EFPN代码解读

标签：kernel 代码 EFPN 2048 解读 stride num Conv2d size

论文

Extended Feature Pyramid Network for Small Object Detection python3 D:/Project/EFPN-detectron2-master/tools/train_net.py --config-file configs/InstanceSegmentation/pointrend_rcnn_R_50_FPN_1x_coco.yaml --num-gpus 1

训练脚本 cfg 中的配置先获取配置文件对象 config ，一旦你获取了配置文件对象 cfg ，你可以通过修改它的属性来自定义模型和训练过程的各种设置。例如，可以通过 cfg.MODEL.WEIGHTS = "path/to/weights.pth" 来设置模型加载的预训练权重路径，或者通过 cfg.SOLVER.BASE_LR = 0.001 来设置学习率。

cfg.merge_from_file() 方法将指定配置文件中的配置选项合并到当前的配置文件对象 cfg 中，以覆盖或添加新的配置选项。这样做的目的是将预定义模型的配置与当前的配置文件对象相结合，以确保模型在训练或推理过程中使用正确的参数和设置。通过合并配置文件，你可以使用预定义模型的默认配置，并根据需要进行修改或覆盖特定的配置选项。这样可以快速配置和使用预训练模型，并进行训练或推理任务。

在这里，将 cfg.MODEL.RESNETS.NUM_GROUPS 设置为 32 表示将使用 ResNeXt 模型，其中输入特征图将被分成32 个组进行卷积操作。如果将其设置为 1 ，则表示使用传统的 ResNet 模型，不进行组卷积。 通过调整 cfg.MODEL.RESNETS.NUM_GROUPS 的值，可以控制 ResNet 或 ResNeXt 模型的架构， 以适应不同的任务和需求。

通过将 cfg.MODEL.BACKBONE.NAME 设置为 "build_resnet_fpn_backbone" ，可以 指定模型使用该函 数构建主干网络 。这意味着在模型的前向传播过程中，输入图像将通过 ResNet 网络提取特征，并与 FPN结构进行融合，以获取多尺度的特征表示。通过设置不同的主干网络名称，可以使用不同的预定义主干网络结构或自定义的主干网络结构来适应不同的任务和数据集。

表示使用 ResNet 的第 2 、 3 、 4 、 5 和 6 层的特征图作为输入。这意味着这些层级的特征将被传递给 FPN 进行融合。通过设置不同的输入特征层，可以根据任务和数据集的需求来选择使用哪些层级的特征图进行特征融合，以获得更好的多尺度表示能力。

cfg.MODEL.RPN.BATCH_SIZE_PER_IMAGE = 128 ：这句代码 设置了区域生成网络（ Region Proposal Network ， RPN ）每张图像的正负样本比例 。在每张图像上， RPN 会生成一系列候选区域，其中一部分是正样本（包含目标），一部分是负样本（不包含目标）。 BATCH_SIZE_PER_IMAGE 表示每张图像中的候选区域的总数，其中正样本和负样本的比例由算法自动调整。 cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 256 ：这句代码 设置了 ROI 头部（ Region of Interest Heads ）每张图像的正负样本比例 。 ROI 头部是用于目标检测中对候选区域进行分类和回归的部分。 BATCH_SIZE_PER_IMAGE 表示每张图像中用于训练 ROI 头部的候选区域的总数，其中 正样本和负样本的比例由算法自动调整 。 cfg.SOLVER.IMS_PER_BATCH = 1 ：这句代码设置了每次训练时用于更新梯度的图像批次大小。 IMS_PER_BATCH 表示每次训练使用的图像数量。在这个例子中，每次训练使用 1 张图像进行梯度更新。

