def __init__(self, n_classes=1000, bn_param=(0.1, 1e-5), dropout_rate=0, depth_list=2, expand_ratio_list=0.25, width_mult_list=1.0): self.depth_list = val2list(depth_list) self.expand_ratio_list = val2list(expand_ratio_list) self.width_mult_list = val2list(width_mult_list) # sort self.depth_list.sort() self.expand_ratio_list.sort() self.width_mult_list.sort() input_channel = [ make_divisible(64 * width_mult, MyNetwork.CHANNEL_DIVISIBLE) for width_mult in self.width_mult_list ] mid_input_channel = [ make_divisible(channel // 2, MyNetwork.CHANNEL_DIVISIBLE) for channel in input_channel ] stage_width_list = ResNets.STAGE_WIDTH_LIST.copy() for i, width in enumerate(stage_width_list): stage_width_list[i] = [ make_divisible(width * width_mult, MyNetwork.CHANNEL_DIVISIBLE) for width_mult in self.width_mult_list ] n_block_list = [base_depth + max(self.depth_list) for base_depth in ResNets.BASE_DEPTH_LIST] stride_list = [1, 2, 2, 2] # build input stem input_stem = [ DynamicConvLayer(val2list(3), mid_input_channel, 3, stride=2, use_bn=True, act_func='relu'), ResidualBlock( DynamicConvLayer(mid_input_channel, mid_input_channel, 3, stride=1, use_bn=True, act_func='relu'), IdentityLayer(mid_input_channel, mid_input_channel) ), DynamicConvLayer(mid_input_channel, input_channel, 3, stride=1, use_bn=True, act_func='relu') ] # blocks blocks = [] for d, width, s in zip(n_block_list, stage_width_list, stride_list): for i in range(d): stride = s if i == 0 else 1 bottleneck_block = DynamicResNetBottleneckBlock( input_channel, width, expand_ratio_list=self.expand_ratio_list, kernel_size=3, stride=stride, act_func='relu', downsample_mode='avgpool_conv', ) blocks.append(bottleneck_block) input_channel = width # classifier classifier = DynamicLinearLayer(input_channel, n_classes, dropout_rate=dropout_rate) super(OFAResNets, self).__init__(input_stem, blocks, classifier) # set bn param self.set_bn_param(*bn_param) # runtime_depth self.input_stem_skipping = 0 self.runtime_depth = [0] * len(n_block_list)
def re_organize_middle_weights(self, expand_ratio_stage=0):
# conv3 -> conv2
importance = torch.sum(torch.abs(self.conv3.conv.conv.weight.data), dim=(0, 2, 3))
if isinstance(self.conv2.bn, DynamicGroupNorm):
channel_per_group = self.conv2.bn.channel_per_group
importance_chunks = torch.split(importance, channel_per_group)
for chunk in importance_chunks:
chunk.data.fill_(torch.mean(chunk))
importance = torch.cat(importance_chunks, dim=0)
if expand_ratio_stage > 0:
sorted_expand_list = copy.deepcopy(self.expand_ratio_list)
sorted_expand_list.sort(reverse=True)
target_width_list = [
make_divisible(round(max(self.out_channel_list) * expand), MyNetwork.CHANNEL_DIVISIBLE)
for expand in sorted_expand_list
]
right = len(importance)
base = - len(target_width_list) * 1e5
for i in range(expand_ratio_stage + 1):
left = target_width_list[i]
importance[left:right] += base
base += 1e5
right = left
sorted_importance, sorted_idx = torch.sort(importance, dim=0, descending=True)
self.conv3.conv.conv.weight.data = torch.index_select(self.conv3.conv.conv.weight.data, 1, sorted_idx)
adjust_bn_according_to_idx(self.conv2.bn.bn, sorted_idx)
self.conv2.conv.conv.weight.data = torch.index_select(self.conv2.conv.conv.weight.data, 0, sorted_idx)
# conv2 -> conv1
importance = torch.sum(torch.abs(self.conv2.conv.conv.weight.data), dim=(0, 2, 3))
if isinstance(self.conv1.bn, DynamicGroupNorm):
channel_per_group = self.conv1.bn.channel_per_group
importance_chunks = torch.split(importance, channel_per_group)
for chunk in importance_chunks:
chunk.data.fill_(torch.mean(chunk))
importance = torch.cat(importance_chunks, dim=0)
if expand_ratio_stage > 0:
sorted_expand_list = copy.deepcopy(self.expand_ratio_list)
