![[Paper Review & Implementation] Rethinking the Inception Architecture for Computer Vision (GoogleNet/Inception Net v2-3, 2015)](/assets/img/papers/cnn/inception_v3.png)
[Paper Review & Implementation] Rethinking the Inception Architecture for Computer Vision (GoogleNet/Inception Net v2-3, 2015)
2023, May 24
Outlines
- References
- Inception v2-v3 Architecture
- Label Smoothing
- Implementation with PyTorch
- Model Summary
- Forward Pass
References
- Rethinking the Inception Architecture for Computer Vision, Christian Szegedy (2015)
- https://hyukppen.modoo.at/?link=5db82s6p
Inception v2-v3 Architecture
Figrue 5 :
Figrue 6:
Figrue 7 :
1. Add alternative ways to factorize convolution and inception
- 5x5 -> 3x3 + 3x3
- avoid representational bottleneck (factorizing large receptive field into multiples of small ones)
- parameter reduction : locally adjacent pixels share parameters
- spatial aggregation : factorizing itself creates strong correlation between adjacent units (less loss of spatial information during reduction)
- nxn -> 1xn + nx1
- effect dramatically increases as n grows
- works better for medium grid-sizes (feature maps)
- 7 is suggested as optimal in the paper
2. Parallel use of stride 2 convolution and pooling to decrease grid-size
- Prevent representational bottlenecks and improve computational cost saving at the same time
- can reduce loss of representational information while decreasing the grid size
3. Inception v2 -> v3 : Use RMSProp, Label Smoothing, BN-auxiliary
- remove lower auxiliary layer (doesn’t contribute to performance at all)
- but network with second (upper) aux classifier does overtake the network without aux near the end of the training (not in the early stages)
- but this effect doesn’t seems to be related to increased gradient signals of lower features as expected in inception v1 paper.
- Instead, it seems to be associated with regularization effect due to the batch normalization added to side branch.
- BN acts as a regularizer by normalizing the activations (making sure subsequent layers to recieve inputs within a constant rage and prevent extreme activation, saturation, and inputs from entering into the non-linear regime of activation function)
Label Smoothing
Implementation with PyTorch
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 InceptionA(nn.Module):
def __init__(self, in_channel, final_F):
super().__init__()
self.final_F = final_F
self.branch_5x5 = nn.Sequential(BasicConv2d(in_channel, 64, 1, 1, padding=0),
BasicConv2d(64, 64, 3, 1, padding=1),
BasicConv2d(64, 96, 3, 1, padding=1))
self.branch_3x3 = nn.Sequential(BasicConv2d(in_channel, 48, 1, 1, padding=0),
BasicConv2d(48, 64, 3, 1, padding=1))
self.branch_pool = nn.Sequential(nn.MaxPool2d(3,1,padding=1),
BasicConv2d(in_channel, 64, 1, 1, padding=0))
self.branch_1x1 = BasicConv2d(in_channel, 64, 1, 1, padding=0)
