def test_lm_bisenet_post_process():
"""Verify LMBiSeNet.post_process() is the same as Bilinear and Softmax post process"""
tf.InteractiveSession()
image_size = [96, 64]
batch_size = 2
classes = Camvid.classes
data_format = "NHWC"
model = LMBiSeNet(
image_size=image_size,
batch_size=batch_size,
classes=classes,
data_format=data_format,
)
post_process = Sequence([
Bilinear(
size=image_size,
data_format=data_format,
compatible_tensorflow_v1=True,
),
Softmax()
])
shape = (batch_size, image_size[0] // 8, image_size[1] // 8, len(classes))
np_output = np.random.uniform(-10., 10., size=shape).astype(np.float32)
output = tf.constant(np_output)
output = model.post_process(output)
expected = post_process(outputs=np_output)["outputs"]
assert np.allclose(output.eval(), expected, atol=1e-5, rtol=1e-5)
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 _build_process(module, processor_config=None):
processors = []
processor_config = processor_config or []
for p in processor_config:
for class_name, kwargs in p.items():
if hasattr(module, class_name):
cls = getattr(module, class_name)
else:
cls = import_from_string(class_name)
processor = cls.__new__(cls)
processor.__dict__.update(kwargs or {})
processors.append(processor)
return Sequence(processors=processors)
def test_yolov2_post_process():
tf.InteractiveSession()
image_size = [96, 64]
batch_size = 2
classes = Pascalvoc2007.classes
anchors = [(0.1, 0.2), (1.2, 1.1)]
data_format = "NHWC"
score_threshold = 0.25
nms_iou_threshold = 0.5
model = YoloV2(
image_size=image_size,
batch_size=batch_size,
classes=classes,
anchors=anchors,
data_format=data_format,
score_threshold=score_threshold,
nms_iou_threshold=nms_iou_threshold,
)
post_process = Sequence([
FormatYoloV2(
image_size=image_size,
classes=classes,
anchors=anchors,
data_format=data_format,
),
ExcludeLowScoreBox(threshold=score_threshold),
NMS(
iou_threshold=nms_iou_threshold,
classes=classes,
),
])
shape = (batch_size, len(anchors) * (len(classes) + 5),
image_size[0] // 32, image_size[1] // 32)
np_output = np.random.uniform(-2., 2., size=shape).astype(np.float32)
output = tf.constant(np_output)
ys = model.post_process(output)
expected_ys = post_process(outputs=np_output)["outputs"]
for y, expected_y in zip(ys, expected_ys):
assert np.allclose(y.eval(), expected_y), (y.eval(), expected_y)
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, profile_step=1)
def test_sequence():
batch_size = 3
image_size = [256, 512]
augmentor = Sequence([
FlipLeftRight(),
FlipTopBottom(),
SSDRandomCrop(),
])
dataset = Pascalvoc2007(
batch_size=batch_size, pre_processor=ResizeWithGtBoxes(image_size),
augmentor=augmentor,
)
dataset = DatasetIterator(dataset)
for _ in range(5):
images, labels = dataset.feed()
_show_images_with_boxes(images, labels)
def build_post_process(post_processor_config):
module_name = "blueoil/post_processor"
f, pathname, description = imp.find_module(module_name)
module = imp.load_module(module_name, f, pathname, description)
processors = []
if post_processor_config is None:
post_processor_config = {}
for p in post_processor_config:
for class_name in p:
class_args = p[class_name]
if class_args is None:
class_args = {}
cls = getattr(module, class_name)
# Create none initialized processor `cls` instance.
processor = cls.__new__(cls)
