🤵 Author ：Horizon Max
✨ 程式設計技巧篇：各種操作小結
🎇 機器視覺篇：會變魔術 OpenCV
💥 深度學習篇：簡單入門 PyTorch
🏆 神經網絡篇：經典網絡模型
💻 算法篇：再忙也别忘了 LeetCode
[ 輕量級網絡 ] 經典網絡模型4——Xception 詳解與複現
- 🚀 Xception
- 🚀 Xception 詳解
-
- 🎨 Xception 網絡結構
-
- 🚩 結構探索
- 🚩 Xception 結構框圖
- 🚩 Xception 實驗測試
- 🚀 Xception 複現
🚀 Xception
Xception 是一種受啟發于 Inception 的新穎深度卷積神經網絡架構 ；
将 Inception 子產品替換為 深度可分離卷積( depthwise separable convolution ) ；
Xception 與 Inception V3 具有相同數量的參數，但由于模型參數的更有效使用，使得性能進一步的提高 ；
🔗 論文位址：Xception: Deep Learning with Depthwise Separable Convolutions
🚀 Xception 詳解
🎨 Xception 網絡結構
🚩 結構探索
（1）初始的 Inception 子產品：
Inception V3 module
（2）簡化 Inception 子產品，僅包含 3×3卷積 去掉簡單的子產品和 平均池化：
A simple Inception module
（3）将所有 1×1卷積 進行拼接 ：
the simplified Inception module
（4）使 3×3卷積 的數量與 1×1卷積 輸出通道的數量相等 ：
An “extreme” version of Inception module
Extreme Inception
也就是作者提出的新結構 ——
Xception
；
🚩 Xception 結構框圖
Xception 結構有36個卷積層，構成網絡特征提取的基礎 ；
每一層後面都接有
邏輯回歸層
，在邏輯回歸層前可能會随機插入
全連接配接層
；
将36個卷積層結構分為14個子產品，除了第一個和最後一個子產品，每個子產品都有
殘差結構
連接配接；
🚩 Xception 實驗測試
Classification performance comparison on ImageNet
Classification performance comparison on ImageNet
Size and training speed comparison
🚀 Xception 複現
# Here is the code :
import torch
import torch.nn as nn
from torchinfo import summary
class SeparableConv(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=1, stride=1, padding=0, dilation=1):
super(SeparableConv, self).__init__()
self.conv = nn.Conv2d(in_channels, in_channels, kernel_size, stride, padding,
dilation, groups=in_channels, bias=False)
self.pointwise = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0,
dilation=1, groups=1, bias=False)
def forward(self, x):
x = self.conv(x)
x = self.pointwise(x)
return x
class Block(nn.Module):
def __init__(self, in_channels, out_channels, num_conv, strides=1, start_with_relu=True, channel_change=True):
super(Block, self).__init__()
if out_channels != in_channels or strides != 1:
self.shutcut = nn.Sequential(
nn.Conv2d(in_channels, out_channels, 1, stride=strides, bias=False),
nn.BatchNorm2d(out_channels)
)
else:
self.shutcut = nn.Sequential()
layers = []
channels = in_channels
if channel_change:
layers.append(nn.ReLU(inplace=True))
layers.append(SeparableConv(in_channels, out_channels, 3, stride=1, padding=1))
layers.append(nn.BatchNorm2d(out_channels))
channels = out_channels
for i in range(num_conv - 1):
layers.append(nn.ReLU(inplace=True))
layers.append(SeparableConv(channels, channels, 3, stride=1, padding=1))
layers.append(nn.BatchNorm2d(channels))
if not channel_change:
layers.append(nn.ReLU(inplace=True))
layers.append(SeparableConv(in_channels, out_channels, 3, stride=1, padding=1))
layers.append(nn.BatchNorm2d(out_channels))
if start_with_relu:
layers = layers
else:
layers = layers[1:]
if strides != 1:
layers.append(nn.MaxPool2d(3, strides, 1))
self.layers = nn.Sequential(*layers)
def forward(self, x):
layer = self.layers(x)
shutcut = self.shutcut(x)
out = layer + shutcut
return out
class Xception(nn.Module):
def __init__(self, num_classes=1000):
super(Xception, self).__init__()
self.num_classes = num_classes
# Entry flow
self.entry_flow = nn.Sequential(
nn.Conv2d(3, 32, 3, 2, bias=False),
nn.BatchNorm2d(32),
nn.ReLU(inplace=True),
nn.Conv2d(32, 64, 3, bias=False),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
Block(64, 128, 2, 2, start_with_relu=False, channel_change=True),
Block(128, 256, 2, 2, start_with_relu=True, channel_change=True),
Block(256, 728, 2, 2, start_with_relu=True, channel_change=True),
)
# Middle flow
self.middle_flow = nn.Sequential(
Block(728, 728, 3, 1, start_with_relu=True, channel_change=True),
Block(728, 728, 3, 1, start_with_relu=True, channel_change=True),
Block(728, 728, 3, 1, start_with_relu=True, channel_change=True),
Block(728, 728, 3, 1, start_with_relu=True, channel_change=True),
Block(728, 728, 3, 1, start_with_relu=True, channel_change=True),
Block(728, 728, 3, 1, start_with_relu=True, channel_change=True),
Block(728, 728, 3, 1, start_with_relu=True, channel_change=True),
Block(728, 728, 3, 1, start_with_relu=True, channel_change=True),
)
# Exit flow
self.exit_flow = nn.Sequential(
Block(728, 1024, 2, 2, start_with_relu=True, channel_change=False),
SeparableConv(1024, 1536, 3, 1, 1),
nn.BatchNorm2d(1536),
