def set_active_subnet(self, wid=None, ks=None, e=None, d=None):
width_mult_id = int2list(wid, 3 + len(self.blocks) - 1)
ks = int2list(ks, len(self.blocks) - 1)
expand_ratio = int2list(e, len(self.blocks) - 1)
depth = int2list(d, len(self.block_group_info))
if len(self.width_mult_list) > 1 and width_mult_id[0] is not None:
# active_out_channel
self.first_conv.active_out_channel = self.first_conv.out_channel_list[
width_mult_id[0]]
self.blocks[0].mobile_inverted_conv.active_out_channel = \
self.blocks[0].mobile_inverted_conv.out_channel_list[width_mult_id[1]]
self.feature_mix_layer.active_out_channel = self.feature_mix_layer.out_channel_list[
width_mult_id[2]]
for block, w, k, e in zip(self.blocks[1:], width_mult_id[3:], ks,
expand_ratio):
if w is not None:
block.mobile_inverted_conv.active_out_channel = block.mobile_inverted_conv.out_channel_list[
w]
if k is not None:
block.mobile_inverted_conv.active_kernel_size = k
if e is not None:
block.mobile_inverted_conv.active_expand_ratio = e
for i, d in enumerate(depth):
if d is not None:
self.runtime_depth[i] = min(len(self.block_group_info[i][0]),
d)
def __init__(self,
in_channel_list,
out_channel_list,
kernel_size_list=3,
expand_ratio_list=6,
stride=1,
act_func='relu6',
use_se=False):
super(DynamicMBConvLayer, self).__init__()
self.in_channel_list = in_channel_list
self.out_channel_list = out_channel_list
self.kernel_size_list = int2list(kernel_size_list, 1)
self.expand_ratio_list = int2list(expand_ratio_list, 1)
self.stride = stride
self.act_func = act_func
self.use_se = use_se
# build modules
max_middle_channel = round(
max(self.in_channel_list) * max(self.expand_ratio_list))
if max(self.expand_ratio_list) == 1:
self.inverted_bottleneck = None
else:
self.inverted_bottleneck = nn.Sequential(
OrderedDict([
('conv',
DynamicPointConv2d(max(self.in_channel_list),
max_middle_channel)),
('bn', DynamicBatchNorm2d(max_middle_channel)),
('act', build_activation(self.act_func, inplace=True)),
]))
self.depth_conv = nn.Sequential(
OrderedDict([('conv',
DynamicSeparableConv2d(max_middle_channel,
self.kernel_size_list,
self.stride)),
('bn', DynamicBatchNorm2d(max_middle_channel)),
('act', build_activation(self.act_func,
inplace=True))]))
if self.use_se:
self.depth_conv.add_module('se', DynamicSE(max_middle_channel))
self.point_linear = nn.Sequential(
OrderedDict([
('conv',
DynamicPointConv2d(max_middle_channel,
max(self.out_channel_list))),
('bn', DynamicBatchNorm2d(max(self.out_channel_list))),
]))
self.active_kernel_size = max(self.kernel_size_list)
self.active_expand_ratio = max(self.expand_ratio_list)
self.active_out_channel = max(self.out_channel_list)
def set_quantization_policy(self, pw_w_bits_setting=None, pw_a_bits_setting=None, dw_w_bits_setting=None,
dw_a_bits_setting=None):
pw_w_bits_setting = int2list(pw_w_bits_setting, len(self.blocks) - 1)
pw_a_bits_setting = int2list(pw_a_bits_setting, len(self.blocks) - 1)
dw_w_bits_setting = int2list(dw_w_bits_setting, len(self.blocks) - 1)
dw_a_bits_setting = int2list(dw_a_bits_setting, len(self.blocks) - 1)
for block, pw_w, pw_a, dw_w, dw_a in zip(self.blocks[1:], pw_w_bits_setting, pw_a_bits_setting,
dw_w_bits_setting,
dw_a_bits_setting):
block.mobile_inverted_conv.set_quantization_policy(pw_w_bit=pw_w, pw_a_bit=pw_a, dw_w_bit=dw_w,
dw_a_bit=dw_a)
