Add cnn model

machine_learning/cnn.py

@@ -0,0 +1,67 @@
+"""
+Convolutional Neural Network (CNN) implementation for image classification.
+
+Reference: https://en.wikipedia.org/wiki/Convolutional_neural_network
+
+>>> import numpy as np
+>>> model = SimpleCNN(input_shape=(1, 28, 28), num_classes=10)
+>>> dummy_input = np.random.rand(1, 28, 28)
+>>> output = model.forward(dummy_input)
+>>> output.shape
+(10,)
+"""
+
+import numpy as np
+
+
+class SimpleCNN:
+    def __init__(self, input_shape: tuple[int, int, int], num_classes: int) -> None:
+        """
+        Initialize a simple CNN model.
+
+        Args:
+            input_shape: Tuple of (channels, height, width)
+            num_classes: Number of output classes
+        """
+        self.input_shape = input_shape
+        self.num_classes = num_classes
+        rng = np.random.default_rng()
+        self.filters = rng.normal(0, 0.1, size=(8, input_shape[0], 3, 3))  # 8 filters
+        self.fc_weights = rng.normal(0, 0.1, size=(8 * 26 * 26, num_classes))
+
+    def relu(self, feature_map: np.ndarray) -> np.ndarray:
+        """Apply ReLU activation to the feature map."""
+        return np.maximum(0, feature_map)
+
+    def convolve(self, input_tensor: np.ndarray, filters: np.ndarray) -> np.ndarray:
+        """Apply convolution operation to the input tensor."""
+        _, height, width = input_tensor.shape
+        num_filters, _, fh, fw = filters.shape
+        output = np.zeros((num_filters, height - fh + 1, width - fw + 1))
+
+        for f in range(num_filters):
+            for i in range(height - fh + 1):
+                for j in range(width - fw + 1):
+                    region = input_tensor[:, i : i + fh, j : j + fw]
+                    output[f, i, j] = np.sum(region * filters[f])
+        return output
+
+    def flatten(self, feature_map: np.ndarray) -> np.ndarray:
+        """Flatten the feature map into a 1D array."""
+        return feature_map.reshape(-1)
+
+    def forward(self, input_tensor: np.ndarray) -> np.ndarray:
+        """
+        Forward pass through the CNN.
+
+        Args:
+            input_tensor: Input image of shape (channels, height, width)
+
+        Returns:
+            Output logits of shape (num_classes,)
+        """
+        conv_out = self.convolve(input_tensor, self.filters)
+        activated = self.relu(conv_out)
+        flattened = self.flatten(activated)
+        logits = flattened @ self.fc_weights
+        return logits