nn.ReLU(inplace=True),
SeparableConv(1536, 2048, 3, 1, 1),
nn.BatchNorm2d(2048),
nn.ReLU(inplace=True),
nn.AdaptiveAvgPool2d((1, 1)),
)
self.fc = nn.Linear(2048, num_classes)
def forward(self, x):
x = self.entry_flow(x)
x = self.middle_flow(x)
x = self.exit_flow(x)
x = x.view(x.size(0), -1)
out = self.fc(x)
return out
def test():
net = Xception()
y = net(torch.randn(1, 3, 224, 224))
print(y.size())
summary(net, (1, 3, 224, 224), depth=5)
if __name__ == '__main__':
test()
輸出結果：
torch.Size([1, 1000])
===============================================================================================
Layer (type:depth-idx) Output Shape Param #
===============================================================================================
Xception -- --
├─Sequential: 1-1 [1, 728, 14, 14] --
│ └─Conv2d: 2-1 [1, 32, 111, 111] 864
│ └─BatchNorm2d: 2-2 [1, 32, 111, 111] 64
│ └─ReLU: 2-3 [1, 32, 111, 111] --
│ └─Conv2d: 2-4 [1, 64, 109, 109] 18,432
│ └─BatchNorm2d: 2-5 [1, 64, 109, 109] 128
│ └─ReLU: 2-6 [1, 64, 109, 109] --
│ └─Block: 2-7 [1, 128, 55, 55] --
│ │ └─Sequential: 3-1 [1, 128, 55, 55] --
│ │ │ └─SeparableConv: 4-1 [1, 128, 109, 109] --
│ │ │ │ └─Conv2d: 5-1 [1, 64, 109, 109] 576
│ │ │ │ └─Conv2d: 5-2 [1, 128, 109, 109] 8,192
│ │ │ └─BatchNorm2d: 4-2 [1, 128, 109, 109] 256
│ │ │ └─ReLU: 4-3 [1, 128, 109, 109] --
│ │ │ └─SeparableConv: 4-4 [1, 128, 109, 109] --
│ │ │ │ └─Conv2d: 5-3 [1, 128, 109, 109] 1,152
│ │ │ │ └─Conv2d: 5-4 [1, 128, 109, 109] 16,384
│ │ │ └─BatchNorm2d: 4-5 [1, 128, 109, 109] 256
│ │ │ └─MaxPool2d: 4-6 [1, 128, 55, 55] --
│ │ └─Sequential: 3-2 [1, 128, 55, 55] --
│ │ │ └─Conv2d: 4-7 [1, 128, 55, 55] 8,192
│ │ │ └─BatchNorm2d: 4-8 [1, 128, 55, 55] 256
│ └─Block: 2-8 [1, 256, 28, 28] --
│ │ └─Sequential: 3-3 [1, 256, 28, 28] --
│ │ │ └─ReLU: 4-9 [1, 128, 55, 55] --
│ │ │ └─SeparableConv: 4-10 [1, 256, 55, 55] --
│ │ │ │ └─Conv2d: 5-5 [1, 128, 55, 55] 1,152
│ │ │ │ └─Conv2d: 5-6 [1, 256, 55, 55] 32,768
│ │ │ └─BatchNorm2d: 4-11 [1, 256, 55, 55] 512
│ │ │ └─ReLU: 4-12 [1, 256, 55, 55] --
│ │ │ └─SeparableConv: 4-13 [1, 256, 55, 55] --
│ │ │ │ └─Conv2d: 5-7 [1, 256, 55, 55] 2,304
│ │ │ │ └─Conv2d: 5-8 [1, 256, 55, 55] 65,536
│ │ │ └─BatchNorm2d: 4-14 [1, 256, 55, 55] 512
│ │ │ └─MaxPool2d: 4-15 [1, 256, 28, 28] --
│ │ └─Sequential: 3-4 [1, 256, 28, 28] --
│ │ │ └─Conv2d: 4-16 [1, 256, 28, 28] 32,768
│ │ │ └─BatchNorm2d: 4-17 [1, 256, 28, 28] 512
│ └─Block: 2-9 [1, 728, 14, 14] --
│ │ └─Sequential: 3-5 [1, 728, 14, 14] --
│ │ │ └─ReLU: 4-18 [1, 256, 28, 28] --
│ │ │ └─SeparableConv: 4-19 [1, 728, 28, 28] --
│ │ │ │ └─Conv2d: 5-9 [1, 256, 28, 28] 2,304
│ │ │ │ └─Conv2d: 5-10 [1, 728, 28, 28] 186,368
│ │ │ └─BatchNorm2d: 4-20 [1, 728, 28, 28] 1,456
│ │ │ └─ReLU: 4-21 [1, 728, 28, 28] --
│ │ │ └─SeparableConv: 4-22 [1, 728, 28, 28] --
│ │ │ │ └─Conv2d: 5-11 [1, 728, 28, 28] 6,552
│ │ │ │ └─Conv2d: 5-12 [1, 728, 28, 28] 529,984
│ │ │ └─BatchNorm2d: 4-23 [1, 728, 28, 28] 1,456
│ │ │ └─MaxPool2d: 4-24 [1, 728, 14, 14] --
│ │ └─Sequential: 3-6 [1, 728, 14, 14] --
│ │ │ └─Conv2d: 4-25 [1, 728, 14, 14] 186,368
│ │ │ └─BatchNorm2d: 4-26 [1, 728, 14, 14] 1,456
├─Sequential: 1-2 [1, 728, 14, 14] --
│ └─Block: 2-10 [1, 728, 14, 14] --
│ │ └─Sequential: 3-7 [1, 728, 14, 14] --
│ │ │ └─ReLU: 4-27 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-28 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-13 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-14 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-29 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-30 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-31 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-15 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-16 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-32 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-33 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-34 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-17 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-18 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-35 [1, 728, 14, 14] 1,456
│ │ └─Sequential: 3-8 [1, 728, 14, 14] --
│ └─Block: 2-11 [1, 728, 14, 14] --
│ │ └─Sequential: 3-9 [1, 728, 14, 14] --