def forward(self, x):
x_concat = [self.branch_5x5(x), self.branch_3x3(x), self.branch_pool(x), self.branch_1x1(x)]
x_concat = torch.cat(x_concat, 1) # N,F,H,W
assert x_concat.shape[1] == self.final_F
return x_concat
class InceptionB(nn.Module):
def __init__(self, in_channel, F, final_F):
super().__init__()
self.final_F = final_F
self.branch_2n = nn.Sequential(BasicConv2d(in_channel, F, 1, 1, padding=0),
BasicConv2d(F, F, (1,7), 1, padding=(0,3)),
BasicConv2d(F, F, (7,1), 1, padding=(3,0)),
BasicConv2d(F, F, (1,7), 1, padding=(0,3)),
BasicConv2d(F, 192, (7,1), 1, padding=(3,0)))
self.branch_n = nn.Sequential(BasicConv2d(in_channel, F, 1, 1, padding=0),
BasicConv2d(F, F, (1,7), 1, padding=(0,3)),
BasicConv2d(F, 192, (7,1), 1, padding=(3,0)))
self.branch_pool = nn.Sequential(nn.MaxPool2d(3,1,padding=1),
BasicConv2d(in_channel, 192, 1, 1, padding=0))
self.branch_1x1 = BasicConv2d(in_channel, 192, 1, 1, padding=0)
def forward(self, x):
x_concat = [self.branch_2n(x), self.branch_n(x), self.branch_pool(x), self.branch_1x1(x)]
x_concat = torch.cat(x_concat, 1) # N,F,H,W
assert x_concat.shape[1] == self.final_F
return x_concat
class GridReduc(nn.Module):
def __init__(self, in_channel, Fs, final_F):
super().__init__()
self.final_F = final_F
F0, F1, F2 = Fs
self.convstride_5x5 = nn.Sequential(BasicConv2d(in_channel, F0, 1, 1, padding=0),
BasicConv2d(F0, F1, 3, 1, padding=1),
BasicConv2d(F1, F1, 3, 2, padding=0))
self.convstride_3x3 = nn.Sequential(BasicConv2d(in_channel, F0, 1, 1, padding=0),
BasicConv2d(F0, F2, 3, 2, padding=0))
self.poolstride = nn.MaxPool2d(3, 2, padding=0)
def forward(self, x):
x_concat = [self.convstride_5x5(x), self.convstride_3x3(x), self.poolstride(x)]
x_concat = torch.cat(x_concat, 1) # N,F,H,W
assert x_concat.shape[1] == self.final_F
return x_concat
class InceptionC(nn.Module):
def __init__(self, in_channel, final_F):
super().__init__()
self.final_F = final_F
self.branch_5x5 = nn.Sequential(BasicConv2d(in_channel, 448, 1, 1, padding=0),
BasicConv2d(448, 384, 3, 1, padding=1))
self.branch_5x5_1 = nn.Sequential(self.branch_5x5,
BasicConv2d(384, 384, (1,3), 1, padding=(0,1)))
self.branch_5x5_2 = nn.Sequential(self.branch_5x5,
BasicConv2d(384, 384, (3,1), 1, padding=(1,0)))
self.branch_3x3 = BasicConv2d(in_channel, 384, 1, 1, padding=0)
self.branch_3x3_1 = nn.Sequential(self.branch_3x3,
BasicConv2d(384, 384, (1,3), 1, padding=(0,1)))
self.branch_3x3_2 = nn.Sequential(self.branch_3x3,
BasicConv2d(384, 384, (3,1), 1, padding=(1,0)))
self.branch_pool = nn.Sequential(nn.MaxPool2d(3,1,padding=1),
BasicConv2d(in_channel, 192, 1, 1, padding=0))
self.branch_1x1 = BasicConv2d(in_channel, 320, 1, 1, padding=0)
def forward(self, x):
branch_5x5_concat = torch.cat([self.branch_5x5_1(x), self.branch_5x5_2(x)], 1)
branch_3x3_concat = torch.cat([self.branch_3x3_1(x), self.branch_3x3_2(x)], 1)