def set_active_subnet(self, wid=None, ks=None, e=None, d=None):
width_mult_id = int2list(wid, 4 + len(self.block_group_info))
ks = int2list(ks, len(self.blocks) - 1)
expand_ratio = int2list(e, len(self.blocks) - 1)
depth = int2list(d, len(self.block_group_info))
for block, k, e in zip(self.blocks[1:], ks, expand_ratio):
if k is not None:
block.mobile_inverted_conv.active_kernel_size = k
if e is not None:
block.mobile_inverted_conv.active_expand_ratio = e
for i, d in enumerate(depth):
if d is not None:
self.runtime_depth[i] = min(len(self.block_group_info[i]), d)
def validate(run_manager, epoch=0, is_test=True, image_size_list=None,
width_mult_list=None, ks_list=None, expand_ratio_list=None, depth_list=None, additional_setting=None):
dynamic_net = run_manager.net
if isinstance(dynamic_net, nn.DataParallel):
dynamic_net = dynamic_net.module
dynamic_net.eval()
if image_size_list is None:
image_size_list = int2list(run_manager.run_config.data_provider.image_size, 1)
if width_mult_list is None:
width_mult_list = [i for i in range(len(dynamic_net.width_mult_list))]
if ks_list is None:
ks_list = dynamic_net.ks_list
if expand_ratio_list is None:
expand_ratio_list = dynamic_net.expand_ratio_list
if depth_list is None:
depth_list = dynamic_net.depth_list
subnet_settings = []
for w in width_mult_list:
for d in depth_list:
for e in expand_ratio_list:
for k in ks_list:
for img_size in image_size_list:
subnet_settings.append([{
'image_size': img_size,
'wid': w,
'd': d,
'e': e,
'ks': k,
}, 'R%s-W%s-D%s-E%s-K%s' % (img_size, w, d, e, k)])
if additional_setting is not None:
subnet_settings += additional_setting
losses_of_subnets, top1_of_subnets, top5_of_subnets = [], [], []
valid_log = ''
for setting, name in subnet_settings:
run_manager.write_log('-' * 30 + ' Validate %s ' % name + '-' * 30, 'train', should_print=False)
run_manager.run_config.data_provider.assign_active_img_size(setting.pop('image_size'))
dynamic_net.set_active_subnet(**setting)
run_manager.write_log(dynamic_net.module_str, 'train', should_print=False)
run_manager.reset_running_statistics(dynamic_net)
loss, top1, top5 = run_manager.validate(epoch=epoch, is_test=is_test, run_str=name, net=dynamic_net)
losses_of_subnets.append(loss)
top1_of_subnets.append(top1)
top5_of_subnets.append(top5)
valid_log += '%s (%.3f), ' % (name, top1)
return list_mean(losses_of_subnets), list_mean(top1_of_subnets), list_mean(top5_of_subnets), valid_log
def __init__(self,
n_classes=1000,
width_mult=1,
bn_param=(0.1, 1e-3),
dropout_rate=0.2,
ks=None,
expand_ratio=None,
depth_param=None,
stage_width_list=None):
if ks is None:
ks = 3
if expand_ratio is None:
expand_ratio = 6
input_channel = 32
last_channel = 1280
input_channel = make_divisible(input_channel * width_mult, 8)
last_channel = make_divisible(last_channel * width_mult,
8) if width_mult > 1.0 else last_channel
inverted_residual_setting = [
# t, c, n, s
[1, 16, 1, 1],
[expand_ratio, 24, 2, 2],
[expand_ratio, 32, 3, 2],
[expand_ratio, 64, 4, 2],
[expand_ratio, 96, 3, 1],
[expand_ratio, 160, 3, 2],
[expand_ratio, 320, 1, 1],
]
if depth_param is not None:
assert isinstance(depth_param, int)
for i in range(1, len(inverted_residual_setting) - 1):
inverted_residual_setting[i][2] = depth_param
if stage_width_list is not None:
for i in range(len(inverted_residual_setting)):
inverted_residual_setting[i][1] = stage_width_list[i]