sorted_expand_list.sort(reverse=True)
target_width_list = [
make_divisible(round(max(self.out_channel_list) * expand), MyNetwork.CHANNEL_DIVISIBLE)
for expand in sorted_expand_list
]
right = len(importance)
base = - len(target_width_list) * 1e5
for i in range(expand_ratio_stage + 1):
left = target_width_list[i]
importance[left:right] += base
base += 1e5
right = left
sorted_importance, sorted_idx = torch.sort(importance, dim=0, descending=True)
self.conv2.conv.conv.weight.data = torch.index_select(self.conv2.conv.conv.weight.data, 1, sorted_idx)
adjust_bn_according_to_idx(self.conv1.bn.bn, sorted_idx)
self.conv1.conv.conv.weight.data = torch.index_select(self.conv1.conv.conv.weight.data, 0, sorted_idx)
return None
def re_organize_middle_weights(self, expand_ratio_stage=0):
importance = torch.sum(torch.abs(self.point_linear.conv.conv.weight.data), dim=(0, 2, 3))
if isinstance(self.depth_conv.bn, DynamicGroupNorm):
channel_per_group = self.depth_conv.bn.channel_per_group
importance_chunks = torch.split(importance, channel_per_group)
for chunk in importance_chunks:
chunk.data.fill_(torch.mean(chunk))
importance = torch.cat(importance_chunks, dim=0)
if expand_ratio_stage > 0:
sorted_expand_list = copy.deepcopy(self.expand_ratio_list)
sorted_expand_list.sort(reverse=True)
target_width_list = [
make_divisible(round(max(self.in_channel_list) * expand), MyNetwork.CHANNEL_DIVISIBLE)
for expand in sorted_expand_list
]
right = len(importance)
base = - len(target_width_list) * 1e5
for i in range(expand_ratio_stage + 1):
left = target_width_list[i]
importance[left:right] += base
base += 1e5
right = left
sorted_importance, sorted_idx = torch.sort(importance, dim=0, descending=True)
self.point_linear.conv.conv.weight.data = torch.index_select(
self.point_linear.conv.conv.weight.data, 1, sorted_idx
)
adjust_bn_according_to_idx(self.depth_conv.bn.bn, sorted_idx)
self.depth_conv.conv.conv.weight.data = torch.index_select(
self.depth_conv.conv.conv.weight.data, 0, sorted_idx
)
if self.use_se:
# se expand: output dim 0 reorganize
se_expand = self.depth_conv.se.fc.expand
se_expand.weight.data = torch.index_select(se_expand.weight.data, 0, sorted_idx)
se_expand.bias.data = torch.index_select(se_expand.bias.data, 0, sorted_idx)
# se reduce: input dim 1 reorganize
se_reduce = self.depth_conv.se.fc.reduce
se_reduce.weight.data = torch.index_select(se_reduce.weight.data, 1, sorted_idx)
# middle weight reorganize
se_importance = torch.sum(torch.abs(se_expand.weight.data), dim=(0, 2, 3))
se_importance, se_idx = torch.sort(se_importance, dim=0, descending=True)
se_expand.weight.data = torch.index_select(se_expand.weight.data, 1, se_idx)
se_reduce.weight.data = torch.index_select(se_reduce.weight.data, 0, se_idx)
se_reduce.bias.data = torch.index_select(se_reduce.bias.data, 0, se_idx)
if self.inverted_bottleneck is not None:
adjust_bn_according_to_idx(self.inverted_bottleneck.bn.bn, sorted_idx)
self.inverted_bottleneck.conv.conv.weight.data = torch.index_select(
self.inverted_bottleneck.conv.conv.weight.data, 0, sorted_idx
)
return None
else:
return sorted_idx
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 = val2list(kernel_size_list)
self.expand_ratio_list = val2list(expand_ratio_list)
self.stride = stride
self.act_func = act_func
self.use_se = use_se
# build modules
max_middle_channel = make_divisible(
round(max(self.in_channel_list) * max(self.expand_ratio_list)),
MyNetwork.CHANNEL_DIVISIBLE)
if max(self.expand_ratio_list) == 1:
self.inverted_bottleneck = None
else:
self.inverted_bottleneck = nn.Sequential(
OrderedDict([
('conv',
DynamicConv2d(max(self.in_channel_list),
max_middle_channel)),
('bn', DynamicBatchNorm2d(max_middle_channel)),