# Fill processor instance member.
for k in class_args:
v = class_args[k]
processor.__dict__[k] = v
processors.append(processor)
seq = Sequence(processors=processors)
return seq
self.prefetcher.terminate = True
self.prefetcher.pool.close()
self.prefetcher.pool.join()
if __name__ == '__main__':
from blueoil.datasets.cifar10 import Cifar10
from blueoil.data_processor import Sequence
from blueoil.data_augmentor import FlipLeftRight, Hue, Blur
cifar10 = Cifar10()
augmentor = Sequence([
FlipLeftRight(0.5),
Hue((-10, 10)),
Blur(),
])
dataset_iterator = DatasetIterator(dataset=cifar10,
enable_prefetch=True,
augmentor=augmentor)
time.sleep(2)
import time
t0 = time.time()
data_batch = next(dataset_iterator)
t1 = time.time()
print("time of prefetch: {}".format(t1 - t0))
dataset_iterator2 = DatasetIterator(dataset=cifar10,
enable_prefetch=False,
SUMMARISE_STEPS = 1000
# pretrain
IS_PRETRAIN = False
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
# for debug
# BATCH_SIZE = 2
# SUMMARISE_STEPS = 1
# IS_DEBUG = True
PRE_PROCESSOR = Sequence([
PerImageStandardization(),
])
POST_PROCESSOR = None
NETWORK = SmartDict()
NETWORK.OPTIMIZER_CLASS = tf.compat.v1.train.AdamOptimizer
NETWORK.OPTIMIZER_KWARGS = {"learning_rate": 0.001}
NETWORK.IMAGE_SIZE = IMAGE_SIZE
NETWORK.BATCH_SIZE = BATCH_SIZE
NETWORK.DATA_FORMAT = DATA_FORMAT
DATASET = SmartDict()
DATASET.BATCH_SIZE = BATCH_SIZE
DATASET.DATA_FORMAT = DATA_FORMAT
DATASET.PRE_PROCESSOR = PRE_PROCESSOR
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
# for debug
# MAX_STEPS = 10
# BATCH_SIZE = 31
# SAVE_CHECKPOINT_STEPS = 2
# KEEP_CHECKPOINT_MAX = 5
# TEST_STEPS = 10
# SUMMARISE_STEPS = 2
# IS_DEBUG = True
PRE_PROCESSOR = Sequence([
Resize(size=IMAGE_SIZE),
PerImageStandardization(),
])
POST_PROCESSOR = None
NETWORK = EasyDict()
NETWORK.OPTIMIZER_CLASS = tf.compat.v1.train.MomentumOptimizer
NETWORK.OPTIMIZER_KWARGS = {"momentum": 0.9}
NETWORK.LEARNING_RATE_FUNC = tf.compat.v1.train.piecewise_constant
step_per_epoch = 50000 // BATCH_SIZE
NETWORK.LEARNING_RATE_KWARGS = {
"values": [0.01, 0.1, 0.01, 0.001, 0.0001],
"boundaries": [step_per_epoch, step_per_epoch * 50, step_per_epoch * 100, step_per_epoch * 198],
}
NETWORK.IMAGE_SIZE = IMAGE_SIZE
NETWORK.BATCH_SIZE = BATCH_SIZE
NETWORK.DATA_FORMAT = DATA_FORMAT
TFDS_PRE_PROCESSOR = TFDSProcessorSequence(
[TFDSResizeWithGtBoxes(IMAGE_SIZE),
TFDSPerImageStandardization()])
anchors = [
(0.5, 0.25),
(1.0, 0.75),
]
POST_PROCESSOR = Sequence([
FormatYoloV2(
image_size=IMAGE_SIZE,
classes=CLASSES,
anchors=anchors,
data_format=DATA_FORMAT,
),
ExcludeLowScoreBox(threshold=0.05),
NMS(