x_concat = [branch_5x5_concat, branch_3x3_concat, self.branch_pool(x), self.branch_1x1(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, 1024, 1, 1, padding=0),
nn.AdaptiveAvgPool2d((1,1)))
self.dropout = nn.Dropout(p)
self.classifier = nn.Linear(1024, num_classes)
def forward(self, x):
x = self.avgpool_conv(x)
x = torch.flatten(x,1)
x = self.dropout(x)
out = self.classifier(x)
return out
class Inception_V3(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.conv1a = BasicConv2d(in_channel, 32, 3, 2, padding=0)
self.conv1b = BasicConv2d(32, 32, 3, 1, padding=0)
self.conv1c = BasicConv2d(32, 64, 3, 1, padding=1)
self.maxpool1 = nn.MaxPool2d(3, 2, padding=0)
self.conv2a = BasicConv2d(64, 80, 3, 1, padding=0)
self.conv2b = BasicConv2d(80, 192, 3, 2, padding=0)
self.conv2c = BasicConv2d(192, 288, 3, 1, padding=1)
# inceptionA x 2 + GridReduc x 1 : input = (35x35x288), output = (17x17x768)
self.inceptionA_1 = InceptionA(288, 288)
self.inceptionA_2 = InceptionA(288, 288)
self.gridreducA = GridReduc(288, (64, 96, 384), 768)
# inceptionB x 4 + GridReduc x 1 : input = (17x17x768), output = (8x8x1280)
self.inceptionB_1 = InceptionB(768, 128, 768)
self.inceptionB_2 = InceptionB(768, 160, 768)
self.inceptionB_3 = InceptionB(768, 160, 768)
self.inceptionB_4 = InceptionB(768, 192, 768)
self.gridreducB = GridReduc(768, (192, 192, 320), 1280)
# inceptionC x 2 : out feature maps 2048
self.inceptionC_1 = InceptionC(1280, 2048)
self.inceptionC_2 = InceptionC(2048, 2048) # 8x8x2048
if use_aux:
self.aux = AuxOut(768, p, num_classes)
else:
self.aux = None
self.gap = nn.AdaptiveAvgPool2d((1,1))
self.dropout = nn.Dropout(p=p)
self.classifier = nn.Linear(2048, 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.conv1a(x)
x = self.conv1b(x)
x = self.conv1c(x)
x = self.maxpool1(x)
x = self.conv2a(x)
x = self.conv2b(x)
x = self.conv2c(x)
x = self.inceptionA_1(x)
x = self.inceptionA_2(x)
x = self.gridreducA(x)
if self.training and self.aux is not None:
aux_out = self.aux(x)
else: aux_out = None
x = self.inceptionB_1(x)
x = self.inceptionB_2(x)
x = self.inceptionB_3(x)
x = self.inceptionB_4(x)
x = self.gridreducB(x)
x = self.inceptionC_1(x)
x = self.inceptionC_2(x)
x = self.gap(x)
x = torch.flatten(x, 1)
x = self.dropout(x)
out = self.classifier(x)
return aux_out, out
Model Summary
model = Inception_V3()
summary(model, input_size=(2, 3, 299, 299), device='cpu')
====================================================================================================
Layer (type:depth-idx) Output Shape Param #
====================================================================================================
Inception_V3 -- 1,814,504
├─BasicConv2d: 1-1 [2, 32, 149, 149] --
│ └─Sequential: 2-1 [2, 32, 149, 149] --
│ │ └─Conv2d: 3-1 [2, 32, 149, 149] 896
│ │ └─BatchNorm2d: 3-2 [2, 32, 149, 149] 64
│ │ └─ReLU: 3-3 [2, 32, 149, 149] --
├─BasicConv2d: 1-2 [2, 32, 147, 147] --