│ │ │ └─ReLU: 4-36 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-37 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-19 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-20 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-38 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-39 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-40 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-21 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-22 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-41 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-42 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-43 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-23 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-24 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-44 [1, 728, 14, 14] 1,456
│ │ └─Sequential: 3-10 [1, 728, 14, 14] --
│ └─Block: 2-12 [1, 728, 14, 14] --
│ │ └─Sequential: 3-11 [1, 728, 14, 14] --
│ │ │ └─ReLU: 4-45 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-46 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-25 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-26 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-47 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-48 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-49 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-27 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-28 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-50 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-51 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-52 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-29 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-30 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-53 [1, 728, 14, 14] 1,456
│ │ └─Sequential: 3-12 [1, 728, 14, 14] --
│ └─Block: 2-13 [1, 728, 14, 14] --
│ │ └─Sequential: 3-13 [1, 728, 14, 14] --
│ │ │ └─ReLU: 4-54 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-55 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-31 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-32 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-56 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-57 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-58 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-33 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-34 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-59 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-60 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-61 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-35 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-36 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-62 [1, 728, 14, 14] 1,456
│ │ └─Sequential: 3-14 [1, 728, 14, 14] --
│ └─Block: 2-14 [1, 728, 14, 14] --
│ │ └─Sequential: 3-15 [1, 728, 14, 14] --
│ │ │ └─ReLU: 4-63 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-64 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-37 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-38 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-65 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-66 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-67 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-39 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-40 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-68 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-69 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-70 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-41 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-42 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-71 [1, 728, 14, 14] 1,456
│ │ └─Sequential: 3-16 [1, 728, 14, 14] --
│ └─Block: 2-15 [1, 728, 14, 14] --
│ │ └─Sequential: 3-17 [1, 728, 14, 14] --
