def test_joints_to_gaussian_heatmap():
image_size = (256, 320)
stride = 2
num_joints = 17
input_joints = np.array([[1, 1, 1], [2, 2, 1], [3, 3, 1], [4, 4, 1],
[5, 5, 1], [6, 6, 1], [7, 7, 1], [8, 8, 1],
[9, 9, 1], [10, 10, 1], [11, 11, 1], [12, 12, 1],
[13, 13, 1], [14, 14, 1], [15, 15, 1],
[16, 16, 1], [17, 17, 0]])
pre_process = JointsToGaussianHeatmap(image_size,
num_joints=num_joints,
stride=stride,
sigma=2,
max_value=10)
heatmap = pre_process(joints=input_joints)["heatmap"]
# It is hard to test semantic correctness of a gaussian heatmap manually.
# That part will be tested jointly with FormatJoints() in test_post_processor.py.
assert isinstance(heatmap, np.ndarray)
assert heatmap.shape[0] == image_size[0] // stride
assert heatmap.shape[1] == image_size[1] // stride
assert heatmap.shape[2] == 17
assert np.max(heatmap) == 10
def test_ytfaces_facial_landmarks_detection():
batch_size = 1
image_size = [256, 320]
stride = 2
pre_processor = Sequence([
ResizeWithJoints(image_size=image_size),
JointsToGaussianHeatmap(image_size=image_size,
num_joints=68,
stride=stride)
])
dataset = YoutubeFacialLandmarks(subset="train",
batch_size=batch_size,
pre_processor=pre_processor)
dataset = DatasetIterator(dataset)
for _ in range(5):
images, labels = dataset.feed()
assert isinstance(images, np.ndarray)
assert images.shape[0] == batch_size
assert images.shape[1] == image_size[0]
assert images.shape[2] == image_size[1]
assert images.shape[3] == 3
assert isinstance(labels, np.ndarray)
assert labels.shape[0] == batch_size
assert labels.shape[1] == image_size[0] // stride
assert labels.shape[2] == image_size[1] // stride
assert labels.shape[3] == 68
dataset = YoutubeFacialLandmarks(subset="validation",
batch_size=batch_size,
pre_processor=pre_processor)
dataset = DatasetIterator(dataset)
for _ in range(5):
images, labels = dataset.feed()
assert isinstance(images, np.ndarray)
assert images.shape[0] == batch_size
assert images.shape[1] == image_size[0]
assert images.shape[2] == image_size[1]
assert images.shape[3] == 3
assert isinstance(labels, np.ndarray)
assert labels.shape[0] == batch_size
assert labels.shape[1] == image_size[0] // stride
assert labels.shape[2] == image_size[1] // stride
assert labels.shape[3] == 68
def test_gaussian_heatmap_to_joints():
from lmnet.pre_processor import JointsToGaussianHeatmap
image_size = (160, 160)
num_dimensions = 2
stride = 1
confidence_threshold = 0.5
num_joints = 17
input_joints = np.array([[1, 1, 1],
[2, 2, 1],
[3, 3, 1],
[4, 4, 1],
[5, 5, 1],
[6, 6, 1],
[7, 7, 1],
[8, 8, 1],
[9, 9, 1],
[10, 10, 1],
[11, 11, 1],
[12, 12, 1],
[13, 13, 1],
[14, 14, 1],
[15, 15, 1],
[16, 16, 1],
[17, 17, 0]])
pre_process = JointsToGaussianHeatmap(image_size, num_joints=num_joints,
stride=1, sigma=3)
post_process = GaussianHeatmapToJoints(num_dimensions=num_dimensions,
stride=stride,
confidence_threshold=confidence_threshold)
heatmap = pre_process(joints=input_joints)["heatmap"]
heatmap = np.expand_dims(heatmap, axis=0)
output_joints = post_process(heatmap)["outputs"][0]
for i in range(num_joints):
if input_joints[i, 2] == 1:
assert np.allclose(output_joints[i], input_joints[i], atol=1e-4, rtol=1e-4)
def test_training():
"""Test only no error raised."""
config = EasyDict()
config.NETWORK_CLASS = LmSinglePoseV1Quantize
config.DATASET_CLASS = MscocoSinglePersonKeypoints
config.IS_DEBUG = False
config.IMAGE_SIZE = [160, 160]
config.BATCH_SIZE = 2
config.TEST_STEPS = 1
config.MAX_STEPS = 2
config.SAVE_CHECKPOINT_STEPS = 1
config.KEEP_CHECKPOINT_MAX = 5
config.SUMMARISE_STEPS = 1
config.IS_PRETRAIN = False
config.IS_DISTRIBUTION = False
config.TASK = Tasks.KEYPOINT_DETECTION
# network model config
config.NETWORK = EasyDict()
config.NETWORK.OPTIMIZER_CLASS = tf.train.AdamOptimizer
config.NETWORK.OPTIMIZER_KWARGS = {"learning_rate": 0.001}
config.NETWORK.IMAGE_SIZE = config.IMAGE_SIZE
config.NETWORK.BATCH_SIZE = config.BATCH_SIZE
config.NETWORK.ACTIVATION_QUANTIZER = linear_mid_tread_half_quantizer
config.NETWORK.ACTIVATION_QUANTIZER_KWARGS = {'bit': 2, 'max_value': 2.0}
config.NETWORK.WEIGHT_QUANTIZER = binary_channel_wise_mean_scaling_quantizer
config.NETWORK.WEIGHT_QUANTIZER_KWARGS = {}
# daasegt config
config.DATASET = EasyDict()
config.DATASET.PRE_PROCESSOR = Sequence([
ResizeWithJoints(image_size=config.IMAGE_SIZE),
JointsToGaussianHeatmap(image_size=config.IMAGE_SIZE, stride=2),
DivideBy255()
])
config.DATASET.BATCH_SIZE = config.BATCH_SIZE
environment.init("test_lm_single_pose_v1")
prepare_dirs(recreate=True)
start_training(config)
IS_PRETRAIN = False
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
# for debug
# BATCH_SIZE = 2
# SUMMARISE_STEPS = 1
# IS_DEBUG = True
# stride of output heatmap. the smaller, the slower.
STRIDE = 8
PRE_PROCESSOR = Sequence([
ResizeWithJoints(image_size=IMAGE_SIZE),
JointsToGaussianHeatmap(image_size=IMAGE_SIZE, stride=STRIDE, sigma=2),
DivideBy255()
])
POST_PROCESSOR = Sequence([
GaussianHeatmapToJoints(num_dimensions=2,
stride=STRIDE,
confidence_threshold=0.1)
])
step_per_epoch = 149813 // BATCH_SIZE
NETWORK = EasyDict()
NETWORK.OPTIMIZER_CLASS = tf.train.AdamOptimizer
NETWORK.OPTIMIZER_KWARGS = {}
NETWORK.LEARNING_RATE_FUNC = tf.train.piecewise_constant
NETWORK.LEARNING_RATE_KWARGS = {