def train(self, data_folder, objects='all',gmm=None, feature_extraction='msc_fv_3', mc_feature=True):
objects_to_train = []
object_directories = np.array(glob.glob(data_folder + '/*'))
if objects == 'all':
for obj_dir in object_directories:
print (obj_dir)
object_name = obj_dir.split('/')[-1]
objects_to_train.append(str(object_name))
else:
objects_to_train = objects
print ("Extracting features....")
n = 0
feature_pool = np.empty([0, 0])
label_pool = []
for obj in objects_to_train:
files = np.array(glob.glob(data_folder + '/' + obj + '/*'))
print (obj, ": ", len(files))
for f in files:
points = self.parse_pcd(f, enable_color=True)
#print ("points",points.shape)
if feature_extraction == 'mc':
features = calculate_feature_vector(points, enable_color=True)
else:
features = calculate_maxrdd_fv_features(points, gmm, feature_extraction, mean_circle=mc_feature)
if n < 1:
feature_pool = np.array(features)
label_pool = [obj]
else:
feature_pool = np.vstack([feature_pool, features])
label_pool.append(obj)
n += 1
mean = np.mean(feature_pool, axis=0)
std = np.std(feature_pool, axis=0)
#feature_pool -= mean
#feature_pool /= std
print (feature_pool.shape)
feature_pool = np.nan_to_num(feature_pool)
print ("Classifying....", feature_pool.shape)
label_encoder = sklearn.preprocessing.LabelEncoder()
label_encoder.fit(label_pool)
encoded_labels = label_encoder.transform(label_pool)[:, np.newaxis]
encoded_labels = np.squeeze(encoded_labels.T)
classifier = sklearn.ensemble.RandomForestClassifier(n_estimators=10)
classifier.fit(feature_pool, encoded_labels)
return SVMObjectClassifier(classifier, label_encoder, mean, std)
def train_using_pgz(self, data_folder, gmm, feature_extraction='msc_fv_3', objects='all', datasize=2):
data = None
labels = None
count = 0
for i in range(datasize):
print ("Processing: {}data{}.pgz".format(data_folder,i))
dataset = self.load_compressed_pickle_file('{}data{}.pgz'.format(data_folder,i))
if i == 0:
data = dataset['data']
labels = dataset['labels']
else:
data = np.concatenate((data, dataset['data']))
labels = np.concatenate((labels, dataset['labels']))
print ("Computing features for datasize: {}".format(data.shape))
print ("Label shape: ", labels.shape)
n = 0
feature_pool = np.empty([0, 0])
label_pool = []
for i,points in enumerate(data):
if feature_extraction == 'mc':
features = calculate_feature_vector(points)
else:
features = calculate_maxrdd_fv_features(points, gmm, feature_extraction, mean_circle=False)
if i < 1:
#if not np.isnan(features.any()):
feature_pool = np.array(features)
label_pool = [labels[i]]
else:
#if not np.isnan(features.any()):
feature_pool = np.vstack([feature_pool, features])
label_pool.append(labels[i])
print ("Done features computation")
# Whitenin, and remember to whiten the input for inference
mean = np.mean(feature_pool, axis=0)
std = np.std(feature_pool, axis=0)
#feature_pool -= mean
#feature_pool /= std
feature_pool = np.nan_to_num(feature_pool)
print ("Classifying....", feature_pool.shape)
label_encoder = sklearn.preprocessing.LabelEncoder()
label_encoder.fit(label_pool)
encoded_labels = label_encoder.transform(label_pool)[:, np.newaxis]
encoded_labels = np.squeeze(encoded_labels.T)
classifier = sklearn.ensemble.RandomForestClassifier(n_estimators=10)
classifier.fit(feature_pool, encoded_labels)
return SVMObjectClassifier(classifier, label_encoder, mean, std)
def evaluate_cls_using_pgz(data_folder,
gmm,
classifier,
