def load_data(session, model_type, feature_type):
"""
Load raw data
if feature_type == FeatureType.MANUAL data is augmented with the magnitude
model_type determines the shape of the data
"""
# loads the data from the given session: 2/3 as training and 1/3 as testing
if MEASUREMENT_PROTOCOL_TYPE == MeasurementProtocol.SAME_DAY:
print('Training data SAME SESSION')
X_train, y_train = \
load_recordings_from_session(session, 1, 154, 1, 5, model_type, feature_type)
print('Testing data SAME SESSION')
X_test, y_test = \
load_recordings_from_session(session, 1, 154, 5, 7, model_type, feature_type)
return X_train, y_train, X_test, y_test
# training data: session_1, testing data: session_2
if MEASUREMENT_PROTOCOL_TYPE == MeasurementProtocol.CROSS_DAY:
print('Training data CROSS SESSION')
X_train, y_train = \
load_recordings_from_session('session_1', 1, 154, 1, 7, model_type, feature_type)
print('Testing data CROSS SESSION')
X_test, y_test = \
load_recordings_from_session('session_2', 1, 154, 1, 7, model_type, feature_type)
return X_train, y_train, X_test, y_test
def extract_manual_features():
modeltype = AUTOENCODER_MODEL_TYPE.LSTM
featuretype = FeatureType.MANUAL
# Session_0
X0, y0 = load_recordings_from_session('session_0', 1, 23, 1, 7, modeltype,
featuretype)
F0 = feature_extraction(X0)
lines = F0.shape[0]
cols = F0.shape[1]
if (CYCLE == True):
csv_file = open(FEAT_DIR + '/' + "session_0_handcrafted_cycles.csv",
mode='w')
else:
csv_file = open(FEAT_DIR + '/' + "session_0_handcrafted_frames.csv",
mode='w')
for i in range(0, lines):
for j in range(0, cols):
# print(F0[i,j])
csv_file.write('%f,' % (F0[i, j]))
csv_file.write('%s\n' % y0[i][0])
# Session_1
X0, y0 = load_recordings_from_session('session_1', 1, 154, 1, 7, modeltype,
featuretype)
F0 = feature_extraction(X0)
lines = F0.shape[0]
cols = F0.shape[1]
if (CYCLE == True):
csv_file = open(FEAT_DIR + '/' + "session_1_handcrafted_cycles.csv",
mode='w')
else:
csv_file = open(FEAT_DIR + '/' + "session_1_handcrafted_frames.csv",
mode='w')
for i in range(0, lines):
for j in range(0, cols):
# print(F0[i,j])
csv_file.write('%f,' % (F0[i, j]))
csv_file.write('%s\n' % y0[i][0])
# Session_2
X0, y0 = load_recordings_from_session('session_2', 1, 154, 1, 7, modeltype,
featuretype)
F0 = feature_extraction(X0)
lines = F0.shape[0]
cols = F0.shape[1]
if (CYCLE == True):
csv_file = open(FEAT_DIR + '/' + "session_2_handcrafted_cycles.csv",
mode='w')
else:
csv_file = open(FEAT_DIR + '/' + "session_2_handcrafted_frames.csv",
mode='w')
for i in range(0, lines):
for j in range(0, cols):
# print(F0[i,j])
csv_file.write('%f,' % (F0[i, j]))
csv_file.write('%s\n' % y0[i][0])
def train_test_autoencoder(num_epochs=10):
# ZJU: session_2
X1, y1 = load_recordings_from_session('session_2', 1, 154, 1, 7,
AUTOENCODER_MODEL_TYPE.CONV1D,
FeatureType.AUTOMATIC)
# IDNet
# X1, y1 = create_idnet_training_dataset(AUTOENCODER_MODEL_TYPE.CONV1D)
# Training from scratch
model_name = 'IDNetModel.h5'
encoder, model = train_FCN_autoencoder(num_epochs,
X1,
y1,
'relu',
update=True,
model_name=model_name)
# print('Saved model: ' + model_name)
# model.save(TRAINED_MODELS_DIR + '/' + model_name)
# encoder_name = 'Encoder_' + model_name
# print('Saved encoder: ' + encoder_name)
# encoder.save(TRAINED_MODELS_DIR + '/' + encoder_name)
X_train, y_train = extract_features(encoder, 'session_1',
AUTOENCODER_MODEL_TYPE.CONV1D, 1, 154,
1, 5)
X_test, y_test = extract_features(encoder, 'session_1',
AUTOENCODER_MODEL_TYPE.CONV1D, 1, 154, 5,
7)
# performs evaluation (train and test a classifier)
evaluation(X_train, y_train, X_test, y_test)
def extract_and_save_features(encoder, modeltype, start_user, stop_user):
mysessions = ['session_1', 'session_2']
feature_type = FeatureType.AUTOMATIC
for session in mysessions:
print('Extract features from ' + session)
data, ydata = load_recordings_from_session(session, start_user,
stop_user, 1, 7, modeltype,
feature_type)
print('data shape: ' + str(data.shape))
# if (modeltype==AUTOENCODER_MODEL_TYPE.CONV1D):
# data = np.array(data)[:, :, np.newaxis]
# Extract features
encoded_frames = encoder.predict(data)