ks = int2list(ks,
sum([n for _, _, n, _ in inverted_residual_setting]) - 1)
_pt = 0
# first conv layer
first_conv = ConvLayer(3,
input_channel,
kernel_size=3,
stride=2,
use_bn=True,
act_func='relu6',
ops_order='weight_bn_act')
# inverted residual blocks
blocks = []
for t, c, n, s in inverted_residual_setting:
output_channel = make_divisible(c * width_mult, 8)
for i in range(n):
if i == 0:
stride = s
else:
stride = 1
if t == 1:
kernel_size = 3
else:
kernel_size = ks[_pt]
_pt += 1
mobile_inverted_conv = MBInvertedConvLayer(
in_channels=input_channel,
out_channels=output_channel,
kernel_size=kernel_size,
stride=stride,
expand_ratio=t,
)
if stride == 1:
if input_channel == output_channel:
shortcut = IdentityLayer(input_channel, input_channel)
else:
shortcut = None
else:
shortcut = None
blocks.append(
MobileInvertedResidualBlock(mobile_inverted_conv,
shortcut))
input_channel = output_channel
# 1x1_conv before global average pooling
feature_mix_layer = ConvLayer(
input_channel,
last_channel,
kernel_size=1,
use_bn=True,
act_func='relu6',
ops_order='weight_bn_act',
)
classifier = LinearLayer(last_channel,
n_classes,
dropout_rate=dropout_rate)
super(MobileNetV2, self).__init__(first_conv, blocks,
feature_mix_layer, classifier)
# set bn param
self.set_bn_param(momentum=bn_param[0], eps=bn_param[1])
def __init__(self, n_classes=1000, bn_param=(0.1, 1e-3), dropout_rate=0.1, base_stage_width=None,
width_mult_list=1.0, ks_list=3, expand_ratio_list=6, depth_list=4,
depth_ensemble_list=None, depth_ensemble_mode='avg'):
self.width_mult_list = int2list(width_mult_list, 1)
self.ks_list = int2list(ks_list, 1)
self.expand_ratio_list = int2list(expand_ratio_list, 1)
self.depth_list = int2list(depth_list, 1)
self.depth_ensemble_list = depth_ensemble_list
self.depth_ensemble_mode = depth_ensemble_mode
self.width_mult_list.sort()
self.ks_list.sort()
self.expand_ratio_list.sort()
self.depth_list.sort()
if base_stage_width == 'v2':
base_stage_width = [32, 16, 24, 32, 64, 96, 160, 320, 1280]
elif base_stage_width == 'old':
base_stage_width = [32, 16, 32, 40, 80, 96, 192, 320, 1280]
else:
# ProxylessNAS Stage Width
base_stage_width = [32, 16, 24, 40, 80, 96, 192, 320, 1280]
input_channel = [make_divisible(base_stage_width[0] * width_mult, 8) for width_mult in self.width_mult_list]
first_block_width = [make_divisible(base_stage_width[1] * width_mult, 8) for width_mult in self.width_mult_list]
last_channel = [
make_divisible(base_stage_width[-1] * width_mult, 8) if width_mult > 1.0 else base_stage_width[-1]
for width_mult in self.width_mult_list
]
# first conv layer
if len(input_channel) == 1:
first_conv = QConvLayer(
3, max(input_channel), kernel_size=3, stride=2, use_bn=True, act_func='relu6',
ops_order='weight_bn_act',
w_bit=8, a_bit=-1, half_wave=False
)
else:
first_conv = DynamicQConvLayer(
in_channel_list=int2list(3, len(input_channel)), out_channel_list=input_channel, kernel_size=3,
stride=2, act_func='relu6', w_bit=8, a_bit=8, half_wave=False
)
# first block
if len(first_block_width) == 1:
first_block_conv = MBInvertedQConvLayer(
in_channels=max(input_channel), out_channels=max(first_block_width), kernel_size=3, stride=1,
expand_ratio=1, act_func='relu6', pw_w_bit=8, pw_a_bit=8, dw_w_bit=8, dw_a_bit=8
)
else:
first_block_conv = DynamicMBQConvLayer(
