def __init__(self,
net=None,
version=None,
target_size=None,
anchor_sizes=[32, 64, 128, 256, 512],
anchor_size_ratios=[1, pow(2, 1 / 3),
pow(2, 2 / 3)],
anchor_aspect_ratios=[0.5, 1, 2],
box_offset=(0.5, 0.5),
anchor_box_clip=True,
ctx=mx.cpu()):
super(AnchorNet, self).__init__()
self._net = net
if version not in [0, 1, 2, 3, 4, 5, 6]:
raise ValueError
feature_sizes = []
bifpn = get_bifpn(version, ctx=mx.cpu(), dummy=True)(
mx.nd.random_uniform(low=0,
high=1,
shape=(1, 3, target_size[0], target_size[1]),
ctx=mx.cpu()))
for bif in bifpn:
feature_sizes.append(bif.shape[2:]) # h, w
# get_fpn_resnet 외부에서 보내지 않으면. 무조건 forward 한번 시켜야 하는데,
# 네트워크를 정확히 정의 해놓지 않아서...(default init) 쓸데 없는 코드를
# 넣어야 한다.
with self.name_scope():
# 아래 두줄은 self.name_scope()안에 있어야 한다. - 새롭게 anchor만드는 네크워크를 생성하는 것이므로.!!!
# self.name_scope() 밖에 있으면 기존의 self._net 과 이름이 겹친다.
self._bifpn = get_bifpn(version, ctx=ctx, dummy=True)
self._bifpn.forward(mx.nd.ones(shape=(1, 3) + target_size,
ctx=ctx))
self._anchor_generators = HybridSequential()
for index, feature_size, anchor_size in zip(
range(len(feature_sizes)), feature_sizes, anchor_sizes):
self._anchor_generators.add(
EfficientAnchorGenerator(
index=index,
input_size=target_size,
feature_size=feature_size,
anchor_size=anchor_size,
anchor_size_ratios=anchor_size_ratios,
anchor_aspect_ratios=anchor_aspect_ratios,
box_offset=box_offset,
box_clip=anchor_box_clip))
self._anchor_generators.initialize(ctx=ctx)
def __init__(self,
version=0,
input_size=(512, 512),
anchor_size_ratios=[1, pow(2, 1 / 3),
pow(2, 2 / 3)],
anchor_aspect_ratios=[0.5, 1, 2],
num_classes=1,
ctx=mx.cpu()):
super(EfficientNet_Except_Anchor, self).__init__()
if version not in [0, 1, 2, 3, 4, 5, 6]:
raise ValueError
feature_sizes = []
fpn_output = get_bifpn(version, ctx=mx.cpu(), dummy=True)(
mx.nd.random_uniform(low=0,
high=1,
shape=(1, 3, input_size[0], input_size[1]),
ctx=mx.cpu()))
for fpn in fpn_output:
feature_sizes.append(fpn.shape[2:]) # h, w
self._bifpn = get_bifpn(version, ctx=ctx)
self._num_classes = num_classes
with self.name_scope():
self._class_subnet = HybridSequential()
self._box_subnet = HybridSequential()
for _ in range(3):
self._class_subnet.add(
ConvPredictor(num_channel=256,
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
activation='relu',
use_bias=True,
in_channels=0,
weight_initializer=mx.init.Normal(0.01),
bias_initializer='zeros'))
self._box_subnet.add(
ConvPredictor(num_channel=256,
kernel=3,
pad=1,
stride=1,
activation='relu',
use_bias=True,
in_channels=0,
weight_initializer=mx.init.Normal(0.01),
bias_initializer='zeros'))
'''
bias_initializer=mx.init.Constant(-np.log((1-0.01)/0.01)?
논문에서,
For the final convlayer of the classification subnet, we set the bias initialization to
b = − log((1 − π)/π), where π specifies that at that at the start of training every anchor should be
labeled as foreground with confidence of ∼π. We use π = .01 in all experiments,
although results are robust to the exact value. As explained in section 3.3,
this initialization prevents the large number of background anchors from generating a large,
destabilizing loss value in the first iteration of training.
-> 초기에 class subnet 마지막 bias를 b = − log((1 − π)/π)로 설정함으로써,
모든 anchor를 0.01 값을 가지는 foreground로 만들어버리는 초기화 방법
거의 대부분인 background anchor가 첫 번째 학습에서 불안정한 loss 값을 가지는 것을 방지해준다함.
