1️⃣ Architecture

• ResNet-34

  Layer	# Filters / neurons	Filter Size	Stride	Padding	Size of Feature Map
  Input	-	-	-	-	224 x 224 x 3
  Convolution 1	64	7 x 7	2	3	112 x 112 x 64
  Max Pool 1	-	3 x 3	2	1	56 x 56 x 64
  Convolution 2_A(Repetition: 6)	64	3 x 3	1	1	56 x 56 x 64
  Convolution 3_A	128	3 x 3	2	1	28 x 28 x 128
  Convolution 3_B(Repetition: 7)	128	3 x 3	1	1	28 x 28 x 128
  Convolution 4_A	256	3 x 3	2	1	14 x 14 x 256
  Convolution 4_B(Repetition: 11)	256	3 x 3	1	1	14 x 14 x 256
  Convolution 5_A	512	3 x 3	2	1	7 x 7 x 512
  Convolution 5_B(Repetition: 5)	512	3 x 3	1	1	7 x 7 x 512
  Average Pool 1	-	7 x 7	1	-	1 x 1 x 512
  Dense(FC) 1	1000	-	-	-	1 x 1000
  Softmax	1000	-	-	-	1 x 1000

• StreamlinedNet

  Layer	# Filters / neurons	Filter Size	Stride	Padding	Size of Feature Map
  Input	-	-	-	-	224 x 224 x 3
  Convolution 1	64	7 x 7	2	3	112 x 112 x 64
  Max Pool 1	-	3 x 3	2	1	56 x 56 x 64
  1xn Convolution A	64	1 x 9	1	(0, 4)	56 x 56 x 64
  3x3 Dilated Convolution A	64	3 x 3	2	2	28 x 28 x 128
  1xn Convolution B
  Shortcuts and Repetition: 3	128	1 x 9	1	(0, 4)	28 x 28 x 128
  3x3 Dilated Convolution B	128	3 x 3	2	2	14 x 14 x 256
  1xn Convolution C
  Shortcuts and Repetition: 4	256	1 x 9	1	(0, 4)	14 x 14 x 256
  3x3 Dilated Convolution C	256	3 x 3	2	2	7 x 7 x 512
  1xn Convolution D
  Shortcuts and Repetition: 5	512	1 x 7	1	(0, 3)	7 x 7 x 512
  Avgerage Pool 1	-	7 x 7	1	-	1 x 1 x 512
  Dense(FC) 1	1000	-	-	-	1 x 1000
  Softmax	1000	-	-	-	1 x 1000

2️⃣ Implement Code

• Model Code

  import torch
  
  class StreamlinedNet(torch.nn.Module):
      """
      Implementation of the StreamlinedNet model.
      > 1xn or nx1 Convolution, Dilated Convolution
      Input: Image tensor (batch_size, 3, 224, 224)
      Output: Class scores (batch_size, 1000)
      """
      def __init__(self, dropout_rate=0.5):
          """
          Args:
              dropout_rate (float): Dropout probability (default=0.5)
          """
          super(StreamlinedNet, self).__init__()
  
          # Convolutional and pooling layers
          # Layer1 > Convolution 1 & Max Pool 1
          self.layer1 = torch.nn.Sequential(
              ConvolutionalLayer.create(3, 64, 7, 2, 3),
              torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
          )
  
          # Layer2 > 1xn Convolution 1
          self.layer2 = ConvolutionalLayer.create(64, 64, (1, 9), 1, (0, 4))
  
          # Layer3 > 3x3 Dilated Convolution 1
          self.layer3 = torch.nn.Sequential(
              torch.nn.Conv2d(64, 64, kernel_size=3, stride=2, padding=2, dilation=2),
              torch.nn.BatchNorm2d(64),
              CustomReLU()
          )
  
          # Layer4 > 1xn Convolution 2; Shortcuts and Repetition: 3
          self.layer4 = torch.nn.Sequential(
              PlainBlock(64, 128, (1, 9), (0, 4)),
              *[PlainBlock(128, 128, (1, 9), (0, 4)) for _ in range(2)]
          )
  