in_channel_list=input_channel, out_channel_list=first_block_width, kernel_size_list=3,
expand_ratio_list=1, stride=1, act_func='relu6',
# pw_w_bit=4, pw_a_bit=4, dw_w_bit=4, dw_a_bit=4
pw_w_bit=8, pw_a_bit=8, dw_w_bit=8, dw_a_bit=8
)
first_block = MobileInvertedResidualBlock(first_block_conv, None)
input_channel = first_block_width
# inverted residual blocks
self.block_group_info = []
blocks = [first_block]
_block_index = 1
stride_stages = [2, 2, 2, 1, 2, 1]
if depth_list is None:
n_block_list = [2, 3, 4, 3, 3, 1]
self.depth_list = [4]
else:
n_block_list = [max(self.depth_list)] * 5 + [1]
width_list = []
for base_width in base_stage_width[2:-1]:
width = [make_divisible(base_width * width_mult, 8) for width_mult in self.width_mult_list]
width_list.append(width)
for width, n_block, s in zip(width_list, n_block_list, stride_stages):
self.block_group_info.append(
([_block_index + i for i in range(n_block)], width)
)
_block_index += n_block
output_channel = width
for i in range(n_block):
if i == 0:
stride = s
else:
stride = 1
mobile_inverted_conv = DynamicMBQConvLayer(
in_channel_list=int2list(input_channel, 1), out_channel_list=int2list(output_channel, 1),
kernel_size_list=ks_list, expand_ratio_list=expand_ratio_list, stride=stride, act_func='relu6',
# pw_w_bit=4, pw_a_bit=4, dw_w_bit=4, dw_a_bit=4
pw_w_bit=8, pw_a_bit=8, dw_w_bit=8, dw_a_bit=8
)
if stride == 1 and input_channel == output_channel:
shortcut = IdentityLayer(input_channel, input_channel)
else:
shortcut = None
mb_inverted_block = MobileInvertedResidualBlock(mobile_inverted_conv, shortcut)
blocks.append(mb_inverted_block)
input_channel = output_channel
# 1x1_conv before global average pooling
if len(last_channel) == 1:
feature_mix_layer = QConvLayer(
max(input_channel), max(last_channel), kernel_size=1, use_bn=True, act_func='relu6',
w_bit=8, a_bit=8, half_wave=False
)
classifier = QLinearLayer(max(last_channel), n_classes, dropout_rate=dropout_rate, w_bit=8, a_bit=8)
else:
feature_mix_layer = DynamicMBQConvLayer(
in_channel_list=input_channel, out_channel_list=last_channel, kernel_size=1, stride=1, act_func='relu6',
# pw_w_bit=4, pw_a_bit=4, dw_w_bit=4, dw_a_bit=4
pw_w_bit=8, pw_a_bit=8, dw_w_bit=8, dw_a_bit=8, half_wave=False
)
classifier = DynamicQLinearLayer(
in_features_list=last_channel, out_features=n_classes, bias=True, dropout_rate=dropout_rate,
w_bit=8, a_bit=8
)
super(DynamicQuantizedProxylessNASNets, self).__init__(first_conv, blocks, feature_mix_layer, classifier)
# set bn param
self.set_bn_param(momentum=bn_param[0], eps=bn_param[1])
# runtime_depth
self.runtime_depth = [
len(block_idx) for block_idx, _ in self.block_group_info
]
if self.depth_ensemble_list is not None:
self.depth_ensemble_list.sort()
def __init__(self,
in_channel_list,
out_channel_list,
kernel_size_list=3,
expand_ratio_list=6,
stride=1,
act_func='relu6',
pw_w_bit=-1,
pw_a_bit=-1,
dw_w_bit=-1,
dw_a_bit=-1,
**kwargs):
super(DynamicMBQConvLayer, self).__init__()
self.in_channel_list = in_channel_list
self.out_channel_list = out_channel_list
self.kernel_size_list = int2list(kernel_size_list, 1)
self.expand_ratio_list = int2list(expand_ratio_list, 1)
self.stride = stride
self.act_func = act_func
# build modules
max_middle_channel = round(
max(self.in_channel_list) * max(self.expand_ratio_list))
if max(self.expand_ratio_list) == 1:
