Uses following API:
Python:
Conv2D(in_channel, o_channel, kernel_size, stride, mode)
[int, 3D FloatTensor] Conv2D.forward(input_image)
Conv2D is a class and it has a forward function as one of its method (apart from its constructor).
-
Did not use zero padding so completed images are somewhat smaller.
-
Saving each output image to workspace directory as:
Task<Task Number>_Image<First, Second, Third, etc. Kernel>_<Original Size>.jpg -
Input images stored in agagneja_HW01/images
- Image1 is 1280 x 720
- Image2 is 1920 x 1080
-
Counted operations as total number of operations including kernel multiplication and addition as well as a divide step where I normalized my pictures by dividing by 3 * kernal_size ^ 2 for the 3 channels and each operation
-
Program took too long so I was unable to include all pictures (especially for task 2 and the larger picture)
-
Could not complete parts B and C because of how long program took to run
-
test.py simply runs all 3 tasks on both images and outputs the files to