'''
prior = 0.01
self._class_subnet.add(
ConvPredictor(
num_channel=num_classes * len(anchor_size_ratios) *
len(anchor_aspect_ratios),
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
activation=None,
use_bias=True,
in_channels=0,
weight_initializer=mx.init.Normal(0.01),
bias_initializer=mx.init.Constant(-np.log((1 - prior) /
prior))))
self._box_subnet.add(
ConvPredictor(num_channel=4 * len(anchor_size_ratios) *
len(anchor_aspect_ratios),
kernel=3,
pad=1,
stride=1,
activation=None,
use_bias=True,
in_channels=0,
weight_initializer=mx.init.Normal(0.01),
bias_initializer='zeros'))
self._class_subnet.initialize(ctx=ctx)
self._box_subnet.initialize(ctx=ctx)
print(f"{self.__class__.__name__} Head weight init 완료")
def __init__(
self,
version=0,
input_size=(512, 512),
anchor_sizes=[32, 64, 128, 256, 512],
anchor_size_ratios=[1, pow(2, 1 / 3), pow(2, 2 / 3)],
anchor_aspect_ratios=[0.5, 1, 2],
num_classes=1, # foreground만
anchor_box_offset=(0.5, 0.5),
anchor_box_clip=True,
alloc_size=[256, 256],
ctx=mx.cpu()):
super(Efficient, self).__init__()
if version not in [0, 1, 2, 3, 4, 5, 6]:
raise ValueError
# Box / Class Layer - 논문에 나온대로
repeat = [3, 3, 3, 4, 4, 4, 5, 5]
channels = [64, 88, 112, 160, 224, 288, 384, 384]
feature_sizes = []
bifpn_output = get_bifpn(version, ctx=mx.cpu(), dummy=True)(
mx.nd.random_uniform(low=0,
high=1,
shape=(1, 3, input_size[0], input_size[1]),
ctx=mx.cpu()))
for bifpn in bifpn_output:
feature_sizes.append(bifpn.shape[2:]) # h, w
self._bifpn = get_bifpn(version, ctx=ctx)
self._num_classes = num_classes
with self.name_scope():
self._class_subnet = HybridSequential()
self._box_subnet = HybridSequential()
self._anchor_generators = HybridSequential()
for _ in range(repeat[version]):
self._class_subnet.add(
ConvPredictor(
num_channel=channels[version],
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
activation='relu',
use_bias=True,
in_channels=0,
))
self._box_subnet.add(
ConvPredictor(
num_channel=channels[version],
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
activation='relu',
use_bias=True,
in_channels=0,
))
'''
bias_initializer=mx.init.Constant(-np.log((1-0.01)/0.01)?
논문에서,
For the final convlayer of the classification subnet, we set the bias initialization to
b = − log((1 − π)/π), where π specifies that at that at the start of training every anchor should be
labeled as foreground with confidence of ∼π. We use π = .01 in all experiments,
although results are robust to the exact value. As explained in section 3.3,
this initialization prevents the large number of background anchors from generating a large,
destabilizing loss value in the first iteration of training.
-> 초기에 class subnet 마지막 bias를 b = − log((1 − π)/π)로 설정함으로써,
모든 anchor를 0.01 값을 가지는 foreground로 만들어버리는 초기화 방법
거의 대부분인 background anchor 가 첫 번째 학습에서 불안정한 loss 값을 가지는 것을 방지해준다함.
'''
prior = 0.01
self._class_subnet.add(
ConvPredictor(
num_channel=num_classes * len(anchor_size_ratios) *
len(anchor_aspect_ratios),
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
activation=None,
use_bias=True,
in_channels=0,
bias_initializer=mx.init.Constant(-np.log((1 - prior) /
prior))))
self._box_subnet.add(
ConvPredictor(
num_channel=4 * len(anchor_size_ratios) *
len(anchor_aspect_ratios),
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
activation=None,
use_bias=True,
in_channels=0,
))
for index, feature_size, anchor_size in zip(
range(len(feature_sizes)), feature_sizes, anchor_sizes):
self._anchor_generators.add(
EfficientAnchorGenerator(
index=index,
input_size=input_size,
feature_size=feature_size,
anchor_size=anchor_size,
anchor_size_ratios=anchor_size_ratios,
anchor_aspect_ratios=anchor_aspect_ratios,
box_offset=anchor_box_offset,
box_clip=anchor_box_clip,
alloc_size=(alloc_size[0] // (2**index),
alloc_size[1] // (2**index))))
self._class_subnet.initialize(ctx=ctx)
self._box_subnet.initialize(ctx=ctx)
self._anchor_generators.initialize(ctx=ctx)
logging.info(
f"{self.__class__.__name__}_{version} Head weight init 완료")