构造训练器

FTT.py

用于将输入的通道数 out_channels 缩放为 out_channels * 4。

import logging
import numpy as np
import fvcore.nn.weight_init as weight_init
import torch
import torch.nn.functional as F
from torch import nn
from detectron2.layers import Conv2d, ShapeSpec, get_norm
import math
from .backbone import Backbone
from .build import BACKBONE_REGISTRY
from .resnet import build_resnet_backbone
# p2, p3 in the paper is p3, p4 for us
# format of p2, p3 is both [bs, channels, height, width] p2和p3都是张量，均表示特征
图
def FTT_get_p3pr(p2, p3, out_channels, norm):
# 1x1卷积，
channel_scaler = Conv2d(
out_channels,
out_channels * 4,
kernel_size=1,
bias=False
#norm=''
)
# 定义两个内部函数
# 用于创建内容特征的函数
# tuple of (conv2d, conv2d, iter)
# 多次应用 1x1 的卷积层和 ReLU 激活函数来提取内容特征。（内容特征也可以通过transformer
来实现）
def create_content_extractor(x, num_channels, iterations=3):
conv1 = Conv2d(
num_channels,
num_channels,
kernel_size=1,
bias=False,
#norm=get_norm(norm, num_channels),
)
conv2 = Conv2d(
num_channels,
num_channels,
kernel_size=1,
bias=False,
#norm=get_norm(norm, num_channels),
)
out = x
# 通过for循环来做
for i in range(iterations):
out = conv1(out)
out = F.relu_(out)
out = conv2(out)
out = F.relu_(out)
return out
# 创建纹理特征的函数
# 最后应用了一个输出通道数为 num_channels/2 的 1x1 卷积层，用于提取纹理特征。
def create_texture_extractor(x, num_channels, iterations=3):
conv1 = Conv2d(
num_channels,
num_channels,
kernel_size=1,
bias=False,
#norm=get_norm(norm, num_channels),
)
conv2 = Conv2d(
num_channels,
num_channels,
kernel_size=1,
bias=False,
#norm=get_norm(norm, num_channels),
)
conv3 = Conv2d(
num_channels,
int(num_channels/2),
kernel_size=1,
bias=False,
)
out = x
for i in range(iterations):
out = conv1(out)
out = F.relu_(out)
out = conv2(out)
out = F.relu_(out)
out = conv3(out)
out = F.relu_(out)
return out
bottom = p3
# 对P3进行通道缩放，通过 channel_scaler 将通道数从 channels 缩放为 channels * 4。
bottom = channel_scaler(bottom)
# 用 create_content_extractor 函数提取内容特征，将缩放后的 p3 作为输入，并将输出存储
在 bottom 变量中。
bottom = create_content_extractor(bottom, out_channels*4)
# 亚像素卷积
# 使用 nn.PixelShuffle(2) 进行像素重排，将 bottom 中的每个像素的特征图尺寸增加两倍。
sub_pixel_conv = nn.PixelShuffle(2)
# 将 p2 和重排后的 bottom 在通道维度上进行连接，形成一个新的张量 top
bottom = sub_pixel_conv(bottom)
#print("\np3 shape: ",bottom.shape,"\n")
# We interpreted "wrap" as concatenating bottom and top
# so the total channels is doubled after (basically place one on top
# of the other)
top = p2
top = torch.cat((bottom, top), axis=1)
# 使用 create_texture_extractor 函数提取纹理特征，将 top 作为输入，并将输出存储在
top 变量中。
top = create_texture_extractor(top, out_channels*2)
#top = top[:,256:]
# 残差连接部分
result = bottom + top
return result