('act', build_activation(self.act_func)),
]))
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))]))
if self.use_se:
self.depth_conv.add_module('se', DynamicSE(max_middle_channel))
self.point_linear = nn.Sequential(
OrderedDict([
('conv',
DynamicConv2d(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 __init__(self, in_channel_list, out_channel_list, expand_ratio_list=0.25,
kernel_size=3, stride=1, act_func='relu', downsample_mode='avgpool_conv'):
super(DynamicResNetBottleneckBlock, self).__init__()
self.in_channel_list = in_channel_list
self.out_channel_list = out_channel_list
self.expand_ratio_list = val2list(expand_ratio_list)
self.kernel_size = kernel_size
self.stride = stride
self.act_func = act_func
self.downsample_mode = downsample_mode
# build modules
max_middle_channel = make_divisible(
round(max(self.out_channel_list) * max(self.expand_ratio_list)), MyNetwork.CHANNEL_DIVISIBLE)
self.conv1 = nn.Sequential(OrderedDict([
('conv', DynamicConv2d(max(self.in_channel_list), max_middle_channel)),
('bn', DynamicBatchNorm2d(max_middle_channel)),
('act', build_activation(self.act_func, inplace=True)),
]))
self.conv2 = nn.Sequential(OrderedDict([
('conv', DynamicConv2d(max_middle_channel, max_middle_channel, kernel_size, stride)),
('bn', DynamicBatchNorm2d(max_middle_channel)),
('act', build_activation(self.act_func, inplace=True))
]))
self.conv3 = nn.Sequential(OrderedDict([
('conv', DynamicConv2d(max_middle_channel, max(self.out_channel_list))),
('bn', DynamicBatchNorm2d(max(self.out_channel_list))),
]))
if self.stride == 1 and self.in_channel_list == self.out_channel_list:
self.downsample = IdentityLayer(max(self.in_channel_list), max(self.out_channel_list))
elif self.downsample_mode == 'conv':
self.downsample = nn.Sequential(OrderedDict([
('conv', DynamicConv2d(max(self.in_channel_list), max(self.out_channel_list), stride=stride)),
('bn', DynamicBatchNorm2d(max(self.out_channel_list))),
]))
elif self.downsample_mode == 'avgpool_conv':
self.downsample = nn.Sequential(OrderedDict([
('avg_pool', nn.AvgPool2d(kernel_size=stride, stride=stride, padding=0, ceil_mode=True)),
('conv', DynamicConv2d(max(self.in_channel_list), max(self.out_channel_list))),
('bn', DynamicBatchNorm2d(max(self.out_channel_list))),
]))
else:
raise NotImplementedError
self.final_act = build_activation(self.act_func, inplace=True)
self.active_expand_ratio = max(self.expand_ratio_list)
self.active_out_channel = max(self.out_channel_list)
def count_flops_given_config(net_config, image_size=224):
flops = 0
# first conv
flops += count_conv_flop((image_size + 1) // 2, 3,
net_config['first_conv']['out_channels'], 3,
1)
# blocks
fsize = (image_size + 1) // 2
for block in net_config['blocks']:
mb_conv = block[
'mobile_inverted_conv'] if 'mobile_inverted_conv' in block else block[
'conv']
if mb_conv is None:
continue
out_fz = int((fsize - 1) / mb_conv['stride'] + 1)
if mb_conv['mid_channels'] is None:
mb_conv['mid_channels'] = round(mb_conv['in_channels'] *
mb_conv['expand_ratio'])
if mb_conv['expand_ratio'] != 1:
# inverted bottleneck
flops += count_conv_flop(fsize, mb_conv['in_channels'],
mb_conv['mid_channels'], 1, 1)
# depth conv
flops += count_conv_flop(out_fz, mb_conv['mid_channels'],
mb_conv['mid_channels'],
mb_conv['kernel_size'],
mb_conv['mid_channels'])
if mb_conv['use_se']:
# SE layer
se_mid = make_divisible(mb_conv['mid_channels'] // 4,
divisor=MyNetwork.CHANNEL_DIVISIBLE)
flops += count_conv_flop(1, mb_conv['mid_channels'], se_mid, 1,
1)
flops += count_conv_flop(1, se_mid, mb_conv['mid_channels'], 1,
1)
# point linear
flops += count_conv_flop(out_fz, mb_conv['mid_channels'],
mb_conv['out_channels'], 1, 1)
fsize = out_fz
# final expand layer
flops += count_conv_flop(
