def load_validation_set(
data_source: DataSources = DataSources.VALIDATION_1) -> [Data]:
root_folder, image_type = data_source.value
data: [Data] = []
with open('./%s/validation_set.csv' % root_folder, newline='') as csv_file:
validation_csv = csv.reader(csv_file, delimiter=',', quotechar='|')
for row in validation_csv:
if row[0]:
image = row[1]
rx, ry, rz, tx, ty, tz = np.array(row[2:8]).astype(np.float)
scale = tz
image_ = './%s/images/%s' % (root_folder, image)
pose = np.array([scale, rx, ry, rz, tx, ty])
with open(
'./%s/images/%s' %
(root_folder, image.replace('.png', '.pts'))) as pts_file:
rows = [rows.strip() for rows in pts_file]
rows = rows[rows.index('{') + 1:rows.index('}')]
landmarks_2d = np.array([row.split(' ')
for row in rows]).astype(np.float)
data.append(Data(image_, landmarks_2d, pose))
return data
def load_validation_dataset2(
data_source: DataSources = DataSources.VALIDATION_2,
recalc_pose=True) -> [Data]:
root_folder, image_type = data_source.value
log.info('load_validation_data2::%s:: started' % data_source.name)
data = []
from GenerateTrainingSet import get_face_model, calc_projection, estimate_pose_from_landmarks
face_model = get_face_model()
with open('%s/valid_set2.csv' % root_folder, newline='') as csv_file:
validation_csv = csv.reader(csv_file, delimiter=',', quotechar='|')
for row in validation_csv:
if row[0]:
image = row[1]
landmarks = np.array([
np.array(re.sub('[\[\]]', '',
re.sub('[ ]+', ' ',
landmark)).strip().split(' '),
dtype=np.float32) for landmark in row[8:]
])
if recalc_pose:
proj_mat_t_land = calc_projection(landmarks,
face_model.model_TD,
face_model)
pose = estimate_pose_from_landmarks(
proj_mat_t_land, face_model)
else:
pose = np.array(row[2:8], dtype=np.float32)
data.append(
Data('%s/%s' % (root_folder, image), landmarks, pose))
return data
def load_test_set(data_source: DataSources = DataSources.TEST_SET) -> [Data]:
root_folder, image_type = data_source.value
data: [Data] = []
for root, dirs, files in os.walk("%s" % root_folder, topdown=False):
for name in files:
if name.endswith('.%s' % image_type):
data.append(Data('%s/%s' % (root_folder, name), None, None))
return data
def main():
### PDI ~ C4.5
# train_x, train_y, test_x, test_y = Data().fetchFromH5('train_catvnoncat.h5', 'test_catvnoncat.h5')
# criterion = "entropy"
# training_data = dt.preProcess(train_x, train_y, normalize=True, gray=True)
# decision_tree = dt.growTree(training_data, criterion)
# dt.plotDiagram(decision_tree, extension=criterion)
# dt.prune(decision_tree, criterion, 0.75)
# dt.plotDiagram(decision_tree, extension=(criterion+"_pruned"))
# test_data = dt.preProcess(test_x, normalize=True, gray=True)
# result_text = dt.classify(decision_tree, test_data, test_y)
# print("\n%s" % "\n".join(result_text))
# Data().saveVariable(name="decision_tree", extension=(criterion+"_classify_result"), value=result_text)
### CI ~ C4.5
train_x, train_y = Data().fetchFromPath('characters', 't0')
test_x, test_y = Data().fetchFromPath('characters', 't1')
criterion = "entropy"
training_data = dt.preProcess(train_x,
train_y,
norm=True,
gray=True,
seg=True)
decision_tree = dt.growTree(training_data, criterion)
dt.plotDiagram(decision_tree, extension=criterion)
# dt.prune(decision_tree, criterion, 0.5)
# dt.plotDiagram(decision_tree, extension=(criterion+"_pruned"))
test_data = dt.preProcess(test_x, norm=True, gray=True, seg=True)
result_text = dt.classify(decision_tree, test_data, test_y)
Data().saveVariable(name="decision_tree",
extension=(criterion + "_classify_result"),
value=result_text)
print("\n%s" % "\n".join(result_text))
def _300w_3d_parser(name: str,
root_folder: str,
image_type: str,
landmarks_fld_name='pt2d') -> Data:
image = '%s/%s' % (root_folder, name)
meta = loadmat('%s/%s' %
(root_folder, name.replace('.%s' % image_type, '.mat')))
rx, ry, rz, tx, ty, tz, scale = meta["Pose_Para"].reshape([-1]).astype(
np.float32) # pitch yaw roll
# convert to rotvec
r = R.from_euler('zxy', [rx, ry, rz], degrees=False)
rx, ry, rz = r.as_rotvec()
pose = np.array([rx, ry, rz, tx, ty, tz, scale])
landmarks_2d = (meta[landmarks_fld_name]).astype(np.float32).transpose()
return Data(image, landmarks_2d, pose)
def load_naive_augmented_dataset(data_source: DataSources, limit=-1) -> [Data]:
root_folder, image_type = data_source.value
log.info('load_naive_augmented_validation_set::%s' % data_source.name)
data: [Data] = []
meta_data = {}
for root, dirs, files in os.walk(root_folder, topdown=False):
for name in files:
if -1 < limit == len(data):
break
if name.endswith('_aug.%s' % image_type):
if name not in meta_data:
meta_file_name = name.split('--')
meta_file_name = '%s.meta' % meta_file_name[0]
with open('%s/%s' % (root_folder, meta_file_name)) as f:
meta_lines = f.readlines()
for meta_line in meta_lines:
file_name, pose, landmarks_2d_ = meta_line.split('|')
meta_data[file_name] = (np.array(
json.loads(pose.replace(' ', ','))),
np.array(
json.loads(
landmarks_2d_.replace(
' ', ' ').replace(
' ', ','))))
rx, ry, rz, tx, ty, tz = meta_data[name][0]
data.append(
Data('%s/%s' % (root_folder, name), meta_data[name][1],
np.array([rx, ry, rz, tx, ty, tz])))
return data
def ae_od(normal_data,
all_data,
ae_model,
EPISODE_MAX=10000,
BATCH_SIZE=64,
verbose=False):
[n_obj, n_f] = all_data.shape
train_data = Data(normal_data)
loss_list = np.zeros(200)
for episode in range(EPISODE_MAX):
batch_x = train_data.next_batch(BATCH_SIZE)
train_loss = ae_model.train_model(batch_x)
loss_list[episode % 200] = train_loss
avg = 0.
std = 0.
if episode % 200 == 0 and episode // 200 != 0:
std = np.std(loss_list)
avg = np.average(loss_list)
if std < 0.05 * avg or avg < 1e-5:
if verbose:
print(
' DeepAE:{}, episode: {}, loss: {:.4}, avg,std: {:.4}, {:.4}'
.format(batch_x.shape, episode, train_loss, avg, std))
break
loss_list = np.zeros(200)
if episode % 2000 == 0 and verbose:
print(
' DeepAE:{}, episode: {}, loss: {:.4}, avg,std: {:.4}, {:.4}'.
format(batch_x.shape, episode, train_loss, avg, std))
anomaly_scores = np.zeros([n_obj])
for i, obj in enumerate(all_data):
anomaly_scores[i] = ae_model.test_model(obj.reshape([1, n_f]))
return anomaly_scores