GeneralizedRCNN(
 (backbone): FPN(
 (fpn_lateral2): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1))
 (fpn_output2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (fpn_lateral3): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1))
 (fpn_output3): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (fpn_lateral4): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1))
 (fpn_output4): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (fpn_lateral5): Conv2d(2048, 256, kernel_size=(1, 1), stride=(1, 1))
 (fpn_output5): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (fpn_lateral6): Conv2d(4096, 256, kernel_size=(1, 1), stride=(1, 1))
 (fpn_output6): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (top_block): LastLevelMaxPool()
 (bottom_up): ResNet(
 (stem): BasicStem(
 (conv1): Conv2d(
 3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False
 (norm): FrozenBatchNorm2d(num_features=64, eps=1e-05)
 )
 )
 (res2): Sequential(
 (0): BottleneckBlock(
 (shortcut): Conv2d(
 64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv1): Conv2d(
 64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv2): Conv2d(
 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv3): Conv2d(
 256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 )
 (1): BottleneckBlock(
 (conv1): Conv2d(
 256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv2): Conv2d(
 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv3): Conv2d(
 256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 )
 (2): BottleneckBlock(
 (conv1): Conv2d(
 256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv2): Conv2d(
 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv3): Conv2d(
 256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 )
 )
 (res3): Sequential(
 (0): SingleDownsampling(
 (conv1): Conv2d(
 256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False
 (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
 )
 )
 )
 (res4): Sequential(
 (0): BottleneckBlock(
 (shortcut): Conv2d(
 512, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv1): Conv2d(
 512, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv2): Conv2d(
 1024, 1024, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv3): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 )
 (1): BottleneckBlock(
 (conv1): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv2): Conv2d(
 1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv3): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 )
 (2): BottleneckBlock(
 (conv1): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv2): Conv2d(
 1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv3): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 )
 (3): BottleneckBlock(
 (conv1): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv2): Conv2d(
 1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv3): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 )
 )
 (res5): Sequential(
 (0): BottleneckBlock(
 (shortcut): Conv2d(
 1024, 2048, kernel_size=(1, 1), stride=(2, 2), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv1): Conv2d(
 1024, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (1): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (2): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (3): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (4): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (5): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (6): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (7): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (8): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (9): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (10): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (11): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (12): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (13): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (14): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (15): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (16): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (17): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (18): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (19): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (20): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (21): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (22): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 )
 (res6): Sequential(
 (0): BottleneckBlock(
 (shortcut): Conv2d(
 2048, 4096, kernel_size=(1, 1), stride=(2, 2), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv1): Conv2d(
 2048, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv2): Conv2d(
 4096, 4096, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv3): Conv2d(
 4096, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 )
 (1): BottleneckBlock(
 (conv1): Conv2d(
 4096, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv2): Conv2d(
 4096, 4096, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv3): Conv2d(
 4096, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 )
 (2): BottleneckBlock(
 (conv1): Conv2d(
 4096, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv2): Conv2d(
 4096, 4096, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv3): Conv2d(
 4096, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 )
 )
 )
 )
 (proposal_generator): RPN(
 (rpn_head): StandardRPNHead(
 (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (objectness_logits): Conv2d(256, 3, kernel_size=(1, 1), stride=(1, 1))
 (anchor_deltas): Conv2d(256, 12, kernel_size=(1, 1), stride=(1, 1))
 )
 (anchor_generator): DefaultAnchorGenerator(
 (cell_anchors): BufferList()
 )
 )
 (roi_heads): StandardROIHeads(
 (box_pooler): ROIPooler(
 (level_poolers): ModuleList(
 (0): ROIAlign(output_size=(7, 7), spatial_scale=0.25, sampling_ratio=0, aligned=True)
 (1): ROIAlign(output_size=(7, 7), spatial_scale=0.125, sampling_ratio=0, aligned=True)
 (2): ROIAlign(output_size=(7, 7), spatial_scale=0.0625, sampling_ratio=0, aligned=True)
 (3): ROIAlign(output_size=(7, 7), spatial_scale=0.03125, sampling_ratio=0, aligned=True)
 )
 )
 (box_head): FastRCNNConvFCHead(
 (flatten): Flatten(start_dim=1, end_dim=-1)
 (fc1): Linear(in_features=12544, out_features=1024, bias=True)
 (fc_relu1): ReLU()
 (fc2): Linear(in_features=1024, out_features=1024, bias=True)
 (fc_relu2): ReLU()
 )
 (box_predictor): FastRCNNOutputLayers(
 (cls_score): Linear(in_features=1024, out_features=81, bias=True)
 (bbox_pred): Linear(in_features=1024, out_features=320, bias=True)
 )
 )
)
GeneralizedRCNN(
 (backbone): FPN(
 (fpn_lateral2): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1))
 (fpn_output2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (fpn_lateral3): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1))
 (fpn_output3): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (fpn_lateral4): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1))
 (fpn_output4): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (fpn_lateral5): Conv2d(2048, 256, kernel_size=(1, 1), stride=(1, 1))
 (fpn_output5): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (fpn_lateral6): Conv2d(4096, 256, kernel_size=(1, 1), stride=(1, 1))
 (fpn_output6): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (top_block): LastLevelMaxPool()
 (bottom_up): ResNet(
 (stem): BasicStem(
 (conv1): Conv2d(
 3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False
 (norm): FrozenBatchNorm2d(num_features=64, eps=1e-05)
 )
 )
 (res2): Sequential(
 (0): BottleneckBlock(
 (shortcut): Conv2d(
 64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv1): Conv2d(
 64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv2): Conv2d(
 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv3): Conv2d(
 256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 )
 (1): BottleneckBlock(
 (conv1): Conv2d(
 256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv2): Conv2d(
 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv3): Conv2d(
 256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 )
 (2): BottleneckBlock(
 (conv1): Conv2d(
 256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv2): Conv2d(
 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 (conv3): Conv2d(
 256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
 )
 )
 )
 (res3): Sequential(
 (0): SingleDownsampling(
 (conv1): Conv2d(
 256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False
 (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
 )
 )
 )
 (res4): Sequential(
 (0): BottleneckBlock(
 (shortcut): Conv2d(
 512, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv1): Conv2d(
 512, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv2): Conv2d(
 1024, 1024, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv3): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 )
 (1): BottleneckBlock(
 (conv1): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv2): Conv2d(
 1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv3): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 )
 (2): BottleneckBlock(
 (conv1): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv2): Conv2d(
 1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv3): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 )
 (3): BottleneckBlock(
 (conv1): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv2): Conv2d(
 1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 (conv3): Conv2d(
 1024, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
 )
 )
 )
 (res5): Sequential(
 (0): BottleneckBlock(
 (shortcut): Conv2d(
 1024, 2048, kernel_size=(1, 1), stride=(2, 2), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv1): Conv2d(
 1024, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (1): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (2): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (3): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (4): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 (5): BottleneckBlock(
 (conv1): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv2): Conv2d(
 2048, 2048, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 (conv3): Conv2d(
 2048, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
 )
 )
 )
 (res6): Sequential(
 (0): BottleneckBlock(
 (shortcut): Conv2d(
 2048, 4096, kernel_size=(1, 1), stride=(2, 2), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv1): Conv2d(
 2048, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv2): Conv2d(
 4096, 4096, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv3): Conv2d(
 4096, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 )
 (1): BottleneckBlock(
 (conv1): Conv2d(
 4096, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv2): Conv2d(
 4096, 4096, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv3): Conv2d(
 4096, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 )
 (2): BottleneckBlock(
 (conv1): Conv2d(
 4096, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv2): Conv2d(
 4096, 4096, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 (conv3): Conv2d(
 4096, 4096, kernel_size=(1, 1), stride=(1, 1), bias=False
 (norm): FrozenBatchNorm2d(num_features=4096, eps=1e-05)
 )
 )
 )
 )
 )
 (proposal_generator): RPN(
 (rpn_head): StandardRPNHead(
 (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
 (objectness_logits): Conv2d(256, 3, kernel_size=(1, 1), stride=(1, 1))
 (anchor_deltas): Conv2d(256, 12, kernel_size=(1, 1), stride=(1, 1))
 )
 (anchor_generator): DefaultAnchorGenerator(
 (cell_anchors): BufferList()
 )
 )
 (roi_heads): StandardROIHeads(
 (box_pooler): ROIPooler(
 (level_poolers): ModuleList(
 (0): ROIAlign(output_size=(7, 7), spatial_scale=0.25, sampling_ratio=0, aligned=True)
 (1): ROIAlign(output_size=(7, 7), spatial_scale=0.125, sampling_ratio=0, aligned=True)
 (2): ROIAlign(output_size=(7, 7), spatial_scale=0.0625, sampling_ratio=0, aligned=True)
 (3): ROIAlign(output_size=(7, 7), spatial_scale=0.03125, sampling_ratio=0, aligned=True)
 )
 )
 (box_head): FastRCNNConvFCHead(
 (flatten): Flatten(start_dim=1, end_dim=-1)
 (fc1): Linear(in_features=12544, out_features=1024, bias=True)
 (fc_relu1): ReLU()
 (fc2): Linear(in_features=1024, out_features=1024, bias=True)
 (fc_relu2): ReLU()
 )
 (box_predictor): FastRCNNOutputLayers(
 (cls_score): Linear(in_features=1024, out_features=81, bias=True)
 (bbox_pred): Linear(in_features=1024, out_features=320, bias=True)
 )
 )
)