│ │ │ └─ReLU: 4-72 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-73 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-43 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-44 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-74 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-75 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-76 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-45 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-46 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-77 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-78 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-79 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-47 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-48 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-80 [1, 728, 14, 14] 1,456
│ │ └─Sequential: 3-18 [1, 728, 14, 14] --
│ └─Block: 2-16 [1, 728, 14, 14] --
│ │ └─Sequential: 3-19 [1, 728, 14, 14] --
│ │ │ └─ReLU: 4-81 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-82 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-49 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-50 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-83 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-84 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-85 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-51 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-52 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-86 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-87 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-88 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-53 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-54 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-89 [1, 728, 14, 14] 1,456
│ │ └─Sequential: 3-20 [1, 728, 14, 14] --
│ └─Block: 2-17 [1, 728, 14, 14] --
│ │ └─Sequential: 3-21 [1, 728, 14, 14] --
│ │ │ └─ReLU: 4-90 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-91 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-55 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-56 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-92 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-93 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-94 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-57 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-58 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-95 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-96 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-97 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-59 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-60 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-98 [1, 728, 14, 14] 1,456
│ │ └─Sequential: 3-22 [1, 728, 14, 14] --
├─Sequential: 1-3 [1, 2048, 1, 1] --
│ └─Block: 2-18 [1, 1024, 7, 7] --
│ │ └─Sequential: 3-23 [1, 1024, 7, 7] --
│ │ │ └─ReLU: 4-99 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-100 [1, 728, 14, 14] --
│ │ │ │ └─Conv2d: 5-61 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-62 [1, 728, 14, 14] 529,984
│ │ │ └─BatchNorm2d: 4-101 [1, 728, 14, 14] 1,456
│ │ │ └─ReLU: 4-102 [1, 728, 14, 14] --
│ │ │ └─SeparableConv: 4-103 [1, 1024, 14, 14] --
│ │ │ │ └─Conv2d: 5-63 [1, 728, 14, 14] 6,552
│ │ │ │ └─Conv2d: 5-64 [1, 1024, 14, 14] 745,472
│ │ │ └─BatchNorm2d: 4-104 [1, 1024, 14, 14] 2,048
│ │ │ └─MaxPool2d: 4-105 [1, 1024, 7, 7] --
│ │ └─Sequential: 3-24 [1, 1024, 7, 7] --
│ │ │ └─Conv2d: 4-106 [1, 1024, 7, 7] 745,472
│ │ │ └─BatchNorm2d: 4-107 [1, 1024, 7, 7] 2,048
│ └─SeparableConv: 2-19 [1, 1536, 7, 7] --
│ │ └─Conv2d: 3-25 [1, 1024, 7, 7] 9,216
│ │ └─Conv2d: 3-26 [1, 1536, 7, 7] 1,572,864
│ └─BatchNorm2d: 2-20 [1, 1536, 7, 7] 3,072
│ └─ReLU: 2-21 [1, 1536, 7, 7] --
│ └─SeparableConv: 2-22 [1, 2048, 7, 7] --
│ │ └─Conv2d: 3-27 [1, 1536, 7, 7] 13,824
│ │ └─Conv2d: 3-28 [1, 2048, 7, 7] 3,145,728
│ └─BatchNorm2d: 2-23 [1, 2048, 7, 7] 4,096
│ └─ReLU: 2-24 [1, 2048, 7, 7] --
│ └─AdaptiveAvgPool2d: 2-25 [1, 2048, 1, 1] --
├─Linear: 1-4 [1, 1000] 2,049,000
===============================================================================================
Total params: 22,855,952
Trainable params: 22,855,952
Non-trainable params: 0
Total mult-adds (G): 4.55
===============================================================================================
Input size (MB): 0.60
Forward/backward pass size (MB): 250.18
Params size (MB): 91.42
Estimated Total Size (MB): 342.20
===============================================================================================