def resnet_unit(cls, data, name, filter, stride, dilate, proj, norm_type,
norm_mom, ndev):
"""
One resnet unit is comprised of 2 or 3 convolutions and a shortcut.
:param data:
:param name:
:param filter:
:param stride:
:param dilate:
:param proj:
:param norm_type:
:param norm_mom:
:param ndev:
:return:
"""
norm = normalizer_factory(type=norm_type, ndev=ndev, mom=norm_mom)
conv1 = conv(data,
name=name + "_conv1",
filter=filter // 4,
stride=stride)
bn1 = norm(data=conv1, name=name + "_bn1")
relu1 = relu(bn1, name=name + "_relu1")
conv2 = conv(relu1,
name=name + "_conv2",
filter=filter // 4,
kernel=3,
dilate=dilate)
bn2 = norm(data=conv2, name=name + "_bn2")
relu2 = relu(bn2, name=name + "_relu2")
conv3 = conv(relu2, name=name + "_conv3", filter=filter)
bn3 = norm(data=conv3, name=name + "_bn3")
if proj:
shortcut = conv(data,
name=name + "_sc",
filter=filter,
stride=stride)
shortcut = norm(data=shortcut, name=name + "_sc_bn")
else:
shortcut = data
eltwise = add(bn3, shortcut, name=name + "_plus")
return relu(eltwise, name=name + "_relu")
def resnet_c1(cls, data, use_3x3_conv0, use_bn_preprocess, norm_type,
norm_mom, ndev):
"""
Resnet C1 is comprised of irregular initial layers.
:param data: image symbol
:param use_3x3_conv0: use three 3x3 convs to replace one 7x7 conv
:param use_bn_preprocess: use batchnorm as the whitening layer, introduced by tornadomeet
:param norm_type: normalization method of activation, could be local, fix, sync, gn, in, ibn
:param norm_mom: normalization momentum, specific to batchnorm
:param ndev: num of gpus for sync batchnorm
:return: C1 symbol
"""
# preprocess
if use_bn_preprocess:
data = whiten(data, name="bn_data")
norm = normalizer_factory(type=norm_type, ndev=ndev, mom=norm_mom)
# C1
if use_3x3_conv0:
data = conv(data, filter=64, kernel=3, stride=2, name="conv0_0")
data = norm(data, name='bn0_0')
data = relu(data, name='relu0_0')
data = conv(data, filter=64, kernel=3, name="conv0_1")
data = norm(data, name='bn0_1')
data = relu(data, name='relu0_1')
data = conv(data, filter=64, kernel=3, name="conv0_2")
data = norm(data, name='bn0_2')
data = relu(data, name='relu0_2')
else:
data = conv(data, filter=64, kernel=7, stride=2, name="conv0")
data = norm(data, name='bn0')
data = relu(data, name='relu0')
data = pool(data, name="pool0", kernel=3, stride=2, pool_type='max')
return data
class NormalizeParam:
# normalizer = normalizer_factory(type="syncbn", ndev=8, wd_mult=1.0)
normalizer = normalizer_factory(type="local")
def __init__(self, pBackbone):
self.p = pBackbone
self.b = Builder()
def get_feature(self, data):
self.symbol = self.b.get_backbone(data, "msra", 50, "c5x2",
self.p.normalizer, self.p.fp16)
return self.symbol
# TODO: hook import with ResNetV1Builder
# import sys
# sys.modules[__name__] = ResNetV1Builder()
if __name__ == "__main__":
#############################################################
# python -m poi.models.common.backbone.resnet_v1
#############################################################
h = Builder()
data = var("data")
sym = h.get_backbone(data,
"msra",
50,
"fpn",
normalizer_factory(type="fixbn"),
fp16=True)
import mxnet as mx
sym = mx.sym.Group(sym)
mx.viz.print_summary(sym)