def set_active_subnet(self, d=None, e=None, w=None, **kwargs):
depth = val2list(d, len(ResNets.BASE_DEPTH_LIST) + 1)
expand_ratio = val2list(e, len(self.blocks))
width_mult = val2list(w, len(ResNets.BASE_DEPTH_LIST) + 2)
for block, e in zip(self.blocks, expand_ratio):
if e is not None:
block.active_expand_ratio = e
if width_mult[0] is not None:
self.input_stem[1].conv.active_out_channel = self.input_stem[0].active_out_channel = \
self.input_stem[0].out_channel_list[width_mult[0]]
if width_mult[1] is not None:
self.input_stem[2].active_out_channel = self.input_stem[
2].out_channel_list[width_mult[1]]
if depth[0] is not None:
self.input_stem_skipping = (depth[0] != max(self.depth_list))
for stage_id, (block_idx, d, w) in enumerate(
zip(self.grouped_block_index, depth[1:], width_mult[2:])):
if d is not None:
self.runtime_depth[stage_id] = max(self.depth_list) - d
if w is not None:
for idx in block_idx:
self.blocks[idx].active_out_channel = self.blocks[
idx].out_channel_list[w]
def set_active_subnet(self, ks=None, e=None, d=None, **kwargs): ks = val2list(ks, len(self.blocks) - 1) expand_ratio = val2list(e, len(self.blocks) - 1) depth = val2list(d, len(self.block_group_info)) for block, k, e in zip(self.blocks[1:], ks, expand_ratio): if k is not None: block.conv.active_kernel_size = k if e is not None: block.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=False, image_size_list=None,
ks_list=None, expand_ratio_list=None, depth_list=None, width_mult_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 = val2list(run_manager.run_config.data_provider.image_size, 1)
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
if width_mult_list is None:
if 'width_mult_list' in dynamic_net.__dict__:
width_mult_list = list(range(len(dynamic_net.width_mult_list)))
else:
width_mult_list = [0]
subnet_settings = []
for d in depth_list:
for e in expand_ratio_list:
for k in ks_list:
for w in width_mult_list:
for img_size in image_size_list:
subnet_settings.append([{
'image_size': img_size,
'd': d,
'e': e,
'ks': k,
'w': w,
}, 'R%s-D%s-E%s-K%s-W%s' % (img_size, d, e, k, w)])
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, 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 __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 __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 __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, 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)