iou_threshold=0.5,
classes=CLASSES,
),
])
NETWORK = EasyDict()
NETWORK.OPTIMIZER_CLASS = tf.train.AdamOptimizer
NETWORK.OPTIMIZER_KWARGS = {"learning_rate": 0.001}
NETWORK.IMAGE_SIZE = IMAGE_SIZE
NETWORK.BATCH_SIZE = BATCH_SIZE
NETWORK.DATA_FORMAT = DATA_FORMAT
NETWORK.ANCHORS = anchors
NETWORK.WEIGHT_DECAY_RATE = 0.0005
MAX_EPOCHS = 400
SAVE_CHECKPOINT_STEPS = 1000
KEEP_CHECKPOINT_MAX = 1
TEST_STEPS = 1000
SUMMARISE_STEPS = 10000
# pretrain
IS_PRETRAIN = False
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
PRE_PROCESSOR = Sequence([
ResizeWithGtBoxes(size=IMAGE_SIZE),
PerImageStandardization()
])
anchors = [
(1.3221, 1.73145), (3.19275, 4.00944), (5.05587, 8.09892), (9.47112, 4.84053), (11.2364, 10.0071)
]
score_threshold = 0.05
nms_iou_threshold = 0.5
nms_max_output_size = 100
POST_PROCESSOR = Sequence([
FormatYoloV2(
image_size=IMAGE_SIZE,
classes=CLASSES,
anchors=anchors,
data_format=DATA_FORMAT,
),
ExcludeLowScoreBox(threshold=score_threshold),
'ship', 'truck'
]
MAX_EPOCHS = 100
SAVE_CHECKPOINT_STEPS = 1000
KEEP_CHECKPOINT_MAX = 1
TEST_STEPS = 1000
SUMMARISE_STEPS = 10000
# pretrain
IS_PRETRAIN = False
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
PRE_PROCESSOR = Sequence([Resize(size=IMAGE_SIZE), PerImageStandardization()])
POST_PROCESSOR = None
NETWORK = EasyDict()
NETWORK.OPTIMIZER_CLASS = tf.compat.v1.train.MomentumOptimizer
NETWORK.OPTIMIZER_KWARGS = {'momentum': 0.9}
NETWORK.LEARNING_RATE_FUNC = tf.compat.v1.train.piecewise_constant
NETWORK.LEARNING_RATE_KWARGS = {
'values': [0.001, 0.0001, 1e-05, 1e-06],
'boundaries': [25781, 51562, 77343]
}
NETWORK.IMAGE_SIZE = IMAGE_SIZE
NETWORK.BATCH_SIZE = BATCH_SIZE
NETWORK.DATA_FORMAT = DATA_FORMAT
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.compat.v1.train.AdamOptimizer
NETWORK.OPTIMIZER_KWARGS = {}
NETWORK.LEARNING_RATE_FUNC = tf.compat.v1.train.piecewise_constant
NETWORK.LEARNING_RATE_KWARGS = {
IS_DEBUG = False IMAGE_SIZE = [448, 448] BATCH_SIZE = 4 DATA_FORMAT = "NHWC" TASK = Tasks.CLASSIFICATION CLASSES = DATASET_CLASS.classes # for debug # MAX_STEPS = 100 # SAVE_CHECKPOINT_STEPS = 100 # TEST_STEPS = 10 # SUMMARISE_STEPS = 100 # IS_PRETRAIN = False # IS_DEBUG = True PRE_PROCESSOR = Sequence([Resize(size=IMAGE_SIZE), DivideBy255()]) POST_PROCESSOR = None NETWORK = SmartDict() NETWORK.IMAGE_SIZE = IMAGE_SIZE NETWORK.BATCH_SIZE = BATCH_SIZE NETWORK.DATA_FORMAT = DATA_FORMAT # dataset DATASET = SmartDict() DATASET.BATCH_SIZE = BATCH_SIZE DATASET.DATA_FORMAT = DATA_FORMAT DATASET.PRE_PROCESSOR = PRE_PROCESSOR
SUMMARISE_STEPS = 1000
# pretrain
IS_PRETRAIN = False
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