          # Layer5 > 3x3 Dilated Convolution 2
          self.layer5 = torch.nn.Sequential(
              torch.nn.Conv2d(128, 128, kernel_size=3, stride=2, padding=2, dilation=2),
              torch.nn.BatchNorm2d(128),
              CustomReLU()
          )
  
          # Layer6 > 1xn Convolution 3; Shortcuts and Repetition: 4
          self.layer6 = torch.nn.Sequential(
              PlainBlock(128, 256, (1, 9), (0, 4)),
              *[PlainBlock(256, 256, (1, 9), (0, 4)) for _ in range(3)]
          )
  
          # Layer7 > 3x3 Dilated Convolution 3
          self.layer7 = torch.nn.Sequential(
              torch.nn.Conv2d(256, 256, kernel_size=3, stride=2, padding=2, dilation=2),
              torch.nn.BatchNorm2d(256),
              CustomReLU()
          )
  
          # Layer8 > 1xn Convolution 4; Shortcuts and Repetition: 5
          self.layer8 = torch.nn.Sequential(
              PlainBlock(256, 512, (1, 7), (0, 3)),
              *[PlainBlock(512, 512, (1, 7), (0, 3)) for _ in range(4)]
          )
  
          # Layer9 > Gloval Average Pooling to reduce spatial dimensions to 1x1
          self.layer9 = torch.nn.AvgPool2d(kernel_size=7, stride=1, padding=0)
  
          
          # Fully Connected layers and dropout
          self.layer_drop = torch.nn.Dropout(p=dropout_rate)
  
          self.layer10 = FullyConnectedLayer.Dense(1 * 1 * 512, 1000)
  
          self.layer11 = FullyConnectedLayer.Dense(1000, 1000)
          
      def forward(self, x: torch.Tensor) -> torch.Tensor:
          """
          Defines the forward pass of the model.
          Args:
              x (torch.Tensor): Input image tensor (batch_size, 3, 224, 224)
          Returns:
              torch.Tensor: Class scores (batch_size, 1000)
          """
          x = self.layer1(x)
          x = self.layer2(x)
          x = self.layer3(x)
          x = self.layer4(x)
          x = self.layer5(x)
          x = self.layer6(x)
          x = self.layer7(x)
          x = self.layer8(x)
          x = self.layer9(x)
  
          x = self.layer_drop(x)
          x = x.view(x.size(0), -1)  # Flatten
  
          x = self.layer10(x)
          x = self.layer11(x)
          return x
  

3️⃣ Use Dataset

• PyTorch Food101

  Food101 — Torchvision main documentation

  • Information

    Food-101 -- Mining Discriminative Components with Random Forests

4️⃣ Training Result (Mar 18, 2025.)

• Epoch 25, Learning Rate 0.001, Batch Size 100
  • Final [Epoch: 25] cost = 0.809117258 | Train Accuracy = 76.28%
  • Top [Epoch: 24] cost = 0.679335654 | Train Accuracy = 79.24%
• 해당 결과 분석
  • 학습이 진행되고 있음을 보여주는 지표인 Cost는 낮아지고, Train Accuracy는 높아지고 있음을 볼 수 있음
  • Gradient를 찾아가고 있음을 Epoch 24 → 25를 통해 확인할 수 있음
  • Epoch를 100까지 늘려서, 최종적으로 Accuracy가 95% 이상 달성하는지 확인하고자 함

5️⃣ Training Result (Mar 24, 2025.)