│ └─Sequential: 2-2 [2, 32, 147, 147] --
│ │ └─Conv2d: 3-4 [2, 32, 147, 147] 9,248
│ │ └─BatchNorm2d: 3-5 [2, 32, 147, 147] 64
│ │ └─ReLU: 3-6 [2, 32, 147, 147] --
├─BasicConv2d: 1-3 [2, 64, 147, 147] --
│ └─Sequential: 2-3 [2, 64, 147, 147] --
│ │ └─Conv2d: 3-7 [2, 64, 147, 147] 18,496
│ │ └─BatchNorm2d: 3-8 [2, 64, 147, 147] 128
│ │ └─ReLU: 3-9 [2, 64, 147, 147] --
├─MaxPool2d: 1-4 [2, 64, 73, 73] --
├─BasicConv2d: 1-5 [2, 80, 71, 71] --
│ └─Sequential: 2-4 [2, 80, 71, 71] --
│ │ └─Conv2d: 3-10 [2, 80, 71, 71] 46,160
│ │ └─BatchNorm2d: 3-11 [2, 80, 71, 71] 160
│ │ └─ReLU: 3-12 [2, 80, 71, 71] --
├─BasicConv2d: 1-6 [2, 192, 35, 35] --
│ └─Sequential: 2-5 [2, 192, 35, 35] --
│ │ └─Conv2d: 3-13 [2, 192, 35, 35] 138,432
│ │ └─BatchNorm2d: 3-14 [2, 192, 35, 35] 384
│ │ └─ReLU: 3-15 [2, 192, 35, 35] --
├─BasicConv2d: 1-7 [2, 288, 35, 35] --
│ └─Sequential: 2-6 [2, 288, 35, 35] --
│ │ └─Conv2d: 3-16 [2, 288, 35, 35] 497,952
│ │ └─BatchNorm2d: 3-17 [2, 288, 35, 35] 576
│ │ └─ReLU: 3-18 [2, 288, 35, 35] --
├─InceptionA: 1-8 [2, 288, 35, 35] --
│ └─Sequential: 2-7 [2, 96, 35, 35] --
│ │ └─BasicConv2d: 3-19 [2, 64, 35, 35] 18,624
│ │ └─BasicConv2d: 3-20 [2, 64, 35, 35] 37,056
│ │ └─BasicConv2d: 3-21 [2, 96, 35, 35] 55,584
│ └─Sequential: 2-8 [2, 64, 35, 35] --
│ │ └─BasicConv2d: 3-22 [2, 48, 35, 35] 13,968
│ │ └─BasicConv2d: 3-23 [2, 64, 35, 35] 27,840
│ └─Sequential: 2-9 [2, 64, 35, 35] --
│ │ └─MaxPool2d: 3-24 [2, 288, 35, 35] --
│ │ └─BasicConv2d: 3-25 [2, 64, 35, 35] 18,624
│ └─BasicConv2d: 2-10 [2, 64, 35, 35] --
│ │ └─Sequential: 3-26 [2, 64, 35, 35] 18,624
├─InceptionA: 1-9 [2, 288, 35, 35] --
│ └─Sequential: 2-11 [2, 96, 35, 35] --
│ │ └─BasicConv2d: 3-27 [2, 64, 35, 35] 18,624
│ │ └─BasicConv2d: 3-28 [2, 64, 35, 35] 37,056
│ │ └─BasicConv2d: 3-29 [2, 96, 35, 35] 55,584
│ └─Sequential: 2-12 [2, 64, 35, 35] --
│ │ └─BasicConv2d: 3-30 [2, 48, 35, 35] 13,968
│ │ └─BasicConv2d: 3-31 [2, 64, 35, 35] 27,840
│ └─Sequential: 2-13 [2, 64, 35, 35] --
│ │ └─MaxPool2d: 3-32 [2, 288, 35, 35] --
│ │ └─BasicConv2d: 3-33 [2, 64, 35, 35] 18,624
│ └─BasicConv2d: 2-14 [2, 64, 35, 35] --
│ │ └─Sequential: 3-34 [2, 64, 35, 35] 18,624
├─GridReduc: 1-10 [2, 768, 17, 17] --
│ └─Sequential: 2-15 [2, 96, 17, 17] --
│ │ └─BasicConv2d: 3-35 [2, 64, 35, 35] 18,624
│ │ └─BasicConv2d: 3-36 [2, 96, 35, 35] 55,584
│ │ └─BasicConv2d: 3-37 [2, 96, 17, 17] 83,232
│ └─Sequential: 2-16 [2, 384, 17, 17] --
│ │ └─BasicConv2d: 3-38 [2, 64, 35, 35] 18,624
│ │ └─BasicConv2d: 3-39 [2, 384, 17, 17] 222,336
│ └─MaxPool2d: 2-17 [2, 288, 17, 17] --
├─InceptionB: 1-11 [2, 768, 17, 17] --
│ └─Sequential: 2-18 [2, 192, 17, 17] --
│ │ └─BasicConv2d: 3-40 [2, 128, 17, 17] 98,688
│ │ └─BasicConv2d: 3-41 [2, 128, 17, 17] 115,072
│ │ └─BasicConv2d: 3-42 [2, 128, 17, 17] 115,072
│ │ └─BasicConv2d: 3-43 [2, 128, 17, 17] 115,072
│ │ └─BasicConv2d: 3-44 [2, 192, 17, 17] 172,608
