Add LeNet Implementation in PyTorch

computer_vision/lenet_pytorch.py

@@ -0,0 +1,82 @@
+"""
+LeNet Network
+
+Paper: http://vision.stanford.edu/cs598_spring07/papers/Lecun98.pdf
+"""
+
+import numpy
+import torch
+import torch.nn as nn
+
+
+class LeNet(nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+
+        self.tanh = nn.Tanh()
+        self.avgpool = nn.AvgPool2d(kernel_size=2, stride=2)
+
+        self.conv1 = nn.Conv2d(
+            in_channels=1,
+            out_channels=6,
+            kernel_size=(5, 5),
+            stride=(1, 1),
+            padding=(0, 0),
+        )
+        self.conv2 = nn.Conv2d(
+            in_channels=6,
+            out_channels=16,
+            kernel_size=(5, 5),
+            stride=(1, 1),
+            padding=(0, 0),
+        )
+        self.conv3 = nn.Conv2d(
+            in_channels=16,
+            out_channels=120,
+            kernel_size=(5, 5),
+            stride=(1, 1),
+            padding=(0, 0),
+        )
+
+        self.linear1 = nn.Linear(120, 84)
+        self.linear2 = nn.Linear(84, 10)
+
+    def forward(self, image_array: numpy.ndarray) -> numpy.ndarray:
+        image_array = self.tanh(self.conv1(image_array))
+        image_array = self.avgpool(image_array)
+        image_array = self.tanh(self.conv2(image_array))
+        image_array = self.avgpool(image_array)
+        image_array = self.tanh(self.conv3(image_array))
+
+        image_array = image_array.reshape(image_array.shape[0], -1)
+        image_array = self.tanh(self.linear1(image_array))
+        image_array = self.linear2(image_array)
+        return image_array
+
+
+def test_model(image_tensor: torch.tensor) -> bool:
+    """
+    Test the model on an input batch of 64 images
+
+    Args:
+        image_tensor (torch.tensor): Batch of Images for the model
+
+    >>> test_model(torch.randn(64, 1, 32, 32))
+    True
+
+    """
+    try:
+        model = LeNet()
+        output = model(image_tensor)
+    except RuntimeError:
+        return False
+
+    return output.shape == torch.zeros([64, 10]).shape
+
+
+if __name__ == "__main__":
+    random_image_1 = torch.randn(64, 1, 32, 32)
+    random_image_2 = torch.randn(1, 32, 32)
+
+    print(f"random_image_1 Model Passed: {test_model(random_image_1)}")
+    print(f"\nrandom_image_2 Model Passed: {test_model(random_image_2)}")

requirements.txt

@@ -15,6 +15,7 @@ statsmodels
 sympy
 tensorflow
 texttable
+torch
 tweepy
 xgboost
 yulewalker