fsize, net_config['final_expand_layer']['in_channels'],
net_config['final_expand_layer']['out_channels'], 1, 1)
# feature mix layer
flops += count_conv_flop(
1, net_config['feature_mix_layer']['in_channels'],
net_config['feature_mix_layer']['out_channels'], 1, 1)
# classifier
flops += count_conv_flop(1, net_config['classifier']['in_features'],
net_config['classifier']['out_features'], 1,
1)
return flops / 1e6 # MFLOPs
def forward(self, x):
in_channel = x.size(1)
if self.inverted_bottleneck is not None:
self.inverted_bottleneck.conv.active_out_channel = \
make_divisible(round(in_channel * self.active_expand_ratio), MyNetwork.CHANNEL_DIVISIBLE)
self.depth_conv.conv.active_kernel_size = self.active_kernel_size
self.point_linear.conv.active_out_channel = self.active_out_channel
if self.inverted_bottleneck is not None:
x = self.inverted_bottleneck(x)
x = self.depth_conv(x)
x = self.point_linear(x)
return x
def count_flops_given_config(net_config, image_size=224):
flops = 0
# input stem
for layer_config in net_config['input_stem']:
if layer_config['name'] != 'ConvLayer':
layer_config = layer_config['conv']
in_channel = layer_config['in_channels']
out_channel = layer_config['out_channels']
out_image_size = int((image_size - 1) / layer_config['stride'] + 1)
flops += count_conv_flop(out_image_size, in_channel, out_channel,
layer_config['kernel_size'],
layer_config.get('groups', 1))
image_size = out_image_size
# max pooling
image_size = int((image_size - 1) / 2 + 1)
# ResNetBottleneckBlocks
for block_config in net_config['blocks']:
in_channel = block_config['in_channels']
out_channel = block_config['out_channels']
out_image_size = int((image_size - 1) / block_config['stride'] + 1)
mid_channel = block_config['mid_channels'] if block_config['mid_channels'] is not None \
else round(out_channel * block_config['expand_ratio'])
mid_channel = make_divisible(mid_channel,
MyNetwork.CHANNEL_DIVISIBLE)
# conv1
flops += count_conv_flop(image_size, in_channel, mid_channel, 1, 1)
# conv2
flops += count_conv_flop(out_image_size, mid_channel, mid_channel,
block_config['kernel_size'],
block_config['groups'])
# conv3
flops += count_conv_flop(out_image_size, mid_channel, out_channel,
1, 1)
# downsample
if block_config['stride'] == 1 and in_channel == out_channel:
pass
else:
flops += count_conv_flop(out_image_size, in_channel,
out_channel, 1, 1)
image_size = out_image_size
# final classifier
flops += count_conv_flop(1, net_config['classifier']['in_features'],
net_config['classifier']['out_features'], 1,
1)
return flops / 1e6 # MFLOPs
def forward(self, x, groups=None): in_channel = x.size(1) num_mid = make_divisible(in_channel // self.reduction, divisor=MyNetwork.CHANNEL_DIVISIBLE) y = x.mean(3, keepdim=True).mean(2, keepdim=True) # reduce reduce_filter = self.get_active_reduce_weight(num_mid, in_channel, groups=groups).contiguous() reduce_bias = self.get_active_reduce_bias(num_mid) y = F.conv2d(y, reduce_filter, reduce_bias, 1, 0, 1, 1) # relu y = self.fc.relu(y) # expand expand_filter = self.get_active_expand_weight(num_mid, in_channel, groups=groups).contiguous() expand_bias = self.get_active_expand_bias(in_channel, groups=groups) y = F.conv2d(y, expand_filter, expand_bias, 1, 0, 1, 1) # hard sigmoid y = self.fc.h_sigmoid(y) return x * y
def get_active_subnet(self, in_channel, preserve_weight=True):
# build the new layer
sub_layer = set_layer_from_config(
self.get_active_subnet_config(in_channel))
sub_layer = sub_layer.to(get_net_device(self))
if not preserve_weight:
return sub_layer
middle_channel = self.active_middle_channel(in_channel)
# copy weight from current layer
if sub_layer.inverted_bottleneck is not None:
sub_layer.inverted_bottleneck.conv.weight.data.copy_(