• Test0(Kaggle)

  [Epoch:    1] cost = 4.66245747 | Train Accuracy = 2.89%
  [Epoch:    2] cost = 4.10329485 | Train Accuracy = 7.53%
  [Epoch:    3] cost = 3.81392169 | Train Accuracy = 11.65%
  [Epoch:    4] cost = 3.59057093 | Train Accuracy = 15.48%
  [Epoch:    5] cost = 3.38048482 | Train Accuracy = 19.35%
  [Epoch:    6] cost = 3.17903066 | Train Accuracy = 23.22%
  [Epoch:    7] cost = 2.99323297 | Train Accuracy = 26.80%
  [Epoch:    8] cost = 2.8377564 | Train Accuracy = 30.25%
  [Epoch:    9] cost = 2.67238164 | Train Accuracy = 33.66%
  [Epoch:   10] cost = 2.52783251 | Train Accuracy = 36.85%
  [Epoch:   11] cost = 2.39527988 | Train Accuracy = 39.49%
  [Epoch:   12] cost = 2.26512456 | Train Accuracy = 42.37%
  [Epoch:   13] cost = 2.14713264 | Train Accuracy = 45.00%
  [Epoch:   14] cost = 2.02744007 | Train Accuracy = 47.44%
  [Epoch:   15] cost = 1.9079479 | Train Accuracy = 50.32%
  [Epoch:   16] cost = 1.78312731 | Train Accuracy = 53.00%
  [Epoch:   17] cost = 1.69548333 | Train Accuracy = 54.94%
  [Epoch:   18] cost = 1.56390786 | Train Accuracy = 57.66%
  [Epoch:   19] cost = 1.42279017 | Train Accuracy = 60.84%
  [Epoch:   20] cost = 1.28580725 | Train Accuracy = 64.03%
  [Epoch:   21] cost = 1.16848469 | Train Accuracy = 66.77%
  [Epoch:   22] cost = 1.04328561 | Train Accuracy = 69.70%
  [Epoch:   23] cost = 0.924164772 | Train Accuracy = 72.77%
  [Epoch:   24] cost = 0.814349592 | Train Accuracy = 75.44%
  [Epoch:   25] cost = 0.729309857 | Train Accuracy = 77.93%
  [Epoch:   26] cost = 0.634472549 | Train Accuracy = 80.33%
  [Epoch:   27] cost = 0.571774662 | Train Accuracy = 82.18%
  [Epoch:   28] cost = 0.516791224 | Train Accuracy = 83.76%
  [Epoch:   29] cost = 0.464977682 | Train Accuracy = 85.20%
  [Epoch:   30] cost = 0.426421076 | Train Accuracy = 86.37%
  [Epoch:   31] cost = 0.408697188 | Train Accuracy = 87.03%
  [Epoch:   32] cost = 0.376045227 | Train Accuracy = 88.16%
  [Epoch:   33] cost = 0.342416942 | Train Accuracy = 88.97%
  [Epoch:   34] cost = 0.328432947 | Train Accuracy = 89.63%
  [Epoch:   35] cost = 0.315804303 | Train Accuracy = 90.03%
  [Epoch:   36] cost = 0.292550355 | Train Accuracy = 90.66%
  [Epoch:   37] cost = 0.284003198 | Train Accuracy = 91.00%
  [Epoch:   38] cost = 0.270740956 | Train Accuracy = 91.42%
  [Epoch:   39] cost = 0.266808569 | Train Accuracy = 91.52%
  [Epoch:   40] cost = 0.24558875 | Train Accuracy = 92.18%
  [Epoch:   41] cost = 0.236012876 | Train Accuracy = 92.39%
  [Epoch:   42] cost = 0.230435148 | Train Accuracy = 92.72%
  [Epoch:   43] cost = 0.244548112 | Train Accuracy = 92.20%
  [Epoch:   44] cost = 0.214064136 | Train Accuracy = 93.21%
  [Epoch:   45] cost = 0.210372195 | Train Accuracy = 93.36%
  [Epoch:   46] cost = 0.205122143 | Train Accuracy = 93.48%
  [Epoch:   47] cost = 0.202804595 | Train Accuracy = 93.64%
  [Epoch:   48] cost = 0.198964417 | Train Accuracy = 93.74%
  [Epoch:   49] cost = 0.185472012 | Train Accuracy = 94.13%
  **[Epoch:   50] cost = 0.184777334 | Train Accuracy = 94.23%**
  