│ └─Sequential: 2-19 [2, 192, 17, 17] --
│ │ └─BasicConv2d: 3-45 [2, 128, 17, 17] 98,688
│ │ └─BasicConv2d: 3-46 [2, 128, 17, 17] 115,072
│ │ └─BasicConv2d: 3-47 [2, 192, 17, 17] 172,608
│ └─Sequential: 2-20 [2, 192, 17, 17] --
│ │ └─MaxPool2d: 3-48 [2, 768, 17, 17] --
│ │ └─BasicConv2d: 3-49 [2, 192, 17, 17] 148,032
│ └─BasicConv2d: 2-21 [2, 192, 17, 17] --
│ │ └─Sequential: 3-50 [2, 192, 17, 17] 148,032
├─InceptionB: 1-12 [2, 768, 17, 17] --
│ └─Sequential: 2-22 [2, 192, 17, 17] --
│ │ └─BasicConv2d: 3-51 [2, 160, 17, 17] 123,360
│ │ └─BasicConv2d: 3-52 [2, 160, 17, 17] 179,680
│ │ └─BasicConv2d: 3-53 [2, 160, 17, 17] 179,680
│ │ └─BasicConv2d: 3-54 [2, 160, 17, 17] 179,680
│ │ └─BasicConv2d: 3-55 [2, 192, 17, 17] 215,616
│ └─Sequential: 2-23 [2, 192, 17, 17] --
│ │ └─BasicConv2d: 3-56 [2, 160, 17, 17] 123,360
│ │ └─BasicConv2d: 3-57 [2, 160, 17, 17] 179,680
│ │ └─BasicConv2d: 3-58 [2, 192, 17, 17] 215,616
│ └─Sequential: 2-24 [2, 192, 17, 17] --
│ │ └─MaxPool2d: 3-59 [2, 768, 17, 17] --
│ │ └─BasicConv2d: 3-60 [2, 192, 17, 17] 148,032
│ └─BasicConv2d: 2-25 [2, 192, 17, 17] --
│ │ └─Sequential: 3-61 [2, 192, 17, 17] 148,032
├─InceptionB: 1-13 [2, 768, 17, 17] --
│ └─Sequential: 2-26 [2, 192, 17, 17] --
│ │ └─BasicConv2d: 3-62 [2, 160, 17, 17] 123,360
│ │ └─BasicConv2d: 3-63 [2, 160, 17, 17] 179,680
│ │ └─BasicConv2d: 3-64 [2, 160, 17, 17] 179,680
│ │ └─BasicConv2d: 3-65 [2, 160, 17, 17] 179,680
│ │ └─BasicConv2d: 3-66 [2, 192, 17, 17] 215,616
│ └─Sequential: 2-27 [2, 192, 17, 17] --
│ │ └─BasicConv2d: 3-67 [2, 160, 17, 17] 123,360
│ │ └─BasicConv2d: 3-68 [2, 160, 17, 17] 179,680
│ │ └─BasicConv2d: 3-69 [2, 192, 17, 17] 215,616
│ └─Sequential: 2-28 [2, 192, 17, 17] --
│ │ └─MaxPool2d: 3-70 [2, 768, 17, 17] --
│ │ └─BasicConv2d: 3-71 [2, 192, 17, 17] 148,032
│ └─BasicConv2d: 2-29 [2, 192, 17, 17] --
│ │ └─Sequential: 3-72 [2, 192, 17, 17] 148,032
├─InceptionB: 1-14 [2, 768, 17, 17] --
│ └─Sequential: 2-30 [2, 192, 17, 17] --
│ │ └─BasicConv2d: 3-73 [2, 192, 17, 17] 148,032
│ │ └─BasicConv2d: 3-74 [2, 192, 17, 17] 258,624
│ │ └─BasicConv2d: 3-75 [2, 192, 17, 17] 258,624
│ │ └─BasicConv2d: 3-76 [2, 192, 17, 17] 258,624
│ │ └─BasicConv2d: 3-77 [2, 192, 17, 17] 258,624
│ └─Sequential: 2-31 [2, 192, 17, 17] --
│ │ └─BasicConv2d: 3-78 [2, 192, 17, 17] 148,032
│ │ └─BasicConv2d: 3-79 [2, 192, 17, 17] 258,624
│ │ └─BasicConv2d: 3-80 [2, 192, 17, 17] 258,624
│ └─Sequential: 2-32 [2, 192, 17, 17] --
│ │ └─MaxPool2d: 3-81 [2, 768, 17, 17] --
│ │ └─BasicConv2d: 3-82 [2, 192, 17, 17] 148,032
│ └─BasicConv2d: 2-33 [2, 192, 17, 17] --
│ │ └─Sequential: 3-83 [2, 192, 17, 17] 148,032
├─GridReduc: 1-15 [2, 1280, 8, 8] --
│ └─Sequential: 2-34 [2, 192, 8, 8] --
│ │ └─BasicConv2d: 3-84 [2, 192, 17, 17] 148,032
│ │ └─BasicConv2d: 3-85 [2, 192, 17, 17] 332,352
│ │ └─BasicConv2d: 3-86 [2, 192, 8, 8] 332,352
│ └─Sequential: 2-35 [2, 320, 8, 8] --
