.Scaler:
# test_X = classifier['scaler'].transform(test_X)
#import pdb; pdb.set_trace()
output = clf.predict(test_X)
# Determine if the true label and classifier prediction match
for val in xrange(len(test_Y)):
true_label = test_Y[val]
predict_label = output[val]
if true_label == 1:
if predict_label == 1:
true_positives += 1.0
#true_positive_list.append(object_names[val])
else:
false_negatives += 1.0
#false_negative_list.append(object_names[val])
else: # label is 0
if predict_label == 1:
false_positives += 1.0
#false_positive_list.append(object_names[val])
else:
true_negatives += 1.0
#true_negative_list.append(object_names[val])
# Compute statistics for the adjective
try:
precision = true_positives / (true_positives + false_positives)
recall = true_positives / (true_positives + false_negatives)
except ZeroDivisionError: # The case when none are found
precision = 0
recall = 0
try:
f1 = 2.0 * precision*recall / (precision + recall)
except ZeroDivisionError:
f1 = 0
print "Precision: %f, Recall: %f, F1: %f \n" % (precision, recall, f1)
print "Alpha = %1.1f" % alpha
adjective_report.write("%s, %1.1f, %f, %f, %f\n" % (classifier['adjective'], alpha, precision, recall, f1))
print "%d False Positive Objects\n" % false_positives
print "%d False Negative Objects\n" % false_negatives
print "%d True Positive Objects\n" % true_positives
print "%d True Negative Objects\n" % true_negatives
return (alpha, precision, recall, f1)
def test_adjective(classifier, adjective_report):
true_positives = 0.0
true_negatives = 0.0
false_positives = 0.0
false_negatives = 0.0
false_positive_list = []
false_negative_list = []
true_positive_list = []
true_negative_list = []
adjective = classifier['adjective']
dynamic_features = utilities.load_adjective_phase(
'/home/imcmahon/Desktop/mkl/dynamic/adjective_phase_set')
static_features = utilities.load_adjective_phase(
'/home/imcmahon/Desktop/mkl/static/adjective_phase_set')
#dynamic_features = utilities.load_adjective_phase(dynamic_path)
#static_features = utilities.load_adjective_phase(static_path)
#import pdb; pdb.set_trace()
#Dynamic Train
dynamic_train = utilities.get_all_train_test_features(adjective,
dynamic_features,
train=True)
dynamic_train_scaler = preprocessing.StandardScaler().fit(dynamic_train[0])
dynamic_train_scaled_X = dynamic_train_scaler.transform(dynamic_train[0])
dynamic_train_kernel = linear_kernel(dynamic_train_scaled_X, -2)
#dynamic_train_kernel = standardize(dynamic_train_kernel)
#Static Train
static_train = utilities.get_all_train_test_features(adjective,
static_features,
train=True)
static_train_scaler = preprocessing.StandardScaler().fit(static_train[0])
static_train_scaled_X = static_train_scaler.transform(static_train[0])
static_train_kernel = linear_kernel(static_train_scaled_X, -2)
#static_train_kernel = standardize(static_train_kernel)
#Recompute the GRAM matrix
#alpha = classifier['alpha'];
#train_X = (alpha)*static_train_kernel + (1-alpha)*dynamic_train_kernel
#import pdb; pdb.set_trace()
dynamic_test = utilities.get_all_train_test_features(adjective,
dynamic_features,
train=False)
dynamic_test_scaled_X = classifier['dynamic_scaler'].transform(
dynamic_test[0])
dynamic_kernel = linear_kernel_test(dynamic_test_scaled_X,
dynamic_train_scaled_X, -2)
#dynamic_kernel = (dynamic_kernel - classifier['dynamic_kernel_mean']) / classifier['dynamic_kernel_std']
static_test = utilities.get_all_train_test_features(adjective,
static_features,
train=False)
static_test_scaled_X = classifier['static_scaler'].transform(
static_test[0])
static_kernel = linear_kernel_test(static_test_scaled_X,
static_train_scaled_X, -2)
#static_kernel = (static_kernel - classifier['static_kernel_mean']) / classifier['static_kernel_std']
alpha = classifier['alpha']
test_X = (alpha) * static_kernel + (1 - alpha) * dynamic_kernel
print '\n \nTesting Adjective: %s' % classifier['adjective']
#Pull out test features/labels
#for phase in phases:
#import pdb; pdb.set_trace()
#test_set = classifier['test']
#test_X.append(test_set['features'])
#test_X.append(test_set['features'])
test_Y = dynamic_test[1]
object_ids = dynamic_test[2]
#object_names = test_set['object_names']
# Pull out the classifier and merge features
#test_X = np.concatenate(test_X, axis=1)
#import pdb; pdb.set_trace()
clf = classifier['classifier']
c_dict[classifier['adjective']] = clf.C
#import pdb; pdb.set_trace()
print clf
# Predict the labels!
#if 'scaler' in classifier:
# if type(classifier['scaler']) == preprocessing.Scaler:
# test_X = classifier['scaler'].transform(test_X)
#import pdb; pdb.set_trace()
output = clf.predict(test_X)
# Determine if the true label and classifier prediction match
for val in xrange(len(test_Y)):
true_label = test_Y[val]
predict_label = output[val]
if true_label == 1:
if predict_label == 1:
true_positives += 1.0
#true_positive_list.append(object_names[val])
else:
false_negatives += 1.0
#false_negative_list.append(object_names[val])
else: # label is 0
if predict_label == 1:
false_positives += 1.0
#false_positive_list.append(object_names[val])
else:
true_negatives += 1.0
#true_negative_list.append(object_names[val])
# Compute statistics for the adjective
try:
precision = true_positives / (true_positives + false_positives)
recall = true_positives / (true_positives + false_negatives)
except ZeroDivisionError: # The case when none are found
precision = 0
recall = 0
try:
f1 = 2.0 * precision * recall / (precision + recall)
except ZeroDivisionError:
f1 = 0
print "Precision: %f, Recall: %f, F1: %f \n" % (precision, recall, f1)
print "Alpha = %1.1f" % alpha
adjective_report.write(
"%s, %1.1f, %f, %f, %f\n" %
(classifier['adjective'], alpha, precision, recall, f1))
print "%d False Positive Objects\n" % false_positives
print "%d False Negative Objects\n" % false_negatives
print "%d True Positive Objects\n" % true_positives
print "%d True Negative Objects\n" % true_negatives
return (alpha, precision, recall, f1)