![[Paper Review & Implementation] Going deeper with convolutions (GoogleNet/Inception Net v1, 2014)](/assets/img/papers/cnn/inception_v1.png)
[Paper Review & Implementation] Going deeper with convolutions (GoogleNet/Inception Net v1, 2014)
2023, May 23
Outlines
References
Inception V1 Architecture
- improves computational efficiency
- reduce dimension of parameters by adding extra 1x1 conv layer before 3x3 and 5x5 conv layers
- take global average pooling before entering into fc layer
- Use different sizes of filters to perform convolution on a single input and concatenate them into one output (output sizes adjusted with padding)
- combination
- recorded lowest error at ImageNet classifcation
- mitigate gradient vanishing problem with auxiliary classifiers
- increase the gradient signal from intermediate or lower layers but with regularization factor 0.3 (loss = final_out_loss + 0.3*(aux1_loss + aux2_loss)
- -> turns out this is not the case. (these branches not help reflecting low-level features) Instead, they work as regularizers with batch noramlization applied
Implementation with PyTorch
from google.colab import drive
drive.mount('/content/drive')
import sys
sys.path.append("/content/drive/MyDrive/Legend13")
import torch
import torch.nn as nn
!pip install torchinfo
from torchinfo import summary
Mounted at /content/drive
Looking in indexes: https://pypi.org/simple, https://us-python.pkg.dev/colab-wheels/public/simple/
Collecting torchinfo
Downloading torchinfo-1.8.0-py3-none-any.whl (23 kB)
Installing collected packages: torchinfo
Successfully installed torchinfo-1.8.0
class BasicConv2d(nn.Module):
def __init__(self, in_channel, F, size, stride, padding=0):
super().__init__()
self.layer = nn.Sequential(nn.Conv2d(in_channel, F, size, stride, padding=padding),
nn.BatchNorm2d(F),
nn.ReLU())
def forward(self, x):
out = self.layer(x)
return out
class Inception(nn.Module):
def __init__(self, in_channel, Fs, final_F):
super().__init__()
self.final_F = final_F
c1_F, c3_red_F, c3_F, c5_red_F, c5_F, poolproj_F = Fs
self.conv1x1 = BasicConv2d(in_channel, c1_F, 1, 1, padding=0)
self.conv3x3 = nn.Sequential(BasicConv2d(in_channel, c3_red_F, 1, 1, padding=0),
BasicConv2d(c3_red_F, c3_F, 3, 1, padding=1))
self.conv5x5 = nn.Sequential(BasicConv2d(in_channel, c5_red_F, 1, 1, padding=0),
BasicConv2d(c5_red_F, c5_F, 5, 1, padding=2))
self.maxpool_conv1x1 = nn.Sequential(nn.MaxPool2d(3, 1, padding=1),
BasicConv2d(in_channel, poolproj_F, 1, 1, padding=0))
def forward(self, x):
x_concat = [self.conv1x1(x), self.conv3x3(x), self.conv5x5(x), self.maxpool_conv1x1(x)]
x_concat = torch.cat(x_concat, 1) # N,F,H,W
# assert x_concat.shape[1] == self.final_F
return x_concat
class AuxOut(nn.Module):
def __init__(self, in_channel, p, num_classes):
super().__init__()
self.avgpool_conv = nn.Sequential(nn.AvgPool2d(5, 3, padding=0),
BasicConv2d(in_channel, 128, 1, 1, padding=0))
self.fc = nn.Sequential(nn.Linear(2048, 1024),
nn.ReLU(),
nn.Dropout(p=p),
nn.Linear(1024, num_classes))
def forward(self, x):