│ │ └─BasicConv2d: 3-87 [2, 192, 17, 17] 148,032
│ │ └─BasicConv2d: 3-88 [2, 320, 8, 8] 553,920
│ └─MaxPool2d: 2-36 [2, 768, 8, 8] --
├─InceptionC: 1-16 [2, 2048, 8, 8] --
│ └─Sequential: 2-37 [2, 384, 8, 8] 2,567,744
│ └─Sequential: 2-38 -- (recursive)
│ │ └─Sequential: 3-89 [2, 384, 8, 8] 2,124,224
│ └─Sequential: 2-39 -- (recursive)
│ │ └─BasicConv2d: 3-90 [2, 384, 8, 8] 443,520
│ └─Sequential: 2-40 [2, 384, 8, 8] 2,124,224
│ │ └─Sequential: 3-91 [2, 384, 8, 8] (recursive)
│ │ └─BasicConv2d: 3-92 [2, 384, 8, 8] 443,520
│ └─Sequential: 2-41 [2, 384, 8, 8] 936,192
│ └─Sequential: 2-42 -- (recursive)
│ │ └─BasicConv2d: 3-93 [2, 384, 8, 8] 492,672
│ └─Sequential: 2-43 -- (recursive)
│ │ └─BasicConv2d: 3-94 [2, 384, 8, 8] 443,520
│ └─Sequential: 2-44 [2, 384, 8, 8] 492,672
│ │ └─BasicConv2d: 3-95 [2, 384, 8, 8] (recursive)
│ │ └─BasicConv2d: 3-96 [2, 384, 8, 8] 443,520
│ └─Sequential: 2-45 [2, 192, 8, 8] --
│ │ └─MaxPool2d: 3-97 [2, 1280, 8, 8] --
│ │ └─BasicConv2d: 3-98 [2, 192, 8, 8] 246,336
│ └─BasicConv2d: 2-46 [2, 320, 8, 8] --
│ │ └─Sequential: 3-99 [2, 320, 8, 8] 410,560
├─InceptionC: 1-17 [2, 2048, 8, 8] --
│ └─Sequential: 2-47 [2, 384, 8, 8] 2,911,808
│ └─Sequential: 2-48 -- (recursive)
│ │ └─Sequential: 3-100 [2, 384, 8, 8] 2,468,288
│ └─Sequential: 2-49 -- (recursive)
│ │ └─BasicConv2d: 3-101 [2, 384, 8, 8] 443,520
│ └─Sequential: 2-50 [2, 384, 8, 8] 2,468,288
│ │ └─Sequential: 3-102 [2, 384, 8, 8] (recursive)
│ │ └─BasicConv2d: 3-103 [2, 384, 8, 8] 443,520
│ └─Sequential: 2-51 [2, 384, 8, 8] 1,231,104
│ └─Sequential: 2-52 -- (recursive)
│ │ └─BasicConv2d: 3-104 [2, 384, 8, 8] 787,584
│ └─Sequential: 2-53 -- (recursive)
│ │ └─BasicConv2d: 3-105 [2, 384, 8, 8] 443,520
│ └─Sequential: 2-54 [2, 384, 8, 8] 787,584
│ │ └─BasicConv2d: 3-106 [2, 384, 8, 8] (recursive)
│ │ └─BasicConv2d: 3-107 [2, 384, 8, 8] 443,520
│ └─Sequential: 2-55 [2, 192, 8, 8] --
│ │ └─MaxPool2d: 3-108 [2, 2048, 8, 8] --
│ │ └─BasicConv2d: 3-109 [2, 192, 8, 8] 393,792
│ └─BasicConv2d: 2-56 [2, 320, 8, 8] --
│ │ └─Sequential: 3-110 [2, 320, 8, 8] 656,320
├─AdaptiveAvgPool2d: 1-18 [2, 2048, 1, 1] --
├─Dropout: 1-19 [2, 2048] --
├─Linear: 1-20 [2, 1000] 2,049,000
====================================================================================================
Total params: 38,345,632
Trainable params: 38,345,632
Non-trainable params: 0
Total mult-adds (G): 11.19
====================================================================================================
Input size (MB): 2.15
Forward/backward pass size (MB): 259.10
Params size (MB): 92.05
Estimated Total Size (MB): 353.29
====================================================================================================
Forward Pass
x = torch.randn(2,3,299,299)
aux_out, out = model(x)
print(aux_out.shape)
print(out.shape)
torch.Size([2, 1000])
torch.Size([2, 1000])