def test_forward(self):
# filter_shape = (3, 3, 3, 3) # (n, c, ih, iw)
# input_shape = (3, 3, 5, 5)
input_shape = (3, 5, 5)
conv = Conv2D(
filter_num=3,
filter_size=(3, 3),
activation="identity",
initialization="he",
pad=0,
stride=1,
input_shape=input_shape,
)
assert conv.biases.all() == 0.0
inputs = np.random.randn(3, 3, 5, 5)
outputs = conv.forward(inputs)
assert outputs.shape == (3, 3, 3, 3) # (n, fn, oh, ow)
outputs = outputs.transpose(0, 2, 3, 1).reshape(3 * 3 * 3, 3)
expansion = im2row(inputs, conv.fh, conv.fw, conv.pad, conv.stride)
assert expansion.shape == (27, 27) # 这里设置为可逆矩阵
weight_reshaped = np.dot(np.linalg.inv(expansion), outputs)
weights = weight_reshaped.T.reshape(3, 3, 3, 3)
assert str(weights) == str(conv.weights)
def test_im2row_with_stride(self):
# input_shape = (1, 5, 5, 1)
input_shape = (1, 1, 5, 5)
img = np.array([
[1, 2, 3, 0, 1],
[0, 1, 2, 3, 2],
[3, 0, 1, 2, 3],
[2, 3, 0, 1, 0],
[1, 0, 3, 2, 1],
])
inputs = np.zeros((input_shape))
# inputs[0, :, :, 0] = img
inputs[0, 0, :, :] = img
fh = fw = 3
pad = 0
stride = 2
expansion = im2row(inputs=inputs, fh=fh, fw=fw, pad=pad, stride=stride)
target = np.array(
[
[1, 2, 3, 0, 1, 2, 3, 0, 1],
[3, 0, 1, 2, 3, 2, 1, 2, 3],
[3, 0, 1, 2, 3, 0, 1, 0, 3],
[1, 2, 3, 0, 1, 0, 3, 2, 1],
],
dtype="float64",
)
assert (expansion == target).all()
def test_im2row_loss_features(self):
# 输入特征图形状为4x4
# 过滤器形状为3x3
# 填充为0， 滑动步长为2
# 此时，过滤器只能在输入特征图上活动一次，输入特征图中局部感受野（即活动窗口）外的像素将会丢失
fh = fw = 3
pad = 0
stride = 2
expansion = im2row(self.inputs, fh, fw, pad, stride)
# print('\n', 'expansion are:')
# print(expansion)
# print('')
# print(expansion.shape)
target = np.array([[1.0, 2.0, 3.0, 0.0, 1.0, 2.0, 3.0, 0.0, 1.0]])
assert (expansion == target).all()
def test_im2row(self):
fh = fw = 3
pad = 0
stride = 1
expansion = im2row(inputs=self.inputs,
fh=fh,
fw=fw,
pad=pad,
stride=stride)
target = np.array([
[1, 2, 3, 0, 1, 2, 3, 0, 1],
[2, 3, 0, 1, 2, 3, 0, 1, 2],
[0, 1, 2, 3, 0, 1, 2, 3, 0],
[1, 2, 3, 0, 1, 2, 3, 0, 1],
])
# print('=' * 80)
# print(expansion)
# print('=' * 80)
assert (expansion == target).all()
def test_im2row_with_batch(self):
# input_shape = (2, 2, 2, 3)
input_shape = (2, 3, 2, 2)
inputs = np.arange(1, 25).reshape(*input_shape)
"""
第一批3个通道上的像素矩阵：
[[1, 2], [[5, 6], [[9, 10],
[3, 4]] [7, 8]] [11, 12]]
第二批3个通道上的像素矩阵：
[[13, 14], [[17, 18], [[21, 22],
[15, 16]] [19, 20]] [23, 24]]
----------------------------------
将统一通道上的局部感受野展开之后：
第一批： 第二批：
[[1, 5, 9], [[13, 17, 21],
[2, 6, 10], [14, 18, 22],
[3, 7, 11], [15, 19, 23],
[4, 8, 12]] [16, 20, 24]]
"""
print("\n", "inputs are:")
print(inputs)
print("")
expansion = im2row(inputs, fh=1, fw=1, pad=0, stride=1)
# print('\n')
# print(expansion)
# print('')
target = np.array([
[1.0, 5.0, 9.0], # 相同通道的像素都扩展到同一行
[2.0, 6.0, 10.0],
[3.0, 7.0, 11.0],
[4.0, 8.0, 12.0],
# ----------- 第一批
[13.0, 17.0, 21.0],
[14.0, 18.0, 22.0],
[15.0, 19.0, 23.0],
[16.0, 20.0, 24.0],
])
# ----------- 第二批
assert (expansion == target).all()
def test_im2row_with_padding(self):
fh = fw = 3
pad = 1
stride = 1
expansion = im2row(self.inputs, fh, fw, pad, stride)
# print('\n', 'expansion are:')
# print(expansion)
# print('')
target = np.array([
[
0.0,
0.0,
0.0,
0.0,
0.0,
1.0,
2.0,
3.0,
0.0,
0.0,
1.0,
2.0,
0.0,
3.0,
0.0,
1.0,
],
[
0.0,
0.0,
0.0,
0.0,
1.0,
2.0,
3.0,
0.0,
0.0,
1.0,
2.0,
3.0,
3.0,
0.0,
1.0,
2.0,
],
[
0.0,
0.0,
0.0,
0.0,
2.0,
3.0,
0.0,
0.0,
1.0,
2.0,
3.0,
0.0,
0.0,
1.0,
2.0,
0.0,
],
[
0.0,
1.0,
2.0,
3.0,
0.0,
0.0,
1.0,
2.0,
0.0,
3.0,
0.0,
1.0,
0.0,
2.0,
3.0,
0.0,
],
[
1.0,
2.0,
3.0,
0.0,
0.0,
1.0,
2.0,
3.0,
3.0,
0.0,
1.0,
2.0,
2.0,
3.0,
0.0,
1.0,
],
[
2.0,
3.0,
0.0,
0.0,
1.0,
2.0,
3.0,
0.0,
0.0,
1.0,
2.0,
0.0,
3.0,
0.0,
1.0,
0.0,
],
[
0.0,
0.0,
1.0,
2.0,
0.0,
3.0,
0.0,
1.0,
0.0,
2.0,
3.0,
0.0,
0.0,
0.0,
0.0,
0.0,
],
[
0.0,
1.0,
2.0,
3.0,
3.0,
0.0,
1.0,
2.0,
2.0,
3.0,
0.0,
1.0,
0.0,
0.0,
0.0,
0.0,
],
[
1.0,
2.0,
3.0,
0.0,
0.0,
1.0,
2.0,
0.0,
3.0,
0.0,
1.0,
0.0,
0.0,
0.0,
0.0,
0.0,
],
]).T
assert (expansion == target).all()