x = self.avgpool_conv(x)
x = torch.flatten(x,1)
out = self.fc(x)
return out
class Inception_V1(nn.Module):
def __init__(self, init_weights=True, p=0.5, use_aux=True, in_channel=3, num_classes=1000):
super().__init__()
self.use_aux = use_aux
self.conv1 = BasicConv2d(in_channel, 64, 7, 2, padding=3)
self.maxpool1 = nn.MaxPool2d(3, 2, padding=1)
self.conv2a = BasicConv2d(64, 64, 1, 1, padding=0)
self.conv2b = BasicConv2d(64, 192, 3, 1, padding=1)
self.maxpool2 = nn.MaxPool2d(3, 2, padding=1)
Fs = (64, 96, 128, 16, 32, 32)
self.inception3a = Inception(192, Fs, 256)
Fs = (128, 128, 192, 32, 96, 64)
self.inception3b = Inception(256, Fs, 480)
self.maxpool3 = nn.MaxPool2d(3, 2, padding=1)
Fs = (192, 96, 208, 16, 48, 64)
self.inception4a = Inception(480, Fs, 512)
Fs = (160, 112, 224, 24, 64, 64)
self.inception4b = Inception(512, Fs, 512)
Fs = (128, 128, 256, 24, 64, 64)
self.inception4c = Inception(512, Fs, 512)
Fs = (112, 144, 288, 32, 64, 64)
self.inception4d = Inception(512, Fs, 528)
Fs = (256, 160, 320, 32, 128, 128)
self.inception4e = Inception(528, Fs, 832)
self.maxpool4 = nn.MaxPool2d(3, 2, padding=1)
Fs = (256, 160, 320, 32, 128, 128)
self.inception5a = Inception(832, Fs, 832)
Fs = (384, 192, 384, 48, 128, 128)
self.inception5b = Inception(832, Fs, 1024)
if use_aux:
self.aux1 = AuxOut(512, p, num_classes)
self.aux2 = AuxOut(528, p, num_classes)
else:
self.aux1, self.aux2 = None, None
self.gap = nn.AdaptiveAvgPool2d((1,1))
self.dropout = nn.Dropout(p=p)
self.classifier = nn.Linear(1024, num_classes)
if init_weights:
for m in self.modules():
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Conv2d):
nn.init.trunc_normal_(m.weight, 0, 1e-2, a=-2, b=2)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def forward(self, x):
x = self.conv1(x)
x = self.maxpool1(x)
x = self.conv2a(x)
x = self.conv2b(x)
x = self.maxpool2(x)
x = self.inception3a(x)
x = self.inception3b(x)
x = self.maxpool3(x)
x = self.inception4a(x)
if self.training and self.aux1 is not None:
aux1_out = self.aux1(x)
else: aux1_out = None
x = self.inception4b(x)
x = self.inception4c(x)
x = self.inception4d(x)
if self.training and self.aux2 is not None:
aux2_out = self.aux2(x)
else: aux2_out = None
x = self.inception4e(x)
x = self.maxpool4(x)
x = self.inception5a(x)
x = self.inception5b(x)
x = self.gap(x)
x = torch.flatten(x, 1)
x = self.dropout(x)
out = self.classifier(x)
return aux1_out, aux2_out, out
Model Summary
model = Inception_V1()
summary(model, input_size=(2, 3, 224, 224), device='cpu')
===============================================================================================
Layer (type:depth-idx) Output Shape Param #
===============================================================================================
Inception_V1 -- 6,380,240
├─BasicConv2d: 1-1 [2, 64, 112, 112] --
│ └─Sequential: 2-1 [2, 64, 112, 112] --
│ │ └─Conv2d: 3-1 [2, 64, 112, 112] 9,472
│ │ └─BatchNorm2d: 3-2 [2, 64, 112, 112] 128
│ │ └─ReLU: 3-3 [2, 64, 112, 112] --
├─MaxPool2d: 1-2 [2, 64, 56, 56] --