self.inverted_bottleneck.conv.get_active_filter(
middle_channel, in_channel).data, )
copy_bn(sub_layer.inverted_bottleneck.bn,
self.inverted_bottleneck.bn.bn)
sub_layer.depth_conv.conv.weight.data.copy_(
self.depth_conv.conv.get_active_filter(
middle_channel, self.active_kernel_size).data)
copy_bn(sub_layer.depth_conv.bn, self.depth_conv.bn.bn)
if self.use_se:
se_mid = make_divisible(middle_channel // SEModule.REDUCTION,
divisor=MyNetwork.CHANNEL_DIVISIBLE)
sub_layer.depth_conv.se.fc.reduce.weight.data.copy_(
self.depth_conv.se.get_active_reduce_weight(
se_mid, middle_channel).data)
sub_layer.depth_conv.se.fc.reduce.bias.data.copy_(
self.depth_conv.se.get_active_reduce_bias(se_mid).data)
sub_layer.depth_conv.se.fc.expand.weight.data.copy_(
self.depth_conv.se.get_active_expand_weight(
se_mid, middle_channel).data)
sub_layer.depth_conv.se.fc.expand.bias.data.copy_(
self.depth_conv.se.get_active_expand_bias(middle_channel).data)
sub_layer.point_linear.conv.weight.data.copy_(
self.point_linear.conv.get_active_filter(self.active_out_channel,
middle_channel).data)
copy_bn(sub_layer.point_linear.bn, self.point_linear.bn.bn)
return sub_layer
def __init__(self,
n_classes=1000,
width_mult=1.0,
bn_param=(0.1, 1e-3),
dropout_rate=0.2,
ks=None,
expand_ratio=None,
depth_param=None,
stage_width_list=None):
ks = 3 if ks is None else ks
expand_ratio = 6 if expand_ratio is None else expand_ratio
input_channel = 32
last_channel = 1280
input_channel = make_divisible(input_channel * width_mult,
MyNetwork.CHANNEL_DIVISIBLE)
last_channel = make_divisible(last_channel * width_mult, MyNetwork.CHANNEL_DIVISIBLE) \
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 = val2list(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,
MyNetwork.CHANNEL_DIVISIBLE)
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 = MBConvLayer(
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(ResidualBlock(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(*bn_param)
def __init__(self,
n_classes=1000,
width_mult=1.0,
bn_param=(0.1, 1e-5),
dropout_rate=0,
expand_ratio=None,
depth_param=None):
expand_ratio = 0.25 if expand_ratio is None else expand_ratio
input_channel = make_divisible(64 * width_mult,
MyNetwork.CHANNEL_DIVISIBLE)
mid_input_channel = make_divisible(input_channel // 2,
MyNetwork.CHANNEL_DIVISIBLE)
stage_width_list = ResNets.STAGE_WIDTH_LIST.copy()
for i, width in enumerate(stage_width_list):
stage_width_list[i] = make_divisible(width * width_mult,
MyNetwork.CHANNEL_DIVISIBLE)
depth_list = [3, 4, 6, 3]
if depth_param is not None:
for i, depth in enumerate(ResNets.BASE_DEPTH_LIST):
depth_list[i] = depth + depth_param
stride_list = [1, 2, 2, 2]
# build input stem
input_stem = [
ConvLayer(3,
mid_input_channel,
3,
stride=2,
use_bn=True,
act_func='relu'),
ResidualBlock(
ConvLayer(mid_input_channel,
mid_input_channel,
3,
stride=1,
use_bn=True,
act_func='relu'),
IdentityLayer(mid_input_channel, mid_input_channel)),
ConvLayer(mid_input_channel,
input_channel,
3,
stride=1,
use_bn=True,
act_func='relu')
]
# blocks
blocks = []
for d, width, s in zip(depth_list, stage_width_list, stride_list):
for i in range(d):
stride = s if i == 0 else 1
bottleneck_block = ResNetBottleneckBlock(
input_channel,
width,
kernel_size=3,
stride=stride,
expand_ratio=expand_ratio,
act_func='relu',
downsample_mode='avgpool_conv',
)
blocks.append(bottleneck_block)
input_channel = width
# classifier
classifier = LinearLayer(input_channel,
n_classes,
dropout_rate=dropout_rate)
super(ResNet50D, self).__init__(input_stem, blocks, classifier)
# set bn param
self.set_bn_param(*bn_param)
def __init__(self,
main_branch,
in_channels,
out_channels,
expand=1.0,
kernel_size=3,