# Normalize data
scaled_data = preprocessing.scale(encoded_frames)
num_features = encoded_frames.shape[1]
# Concatenate features(encoded_frames) with labels (ydata)
df1 = pd.DataFrame(data=scaled_data)
df2 = pd.DataFrame(data=ydata)
df2[0] = df2[0].apply(lambda x: x.replace('subj_', 'u'))
df3 = pd.concat([df1, df2], axis=1)
# Save data into a CSV file
df3.to_csv('./' + FEAT_DIR + '/' + session + ".csv",
header=False,
index=False)
def load_data(session, model_type, feature_type):
# loads the data from the given session: 2/3 as training and 1/3 as testing
if MEASUREMENT_PROTOCOL_TYPE == MEASUREMENT_PROTOCOL.SAME_DAY:
print('Training data SAME SESSION')
X_train, y_train = load_recordings_from_session(
session, 1, 154, 1, 5, model_type, feature_type)
print('Testing data SAME SESSION')
X_test, y_test = load_recordings_from_session(session, 1, 154, 5, 7,
model_type, feature_type)
return X_train, y_train, X_test, y_test
# training data: session_1, testing data: session_2
if MEASUREMENT_PROTOCOL_TYPE == MEASUREMENT_PROTOCOL.CROSS_DAY:
print('Training data CROSS SESSION')
X_train, y_train = load_recordings_from_session(
'session_1', 1, 154, 1, 7, model_type, feature_type)
print('Testing data CROSS SESSION')
X_test, y_test = load_recordings_from_session('session_2', 1, 154, 1,
7, model_type,
feature_type)
return X_train, y_train, X_test, y_test
def create_zju_training_dataset(augmentation, modeltype):
# modeltype=AUTOENCODER_MODEL_TYPE.DENSE
print('Create training dataset for autoencoder')
feature_type = FeatureType.AUTOMATIC
if (EVALUATION_DATA == EvaluationData.INDEPENDENT):
data1, ydata1 = load_recordings_from_session('session_1', 1, 103, 1, 7,
modeltype, feature_type)
data2, ydata2 = load_recordings_from_session('session_2', 1, 103, 1, 7,
modeltype, feature_type)
if (EVALUATION_DATA == EvaluationData.ALL):
data1, ydata1 = load_recordings_from_session('session_1', 1, 153, 1, 7,
modeltype, feature_type)
data2, ydata2 = load_recordings_from_session('session_2', 1, 153, 1, 7,
modeltype, feature_type)
if (EVALUATION_DATA == EvaluationData.MSTHESIS):
data1, ydata1 = load_recordings_from_session('session_1', 1, 153, 1, 4,
modeltype, feature_type)
data2, ydata2 = load_recordings_from_session('session_0', 1, 23, 1, 7,
modeltype, feature_type)
data = np.concatenate((data1, data2), axis=0)
ydata = np.concatenate((ydata1, ydata2), axis=0)
if augmentation == True:
data, ydata = data_augmentation(data, ydata)
print('ZJU GaitAcc dataset - Data shape: ' + str(data.shape) +
' augmentation: ' + str(augmentation))
return data, ydata
def create_zju_training_dataset(augmentation=False):
"""
Create training dataset for autoencoder using the ZJU-GaitAcc dataset
Users [1, 103)
Recordings [1, 7)
"""
modeltype = const.AutoencoderModelType.DENSE
print('Create training dataset for autoencoder')
feature_type = const.FeatureType.AUTOMATIC
data1, ydata1 = load_recordings_from_session('session_1', 1, 103, 1, 7,
modeltype, feature_type)
data2, ydata2 = load_recordings_from_session('session_2', 1, 103, 1, 7,
modeltype, feature_type)
data = np.concatenate((data1, data2), axis=0)
ydata = np.concatenate((ydata1, ydata2), axis=0)
if augmentation is True:
data, ydata = data_augmentation(data, ydata)
print('ZJU GaitAcc dataset - Data shape: ' + str(data.shape) +
' augmentation: ' + str(augmentation))
return data, ydata
def extract_features(encoder, session, modeltype, start_user, stop_user,
start_recording, stop_recording):
feature_type = FeatureType.AUTOMATIC
data, ydata = load_recordings_from_session(session, start_user, stop_user,
start_recording, stop_recording,
modeltype, feature_type)
# print('data shape: '+ str(data.shape))
# Extract features
encoded_frames = encoder.predict(data)
# Normalize data
print(encoded_frames.shape)
scaled_data = preprocessing.scale(encoded_frames)
#print('Scaling')
#print(scaled_data)
#print('features shape: ' + str(encoded_frames.shape))
return scaled_data, ydata
def extract_and_save_features(encoder, modeltype, start_user, stop_user):
