implement_cnn

합성곱 연산을 사용하는 신경망

row = []
matrix = []

for matrix_row in range(input.shape[0]-w.shape[0]+1):
  for matrix_col in range(input.shape[1]-w.shape[1]+1):
    matrix = []
    for w_row in range(w.shape[0]):
      row = []
      for w_col in range(w.shape[0]):
        row.append(input[matrix_row+w_row,matrix_col+w_col])
      matrix.append(row)
    print(matrix)

각 커널에 대한 합성곱 연산을 할 행렬, matrix의 생성, 이는 가중치 행렬과 동일한 크기다.

row = []
matrix = []
result = []

for matrix_row in range(input.shape[0]-w.shape[0]+1):
  for matrix_col in range(input.shape[1]-w.shape[1]+1):
    matrix = []
    for w_row in range(w.shape[0]):
      row = []
      for w_col in range(w.shape[0]):
        row.append(input[matrix_row+w_row,matrix_col+w_col])
      matrix.append(row)
    result.append(matrix*w)

result = np.array(result)
result.reshape(9,-1)
result
>>>
array([[[0.08696687, 0.78912802],
        [0.55418332, 0.44576419]],

       [[0.9076278 , 0.36782108],
        [0.29988894, 0.06308654]],

       [[0.4230551 , 0.67371227],
        [0.04244163, 0.79072746]],

       [[0.88224343, 0.41508316],
        [0.23808613, 0.5482694 ]],

       [[0.47741432, 0.05874443],
        [0.36884957, 0.88970222]],

       [[0.06756581, 0.73630332],
        [0.59854934, 0.82270165]],

       [[0.37902607, 0.51053315],
        [0.39487136, 0.31777099]],

       [[0.5871976 , 0.82846585],
        [0.21378121, 0.6028543 ]],

       [[0.95287281, 0.76607679],
        [0.4055717 , 0.72192621]]])

matrix의 크기에 따라 용량을 많이 요구할 수도 있기 때문에 생성과 동시에 연산을 수행하는 것이 더 나을 것

result_arr = []

for matrix_row in range(input.shape[0]-w.shape[0]+1):
  for matrix_col in range(input.shape[1]-w.shape[1]+1):
    result = []
    for w_row in range(w.shape[0]):
      for w_col in range(w.shape[1]):
        result.append(input[matrix_row+w_row,matrix_col+w_col] * w[w_row, w_col])
    result_arr.append(np.sum(result))

result_arr
>>>
[1.8760423945638465,
 1.6384243652554065,
 1.9299364610860785,
 2.083682123609485,
 1.7947105381568158,
 2.225120122119695,
 1.6022015709429518,
 2.2322989528570516,
 2.8464475164067666]

동일한 연산 결과 확인, 불필요한 반복의 제거

result = np.array(result_arr).reshape(input.shape[0]-w.shape[0]+1,-1)
result
>>>
array([[1.87604239, 1.63842437, 1.92993646],
       [2.08368212, 1.79471054, 2.22512012],
       [1.60220157, 2.23229895, 2.84644752]])

리스트의 행렬 변환 수행,