feature_extraction='mc',
idx=[5, 6],
objects='all',
color=False):
data = None
labels = None
count = 0
for i in range(idx[0], idx[1]):
print("Loading:...", '{}data{}.pgz'.format(data_folder, i))
dataset = load_compressed_pickle_file('{}data{}.pgz'.format(
data_folder, i))
print(dataset['data'].shape)
if count == 0:
data = dataset['data']
labels = dataset['labels']
else:
data = np.concatenate((data, dataset['data']))
labels = np.concatenate((labels, dataset['labels']))
count += 1
print(labels.shape)
seen = np.zeros((len(labels)))
correct = np.zeros((len(labels)))
for i, points in enumerate(data):
#print ("ITERA",i)
#files = np.array(glob.glob(self.data_folder + '/' + obj + '/*'))
if feature_extraction == 'mc':
features = calculate_feature_vector(points, False)
else:
features = calculate_maxrdd_fv_features(points, gmm,
feature_extraction)
#features = np.nan_to_num(features)
features = np.reshape(features, (1, -1))
pred, probability = classifier.classify(features)
seen[i] += 1
if pred == labels[i]:
correct[i] += 1
print("True: {}, Pred: {}".format(pred, labels[i]))
print("Seen: ", seen)
print("Correct: ", correct)
print("Avg class acc: ", correct / seen)
print("Accuracy: ", np.mean(correct, axis=0))
return np.mean(correct, axis=0)
def evaluate_maxrdd_fv_cls(data_folder,
classifier,
objects='all',
color=False,
gmm=None,
feature_extraction='msc_fv_3',
mc_feature=True):
objects_to_eval = []
object_directories = np.array(glob.glob(data_folder + '/*'))
for obj_dir in object_directories:
print(obj_dir)
object_name = obj_dir.split('/')[-1]
objects_to_eval.append(str(object_name))
seen = np.zeros((len(objects_to_eval)))
correct = np.zeros((len(objects_to_eval)))
print(objects_to_eval)
for i, obj in enumerate(objects_to_eval):
files = np.array(glob.glob(data_folder + '/' + obj + '/*'))
for f in files:
points = parse_pcd(f, enable_color=True)
if feature_extraction == 'mc':
features = calculate_feature_vector(points, enable_color=True)
else:
features = calculate_maxrdd_fv_features(points,
gmm,
feature_extraction,
mean_circle=mc_feature)
#xyzhsv = parse_pcd(f)
#features = calculate_feature_vector(xyzhsv, False)
features = np.reshape(features, (1, -1))
print("Shape: ", features.shape)
label, probability = classifier.classify(features)
seen[i] += 1
if label == str(obj):
correct[i] += 1
print("True: {}, Pred: {}".format(str(obj), label))
print("Seen: ", seen)
print("Correct: ", correct)
print("Avg class acc: ", correct / seen)
print("Accuracy: ", np.sum(correct / seen) / len(objects_to_eval))
return np.mean(np.sum(correct / seen) / len(objects_to_eval))
def evaluate_cls(data_folder, classifier, objects='all', color=False):
objects_to_test = os.listdir(os.path.abspath(data_folder))
print("Training classifer for objects: ", objects_to_test)
seen = np.zeros((len(objects_to_test)))
correct = np.zeros((len(objects_to_test)))
for i, obj in enumerate(objects_to_test):
class_folder = os.path.join(data_folder, obj, "")
files = glob.glob(class_folder + '**/*.pcd', recursive=True)
for f in files:
#xyzhsv = convert_to_xyzhsv(f)
xyzhsv = parse_pcd(f)
features = calculate_feature_vector(xyzhsv, False)
features = np.reshape(features, (1, -1))
label, probability = classifier.classify(features)
seen[i] += 1
if label == str(obj):
correct[i] += 1
print("Seen: ", seen)
print("Correct: ", correct)
print("Avg class acc: ", correct / seen)
print("Accuracy: ", np.mean(correct, axis=0))