self.inverted_bottleneck = None
else:
self.inverted_bottleneck = nn.Sequential(
OrderedDict([
('conv',
DynamicPointQConv2d(max(self.in_channel_list),
max_middle_channel,
w_bit=pw_w_bit,
a_bit=pw_a_bit,
half_wave=False)),
('bn', DynamicBatchNorm2d(max_middle_channel)),
('act', build_activation(self.act_func, inplace=True)),
]))
self.depth_conv = nn.Sequential(
OrderedDict([('conv',
DynamicSeparableQConv2d(max_middle_channel,
self.kernel_size_list,
self.stride,
w_bit=dw_w_bit,
a_bit=dw_a_bit)),
('bn', DynamicBatchNorm2d(max_middle_channel)),
('act', build_activation(self.act_func,
inplace=True))]))
self.point_linear = nn.Sequential(
OrderedDict([
('conv',
DynamicPointQConv2d(max_middle_channel,
max(self.out_channel_list),
w_bit=pw_w_bit,
a_bit=pw_a_bit)),
('bn', DynamicBatchNorm2d(max(self.out_channel_list))),
]))
self.active_kernel_size = max(self.kernel_size_list)
self.active_expand_ratio = max(self.expand_ratio_list)
self.active_out_channel = max(self.out_channel_list)
def __init__(self,
n_classes=1000,
bn_param=(0.1, 1e-5),
dropout_rate=0.1,
base_stage_width=None,
width_mult_list=1.0,
ks_list=3,
expand_ratio_list=6,
depth_list=4):
self.width_mult_list = int2list(width_mult_list, 1)
self.ks_list = int2list(ks_list, 1)
self.expand_ratio_list = int2list(expand_ratio_list, 1)
self.depth_list = int2list(depth_list, 1)
self.base_stage_width = base_stage_width
self.width_mult_list.sort()
self.ks_list.sort()
self.expand_ratio_list.sort()
self.depth_list.sort()
base_stage_width = [16, 24, 40, 80, 112, 160, 960, 1280]
final_expand_width = [
make_divisible(base_stage_width[-2] * max(self.width_mult_list), 8)
for _ in self.width_mult_list
]
self.final_expand_width = final_expand_width
last_channel = [
make_divisible(base_stage_width[-1] * max(self.width_mult_list), 8)
for _ in self.width_mult_list
]
self.last_channel = last_channel
# stride_stages = [1, 2, 2, 2, 1, 2]
stride_stages = [1, 2, 2, 2, 1, 1]
act_stages = ['relu', 'relu', 'relu', 'h_swish', 'h_swish', 'h_swish']
se_stages = [False, False, True, False, True, True]
if depth_list is None:
n_block_list = [1, 2, 3, 4, 2, 3]
self.depth_list = [4, 4]
print('Use MobileNetV3 Depth Setting')
else:
n_block_list = [1] + [max(self.depth_list)] * 5
width_list = []
for base_width in base_stage_width[:-2]:
width = [
make_divisible(base_width * width_mult, 8)
for width_mult in self.width_mult_list
]
width_list.append(width)
input_channel = width_list[0]
# first conv layer
# if width_mult_list has only one elem
if len(set(input_channel)) == 1:
first_conv = ConvLayer(3,
max(input_channel),
kernel_size=3,
stride=2,
act_func='h_swish')
first_block_conv = MBInvertedConvLayer(
in_channels=max(input_channel),
out_channels=max(input_channel),
kernel_size=3,
stride=stride_stages[0],
expand_ratio=1,
act_func=act_stages[0],
use_se=se_stages[0],
)
else:
first_conv = DynamicConvLayer(
in_channel_list=int2list(3, len(input_channel)),
out_channel_list=input_channel,
kernel_size=3,
stride=2,
act_func='h_swish',
)
first_block_conv = DynamicMBConvLayer(
in_channel_list=input_channel,
out_channel_list=input_channel,
kernel_size_list=3,
expand_ratio_list=1,
stride=stride_stages[0],
act_func=act_stages[0],
use_se=se_stages[0],
)
first_block = MobileInvertedResidualBlock(
first_block_conv, IdentityLayer(input_channel, input_channel))