# for debug
# BATCH_SIZE = 2
# SUMMARISE_STEPS = 1
# IS_DEBUG = True
PRE_PROCESSOR = Sequence([
DivideBy255(),
])
POST_PROCESSOR = None
NETWORK = EasyDict()
NETWORK.OPTIMIZER_CLASS = tf.compat.v1.train.AdamOptimizer
NETWORK.OPTIMIZER_KWARGS = {"learning_rate": 0.001}
NETWORK.IMAGE_SIZE = IMAGE_SIZE
NETWORK.BATCH_SIZE = BATCH_SIZE
NETWORK.DATA_FORMAT = DATA_FORMAT
NETWORK.ACTIVATION_QUANTIZER = linear_mid_tread_half_quantizer
NETWORK.ACTIVATION_QUANTIZER_KWARGS = {
'bit': 2,
'max_value': 2
}
NETWORK.WEIGHT_QUANTIZER = binary_mean_scaling_quantizer
IMAGE_SIZE = [416, 416]
BATCH_SIZE = 8
DATA_FORMAT = "NCHW"
TASK = Tasks.OBJECT_DETECTION
CLASSES = DATASET_CLASS.classes
# for debug
# MAX_STEPS = 100
# SAVE_CHECKPOINT_STEPS = 100
# TEST_STEPS = 10
# SUMMARISE_STEPS = 100
# IS_PRETRAIN = False
# IS_DEBUG = True
PRE_PROCESSOR = Sequence([
ResizeWithGtBoxes(size=IMAGE_SIZE),
DivideBy255(),
])
anchors = [(1.3221, 1.73145), (3.19275, 4.00944), (5.05587, 8.09892),
(9.47112, 4.84053), (11.2364, 10.0071)]
score_threshold = 0.005
nms_iou_threshold = 0.55
nms_max_output_size = 100
POST_PROCESSOR = Sequence([
FormatYoloV2(
image_size=IMAGE_SIZE,
classes=CLASSES,
anchors=anchors,
data_format=DATA_FORMAT,
),
ExcludeLowScoreBox(threshold=score_threshold),
NMS(
SUMMARISE_STEPS = 1000
# pretrain
IS_PRETRAIN = False
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
# for debug
# BATCH_SIZE = 2
# SUMMARISE_STEPS = 1
# IS_DEBUG = True
PRE_PROCESSOR = Sequence([
Resize(size=IMAGE_SIZE),
PerImageStandardization(),
])
POST_PROCESSOR = Sequence([
Bilinear(size=IMAGE_SIZE, data_format=DATA_FORMAT, compatible_tensorflow_v1=True),
Softmax(),
])
NETWORK = SmartDict()
NETWORK.OPTIMIZER_CLASS = tf.compat.v1.train.AdamOptimizer
NETWORK.OPTIMIZER_KWARGS = {"learning_rate": 0.001}
NETWORK.IMAGE_SIZE = IMAGE_SIZE
NETWORK.BATCH_SIZE = BATCH_SIZE
NETWORK.DATA_FORMAT = DATA_FORMAT
NETWORK.WEIGHT_DECAY_RATE = 0.
NETWORK.AUXILIARY_LOSS_WEIGHT = 0.5
"conv2/kernel:",
"conv2/bias:",
"conv3/kernel:",
"conv3/bias:",
"conv4/kernel:",
"conv4/bias:",
"conv5/kernel:",
"conv5/bias:",
"conv6/kernel:",
"conv6/bias:",
]
PRETRAIN_DIR = "saved/lmnet_0.01_caltech101/checkpoints"
PRETRAIN_FILE = "save.ckpt-99001"
PRE_PROCESSOR = Sequence([
Resize(size=IMAGE_SIZE),
PerImageStandardization()
])
POST_PROCESSOR = None
NETWORK = SmartDict()
NETWORK.OPTIMIZER_CLASS = tf.compat.v1.train.AdamOptimizer
NETWORK.OPTIMIZER_KWARGS = {"learning_rate": 0.001}
NETWORK.IMAGE_SIZE = IMAGE_SIZE
NETWORK.BATCH_SIZE = BATCH_SIZE
NETWORK.DATA_FORMAT = DATA_FORMAT
NETWORK.WEIGHT_DECAY_RATE = 0.0005