├─BasicConv2d: 1-3 [2, 64, 56, 56] --
│ └─Sequential: 2-2 [2, 64, 56, 56] --
│ │ └─Conv2d: 3-4 [2, 64, 56, 56] 4,160
│ │ └─BatchNorm2d: 3-5 [2, 64, 56, 56] 128
│ │ └─ReLU: 3-6 [2, 64, 56, 56] --
├─BasicConv2d: 1-4 [2, 192, 56, 56] --
│ └─Sequential: 2-3 [2, 192, 56, 56] --
│ │ └─Conv2d: 3-7 [2, 192, 56, 56] 110,784
│ │ └─BatchNorm2d: 3-8 [2, 192, 56, 56] 384
│ │ └─ReLU: 3-9 [2, 192, 56, 56] --
├─MaxPool2d: 1-5 [2, 192, 28, 28] --
├─Inception: 1-6 [2, 256, 28, 28] --
│ └─BasicConv2d: 2-4 [2, 64, 28, 28] --
│ │ └─Sequential: 3-10 [2, 64, 28, 28] 12,480
│ └─Sequential: 2-5 [2, 128, 28, 28] --
│ │ └─BasicConv2d: 3-11 [2, 96, 28, 28] 18,720
│ │ └─BasicConv2d: 3-12 [2, 128, 28, 28] 110,976
│ └─Sequential: 2-6 [2, 32, 28, 28] --
│ │ └─BasicConv2d: 3-13 [2, 16, 28, 28] 3,120
│ │ └─BasicConv2d: 3-14 [2, 32, 28, 28] 12,896
│ └─Sequential: 2-7 [2, 32, 28, 28] --
│ │ └─MaxPool2d: 3-15 [2, 192, 28, 28] --
│ │ └─BasicConv2d: 3-16 [2, 32, 28, 28] 6,240
├─Inception: 1-7 [2, 480, 28, 28] --
│ └─BasicConv2d: 2-8 [2, 128, 28, 28] --
│ │ └─Sequential: 3-17 [2, 128, 28, 28] 33,152
│ └─Sequential: 2-9 [2, 192, 28, 28] --
│ │ └─BasicConv2d: 3-18 [2, 128, 28, 28] 33,152
│ │ └─BasicConv2d: 3-19 [2, 192, 28, 28] 221,760
│ └─Sequential: 2-10 [2, 96, 28, 28] --
│ │ └─BasicConv2d: 3-20 [2, 32, 28, 28] 8,288
│ │ └─BasicConv2d: 3-21 [2, 96, 28, 28] 77,088
│ └─Sequential: 2-11 [2, 64, 28, 28] --
│ │ └─MaxPool2d: 3-22 [2, 256, 28, 28] --
│ │ └─BasicConv2d: 3-23 [2, 64, 28, 28] 16,576
├─MaxPool2d: 1-8 [2, 480, 14, 14] --
├─Inception: 1-9 [2, 512, 14, 14] --
│ └─BasicConv2d: 2-12 [2, 192, 14, 14] --
│ │ └─Sequential: 3-24 [2, 192, 14, 14] 92,736
│ └─Sequential: 2-13 [2, 208, 14, 14] --
│ │ └─BasicConv2d: 3-25 [2, 96, 14, 14] 46,368
│ │ └─BasicConv2d: 3-26 [2, 208, 14, 14] 180,336
│ └─Sequential: 2-14 [2, 48, 14, 14] --
│ │ └─BasicConv2d: 3-27 [2, 16, 14, 14] 7,728
│ │ └─BasicConv2d: 3-28 [2, 48, 14, 14] 19,344
│ └─Sequential: 2-15 [2, 64, 14, 14] --
│ │ └─MaxPool2d: 3-29 [2, 480, 14, 14] --
│ │ └─BasicConv2d: 3-30 [2, 64, 14, 14] 30,912
├─Inception: 1-10 [2, 512, 14, 14] --
│ └─BasicConv2d: 2-16 [2, 160, 14, 14] --
│ │ └─Sequential: 3-31 [2, 160, 14, 14] 82,400
│ └─Sequential: 2-17 [2, 224, 14, 14] --
│ │ └─BasicConv2d: 3-32 [2, 112, 14, 14] 57,680
│ │ └─BasicConv2d: 3-33 [2, 224, 14, 14] 226,464
│ └─Sequential: 2-18 [2, 64, 14, 14] --
│ │ └─BasicConv2d: 3-34 [2, 24, 14, 14] 12,360
│ │ └─BasicConv2d: 3-35 [2, 64, 14, 14] 38,592
│ └─Sequential: 2-19 [2, 64, 14, 14] --
│ │ └─MaxPool2d: 3-36 [2, 512, 14, 14] --
│ │ └─BasicConv2d: 3-37 [2, 64, 14, 14] 32,960
├─Inception: 1-11 [2, 512, 14, 14] --
│ └─BasicConv2d: 2-20 [2, 128, 14, 14] --
│ │ └─Sequential: 3-38 [2, 128, 14, 14] 65,920
│ └─Sequential: 2-21 [2, 256, 14, 14] --
│ │ └─BasicConv2d: 3-39 [2, 128, 14, 14] 65,920
│ │ └─BasicConv2d: 3-40 [2, 256, 14, 14] 295,680
│ └─Sequential: 2-22 [2, 64, 14, 14] --
│ │ └─BasicConv2d: 3-41 [2, 24, 14, 14] 12,360
│ │ └─BasicConv2d: 3-42 [2, 64, 14, 14] 38,592
│ └─Sequential: 2-23 [2, 64, 14, 14] --