• Test1(Google Colab)

  [Epoch:    1] cost = 4.63286686 | Train Accuracy = 3.22%
  [Epoch:    2] cost = 4.10520792 | Train Accuracy = 7.72%
  [Epoch:    3] cost = 3.81570077 | Train Accuracy = 12.10%
  [Epoch:    4] cost = 3.55455017 | Train Accuracy = 16.29%
  [Epoch:    5] cost = 3.31452513 | Train Accuracy = 20.62%
  [Epoch:    6] cost = 3.10248852 | Train Accuracy = 24.72%
  [Epoch:    7] cost = 2.91098142 | Train Accuracy = 28.83%
  [Epoch:    8] cost = 2.7413168 | Train Accuracy = 32.28%
  [Epoch:    9] cost = 2.58862662 | Train Accuracy = 35.32%
  [Epoch:   10] cost = 2.43815517 | Train Accuracy = 38.63%
  [Epoch:   11] cost = 2.30618572 | Train Accuracy = 41.53%
  [Epoch:   12] cost = 2.35105157 | Train Accuracy = 41.13%
  [Epoch:   13] cost = 2.18907833 | Train Accuracy = 44.19%
  [Epoch:   14] cost = 1.98132408 | Train Accuracy = 48.76%
  [Epoch:   15] cost = 1.86699438 | Train Accuracy = 51.35%
  [Epoch:   16] cost = 1.73786211 | Train Accuracy = 53.61%
  [Epoch:   17] cost = 1.6146971 | Train Accuracy = 56.45%
  **[Epoch:   18] cost = 1.49757981 | Train Accuracy = 59.05%**
  

• Test2(Google Colab)

  [Epoch:    1] cost = 4.64463949 | Train Accuracy = 3.45%
  [Epoch:    2] cost = 4.05979538 | Train Accuracy = 8.33%
  [Epoch:    3] cost = 3.75822496 | Train Accuracy = 12.79%
  [Epoch:    4] cost = 3.52590299 | Train Accuracy = 16.75%
  [Epoch:    5] cost = 3.3062396 | Train Accuracy = 20.89%
  [Epoch:    6] cost = 3.10596752 | Train Accuracy = 24.81%
  [Epoch:    7] cost = 2.90648746 | Train Accuracy = 28.83%
  [Epoch:    8] cost = 2.7514317 | Train Accuracy = 32.24%
  [Epoch:    9] cost = 2.5767386 | Train Accuracy = 35.91%
  [Epoch:   10] cost = 2.55877733 | Train Accuracy = 36.30%
  [Epoch:   11] cost = 2.3319118 | Train Accuracy = 40.91%
  [Epoch:   12] cost = 2.20079851 | Train Accuracy = 43.85%
  [Epoch:   13] cost = 2.08096933 | Train Accuracy = 46.65%
  [Epoch:   14] cost = 1.96262825 | Train Accuracy = 49.00%
  [Epoch:   15] cost = 1.85471165 | Train Accuracy = 51.51%
  [Epoch:   16] cost = 1.71026862 | Train Accuracy = 54.73%
  [Epoch:   17] cost = 1.58452141 | Train Accuracy = 57.11%
  **[Epoch:   18] cost = 1.4524734 | Train Accuracy = 60.25%**
  

• Test3(Google Colab)

