def test_dataset_val(self):
config = LocalConfig()
dataset = Dataset().get_val()
for batch in dataset.take(1):
image = ((batch[0].numpy()[0, :, :, :] + 1) / 2 * 255).astype(
np.int32)
output = np.sum(image)
expected = np.array(8123856, dtype=np.int32)
self.assertAllCloseAccordingToType(expected, output)
self.assertEqual(config.ENCODER_INPUT_SHAPE, image.shape)
kp2d = ((batch[1].numpy()[0, :, :2] + 1) / 2 *
image.shape[:2]).astype(np.int32)
output = np.sum(kp2d)
expected = np.array(4040, dtype=np.int32)
self.assertAllCloseAccordingToType(expected, output)
self.assertEqual((config.NUM_KP2D, 2), kp2d.shape)
vis = batch[1].numpy()[0, :, 2].astype(np.int32)
output = np.sum(vis)
expected = np.array(17, dtype=np.int32)
self.assertAllCloseAccordingToType(expected, output)
self.assertEqual((config.NUM_KP2D, ), vis.shape)
kp3d = batch[2].numpy()[0, :, :]
output = np.sum(kp3d)
expected = np.array(56.9217948, dtype=np.float32)
self.assertAllCloseAccordingToType(expected, output)
self.assertEqual((config.NUM_KP3D, 3), kp3d.shape)
# check if has3d flag is included in correct shape
has3d = batch[3].numpy()[0]
self.assertEqual(tf.constant(1, tf.int64), has3d)
self.assertEqual((config.BATCH_SIZE, ), batch[3].shape)
def train(self):
# Place tensors on the CPU
with tf.device('/CPU:0'):
dataset = Dataset()
ds_train = dataset.get_train()
ds_smpl = dataset.get_smpl()
ds_val = dataset.get_val()
start = 1
if self.config.RESTORE_EPOCH:
start = self.config.RESTORE_EPOCH
for epoch in range(start, self.config.EPOCHS + 1):
start = time.time()
print('Start of Epoch {}'.format(epoch))
dataset_train = ExceptionHandlingIterator(
tf.data.Dataset.zip((ds_train, ds_smpl)))
total = int(self.config.NUM_TRAINING_SAMPLES /
self.config.BATCH_SIZE)
for image_data, theta in tqdm(dataset_train,
total=total,
position=0,
desc='training'):
images, kp2d, kp3d, has3d = image_data[0], image_data[
1], image_data[2], image_data[3]
self._train_step(images, kp2d, kp3d, has3d, theta)
self._log_train(epoch=epoch)
total = int(self.config.NUM_VALIDATION_SAMPLES /
self.config.BATCH_SIZE)
for image_data in tqdm(ds_val,
total=total,
position=0,
desc='validate'):
images, kp2d, kp3d, has3d = image_data[0], image_data[
1], image_data[2], image_data[3]
self._val_step(images, kp2d, kp3d, has3d)
self._log_val(epoch=epoch)
print('Time taken for epoch {} is {} sec\n'.format(
epoch,
time.time() - start))
# saving (checkpoint) the model every 5 epochs
if epoch % 5 == 0:
print('saving checkpoint\n')
self.checkpoint_manager.save(epoch)
self.summary_writer.flush()
self.checkpoint_manager.save(self.config.EPOCHS + 1)
def test(self):
"""Run evaluation of the model
Specify LOG_DIR to point to the saved checkpoint directory
"""
if self.restore_check is None:
raise RuntimeError(
'restore did not succeed, pleas check if you set config.LOG_DIR correctly'
)
self.restore_check.assert_existing_objects_matched(
).assert_nontrivial_match()
# Place tensors on the CPU
with tf.device('/CPU:0'):
dataset = Dataset()
ds_test = dataset.get_test()
start = time.time()
print('Start of Testing')
mpjpe, mpjpe_aligned, sequences = [], [], []
total = int(self.config.NUM_TEST_SAMPLES / self.config.BATCH_SIZE)
for image_data in tqdm(ds_test,
total=total,
position=0,
desc='testing'):
image, kp3d, sequence = image_data[0], image_data[1], image_data[2]
kp3d_mpjpe, kp3d_mpjpe_aligned = self._test_step(image, kp3d)
mpjpe.append(kp3d_mpjpe)
mpjpe_aligned.append(kp3d_mpjpe_aligned)
sequences.append(sequence)
print('Time taken for testing {} sec\n'.format(time.time() - start))
def convert(tensor, num=None):
if num is None:
num = self.config.NUM_KP3D
return tf.squeeze(tf.reshape(tf.stack(tensor), [-1, num]))
mpjpe, mpjpe_aligned, sequences = convert(mpjpe), convert(
mpjpe_aligned), convert(sequences, 1)
result_dict = {
"kp3d_mpjpe": mpjpe,
"kp3d_mpjpe_aligned": mpjpe_aligned,
"seq": sequences,
}
return result_dict
def __init__(self):
self.is_data_posted = False
self.log = Log() # log file to test and debug (writes debug messages)
self.log_data = LogData() # sensor data log (if no internet available)
self.api = SamsApi() # https://sams.science.itf.llu.lv/ Data Warehouse Plugin to send the data
self.config = Config() # Configurations (/config/config.ini)
self.config_data = self.config.get_config_data()
self.repost_seconds = int(self.config_data['INTERVAL']['repost_seconds'])
self.app_wait_time = int(self.config_data['INTERVAL']['app_wait_seconds'])
self.dataset_taken = False
