def resnet_v2_50(inputs,
config,
is_training=True,
scope='resnet_v2_50'):
"""Modified ResNet-50 model."""
blocks = [
resnet_v2.resnet_v2_block('block1',
base_depth=config.block1_depth,
num_units=config.block1_units,
stride=config.block1_stride),
resnet_v2.resnet_v2_block('block2',
base_depth=config.block2_depth,
num_units=config.block2_units,
stride=config.block2_stride),
resnet_v2.resnet_v2_block('block3',
base_depth=config.block3_depth,
num_units=config.block3_units,
stride=config.block3_stride),
resnet_v2.resnet_v2_block('block4',
base_depth=config.block4_depth,
num_units=config.block4_units,
stride=config.block4_stride),
]
return resnet_v2.resnet_v2(inputs,
blocks,
is_training=is_training,
global_pool=False,
include_root_block=True,
scope=scope)
def resnet_12(inputs, num_classes, scope='resnet_12'):
blocks = [
resnet_v2.resnet_v2_block('block1',
base_depth=64,
num_units=2,
stride=1),
resnet_v2.resnet_v2_block('block2',
base_depth=64,
num_units=2,
stride=1),
resnet_v2.resnet_v2_block('block3',
base_depth=64,
num_units=2,
stride=1),
resnet_v2.resnet_v2_block('block4',
base_depth=64,
num_units=2,
stride=1),
resnet_v2.resnet_v2_block('block5',
base_depth=64,
num_units=2,
stride=1)
]
return resnet_v2.resnet_v2(inputs,
blocks,
num_classes,
is_training=True,
global_pool=True,
output_stride=None,
include_root_block=True,
reuse=None,
scope=scope)
def resnet_small(inputs,
num_classes=None,
is_training=True,
global_pool=True,
output_stride=None,
include_root_block=True,
reuse=None,
scope='resnet_small'):
blocks = [
resnet_v2.resnet_v2_block('block1',
base_depth=32,
num_units=2,
stride=2),
resnet_v2.resnet_v2_block('block2',
base_depth=64,
num_units=2,
stride=2),
resnet_v2.resnet_v2_block('block3',
base_depth=128,
num_units=2,
stride=2),
resnet_v2.resnet_v2_block('block4',
base_depth=256,
num_units=2,
stride=2),
]
return resnet_v2.resnet_v2(inputs,
blocks,
num_classes,
is_training=is_training,
global_pool=global_pool,
output_stride=output_stride,
include_root_block=include_root_block,
reuse=reuse,
scope=scope)
def fr_v2(x,
output_neurons,
inside_neurons,
is_training,
name='fr',
wt_decay=0.0001,
stride=1,
updates_collections=tf.GraphKeys.UPDATE_OPS):
"""Performs fusion of information between the map and the reward map.
Inputs
x: NxHxWxC1
Outputs
fr map: NxHxWx(output_neurons)
"""
if type(stride) != list:
stride = [stride]
with slim.arg_scope(
resnet_v2.resnet_utils.resnet_arg_scope(weight_decay=wt_decay)):
with slim.arg_scope([slim.batch_norm],
updates_collections=updates_collections) as arg_sc:
# Change the updates_collections for the conv normalizer_params to None
for i in range(len(arg_sc.keys())):
if 'convolution' in arg_sc.keys()[i]:
arg_sc.values()[i]['normalizer_params'][
'updates_collections'] = updates_collections
with slim.arg_scope(arg_sc):
bottleneck = resnet_v2.bottleneck
blocks = []
for i, s in enumerate(stride):
b = resnet_v2.resnet_utils.Block(
'block{:d}'.format(i + 1), bottleneck,
[{
'depth': output_neurons,
'depth_bottleneck': inside_neurons,
'stride': stride[i]
}])
blocks.append(b)
x, outs = resnet_v2.resnet_v2(x,
blocks,
num_classes=None,
is_training=is_training,
global_pool=False,
output_stride=None,
include_root_block=False,
reuse=False,
scope=name)
return x, outs
def _resnet_small(self,
inputs,
num_classes=None,
global_pool=True,
output_stride=None,
include_root_block=True,
reuse=None,
scope='resnet_v2_small'):
"""A shallow and thin ResNet v2 for faster tests."""
bottleneck = resnet_v2.bottleneck
blocks = [
resnet_utils.Block(
'block1', bottleneck, [(4, 1, 1)] * 2 + [(4, 1, 2)]),
resnet_utils.Block(
'block2', bottleneck, [(8, 2, 1)] * 2 + [(8, 2, 2)]),
resnet_utils.Block(
'block3', bottleneck, [(16, 4, 1)] * 2 + [(16, 4, 2)]),
resnet_utils.Block(
'block4', bottleneck, [(32, 8, 1)] * 2)]
return resnet_v2.resnet_v2(inputs, blocks, num_classes, global_pool,
output_stride, include_root_block, reuse, scope)
def _resnet_small(self,
inputs,
num_classes=None,
global_pool=True,
output_stride=None,
include_root_block=True,
reuse=None,
scope='resnet_v2_small'):
"""A shallow and thin ResNet v2 for faster tests."""
bottleneck = resnet_v2.bottleneck
blocks = [
resnet_utils.Block(
'block1', bottleneck, [(4, 1, 1)] * 2 + [(4, 1, 2)]),
resnet_utils.Block(
'block2', bottleneck, [(8, 2, 1)] * 2 + [(8, 2, 2)]),
resnet_utils.Block(
'block3', bottleneck, [(16, 4, 1)] * 2 + [(16, 4, 2)]),
resnet_utils.Block(
'block4', bottleneck, [(32, 8, 1)] * 2)]
return resnet_v2.resnet_v2(inputs, blocks, num_classes, global_pool,
output_stride, include_root_block, reuse, scope)
def resnet_v2_light(inputs,
num_classes=None,
is_training=True,
global_pool=True,
output_stride=None,
spatial_squeeze=True,
reuse=None,
scope='resnet_v2_light'):
"""ResNet-light model of AIlab. See resnet_v2() for arg and return description."""
blocks = [
resnet_v2.resnet_v2_block('block1',
base_depth=16,
num_units=3,
stride=2),
resnet_v2.resnet_v2_block('block2',
base_depth=32,
num_units=4,
stride=2),
resnet_v2.resnet_v2_block('block3',
base_depth=64,
num_units=8,
stride=2),
resnet_v2.resnet_v2_block('block4',
base_depth=128,
num_units=3,
stride=1),
]
return resnet_v2.resnet_v2(inputs,
blocks,
num_classes,
is_training=is_training,
global_pool=global_pool,
output_stride=output_stride,
include_root_block=True,
reuse=reuse,
scope=scope)