│ │ └─MaxPool2d: 3-43 [2, 512, 14, 14] --
│ │ └─BasicConv2d: 3-44 [2, 64, 14, 14] 32,960
├─Inception: 1-12 [2, 528, 14, 14] --
│ └─BasicConv2d: 2-24 [2, 112, 14, 14] --
│ │ └─Sequential: 3-45 [2, 112, 14, 14] 57,680
│ └─Sequential: 2-25 [2, 288, 14, 14] --
│ │ └─BasicConv2d: 3-46 [2, 144, 14, 14] 74,160
│ │ └─BasicConv2d: 3-47 [2, 288, 14, 14] 374,112
│ └─Sequential: 2-26 [2, 64, 14, 14] --
│ │ └─BasicConv2d: 3-48 [2, 32, 14, 14] 16,480
│ │ └─BasicConv2d: 3-49 [2, 64, 14, 14] 51,392
│ └─Sequential: 2-27 [2, 64, 14, 14] --
│ │ └─MaxPool2d: 3-50 [2, 512, 14, 14] --
│ │ └─BasicConv2d: 3-51 [2, 64, 14, 14] 32,960
├─Inception: 1-13 [2, 832, 14, 14] --
│ └─BasicConv2d: 2-28 [2, 256, 14, 14] --
│ │ └─Sequential: 3-52 [2, 256, 14, 14] 135,936
│ └─Sequential: 2-29 [2, 320, 14, 14] --
│ │ └─BasicConv2d: 3-53 [2, 160, 14, 14] 84,960
│ │ └─BasicConv2d: 3-54 [2, 320, 14, 14] 461,760
│ └─Sequential: 2-30 [2, 128, 14, 14] --
│ │ └─BasicConv2d: 3-55 [2, 32, 14, 14] 16,992
│ │ └─BasicConv2d: 3-56 [2, 128, 14, 14] 102,784
│ └─Sequential: 2-31 [2, 128, 14, 14] --
│ │ └─MaxPool2d: 3-57 [2, 528, 14, 14] --
│ │ └─BasicConv2d: 3-58 [2, 128, 14, 14] 67,968
├─MaxPool2d: 1-14 [2, 832, 7, 7] --
├─Inception: 1-15 [2, 832, 7, 7] --
│ └─BasicConv2d: 2-32 [2, 256, 7, 7] --
│ │ └─Sequential: 3-59 [2, 256, 7, 7] 213,760
│ └─Sequential: 2-33 [2, 320, 7, 7] --
│ │ └─BasicConv2d: 3-60 [2, 160, 7, 7] 133,600
│ │ └─BasicConv2d: 3-61 [2, 320, 7, 7] 461,760
│ └─Sequential: 2-34 [2, 128, 7, 7] --
│ │ └─BasicConv2d: 3-62 [2, 32, 7, 7] 26,720
│ │ └─BasicConv2d: 3-63 [2, 128, 7, 7] 102,784
│ └─Sequential: 2-35 [2, 128, 7, 7] --
│ │ └─MaxPool2d: 3-64 [2, 832, 7, 7] --
│ │ └─BasicConv2d: 3-65 [2, 128, 7, 7] 106,880
├─Inception: 1-16 [2, 1024, 7, 7] --
│ └─BasicConv2d: 2-36 [2, 384, 7, 7] --
│ │ └─Sequential: 3-66 [2, 384, 7, 7] 320,640
│ └─Sequential: 2-37 [2, 384, 7, 7] --
│ │ └─BasicConv2d: 3-67 [2, 192, 7, 7] 160,320
│ │ └─BasicConv2d: 3-68 [2, 384, 7, 7] 664,704
│ └─Sequential: 2-38 [2, 128, 7, 7] --
│ │ └─BasicConv2d: 3-69 [2, 48, 7, 7] 40,080
│ │ └─BasicConv2d: 3-70 [2, 128, 7, 7] 153,984
│ └─Sequential: 2-39 [2, 128, 7, 7] --
│ │ └─MaxPool2d: 3-71 [2, 832, 7, 7] --
│ │ └─BasicConv2d: 3-72 [2, 128, 7, 7] 106,880
├─AdaptiveAvgPool2d: 1-17 [2, 1024, 1, 1] --
├─Dropout: 1-18 [2, 1024] --
├─Linear: 1-19 [2, 1000] 1,025,000
===============================================================================================
Total params: 13,393,352
Trainable params: 13,393,352
Non-trainable params: 0
Total mult-adds (G): 3.17
===============================================================================================
Input size (MB): 1.20
Forward/backward pass size (MB): 103.25
Params size (MB): 28.05
Estimated Total Size (MB): 132.51
===============================================================================================
Forward Pass
x = torch.randn(2,3,224,224)
aux1_out, aux2_out, out = model(x)
print(aux1_out.shape)
print(aux2_out.shape)
print(out.shape)
torch.Size([2, 1000])
torch.Size([2, 1000])
torch.Size([2, 1000])