"""
Automatic features using the autoencoder
ZJU-GaitAcc
Extract and save features for session_1 and session_2
encoder: the encoder part of a trained autoencoder
modeltype: used for data shape
users: [start_user, stop_user)
Output: session_1.csv, session_2.csv
"""
mysessions = ['session_1', 'session_2']
feature_type = const.FeatureType.AUTOMATIC
for session in mysessions:
print('Extract features from ' + session)
data, ydata = load_recordings_from_session(session, start_user,
stop_user, 1, 7, modeltype,
feature_type)
print('data shape: ' + str(data.shape))
if (modeltype == const.AutoencoderModelType.CONV1D):
data = np.array(data)[:, :, np.newaxis]
# Extract features
encoded_frames = encoder.predict(data)
# Normalize data
scaled_data = preprocessing.scale(encoded_frames)
num_features = encoded_frames.shape[1]
# Concatenate features (encoded_frames) with labels (ydata)
df1 = pd.DataFrame(data=scaled_data)
df2 = pd.DataFrame(data=ydata)
df2[0] = df2[0].apply(lambda x: x.replace('subj_', 'u'))
df3 = pd.concat([df1, df2], axis=1)
# Save data into a CSV file
df3.to_csv('./' + const.FEAT_DIR + '/' + session + ".csv",
header=False,
index=False)
def extract_handcrafted_features(session, start_user, stop_user, output_file):
# create $FEAT_DIR if does not exist
create_directory('./' + FEAT_DIR)
modeltype = AUTOENCODER_MODEL_TYPE.NONE
featuretype = FeatureType.MANUAL
X, y = load_recordings_from_session(session, start_user, stop_user, 1, 7,
modeltype, featuretype)
# print(X.shape)
# print(y.shape)
features = feature_extraction(X)
num_segments = features.shape[0]
num_features = features.shape[1]
csv_file = open(FEAT_DIR + '/' + output_file, mode='w+')
for i in range(0, num_segments):
for j in range(0, num_features):
csv_file.write('%f,' % (features[i, j]))
user = y[i, 0].replace("subj_", "u")
csv_file.write('%s\n' % user)
csv_file.close()
def load_session_data(session, model_type, feature_type):
X, y = load_recordings_from_session(session, 1, 154, 1, 7, model_type,
feature_type)
return X, y
def ZJU_feature_extraction(encoder_name, modeltype):
# loads the encoder part of the model - it is needed for feature extraction
encoder_name = TRAINED_MODELS_DIR + '/' + encoder_name
encoder = load_model(encoder_name)
print('Loaded model: ' + encoder_name)
feature_type = FeatureType.AUTOMATIC
# Extract features from session_1 and session_2
mysessions = ['session_1', 'session_2']
start_user = 1
stop_user = 154
for session in mysessions:
print('Extract features from ' + session)
data, ydata = load_recordings_from_session(session, start_user,
stop_user, 1, 7, modeltype,
feature_type)
print('data shape: ' + str(data.shape))
# if (modeltype==AUTOENCODER_MODEL_TYPE.CONV1D):
# data = np.array(data)[:, :, np.newaxis]
# Extract features
encoded_frames = encoder.predict(data)
# Normalize data
scaled_data = preprocessing.scale(encoded_frames)
num_features = encoded_frames.shape[1]
# Concatenate features(encoded_frames) with labels (ydata)
df1 = pd.DataFrame(data=scaled_data)
df2 = pd.DataFrame(data=ydata)
df2[0] = df2[0].apply(lambda x: x.replace('subj_', 'u'))
df3 = pd.concat([df1, df2], axis=1)
# Save data into a CSV file
df3.to_csv('./' + FEAT_DIR + '/' + session + ".csv",
header=False,
index=False)
print('Extract features from session_0')
session = 'session_0'
start_user = 1
stop_user = 23
data, ydata = load_recordings_from_session(session, start_user, stop_user,
1, 7, modeltype, feature_type)
print('data shape: ' + str(data.shape))
# if (modeltype==AUTOENCODER_MODEL_TYPE.CONV1D):
# data = np.array(data)[:, :, np.newaxis]
# Extract features
encoded_frames = encoder.predict(data)
# Normalize data
scaled_data = preprocessing.scale(encoded_frames)
num_features = encoded_frames.shape[1]
# Concatenate features(encoded_frames) with labels (ydata)
df1 = pd.DataFrame(data=scaled_data)
df2 = pd.DataFrame(data=ydata)
df2[0] = df2[0].apply(lambda x: x.replace('subj_', 'u'))
df3 = pd.concat([df1, df2], axis=1)
# Save data into a CSV file
df3.to_csv('./' + FEAT_DIR + '/' + session + ".csv",
header=False,
index=False)