EfficientNet Backbone结构解析 -- 以EfficientNet-B0为例说明

骑猪看日落 2022-09-15 13:53 189阅读 0赞

一般而言，不论我们是否要将该主干结构用于何种模型，一般都是在这个基础上进行的。例如，将在Yolact中用EfficientNet替换Resnet，网上有说可以在同等效果下让模型数据量降为原来的大约1/4到1/5左右。

下面我列出一下主干结构。

## self.\_conv\_stem ##

Conv2dStaticSamePadding(
      (conv): Conv2d(3, 32, kernel_size=(3, 3), stride=(2, 2), bias=False)
    )

## self.\_blocks ##

模型参数参考utils.py -- > efficientnet -- > block\_args，所有的blocks都是基于这个参数自动派生出来的，

blocks_args = [
            'r1_k3_s11_e1_i32_o16_se0.25', 'r2_k3_s22_e6_i16_o24_se0.25',
            'r2_k5_s22_e6_i24_o40_se0.25', 'r3_k3_s22_e6_i40_o80_se0.25',
            'r3_k5_s11_e6_i80_o112_se0.25', 'r4_k5_s22_e6_i112_o192_se0.25',
            'r1_k3_s11_e6_i192_o320_se0.25',
        ]

其中的参数意义如下

args = [
      'r%d' % block.num_repeat,
      'k%d' % block.kernel_size,
      's%d%d' % (block.strides[0], block.strides[1]),
      'e%s' % block.expand_ratio,
      'i%d' % block.input_filters,
      'o%d' % block.output_filters
      ]

在生成模型时，根据输入的模型名称（utils.py)，选择不同的（width, depth, resolution, dropout）参数，例如efficientnet-b0的参数就是(1.0, 1.0, 224, 0.2)，

def efficientnet_params(model_name):
        """ Map EfficientNet model name to parameter coefficients. """
        params_dict = {
            # Coefficients:   width,depth,res,dropout
            'efficientnet-b0': (1.0, 1.0, 224, 0.2),
            'efficientnet-b1': (1.0, 1.1, 240, 0.2),
            'efficientnet-b2': (1.1, 1.2, 260, 0.3),
            'efficientnet-b3': (1.2, 1.4, 300, 0.3),
            'efficientnet-b4': (1.4, 1.8, 380, 0.4),
            'efficientnet-b5': (1.6, 2.2, 456, 0.4),
            'efficientnet-b6': (1.8, 2.6, 528, 0.5),
            'efficientnet-b7': (2.0, 3.1, 600, 0.5),
            'efficientnet-b8': (2.2, 3.6, 672, 0.5),
            'efficientnet-l2': (4.3, 5.3, 800, 0.5),
        }
        return params_dict[model_name]

在后面生成blocks时，这个depth会和前面提到的 block\_args中的num\_repeat相乘，以最终决定生成多少个子block；width则会和filters(也就是channel)相乘，以决定输出的filter数目。具体可参考下面两个函数：

def round_filters(filters, global_params):
        """ Calculate and round number of filters based on depth multiplier. """
        multiplier = global_params.width_coefficient
        if not multiplier:
            return filters
        divisor = global_params.depth_divisor
        min_depth = global_params.min_depth
        filters *= multiplier
        min_depth = min_depth or divisor
        new_filters = max(min_depth, int(filters + divisor / 2) // divisor * divisor)
        if new_filters < 0.9 * filters:  # prevent rounding by more than 10%
            new_filters += divisor
        return int(new_filters)
    
    
    def round_repeats(repeats, global_params):
        """ Round number of filters based on depth multiplier. """
        multiplier = global_params.depth_coefficient
        if not multiplier:
            return repeats
        return int(math.ceil(multiplier * repeats))

