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_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 set_active_subnet(self, wid=None, ks=None, e=None, d=None, pixel_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) - 2) # 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) ks = int2list(ks, len(self.blocks) - 4) expand_ratio = int2list(e, len(self.blocks) - 4) depth = int2list(d, len(self.block_group_info) - 2) pixelshuffle_depth = int2list(pixel_d, 2) depth.insert(0, pixelshuffle_depth[0]) depth.insert(-1, pixelshuffle_depth[0]) for block, k, e in zip(self.blocks[2:-2], 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 __init__(self, 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, pixelshuffle_depth_list=2): self.width_mult_list = int2list(width_mult_list, 1) # 이게 output width 조절하는 변수 self.ks_list = int2list(ks_list, 1) self.expand_ratio_list = int2list(expand_ratio_list, 1) self.depth_list = int2list(depth_list, 1) self.pixelshuffle_depth_list = int2list(pixelshuffle_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() self.pixelshuffle_depth_list.sort() # FROM [3,64 64, 64, 64, 64, 64, 64, 64,64, 64, 64, 64, 64, 64, 64, 64, 64] base_stage_width = [ 16, 64, 64, 64, 64, 64, 64, 3, 64, 64, 64, 64, 64, 64, 64, 256, 3 ] # [Unshu ResBlock ResCon ResBlock ResCon Shu] # [2, 4, 4, 4, 4, 1, 1, 1, 1, 4, 4, 4, 4, 1, 1, 2, 1] # [ Skip, Con, Skip, Con] # 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 # ] stride_stages = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] act_stages = [ 'pixelunshuffle', 'relu6', 'relu6', 'relu6', 'relu6', None, None, None, None, 'relu6', 'relu6', 'relu6', 'relu6', None, None, 'pixelshuffle', None ] se_stages = [ False, False, False, False, False, False, False, False, False, False, False, False, False, False, False, False, False ] 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 = [ max(self.pixelshuffle_depth_list) ] + [max(self.depth_list)] * 4 + [1] * 4 + [max( self.depth_list)] * 4 + [1] * 2 + [ max(self.pixelshuffle_depth_list) ] + [1] # 2는 pixelshuffle, pixelunshuffle의 depth # [2, 4, 4, 1, 1, 1, 1, 1, 4, 4, 1, 1, 2, 1] width_list = [] for base_width in base_stage_width: # width = [make_divisible(base_width * width_mult, 8) for width_mult in self.width_mult_list] width = [ make_divisible(base_width * width_mult, 1) for width_mult in self.width_mult_list ] width_list.append(width) #################################################################################################### encoder unshuffle input_channel = width_list[0] enc_first_pixelunshuffle = ConvLayer(3, max(input_channel), kernel_size=3, stride=stride_stages[0], act_func=act_stages[0], use_bn=True) enc_second_pixelunshuffle = ConvLayer(max(input_channel) * 4, max(input_channel), kernel_size=3, stride=stride_stages[0], act_func=act_stages[0], use_bn=True) #################################################################################################### encoder inverted residual blocks self.block_group_info = [[0, 1]] blocks = [enc_first_pixelunshuffle, enc_second_pixelunshuffle] _block_index = 2 feature_dim = width_list[1] # pixelunshuffle 해서 x4 되기때문에 그냥 이렇게함 for width, n_block, s, act_func, use_se in zip(width_list[1:5], n_block_list[1:5], stride_stages[1:5], act_stages[1:5], se_stages[1:5]): 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, ) shortcut = IdentityLayer(feature_dim, feature_dim) blocks.append( MobileInvertedResidualBlock(mobile_inverted_conv, shortcut)) feature_dim = output_channel #################################################################################################### encoder final conv blocks enc_final_conv_blocks = [] for width, n_block, s, act_func, use_se in zip(width_list[5:8], n_block_list[5:8], stride_stages[5:8], act_stages[5:8], se_stages[5:8]): # 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 enc_final_conv_blocks.append( ConvLayer(max(feature_dim), max(output_channel), kernel_size=3, stride=s, act_func=act_func, use_bn=True)) feature_dim = output_channel #################################################################################################### decoder first conv block dec_first_conv_block = ConvLayer(max(feature_dim), max(width_list[8]), kernel_size=3, stride=stride_stages[8], act_func=act_stages[8], use_bn=True) #################################################################################################### decoder inverted residual blocks feature_dim = width_list[6] for width, n_block, s, act_func, use_se in zip(width_list[9:13], n_block_list[9:13], stride_stages[9:13], act_stages[9:13], se_stages[9:13]): 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, ) shortcut = IdentityLayer(feature_dim, feature_dim) blocks.append( MobileInvertedResidualBlock(mobile_inverted_conv, shortcut)) feature_dim = output_channel #################################################################################################### decoder final conv blocks dec_final_conv_blocks = [] for width, n_block, s, act_func, use_se in zip(width_list[13:15], n_block_list[13:15], stride_stages[13:15], act_stages[13:15], se_stages[13:15]): # 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 dec_final_conv_blocks.append( ConvLayer(max(feature_dim), max(output_channel), kernel_size=3, stride=s, act_func=act_func, use_bn=True)) feature_dim = output_channel #################################################################################################### decoder shuffle # for width, n_block, s, act_func, use_se in zip(width_list[11], n_block_list[11], # stride_stages[11], act_stages[11], se_stages[11]): self.block_group_info.append( [_block_index + i for i in range(n_block_list[15])]) _block_index += n_block_list[15] output_channel = width_list[15] for i in range(n_block_list[15]): if i == 0: stride = stride_stages[15] else: stride = 1 blocks.append( ConvLayer(max(feature_dim), max(output_channel), kernel_size=3, stride=s, act_func=act_stages[15], use_bn=True)) #################################################################################################### decoder final output conv block dec_final_output_conv_block = ConvLayer(max(feature_dim), max(width_list[16]), kernel_size=3, stride=stride_stages[16], act_func=act_stages[16], use_bn=True) #################################################################################################### # runtime_depth self.runtime_depth = [ len(block_idx) for block_idx in self.block_group_info ] super(OFAMobileNetX4, self).__init__(blocks, enc_final_conv_blocks, dec_first_conv_block, dec_final_conv_blocks, dec_final_output_conv_block, self.runtime_depth) # set bn param self.set_bn_param(momentum=bn_param[0], eps=bn_param[1])
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, pixelshuffle_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 if pixelshuffle_depth_list is None: pixelshuffle_depth_list = dynamic_net.pixelshuffle_depth_list subnet_settings = [] for pixel_d in pixelshuffle_depth_list: 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, 'pixel_d': pixel_d, 'wid': w, 'd': d, 'e': e, 'ks': k, }, 'PD%s-W%s-D%s-E%s-K%s' % (pixel_d, w, d, e, k) ]) if additional_setting is not None: subnet_settings += additional_setting # losses_of_subnets, top1_of_subnets, top5_of_subnets = [], [], [] losses_of_subnets, psnr_of_subnets = [], [] valid_log = '' for setting, name in subnet_settings: #################### Validation Architecture 정하는 부분인데, Single Architecture Overfitting 혹은 뭐 빠르게 테스트 해볼일 있으면 여기서 그냥 스킵하면됨 # if name.find('PD1-W0-D2-E3-K7') == -1: # continue 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')) #################### Random Sampling과 Structured Sampling중에 주석 바꿔가면서 고르면 됨 # dynamic_net.sample_active_subnet() dynamic_net.set_active_subnet(**setting) run_manager.write_log(dynamic_net.module_str, 'train', should_print=False) #################### Oracle Training 시에는 Batch Mean/Variance 현재 데이터로 업데이트하면 망함. # run_manager.reset_running_statistics(dynamic_net) loss, psnr = 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) psnr_of_subnets.append(psnr) valid_log += '%s (%.3f), ' % (name, psnr) return list_mean(losses_of_subnets), list_mean(psnr_of_subnets), valid_log
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] base_stage_width = [16, 24, 40, 80, 112, 160, 192, 224, 256, 320, 480, 960, 1280] stride_stages = [1, 2, 2, 2, 1, 2, 2, 2, 2, 1, 2] act_stages = ['relu', 'relu', 'relu', 'h_swish', 'h_swish', 'h_swish', 'relu', 'relu', 'h_swish', 'h_swish', 'h_swish'] se_stages = [False, False, True, False, True, True, False, True, False, True, True] 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 ] # 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] 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)] * 10 # depth_list = [,12,3,4] # [1, 4,4,4,4,..........] 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 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_depth, 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-3), 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() if base_stage_width == 'google': 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] * 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 = ConvLayer(3, max(input_channel), kernel_size=3, stride=2, use_bn=True, act_func='relu6', ops_order='weight_bn_act') else: first_conv = DynamicConvLayer(in_channel_list=int2list( 3, len(input_channel)), out_channel_list=input_channel, kernel_size=3, stride=2, act_func='relu6') # first block if len(first_block_width) == 1: first_block_conv = MBInvertedConvLayer( in_channels=max(input_channel), out_channels=max(first_block_width), kernel_size=3, stride=1, expand_ratio=1, act_func='relu6', ) else: first_block_conv = DynamicMBConvLayer( in_channel_list=input_channel, out_channel_list=first_block_width, kernel_size_list=3, expand_ratio_list=1, stride=1, act_func='relu6', ) 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, 4] print('Use MobileNetV2 Depth Setting') 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)]) _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=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', ) 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 = ConvLayer( max(input_channel), max(last_channel), kernel_size=1, use_bn=True, act_func='relu6', ) classifier = LinearLayer(max(last_channel), n_classes, dropout_rate=dropout_rate) else: feature_mix_layer = DynamicConvLayer( in_channel_list=input_channel, out_channel_list=last_channel, kernel_size=1, stride=1, act_func='relu6', ) classifier = DynamicLinearLayer(in_features_list=last_channel, out_features=n_classes, bias=True, 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, 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])