def mnasnet_d1_320(depth_multiplier=None):
"""Creates a jointly searched mnasnet backbone for 320x320 input size.
Args:
depth_multiplier: multiplier to number of filters per layer.
Returns:
blocks_args: a list of BlocksArgs for internal MnasNet blocks.
global_params: GlobalParams, global parameters for the model.
"""
blocks_args = [
'r3_k5_s11_e6_i32_o24', 'r4_k7_s22_e9_i24_o36', 'r5_k5_s22_e9_i36_o48',
'r5_k7_s22_e6_i48_o96', 'r5_k3_s11_e9_i96_o144',
'r5_k5_s22_e6_i144_o160', 'r1_k7_s11_e9_i160_o320'
]
global_params = mnasnet_model.GlobalParams(
batch_norm_momentum=0.99,
batch_norm_epsilon=1e-3,
dropout_rate=0.2,
data_format='channels_last',
num_classes=1000,
depth_multiplier=depth_multiplier,
depth_divisor=8,
min_depth=None,
stem_size=32,
use_keras=False)
decoder = MnasNetDecoder()
return decoder.decode(blocks_args), global_params
def mnasnet_a1(depth_multiplier=None):
"""Creates a mnasnet-a1 model.
Args:
depth_multiplier: multiplier to number of filters per layer.
Returns:
blocks_args: a list of BlocksArgs for internal MnasNet blocks.
global_params: GlobalParams, global parameters for the model.
"""
blocks_args = [
'r1_k3_s11_e1_i32_o16_noskip', 'r2_k3_s22_e6_i16_o24',
'r3_k5_s22_e3_i24_o40_se0.25', 'r4_k3_s22_e6_i40_o80',
'r2_k3_s11_e6_i80_o112_se0.25', 'r3_k5_s22_e6_i112_o160_se0.25',
'r1_k3_s11_e6_i160_o320'
]
global_params = mnasnet_model.GlobalParams(
batch_norm_momentum=0.99,
batch_norm_epsilon=1e-3,
dropout_rate=0.2,
data_format='channels_last',
num_classes=1000,
depth_multiplier=depth_multiplier,
depth_divisor=8,
min_depth=None)
decoder = MnasNetDecoder()
return decoder.decode(blocks_args), global_params
def legrnet70(depth_multiplier=None):
"""Creates a legrnet model.
Args:
depth_multiplier: multiplier to number of filters per layer.
Returns:
blocks_args: a list of BlocksArgs for internal MnasNet blocks.
global_params: GlobalParams, global parameters for the model.
"""
blocks_args = [
'r1_k3_s11_m17_i17_o16_noskip', 'r1_k3_s22_m20_i16_o24_noskip',
'r1_k3_s11_m29_i24_o24', 'r1_k3_s22_m144_i24_o32_noskip',
'r1_k3_s11_m191_i32_o32', 'r1_k3_s11_m39_i32_o32',
'r1_k3_s22_m192_i32_o64_noskip', 'r1_k3_s11_m268_i64_o64',
'r1_k3_s11_m80_i64_o64', 'r1_k3_s11_m240_i64_o64',
'r1_k3_s11_m384_i64_o96_noskip', 'r1_k3_s11_m572_i96_o96',
'r1_k3_s11_m574_i96_o96', 'r1_k3_s22_m574_i96_o160_noskip',
'r1_k3_s11_m192_i160_o160', 'r1_k3_s11_m951_i160_o160',
'r1_k3_s11_m959_i160_o320_noskip'
]
decoder = LeGRNetDecoder()
global_params = mnasnet_model.GlobalParams(
batch_norm_momentum=0.99,
batch_norm_epsilon=1e-3,
dropout_rate=0.2,
data_format='channels_last',
num_classes=1000,
depth_multiplier=depth_multiplier,
depth_divisor=1,
stem_filters=17,
last_filters=1280,
mask=False,
min_depth=None)
return decoder.decode(blocks_args), global_params
def mnasnet_small(depth_multiplier=None,
output_stride=32,
squeeze=False,
more=False):
"""Creates a mnasnet-a1 model.
Args:
depth_multiplier: multiplier to number of filters per layer.
Returns:
blocks_args: a list of BlocksArgs for internal MnasNet blocks.
global_params: GlobalParams, global parameters for the model.
"""
blocks_args = [
'r1_k3_s11_e1_i16_o8',
'r1_k3_s22_e3_i8_o16',
'r2_k3_s22_e6_i16_o16',
'r4_k5_s22_e6_i16_o32_se0.25',
]
if squeeze:
if more:
blocks_args.append('r3_k3_s11_e6_i32_o16_se0.25')
if output_stride == 16:
blocks_args.append('r3_k5_s11_e6_i16_o44_se0.25')
else:
blocks_args.append('r3_k5_s22_e6_i16_o44_se0.25')
blocks_args.append('r1_k3_s11_e6_i44_o72')
else:
blocks_args.append('r3_k3_s11_e6_i32_o32_se0.25')
if output_stride == 16:
blocks_args.append('r3_k5_s11_e6_i32_o44_se0.25')
else:
blocks_args.append('r3_k5_s22_e6_i32_o44_se0.25')
blocks_args.append('r1_k3_s11_e6_i44_o72')
else:
blocks_args.append('r3_k3_s11_e6_i32_o32_se0.25')
if output_stride == 16:
blocks_args.append('r3_k5_s11_e6_i32_o88_se0.25')
else:
blocks_args.append('r3_k5_s22_e6_i32_o88_se0.25')
blocks_args.append('r1_k3_s11_e6_i88_o144')
global_params = mnasnet_model.GlobalParams(
batch_norm_momentum=0.99,
batch_norm_epsilon=1e-3,
dropout_rate=0,
data_format='channels_last',
num_classes=1000,
depth_multiplier=depth_multiplier,
depth_divisor=8,
min_depth=None,
stem_size=8,
use_keras=True)
decoder = MnasNetDecoder()
return decoder.decode(blocks_args), global_params
