Custom Multidimensional Convolution | Dataford Interview Questions

Implement a custom multidimensional convolution operator in Python without relying on high-level deep learning frameworks.

Implement custom_conv_nd(input_tensor, kernel, stride, padding, dilation). Both tensors are rectangular nested lists with the same number of dimensions. Compute mathematical convolution by reversing the kernel along every axis, using zero-padding, per-axis stride, and dilation. Return the result as a nested list with the appropriate output shape. Do not use NumPy or deep learning libraries.

Constraints

1 <= number of dimensions <= 5
Every tensor is a non-empty rectangular nested list
input_tensor and kernel have the same number of dimensions
1 <= input_tensor dimension size <= 32
1 <= kernel dimension size <= 8
1 <= stride[i] and dilation[i] <= 4
0 <= padding[i] <= 8
The computed output size is positive on every axis
Tensor values are integers or floating-point numbers

Function Signature

def custom_conv_nd(input_tensor, kernel, stride, padding, dilation):

Your question is Custom Multidimensional Convolution. Start with the requirements in the Question tab.

Run and submit as often as you like. When you're ready, talk me through your approach or go straight to the code.

Problem

Implement a custom multidimensional convolution operator in Python without relying on high-level deep learning frameworks.

Constraints

1 <= number of dimensions <= 5
Every tensor is a non-empty rectangular nested list
input_tensor and kernel have the same number of dimensions
1 <= input_tensor dimension size <= 32
1 <= kernel dimension size <= 8
1 <= stride[i] and dilation[i] <= 4
0 <= padding[i] <= 8
The computed output size is positive on every axis
Tensor values are integers or floating-point numbers

Function Signature

def custom_conv_nd(input_tensor, kernel, stride, padding, dilation):

Your question is Custom Multidimensional Convolution. Start with the requirements in the Question tab.

Run and submit as often as you like. When you're ready, talk me through your approach or go straight to the code.