NETWORK.ACTIVATION_QUANTIZER = linear_mid_tread_half_quantizer
NETWORK.ACTIVATION_QUANTIZER_KWARGS = {
'bit': 2,
'max_value': 2
}
CLASSES = DATASET_CLASS.classes
MAX_STEPS = 100000
SAVE_CHECKPOINT_STEPS = 5000
KEEP_CHECKPOINT_MAX = 5
TEST_STEPS = 1000
SUMMARISE_STEPS = 100
# pretrain
IS_PRETRAIN = False
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
PRE_PROCESSOR = Sequence([
Resize(size=IMAGE_SIZE),
DivideBy255(),
])
POST_PROCESSOR = None
STEP_PER_EPOCH = 50000 // BATCH_SIZE
TUNE_SPEC = {
'run': 'tunable',
'resources_per_trial': {
"cpu": 2,
"gpu": 0.5
},
'stop': {
'mean_accuracy': 0.87,
'training_iteration': 200,
},
CLASSES = DATASET_CLASS.classes
MAX_STEPS = 2
SAVE_CHECKPOINT_STEPS = 1
KEEP_CHECKPOINT_MAX = 5
TEST_STEPS = 100
SUMMARISE_STEPS = 100
# pretrain
IS_PRETRAIN = False
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
PRE_PROCESSOR = Sequence(
[ResizeWithGtBoxes(IMAGE_SIZE),
PerImageStandardization()])
anchors = [
(0.5, 0.25),
(1.0, 0.75),
]
POST_PROCESSOR = Sequence([
FormatYoloV2(
image_size=IMAGE_SIZE,
classes=CLASSES,
anchors=anchors,
data_format=DATA_FORMAT,
),
ExcludeLowScoreBox(threshold=0.05),
NMS(
TASK = Tasks.OBJECT_DETECTION
CLASSES = DATASET_CLASS.classes
KEEP_CHECKPOINT_MAX = 5
MAX_EPOCHS = 100
SAVE_CHECKPOINT_STEPS = 100
TEST_STEPS = 100
SUMMARISE_STEPS = 10
# pretrain
IS_PRETRAIN = False
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
PRE_PROCESSOR = Sequence([ResizeWithGtBoxes(size=IMAGE_SIZE), DivideBy255()])
anchors = [(1.3221, 1.73145), (3.19275, 4.00944), (5.05587, 8.09892),
(9.47112, 4.84053), (11.2364, 10.0071)]
score_threshold = 0.05
nms_iou_threshold = 0.5
nms_max_output_size = 100
POST_PROCESSOR = Sequence([
FormatYoloV2(
image_size=IMAGE_SIZE,
classes=CLASSES,
anchors=anchors,
data_format=DATA_FORMAT,
),
ExcludeLowScoreBox(threshold=score_threshold),
NMS(
iou_threshold=nms_iou_threshold,
# pretrain
IS_PRETRAIN = False
PRETRAIN_VARS = []
PRETRAIN_DIR = ""
PRETRAIN_FILE = ""
# for debug
# MAX_STEPS = 10
# BATCH_SIZE = 31
# SAVE_CHECKPOINT_STEPS = 2
# TEST_STEPS = 10
# SUMMARISE_STEPS = 2
# IS_DEBUG = True
PRE_PROCESSOR = Sequence([Resize(size=IMAGE_SIZE), DivideBy255()])
POST_PROCESSOR = None
NETWORK = EasyDict()
NETWORK.OPTIMIZER_CLASS = tf.compat.v1.train.MomentumOptimizer
NETWORK.OPTIMIZER_KWARGS = {"momentum": 0.9}
NETWORK.LEARNING_RATE_FUNC = tf.compat.v1.train.polynomial_decay
# TODO(wakiska): It is same as original yolov2 paper (batch size = 128).
NETWORK.LEARNING_RATE_KWARGS = {
"learning_rate": 1e-1,
"decay_steps": 1600000,
"power": 4.0,
"end_learning_rate": 0.0
}
NETWORK.IMAGE_SIZE = IMAGE_SIZE
NETWORK.BATCH_SIZE = BATCH_SIZE