act_func='relu',
n_groups=2,
downsample_ratio=2,
upsample_type='bilinear',
stride=1):
super(LiteResidualModule, self).__init__()
self.main_branch = main_branch
self.lite_residual_config = {
'in_channels': in_channels,
'out_channels': out_channels,
'expand': expand,
'kernel_size': kernel_size,
'act_func': act_func,
'n_groups': n_groups,
'downsample_ratio': downsample_ratio,
'upsample_type': upsample_type,
'stride': stride,
}
kernel_size = 1 if downsample_ratio is None else kernel_size
padding = get_same_padding(kernel_size)
if downsample_ratio is None:
pooling = MyGlobalAvgPool2d()
else:
pooling = nn.AvgPool2d(downsample_ratio, downsample_ratio, 0)
num_mid = make_divisible(int(in_channels * expand),
divisor=MyNetwork.CHANNEL_DIVISIBLE)
self.lite_residual = nn.Sequential(
OrderedDict({
'pooling':
pooling,
'conv1':
nn.Conv2d(in_channels,
num_mid,
kernel_size,
stride,
padding,
groups=n_groups,
bias=False),
'bn1':
nn.BatchNorm2d(num_mid),
'act':
build_activation(act_func),
'conv2':
nn.Conv2d(num_mid, out_channels, 1, 1, 0, bias=False),
'final_bn':
nn.BatchNorm2d(out_channels),
}))
# initialize
init_models(self.lite_residual)
self.lite_residual.final_bn.weight.data.zero_()
#for i, stage in enumerate(stages):
# depth = sample['d'][i]
# kernels = sample['ks'][4*i: 4*(i+1)]
# expand_ratios = sample['e'][4*i: 4*(i+1)]
out_channels = [model.blocks[i].mobile_inverted_conv.point_linear.conv.conv.out_channels for i in range(1, 21)]
out_channels = [model.blocks[0].mobile_inverted_conv.point_linear.conv.out_channels] + out_channels
l2_squared = np.zeros([20, 3, 3], dtype=np.float)
for i in range(20):
for k in [3, 5, 7]:
for e in [3, 4, 6]:
l2 = 0.0
module = model.blocks[i+1].mobile_inverted_conv
in_channel = out_channels[i]
mid_channels = make_divisible(round(in_channel * e), 8)
if module.inverted_bottleneck is not None:
l2 += torch.norm(module.inverted_bottleneck.conv.conv.weight[:mid_channels, :in_channel, :, :], p=1)
l2 += torch.norm(module.inverted_bottleneck.bn.bn.weight[:mid_channels], p=1)
if module.inverted_bottleneck.bn.bn.bias is not None:
l2 += torch.norm(module.inverted_bottleneck.bn.bn.bias[:mid_channels], p=1)
l2 += torch.norm(module.depth_conv.conv.get_active_filter(mid_channels, k), p=1)
l2 += torch.norm(module.depth_conv.bn.bn.weight[:mid_channels], p=1)
if module.depth_conv.bn.bn.bias is not None:
l2 += torch.norm(module.depth_conv.bn.bn.bias[:mid_channels], p=1)
if hasattr(module.depth_conv, 'se'):
se_channel = make_divisible(mid_channels // module.depth_conv.se.reduction, divisor=8)
l2 += torch.norm(module.depth_conv.se.fc.reduce.weight[:se_channel, :mid_channels, :, :], p=1)
if module.depth_conv.se.fc.reduce.bias is not None:
def __init__(self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
expand_ratio=0.25,
mid_channels=None,
act_func='relu',
groups=1,
downsample_mode='avgpool_conv'):
super(ResNetBottleneckBlock, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.expand_ratio = expand_ratio
self.mid_channels = mid_channels
self.act_func = act_func
self.groups = groups
self.downsample_mode = downsample_mode
if self.mid_channels is None:
feature_dim = round(self.out_channels * self.expand_ratio)
else:
feature_dim = self.mid_channels
feature_dim = make_divisible(feature_dim, MyNetwork.CHANNEL_DIVISIBLE)
self.mid_channels = feature_dim
# build modules
self.conv1 = nn.Sequential(
OrderedDict([
('conv',
nn.Conv2d(self.in_channels, feature_dim, 1, 1, 0,
bias=False)),
('bn', nn.BatchNorm2d(feature_dim)),
('act', build_activation(self.act_func, inplace=True)),
]))
pad = get_same_padding(self.kernel_size)
self.conv2 = nn.Sequential(
OrderedDict([('conv',
nn.Conv2d(feature_dim,
feature_dim,
kernel_size,
stride,
pad,
groups=groups,
