def __init__(self, net=None, version=None, anchors=None, target_size=None,
ctx=mx.cpu()):
super(AnchorOffstNet, self).__init__()
self._net = net
features = []
strides = []
darknet_output = get_darknet(version, pretrained=False, 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 out in darknet_output: # feature_14, feature_24, feature_28
out_height, out_width = out.shape[2:]
features.append([out_width, out_height])
strides.append([target_size[1] // out_width, target_size[0] // out_height])
features = features[::-1]
strides = strides[::-1] # deep -> middle -> shallow 순으로 !!!
anchors = OrderedDict(anchors)
anchors = list(anchors.values())[::-1]
self._numoffst = len(anchors)
with self.name_scope():
self._anchor_generators = HybridSequential()
for i, anchor, feature, stride in zip(range(len(features)), anchors, features, strides):
self._anchor_generators.add(YoloAnchorGenerator(i, anchor, feature, stride))
self._anchor_generators.initialize(ctx=ctx)
def __init__(self, Darknetlayer=53,
anchors={"shallow": [(10, 13), (16, 30), (33, 23)],
"middle": [(30, 61), (62, 45), (59, 119)],
"deep": [(116, 90), (156, 198), (373, 326)]},
num_classes=1, # foreground만
pretrained=True,
pretrained_path="modelparam",
ctx=mx.cpu()):
super(YoloV3output, self).__init__()
if Darknetlayer not in [53]:
raise ValueError
anchors = OrderedDict(anchors)
anchors = list(anchors.values())[::-1]
# 각 레이어에서 나오는 anchor갯수가 바뀔수도 있으니!!
self._num_anchors = []
for anchor in anchors:
self._num_anchors.append(len(anchor)) # 변화 가능
self._darkenet = get_darknet(Darknetlayer, pretrained=pretrained, ctx=ctx, root=pretrained_path)
self._num_classes = num_classes
self._num_pred = 5 + num_classes # 고정
with self.name_scope():
head_init_num_channel = 512
trans_init_num_channel = 256
self._head = HybridSequential()
self._transition = HybridSequential()
self._upsampleconv = HybridSequential()
# output
for j in range(len(anchors)):
if j == 0:
factor = 1
else:
factor = 2
head_init_num_channel = head_init_num_channel // factor
for _ in range(3):
self._head.add(Conv2D(channels=head_init_num_channel,
kernel_size=(1, 1),
strides=(1, 1),
padding=(0, 0),
use_bias=True,
in_channels=0,
weight_initializer=mx.init.Normal(0.01),
bias_initializer='zeros'
))
self._head.add(BatchNorm(epsilon=1e-5, momentum=0.9))
self._head.add(LeakyReLU(0.1))
self._head.add(Conv2D(channels=head_init_num_channel * 2,
kernel_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
use_bias=True,
in_channels=0,
weight_initializer=mx.init.Normal(0.01),
bias_initializer='zeros'
))
self._head.add(BatchNorm(epsilon=1e-5, momentum=0.9))
self._head.add(LeakyReLU(0.1))
self._head.add(Conv2D(channels=len(anchors[j]) * self._num_pred,
kernel_size=(1, 1),
strides=(1, 1),
padding=(0, 0),
use_bias=True,
in_channels=0,
weight_initializer=mx.init.Normal(0.01),
bias_initializer='zeros'
))
# for upsample - transition
for i in range(len(anchors) - 1):
if i == 0:
factor = 1
else:
factor = 2
trans_init_num_channel = trans_init_num_channel // factor
self._transition.add(Conv2D(channels=trans_init_num_channel,
kernel_size=(1, 1),
strides=(1, 1),
padding=(0, 0),
use_bias=True,
in_channels=0,
weight_initializer=mx.init.Normal(0.01),
bias_initializer='zeros'
))
self._transition.add(BatchNorm(epsilon=1e-5, momentum=0.9))
self._transition.add(LeakyReLU(0.1))
# for deconvolution upsampleing
for i in range(len(anchors) - 1):
if i == 0:
factor = 1
else:
factor = 2
trans_init_num_channel = trans_init_num_channel // factor
self._upsampleconv.add(Conv2DTranspose(trans_init_num_channel, kernel_size=3, strides=2, padding=1,
output_padding=1, use_bias=True, in_channels=0))
self._upsampleconv.add(BatchNorm(epsilon=1e-5, momentum=0.9))
self._upsampleconv.add(LeakyReLU(0.1))
self._head.initialize(ctx=ctx)
self._transition.initialize(ctx=ctx)
self._upsampleconv.initialize(ctx=ctx)
print(f"{self.__class__.__name__} Head weight init 완료")