  [Epoch:    1] cost = 4.64526272 | Train Accuracy = 2.85%
  [Epoch:    2] cost = 4.0693264 | Train Accuracy = 8.35%
  [Epoch:    3] cost = 3.73745489 | Train Accuracy = 13.18%
  [Epoch:    4] cost = 3.50912642 | Train Accuracy = 17.11%
  [Epoch:    5] cost = 3.37433243 | Train Accuracy = 19.90%
  [Epoch:    6] cost = 3.1333003 | Train Accuracy = 24.31%
  [Epoch:    7] cost = 2.93988013 | Train Accuracy = 28.06%
  [Epoch:    8] cost = 2.7723 | Train Accuracy = 31.56%
  [Epoch:    9] cost = 2.621418 | Train Accuracy = 34.82%
  [Epoch:   10] cost = 2.46938396 | Train Accuracy = 37.92%
  [Epoch:   11] cost = 2.33324432 | Train Accuracy = 40.95%
  [Epoch:   12] cost = 2.19547772 | Train Accuracy = 43.85%
  [Epoch:   13] cost = 2.06170535 | Train Accuracy = 46.78%
  [Epoch:   14] cost = 1.93124735 | Train Accuracy = 49.62%
  [Epoch:   15] cost = 1.80505764 | Train Accuracy = 52.55%
  [Epoch:   16] cost = 1.66252995 | Train Accuracy = 55.61%
  **[Epoch:   17] cost = 1.52216458 | Train Accuracy = 58.56%**
  

• Test4(Google Colab)

  [Epoch:    1] cost = 4.69737673 | Train Accuracy = 2.43%
  [Epoch:    2] cost = 4.15832233 | Train Accuracy = 6.93%
  [Epoch:    3] cost = 3.83768678 | Train Accuracy = 11.54%
  [Epoch:    4] cost = 3.62141037 | Train Accuracy = 15.13%
  [Epoch:    5] cost = 3.3984046 | Train Accuracy = 19.00%
  [Epoch:    6] cost = 3.1942234 | Train Accuracy = 22.97%
  [Epoch:    7] cost = 3.00066614 | Train Accuracy = 26.85%
  [Epoch:    8] cost = 2.81827188 | Train Accuracy = 30.64%
  [Epoch:    9] cost = 2.65976906 | Train Accuracy = 33.93%
  [Epoch:   10] cost = 2.51142859 | Train Accuracy = 37.03%
  [Epoch:   11] cost = 2.35400581 | Train Accuracy = 40.38%
  [Epoch:   12] cost = 2.20701766 | Train Accuracy = 43.46%
  [Epoch:   13] cost = 2.07967377 | Train Accuracy = 46.32%
  [Epoch:   14] cost = 1.95324934 | Train Accuracy = 49.08%
  [Epoch:   15] cost = 1.80599356 | Train Accuracy = 52.29%
  [Epoch:   16] cost = 1.65747666 | Train Accuracy = 55.50%
  **[Epoch:   17] cost = 1.51262736 | Train Accuracy = 58.78%**
  

• Test5(Google Colab)

  [Epoch:    1] cost = 4.65966272 | Train Accuracy = 3.03%
  [Epoch:    2] cost = 4.13851261 | Train Accuracy = 7.32%
  [Epoch:    3] cost = 3.84493947 | Train Accuracy = 11.52%
  [Epoch:    4] cost = 3.59627366 | Train Accuracy = 15.47%
  [Epoch:    5] cost = 3.37425065 | Train Accuracy = 19.39%
  [Epoch:    6] cost = 3.16760945 | Train Accuracy = 23.31%
  [Epoch:    7] cost = 2.9613378 | Train Accuracy = 27.81%
  [Epoch:    8] cost = 2.77564049 | Train Accuracy = 31.64%
  [Epoch:    9] cost = 2.61471176 | Train Accuracy = 34.73%
  [Epoch:   10] cost = 2.46067524 | Train Accuracy = 38.14%
  [Epoch:   11] cost = 2.31782818 | Train Accuracy = 41.05%
  [Epoch:   12] cost = 2.17893553 | Train Accuracy = 44.09%
  [Epoch:   13] cost = 2.04192352 | Train Accuracy = 47.07%
  [Epoch:   14] cost = 1.91332364 | Train Accuracy = 49.98%
  [Epoch:   15] cost = 1.78040743 | Train Accuracy = 52.89%
  [Epoch:   16] cost = 1.64149928 | Train Accuracy = 55.90%
  [Epoch:   17] cost = 1.48815262 | Train Accuracy = 59.15%
  **[Epoch:   18] cost = 1.33682942 | Train Accuracy = 62.64%**