bias=False)),
('bn', nn.BatchNorm2d(feature_dim)),
('act', build_activation(self.act_func,
inplace=True))]))
self.conv3 = nn.Sequential(
OrderedDict([
('conv',
nn.Conv2d(feature_dim, self.out_channels, 1, 1, 0,
bias=False)),
('bn', nn.BatchNorm2d(self.out_channels)),
]))
if stride == 1 and in_channels == out_channels:
self.downsample = IdentityLayer(in_channels, out_channels)
elif self.downsample_mode == 'conv':
self.downsample = nn.Sequential(
OrderedDict([
('conv',
nn.Conv2d(in_channels,
out_channels,
1,
stride,
0,
bias=False)),
('bn', nn.BatchNorm2d(out_channels)),
]))
elif self.downsample_mode == 'avgpool_conv':
self.downsample = nn.Sequential(
OrderedDict([
('avg_pool',
nn.AvgPool2d(kernel_size=stride,
stride=stride,
padding=0,
ceil_mode=True)),
('conv',
nn.Conv2d(in_channels, out_channels, 1, 1, 0,
bias=False)),
('bn', nn.BatchNorm2d(out_channels)),
]))
else:
raise NotImplementedError
self.final_act = build_activation(self.act_func, inplace=True)
def __init__(self, n_classes=1000, bn_param=(0.1, 1e-3), dropout_rate=0.1, base_stage_width=None, width_mult=1.0, ks_list=3, expand_ratio_list=6, depth_list=4): self.width_mult = width_mult self.ks_list = val2list(ks_list, 1) self.expand_ratio_list = val2list(expand_ratio_list, 1) self.depth_list = val2list(depth_list, 1) self.ks_list.sort() self.expand_ratio_list.sort() self.depth_list.sort() if base_stage_width == 'google': # MobileNetV2 Stage Width base_stage_width = [32, 16, 24, 32, 64, 96, 160, 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] * self.width_mult, MyNetwork.CHANNEL_DIVISIBLE) first_block_width = make_divisible(base_stage_width[1] * self.width_mult, MyNetwork.CHANNEL_DIVISIBLE) last_channel = make_divisible(base_stage_width[-1] * self.width_mult, MyNetwork.CHANNEL_DIVISIBLE) # 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' ) # first block first_block_conv = MBConvLayer( in_channels=input_channel, out_channels=first_block_width, kernel_size=3, stride=1, expand_ratio=1, act_func='relu6', ) first_block = ResidualBlock(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] 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 * self.width_mult, MyNetwork.CHANNEL_DIVISIBLE) 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)]) _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=val2list(input_channel, 1), out_channel_list=val2list(output_channel, 1), kernel_size_list=ks_list, expand_ratio_list=expand_ratio_list, stride=stride, act_func='relu6', ) if stride == 1 and input_channel == output_channel: shortcut = IdentityLayer(input_channel, input_channel) else: shortcut = None mb_inverted_block = ResidualBlock(mobile_inverted_conv, shortcut) blocks.append(mb_inverted_block) 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', ) classifier = LinearLayer(last_channel, n_classes, dropout_rate=dropout_rate) super(OFAProxylessNASNets, 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]
def __init__(self,
n_classes=1000,
bn_param=(0.1, 1e-5),
dropout_rate=0.1,
base_stage_width=None,
width_mult=1.0,
ks_list=3,
expand_ratio_list=6,
depth_list=4):
self.width_mult = width_mult
self.ks_list = val2list(ks_list, 1)
self.expand_ratio_list = val2list(expand_ratio_list, 1)
self.depth_list = val2list(depth_list, 1)
self.ks_list.sort()
self.expand_ratio_list.sort()
self.depth_list.sort()
base_stage_width = [16, 16, 24, 40, 80, 112, 160, 960, 1280]
final_expand_width = make_divisible(
base_stage_width[-2] * self.width_mult,
MyNetwork.CHANNEL_DIVISIBLE)
last_channel = make_divisible(base_stage_width[-1] * self.width_mult,
MyNetwork.CHANNEL_DIVISIBLE)
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]
n_block_list = [1] + [max(self.depth_list)] * 5
width_list = []
