ConvFilter Kernel

readme.md

Overview

In this simple exercise, a simple ConvFilter in Numpy is created

conv.py

# Defines a WxHxC ConvFilter 
# Conventional Internal Representation is C,H,W 
class Conv: 
  # w, h = Kernel Spatial Domain 
  # in_c = Input Tensor Channels 
  # out_c = Output Tensor Channels 
  def __init__(self, w, h, in_c, out_c): 
    self.w = w 
    self.h = h 
    self.in_c = in_c
    self.out_c = out_c
    
    # Choose the preferred initialization: ones or random 
    self.filters = np.ones((out_c, in_c,h,w)) 
    #self.filters = np.random.randn(out_c, in_c,h,w) / 9 
    
  # Image Convention is CxHxW 
  def apply(self, image): 
    if(self.in_c != image.shape[0]): 
      raise("Channel Size Mismatch")
      
    out_h = image.shape[1] - 2*(self.h//2)
    out_w = image.shape[2] - 2*(self.w//2)
    res = np.zeros((self.out_c, out_h, out_w))
    # Spatial Loop 
    for _c in range(self.out_c): 
      for _h in range( out_h ): 
        for _w in range( out_w ): 
          res[_c, _h, _w] = np.sum(image[:, _h:(_h+self.h), _w:(_w+self.w)] * self.filters[_c,:,:,:] )
          
    return res

test.py

conv1 = Conv(3,3,1,6)
test_image = np.ones((1, 10, 10))
res = conv1.apply(test_image)