HAT代码

CAB

由于基于 Transformer 的结构通常需要大量的通道来嵌入令牌，因此直接使用具有恒定宽度的卷积会产生很大的计算成本。因此，我们用常数β 压缩两个卷积层的通道数。对于具有 C 个通道的输入特征，第一个卷积层之后的输出特征的通道数被压缩为C/ β ，然后通过第二层将特征扩展到 C 个通道。接下来，利用标准CA 模块 [68] 自适应地重新缩放信道特征。

HAB

W-MSA

窗口划分

Linear

forward的过程

PatchMergin

在视觉注意力机制中 引入更大感受野的上下文信息 ，以帮助模型更好地理解图像。通过将输入特征划分为四个子区域并进行合并

OCAB

具体来说， nn.Linear(dim, dim * 3, bias=qkv_bias) 创建了一个线性变换层，它接受维度为 dim 的输入特征，并将其映射到维度为 dim * 3 的输出。这里的 dim * 3 是因为输出包含了查询（ q ）、键（ k ）和值（ v ）三个部分。该线性变换层的权重矩阵的形状为 (dim * 3, dim) ，表示将输入特征的每个元素与权重矩阵相乘，然后进行偏置项的加和。 bias=qkv_bias 参数用于控制是否包含偏置项。通过这个线性变换层，输入特征经过映射后可以分别得到查询（ q ）、键（ k ）和值（ v ）的表示，用于后续的注意力计算。