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)