# inverted residual blocks
self.block_group_info = []
blocks = [first_block]
_block_index = 1
feature_dim = input_channel
for width, n_block, s, act_func, use_se in zip(width_list[1:],
n_block_list[1:],
stride_stages[1:],
act_stages[1:],
se_stages[1:]):
self.block_group_info.append(
[_block_index + i for i in range(n_block)])
_block_index += n_block
output_channel = width
for i in range(n_block):
if i == 0:
stride = s
else:
stride = 1
mobile_inverted_conv = DynamicMBConvLayer(
in_channel_list=feature_dim,
out_channel_list=output_channel,
kernel_size_list=ks_list,
expand_ratio_list=expand_ratio_list,
stride=stride,
act_func=act_func,
use_se=use_se,
)
if stride == 1 and feature_dim == output_channel:
shortcut = IdentityLayer(feature_dim, feature_dim)
else:
shortcut = None
blocks.append(
MobileInvertedResidualBlock(mobile_inverted_conv,
shortcut))
feature_dim = output_channel
# final expand layer, feature mix layer & classifier
if len(final_expand_width) == 1:
final_expand_layer = ConvLayer(max(feature_dim),
max(final_expand_width),
kernel_size=1,
act_func='h_swish')
feature_mix_layer = ConvLayer(
max(final_expand_width),
max(last_channel),
kernel_size=1,
bias=False,
use_bn=False,
act_func='h_swish',
)
else:
final_expand_layer = DynamicConvLayer(
in_channel_list=feature_dim,
out_channel_list=final_expand_width,
kernel_size=1,
act_func='h_swish')
feature_mix_layer = DynamicConvLayer(
in_channel_list=final_expand_width,
out_channel_list=last_channel,
kernel_size=1,
use_bn=False,
act_func='h_swish',
)
if len(set(last_channel)) == 1:
classifier = LinearLayer(max(last_channel),
n_classes,
dropout_rate=dropout_rate)
else:
classifier = DynamicLinearLayer(in_features_list=last_channel,
out_features=n_classes,
bias=True,
dropout_rate=dropout_rate)
super(OFAMobileNetV3,
self).__init__(first_conv, blocks, final_expand_layer,
feature_mix_layer, classifier)
# set bn param
self.set_bn_param(momentum=bn_param[0], eps=bn_param[1])
# runtime_depth
self.runtime_depth = [
len(block_idx) for block_idx in self.block_group_info
]
def __init__(self,
n_classes=1000,
bn_param=(0.1, 1e-5),
dropout_rate=0.1,
base_stage_width=None,
width_mult_list=1.0,
ks_list=3,
expand_ratio_list=6,
depth_list=4):
"""
Args:
n_classes: 分类类数
bn_param: bn参数
dropout_rate: 用在哪些层里面呢
width_mult_list: 在单层layer重复一些操作[~~网络基础宽度缩放 X 并不是~~]
ks_list: 卷积核的候选大小
expand_ratio_list: 网络宽度/channel数的扩大倍数
depth_list: 网络深度/layer的重复/堆叠次数
"""
# int2list 将列表，元组，整数都变为一个列表
self.width_mult_list = int2list(width_mult_list, 1)
self.ks_list = int2list(ks_list, 1)
self.expand_ratio_list = int2list(expand_ratio_list, 1)
self.depth_list = int2list(depth_list, 1)
self.base_stage_width = base_stage_width
self.width_mult_list.sort()
self.ks_list.sort()
self.expand_ratio_list.sort()
self.depth_list.sort()
base_stage_width = [16, 24, 40, 80, 112, 160, 960, 1280]
# make_divisible 使得卷积channel数为8的倍数，并以8为基底3舍4入
final_expand_width = [
make_divisible(base_stage_width[-2] * max(self.width_mult_list), 8)
for _ in self.width_mult_list
]
last_channel = [
make_divisible(base_stage_width[-1] * max(self.width_mult_list), 8)
for _ in self.width_mult_list
]
# 步长，决定下采样; 激活函数; se指的是,难道是self-attention
stride_stages = [1, 2, 2, 2, 1, 2]
act_stages = ['relu', 'relu', 'relu', 'h_swish', 'h_swish', 'h_swish']
se_stages = [False, False, True, False, True, True]
# 深度的配置除了第一个卷积，其他五层都可能expand
if depth_list is None:
n_block_list = [1, 2, 3, 4, 2, 3]
self.depth_list = [4, 4]
print('Use MobileNetV3 Depth Setting')
else:
n_block_list = [1] + [max(self.depth_list)] * 5
# 宽度/channel数配置
width_list = []
for base_width in base_stage_width[:-2]:
width = [
make_divisible(base_width * width_mult, 8)
for width_mult in self.width_mult_list
]
width_list.append(width)
# width_list好想和我想象的功能不太一样，我以为是初始channel的expand倍数
input_channel = width_list[0]
# first conv layer
if len(set(input_channel)) == 1:
first_conv = ConvLayer(3,
max(input_channel),
kernel_size=3,
stride=2,
act_func='h_swish')
first_block_conv = MBInvertedConvLayer(
in_channels=max(input_channel),
out_channels=max(input_channel),
kernel_size=3,
stride=stride_stages[0],
expand_ratio=1,
act_func=act_stages[0],
use_se=se_stages[0],
)
else:
first_conv = DynamicConvLayer(
in_channel_list=int2list(3, len(input_channel)),
out_channel_list=input_channel,
kernel_size=3,
stride=2,
act_func='h_swish',
)
first_block_conv = DynamicMBConvLayer(
in_channel_list=input_channel,
out_channel_list=input_channel,
kernel_size_list=3,
expand_ratio_list=1,
stride=stride_stages[0],
act_func=act_stages[0],
use_se=se_stages[0],
)
first_block = MobileInvertedResidualBlock(
first_block_conv, IdentityLayer(input_channel, input_channel))
# inverted residual blocks
self.block_group_info = []
blocks = [first_block]
_block_index = 1
feature_dim = input_channel
for width, n_block, s, act_func, use_se in zip(width_list[1:],
n_block_list[1:],
stride_stages[1:],
act_stages[1:],
se_stages[1:]):
self.block_group_info.append(
[_block_index + i for i in range(n_block)])
_block_index += n_block
output_channel = width
for i in range(n_block):
if i == 0:
stride = s
else:
stride = 1
mobile_inverted_conv = DynamicMBConvLayer(
in_channel_list=feature_dim,
out_channel_list=output_channel,
kernel_size_list=ks_list,
expand_ratio_list=expand_ratio_list,
stride=stride,
act_func=act_func,
use_se=use_se,
)
if stride == 1 and feature_dim == output_channel:
shortcut = IdentityLayer(feature_dim, feature_dim)
else:
shortcut = None
blocks.append(
MobileInvertedResidualBlock(mobile_inverted_conv,
shortcut))
feature_dim = output_channel
# final expand layer, feature mix layer & classifier
if len(final_expand_width) == 1:
final_expand_layer = ConvLayer(max(feature_dim),
max(final_expand_width),
kernel_size=1,
act_func='h_swish')
feature_mix_layer = ConvLayer(
max(final_expand_width),
max(last_channel),
kernel_size=1,
bias=False,
use_bn=False,
act_func='h_swish',
)
else:
final_expand_layer = DynamicConvLayer(
in_channel_list=feature_dim,
out_channel_list=final_expand_width,
kernel_size=1,
act_func='h_swish')
feature_mix_layer = DynamicConvLayer(
in_channel_list=final_expand_width,
out_channel_list=last_channel,
kernel_size=1,
use_bn=False,
act_func='h_swish',
)
if len(set(last_channel)) == 1:
classifier = LinearLayer(max(last_channel),
n_classes,
dropout_rate=dropout_rate)
else:
classifier = DynamicLinearLayer(in_features_list=last_channel,
out_features=n_classes,
bias=True,
dropout_rate=dropout_rate)
super(OFAMobileNetV3,
self).__init__(first_conv, blocks, final_expand_layer,
feature_mix_layer, classifier)
# set bn param
self.set_bn_param(momentum=bn_param[0], eps=bn_param[1])
# runtime_depth
self.runtime_depth = [
len(block_idx) for block_idx in self.block_group_info
]