其ModuleList共计有(r1+r2+r2+r3+r3+r4+r1) = 16层，如下，

ModuleList(
    (0): MBConvBlock(
    (_depthwise_conv): Conv2dStaticSamePadding( (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1,), groups=32, bias=False) )
    (_bn1): BatchNorm2d(32, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(32, 8, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(8, 32, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(32, 16, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(16, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (1): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(     (conv): Conv2d(16, 96, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(96, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding(  (conv): Conv2d(96, 96, kernel_size=(3, 3), stride=(2,), groups=96, bias=False) )
    (_bn1): BatchNorm2d(96, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(       (conv): Conv2d(96, 4, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(       (conv): Conv2d(4, 96, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(    (conv): Conv2d(96, 24, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(24, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (2): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(     (conv): Conv2d(24, 144, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(144, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding(  (conv): Conv2d(144,144, kernel_size=(3,3), stride=(1,1), groups=144,bias=False))
    (_bn1): BatchNorm2d(144, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(       (conv): Conv2d(144, 6, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(       (conv): Conv2d(6, 144, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(    (conv): Conv2d(144, 24, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(24, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (3): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(     (conv): Conv2d(24, 144, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(144, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding(  (conv): Conv2d(144,144,kernel_size=(5,5), stride=(2,), groups=144, bias=False) )
    (_bn1): BatchNorm2d(144, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(       (conv): Conv2d(144, 6, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(       (conv): Conv2d(6, 144, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(    (conv): Conv2d(144, 40, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(40, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (4): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(    (conv): Conv2d(40, 240, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(240, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding( (conv): Conv2d(240,240, kernel_size=(5,5), stride=(1,1), groups=240,bias=False))
    (_bn1): BatchNorm2d(240, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(240, 10, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(10, 240, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(240, 40, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(40, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (5): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(    (conv): Conv2d(40, 240, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(240, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding( (conv): Conv2d(240, 240, kernel_size=(3, 3), stride=(2,), groups=240, bias=False)  )
    (_bn1): BatchNorm2d(240, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(240, 10, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(10, 240, kernel_size=(1, 1), stride=(1, 1))   )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(240, 80, kernel_size=(1, 1), stride=(1, 1), bias=False) )
    (_bn2): BatchNorm2d(80, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (6): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(   (conv): Conv2d(80, 480, kernel_size=(1, 1), stride=(1, 1), bias=False)   )
    (_bn0): BatchNorm2d(480, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding((conv): Conv2d(480,480, kernel_size=(3,3), stride=(1,1), groups=480,bias=False))
    (_bn1): BatchNorm2d(480, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(     (conv): Conv2d(480, 20, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(     (conv): Conv2d(20, 480, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(  (conv): Conv2d(480, 80, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(80, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (7): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(    (conv): Conv2d(80, 480, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(480, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding((conv):  Conv2d(480,480, kernel_size=(3,3), stride=(1,1), groups=480, bias=False))
    (_bn1): BatchNorm2d(480, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(480, 20, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(20, 480, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(480, 80, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(80, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (8): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(     (conv): Conv2d(80, 480, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(480, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding(
    (conv): Conv2d(480, 480, kernel_size=(5, 5), stride=(1,), groups=480, bias=False)
    )
    (_bn1): BatchNorm2d(480, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(480, 20, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(20, 480, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(480, 112, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(112, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (9): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(    (conv): Conv2d(112, 672, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(672, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding((conv):  Conv2d(672,672, kernel_size=(5,5), stride=(1,1), groups=672, bias=False))
    (_bn1): BatchNorm2d(672, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(672, 28, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(28, 672, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(672, 112, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(112, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (10): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(    (conv): Conv2d(112, 672, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(672, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding((conv):  Conv2d(672,672, kernel_size=(5,5), stride=(1,1), groups=672, bias=False))
    (_bn1): BatchNorm2d(672, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(672, 28, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(28, 672, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(672, 112, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(112, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (11): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(    (conv): Conv2d(112, 672, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(672, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding( (conv): Conv2d(672, 672, kernel_size=(5,5), stride=(2,), groups=672, bias=False))
    (_bn1): BatchNorm2d(672, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(672, 28, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(28, 672, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(672, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(192, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (12): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(    (conv): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(1152, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding( (conv):Conv2d(1152, 1152, kernel_size=(5,5), stride=(1,1), groups=1152,bias=False)
    )
    (_bn1): BatchNorm2d(1152, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(1152, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(192, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (13): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(    (conv): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(1152, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding((conv):  Conv2d(1152,1152, kernel_size=(5,5), stride=(1,1), groups=1152,bias=False))
    (_bn1): BatchNorm2d(1152, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(1152, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(192, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (14): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(    (conv): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(1152, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding((conv):  Conv2d(1152,1152, kernel_size=(5,5), stride=(1,1), groups=1152,bias=False))
    (_bn1): BatchNorm2d(1152, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(1152, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(192, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    (15): MBConvBlock(
    (_expand_conv): Conv2dStaticSamePadding(    (conv): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn0): BatchNorm2d(1152, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_depthwise_conv): Conv2dStaticSamePadding( (conv): Conv2d(1152,1152, kernel_size=(3,3), stride=(1,),groups=1152,bias=False))
    (_bn1): BatchNorm2d(1152, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_se_reduce): Conv2dStaticSamePadding(      (conv): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))    )
    (_se_expand): Conv2dStaticSamePadding(      (conv): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))    )
    (_project_conv): Conv2dStaticSamePadding(   (conv): Conv2d(1152, 320, kernel_size=(1, 1), stride=(1, 1), bias=False)    )
    (_bn2): BatchNorm2d(320, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
    (_swish): MemoryEfficientSwish()
    )
    )

这里，注意到1, 3, 5, 11 层的结果其stride都是[2, 2]，一般在这里输出 feature map，加上最后一层（第15层），共计5层，也就是FPN和各层。EfficientDet就是利用这5层作为Feature Map的输入，事实上EfficientDet只用到了最后3层（p3, p4, p5)，p5上采样生成p6，p7，这5层通过BiFPN操作后，最终生成class和box prediction layer，这些就不在这里具体展开了，可参考backbone.py：GitHub - zylo117/Yet-Another-EfficientDet-Pytorch: The pytorch re-implement of the official efficientdet with SOTA performance in real time and pretrained weights.。 
    
    后面的_conv_head，_fc，还有一些没列出来的averagepool等在EfficientDet应用中没有用到，直接删除掉了。

## self.\_conv\_head ##

Conv2dStaticSamePadding(
      (conv): Conv2d(320, 1280, kernel_size=(1, 1), stride=(1, 1), bias=False)
    )

## self.\_fc ##

Linear(in_features=1280, out_features=1000, bias=True)

主要参考模型：

[GitHub - zylo117/Yet-Another-EfficientDet-Pytorch: The pytorch re-implement of the official efficientdet with SOTA performance in real time and pretrained weights.][GitHub - zylo117_Yet-Another-EfficientDet-Pytorch_ The pytorch re-implement of the official efficientdet with SOTA performance in real time and pretrained weights.]

[GitHub - lukemelas/EfficientNet-PyTorch: A PyTorch implementation of EfficientNet and EfficientNetV2 (coming soon!)][GitHub - lukemelas_EfficientNet-PyTorch_ A PyTorch implementation of EfficientNet and EfficientNetV2 _coming soon]

[GitHub - zylo117_Yet-Another-EfficientDet-Pytorch_ The pytorch re-implement of the official efficientdet with SOTA performance in real time and pretrained weights.]: https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch
[GitHub - lukemelas_EfficientNet-PyTorch_ A PyTorch implementation of EfficientNet and EfficientNetV2 _coming soon]: https://github.com/lukemelas/EfficientNet-PyTorch