for base_width in base_stage_width[:-2]:
width = make_divisible(base_width * self.width_mult,
MyNetwork.CHANNEL_DIVISIBLE)
width_list.append(width)
input_channel, first_block_dim = width_list[0], width_list[1]
# first conv layer
first_conv = ConvLayer(3,
input_channel,
kernel_size=3,
stride=2,
act_func='h_swish')
first_block_conv = MBConvLayer(
in_channels=input_channel,
out_channels=first_block_dim,
kernel_size=3,
stride=stride_stages[0],
expand_ratio=1,
act_func=act_stages[0],
use_se=se_stages[0],
)
first_block = ResidualBlock(
first_block_conv,
IdentityLayer(first_block_dim, first_block_dim)
if input_channel == first_block_dim else None,
)
# inverted residual blocks
self.block_group_info = []
blocks = [first_block]
_block_index = 1
feature_dim = first_block_dim
for width, n_block, s, act_func, use_se in zip(width_list[2:],
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=val2list(feature_dim),
out_channel_list=val2list(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(ResidualBlock(mobile_inverted_conv, shortcut))
feature_dim = output_channel
# final expand layer, feature mix layer & classifier
final_expand_layer = ConvLayer(feature_dim,
final_expand_width,
kernel_size=1,
act_func='h_swish')
feature_mix_layer = ConvLayer(
final_expand_width,
last_channel,
kernel_size=1,
bias=False,
use_bn=False,
act_func='h_swish',
)
classifier = LinearLayer(last_channel,
n_classes,
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 active_middle_channel(self, in_channel):
return make_divisible(round(in_channel * self.active_expand_ratio), MyNetwork.CHANNEL_DIVISIBLE)
def __init__(self,
n_classes=1000,
width_mult=1.0,
bn_param=(0.1, 1e-5),
dropout_rate=0.2,
ks=None,
expand_ratio=None,
depth_param=None,
stage_width_list=None):
input_channel = 16
last_channel = 1280
input_channel = make_divisible(input_channel * width_mult,
MyNetwork.CHANNEL_DIVISIBLE)
last_channel = make_divisible(last_channel * width_mult, MyNetwork.CHANNEL_DIVISIBLE) \
if width_mult > 1.0 else last_channel
cfg = {
# k, exp, c, se, nl, s, e,
'0': [
[3, 16, 16, False, 'relu', 1, 1],
],
'1': [
[3, 64, 24, False, 'relu', 2, None], # 4
[3, 72, 24, False, 'relu', 1, None], # 3
],
'2': [
[5, 72, 40, True, 'relu', 2, None], # 3
[5, 120, 40, True, 'relu', 1, None], # 3
[5, 120, 40, True, 'relu', 1, None], # 3
],
'3': [
[3, 240, 80, False, 'h_swish', 2, None], # 6
[3, 200, 80, False, 'h_swish', 1, None], # 2.5
[3, 184, 80, False, 'h_swish', 1, None], # 2.3
[3, 184, 80, False, 'h_swish', 1, None], # 2.3
],
'4': [
[3, 480, 112, True, 'h_swish', 1, None], # 6
[3, 672, 112, True, 'h_swish', 1, None], # 6
],
'5': [
[5, 672, 160, True, 'h_swish', 2, None], # 6
[5, 960, 160, True, 'h_swish', 1, None], # 6
[5, 960, 160, True, 'h_swish', 1, None], # 6
]
}
cfg = self.adjust_cfg(cfg, ks, expand_ratio, depth_param,
stage_width_list)
# width multiplier on mobile setting, change `exp: 1` and `c: 2`
for stage_id, block_config_list in cfg.items():
for block_config in block_config_list:
if block_config[1] is not None:
block_config[1] = make_divisible(
block_config[1] * width_mult,
MyNetwork.CHANNEL_DIVISIBLE)
block_config[2] = make_divisible(block_config[2] * width_mult,
MyNetwork.CHANNEL_DIVISIBLE)
first_conv, blocks, final_expand_layer, feature_mix_layer, classifier = self.build_net_via_cfg(
cfg, input_channel, last_channel, n_classes, dropout_rate)
super(MobileNetV3Large,
self).__init__(first_conv, blocks, final_expand_layer,
feature_mix_layer, classifier)
# set bn param
self.set_bn_param(*bn_param)
def active_middle_channels(self):
feature_dim = round(self.active_out_channel * self.active_expand_ratio)
feature_dim = make_divisible(feature_dim, MyNetwork.CHANNEL_DIVISIBLE)
return feature_dim