self.dataset_taken_counter = 0
self.data = Dataset() # collect all the data from sensors
self.dataset = ""
def __init__(self):
self.token_handler = TokenHandler()
self.app_config = ApplicationConfig()
self.mic = MicrophoneHelper()
self.dataset = Dataset()
self.color_print = Color()
self.sensors = []
if self.app_config.local_config.is_dht22:
self.sensors.append("dht22")
if self.app_config.local_config.is_ds18b20:
self.sensors.append("ds18b20")
if self.app_config.local_config.is_scale:
self.sensors.append("scale")
def test_dataset_smpl(self):
config = LocalConfig()
dataset = Dataset().get_smpl()
for batch in dataset.take(1):
shape = (config.BATCH_SIZE * config.ITERATIONS, (config.NUM_POSE_PARAMS + config.NUM_SHAPE_PARAMS))
self.assertEqual(shape, batch.shape)
pose = batch[0].numpy()[:config.NUM_POSE_PARAMS:]
mean = tf.reduce_mean(pose)
expected = np.array(0.0411809, dtype=np.float32)
self.assertAllCloseAccordingToType(expected, mean)
shape = batch[0].numpy()[-config.NUM_SHAPE_PARAMS:]
mean = tf.reduce_mean(shape)
expected = np.array(0.12554605, dtype=np.float32)
self.assertAllCloseAccordingToType(expected, mean)
def test_dataset_test(self):
config = LocalConfig()
dataset = Dataset().get_test()
for batch in dataset.take(1):
image = ((batch[0].numpy()[0, :, :, :] + 1) / 2 * 255).astype(np.int32)
output = np.sum(image)
expected = np.array(10050903, dtype=np.int32)
self.assertAllCloseAccordingToType(expected, output)
self.assertEqual(config.ENCODER_INPUT_SHAPE, image.shape)
kp3d = batch[1].numpy()[0, :, :]
output = np.sum(kp3d)
expected = np.array(38780.2031, dtype=np.float32)
self.assertAllCloseAccordingToType(expected, output)
self.assertEqual((config.NUM_KP3D, 3), kp3d.shape)
# check if sequence flag is included in correct shape
sequence = batch[2].numpy()[0].decode("utf-8")
self.assertEqual('TS1', sequence)
def __init__(self):
self.dataset_helper = DatasetLogHelper(
) # saving and sends the dataset
self.dataset = Dataset() # dataset for take sensor data
self.app_config = ApplicationConfig(
) # configuration data (on- and offline)
self.wifi_helper = WifiHelper(
) # gets signal strength for debug purpose
self.attempts = 0
self.dwh_api = DataApi()
self.token_handler = TokenHandler()
self.error_helper = ErrorHelper()
self.failed_sensor = ""
self.handle_online_status()
self.checker = SelfChecker()
# send status:
try:
send_log(
f'Start Application: {self.app_config.local_config.version}',
"debug")
send_log(f'Config Name: {self.app_config.local_config.group}',
"debug")
send_log(
f'Signal Strength: {self.wifi_helper.get_signal_strength()}',
"debug")
if self.current_volt():
send_log(f'Voltage: {self.current_volt()}', "debug")
set_timezone(self.app_config.local_config.timezone)
for file, status in self.checker.check_files().items():
send_log(f"created file: {file}.", "warning")
for failed_sensor in self.error_helper.get_sensors_with_errors():
send_log(
f'Please check {str(failed_sensor)} and reset all errors to reactivate the sensor.',
"warning")
except Exception as e:
print(e)
def test_dataset_train(self):
config = LocalConfig()
dataset = Dataset().get_train()
for batch in dataset.take(1):
image = ((batch[0].numpy()[0, :, :, :] + 1) / 2 * 255).astype(
np.int32)
output = np.sum(image)
expected = np.array(6991299, dtype=np.int32)
self.assertAllCloseAccordingToType(
expected, output) # this can sometimes fail with output=66
self.assertEqual(config.BATCH_SIZE, batch[0].shape[0])
self.assertEqual(config.ENCODER_INPUT_SHAPE, batch[0].shape[1:])
kp2d = ((batch[1].numpy()[0, :, :2] + 1) / 2 *
image.shape[:2]).astype(np.int32)
output = np.sum(kp2d)
expected = np.array(3818, dtype=np.int32)
self.assertAllCloseAccordingToType(expected, output)
self.assertEqual((config.BATCH_SIZE, config.NUM_KP2D, 3),
batch[1].shape)
vis = batch[1].numpy()[0, :, 2].astype(np.int32)
output = np.sum(vis)
expected = np.array(17, dtype=np.int32)
self.assertAllCloseAccordingToType(expected, output)
kp3d = batch[2].numpy()[0, :, :]
output = np.sum(kp3d)
expected = np.array(4.11272e-06, dtype=np.float32)
self.assertAllCloseAccordingToType(expected, output)
self.assertEqual((config.BATCH_SIZE, config.NUM_KP3D, 3),
batch[2].shape)
# check if has3d flag is included in correct shape
has3d = batch[3].numpy()[0]
self.assertEqual(tf.constant(1, tf.int64), has3d)
self.assertEqual((config.BATCH_SIZE, ), batch[3].shape)
from mpl_toolkits.mplot3d import Axes3D # noqa: F401 unused import
class DastasetConfig(LocalConfig):
