transpose_conv.py

"""
Example showing that backward pass of conv layer can be done using transposed convolution
"""
import numpy as np

a_prev = np.array([
    [1,1,1,1],
    [2,2,2,2],
    [3,3,3,3],
    [4,4,4,4]
], dtype=np.float32).reshape(16, 1)

w = np.array([
    [1,0,0],
    [0,1,0],
    [0,0,1]
    ], dtype=np.float32)

m = np.array([
    [1,0,0,0,0,1,0,0,0,0,1,0,0,0,0,0],
    [0,1,0,0,0,0,1,0,0,0,0,1,0,0,0,0],
    [0,0,0,0,1,0,0,0,0,1,0,0,0,0,1,0],
    [0,0,0,0,0,1,0,0,0,0,1,0,0,0,0,1]
], dtype=np.float32)

a = np.dot(m, a_prev).reshape(2,2)

da_prev = np.zeros((4, 4))

# backward pass of convolution
for i in range(2):
    v_start = 1 * i
    v_end = v_start + 3

    for j in range(2):
        h_start = 1 * j
        h_end = h_start + 3

        da_prev[v_start:v_end, h_start:h_end] += w

print("Using backward:\n", da_prev)

# backward pass using transposed conv
print("Using transposed conv:\n", np.dot(m.T, tf.ones_like(a.reshape(4, 1))).reshape(4,4))

np.testing.assert_array_almost_equal(da_prev, (np.dot(m.T, tf.ones_like(a.reshape(4, 1))).reshape(4,4)))