# DATA_DIR = join('/', 'data', 'ssd1', 'russales', 'new_records')
# DATASETS = ['coco'] #['lsp', 'lsp_ext', 'mpii', 'coco', 'mpii_3d', 'h36m']
# SMPL_DATASETS = ['cmu', 'joint_lim']
TRANS_MAX = 20
# class Config is implemented as singleton, inizialize subclass first!
config = DastasetConfig()
import tensorflow as tf
# Place tensors on the CPU
with tf.device('/CPU:0'):
dataset = Dataset()
ds_train = dataset.get_train()
ds_smpl = dataset.get_smpl()
ds_val = dataset.get_val()
import matplotlib.pyplot as plt
for images, kp2d, kp3d, has3d in ds_train.take(1):
fig = plt.figure(figsize=(9.6, 5.4))
image_orig = tf.image.decode_jpeg(images[0], channels=3)
image_orig = image_orig.numpy()
kp2d = kp2d[0].numpy()
ax0 = fig.add_subplot(111)
image_2d = draw_2d_on_image(image_orig, kp2d[:, :2], vis=kp2d[:, 2])
ax0.imshow(image_2d)
def test(self, return_kps=False):
"""Run evaluation of the model
Specify LOG_DIR to point to the saved checkpoint directory
Args:
return_kps: set to return keypoints - default = False
"""
if self.restore_check is None:
raise RuntimeError(
'restore did not succeed, pleas check if you set config.LOG_DIR correctly'
)
if self.config.INITIALIZE_CUSTOM_REGRESSOR:
self.restore_check.assert_nontrivial_match()
else:
self.restore_check.assert_existing_objects_matched(
).assert_nontrivial_match()
# Place tensors on the CPU
with tf.device('/CPU:0'):
dataset = Dataset()
ds_test = dataset.get_test()
start = time.time()
print('Start of Testing')
mpjpe, mpjpe_aligned, sequences, kps3d_pred, kps3d_real = [], [], [], [], []
total = int(self.config.NUM_TEST_SAMPLES / self.config.BATCH_SIZE)
for image_data in tqdm(ds_test,
total=total,
position=0,
desc='testing'):
image, kp3d, sequence = image_data[0], image_data[1], image_data[2]
kp3d_mpjpe, kp3d_mpjpe_aligned, predict_kp3d, real_kp3d = self._test_step(
image, kp3d, return_kps=return_kps)
if return_kps:
kps3d_pred.append(predict_kp3d)
kps3d_real.append(real_kp3d)
mpjpe.append(kp3d_mpjpe)
mpjpe_aligned.append(kp3d_mpjpe_aligned)
sequences.append(sequence)
print('Time taken for testing {} sec\n'.format(time.time() - start))
def convert(tensor, num=None, is_kp=False):
if num is None:
num = self.config.NUM_KP3D
if is_kp:
return tf.squeeze(tf.reshape(tf.stack(tensor), [-1, num, 3]))
return tf.squeeze(tf.reshape(tf.stack(tensor), [-1, num]))
mpjpe, mpjpe_aligned, sequences = convert(mpjpe), convert(
mpjpe_aligned), convert(sequences, 1)
result_dict = {
"kp3d_mpjpe": mpjpe,
"kp3d_mpjpe_aligned": mpjpe_aligned,
"seq": sequences,
}
if return_kps:
kps3d_pred, kps3d_real = convert(kps3d_pred,
is_kp=True), convert(kps3d_real,
is_kp=True)
result_dict.update({
'kps3d_pred': kps3d_pred,
'kps3d_real': kps3d_real
})
return result_dict
import numpy as np
if __name__ == '__main__':
text = '花菇历来被国人作为延年益寿的补品，是香菇中的上品，含有丰富的营养价值，可帮助调节人体新陈代谢，助消化'
text = '冬天到了，吃什么比较好'
text = '天气冷了吃什么比较好'
text = '天气冷了'
text = '小孩'
text = '美容'
text = '老公'
conf = configparser.ConfigParser()
conf.read("./conf.ini")
ds = Dataset(conf)
# # prepro
# ds.segment_data('./data/baidu_recipe.json',
# './data/baidu_recipe_segment.json',
# ["title", "illustration", "practice", "materials"],
# True)
# ds.ngram_data('./data/baidu_recipe_segment.json',
# './data/baidu_recipe_ngram.json',
# ["title", "illustration", "practice", "materials"])
#
# ds.segment_data('./data/recipe_label_data.json',
# './data/recipe_segment.json',
# ["title", "illustration", "practice", "materials"],
# True)