def main():
z = np.ones((5, 5))
k = np.ones((3, 3))
b = 3
# print(_single_channel_conv(z, k,padding=(1,1)))
# print(_single_channel_conv(z, k, strides=(2, 2)))
assert _single_channel_conv(z, k).shape == (3, 3)
assert _single_channel_conv(z, k, padding=(1, 1)).shape == (5, 5)
assert _single_channel_conv(z, k, strides=(2, 2)).shape == (2, 2)
assert _single_channel_conv(z, k, strides=(2, 2),
padding=(1, 1)).shape == (3, 3)
assert _single_channel_conv(z, k, strides=(2, 2),
padding=(1, 0)).shape == (3, 2)
assert _single_channel_conv(z, k, strides=(2, 1),
padding=(1, 1)).shape == (3, 5)
dz = np.ones((1, 1, 3, 3))
assert _insert_zeros(dz, (1, 1)).shape == (1, 1, 3, 3)
print(_insert_zeros(dz, (3, 2)))
assert _insert_zeros(dz, (1, 2)).shape == (1, 1, 3, 5)
assert _insert_zeros(dz, (2, 2)).shape == (1, 1, 5, 5)
assert _insert_zeros(dz, (2, 4)).shape == (1, 1, 5, 9)
z = np.ones((8, 16, 5, 5))
k = np.ones((16, 32, 3, 3))
b = np.ones((32))
assert conv_forward(z, k, b).shape == (8, 32, 3, 3)
print(conv_forward(z, k, b)[0, 0])
print(np.argmax(np.array([[1, 2], [3, 4]])))
def test_conv_and_max_pooling():
# 测试卷积和最大池化
z = np.random.randn(3, 3, 28, 28).astype(np.float64)
K = np.random.randn(3, 4, 3, 3).astype(np.float64) * 1e-3
b = np.zeros(4).astype(np.float64)
next_z = conv_forward(z, K, b)
y_pred = max_pooling_forward_bak(next_z, pooling=(2, 2))
y_true = np.ones_like(y_pred)
from nn.losses import mean_squared_loss
for i in range(10000):
# 前向
next_z = conv_forward(z, K, b)
y_pred = max_pooling_forward_bak(next_z, pooling=(2, 2))
# 反向
loss, dy = mean_squared_loss(y_pred, y_true)
next_dz = max_pooling_backward_bak(dy, next_z, pooling=(2, 2))
dK, db, _ = conv_backward(next_dz, K, z)
K -= 0.001 * dK
b -= 0.001 * db
if i % 10 == 0:
print("i:{},loss:{},mindy:{},maxdy:{}".format(
i, loss, np.mean(dy), np.max(dy)))
if np.allclose(y_true, y_pred):
print("yes")
break
def conv_backward(next_dz, K, z, padding=(0, 0), strides=(1, 1)):
"""
多通道卷积层的反向过程
:param next_dz: 卷积输出层的梯度,(N,D,H,W),H,W为卷积输出层的高度和宽度
:param K: 当前层卷积核，(C,D,k1,k2)
:param z: 卷积层矩阵,形状(N,C,H,W)，N为batch_size，C为通道数
:param padding: padding
:param strides: 步长
:return:
"""
N, C, H, W = z.shape
C, D, k1, k2 = K.shape
# 卷积核梯度
# dK = np.zeros((C, D, k1, k2))
padding_next_dz = _insert_zeros(next_dz, strides)
# 卷积核高度和宽度翻转180度
flip_K = np.flip(K, (2, 3))
# 交换C,D为D,C；D变为输入通道数了，C变为输出通道数了
swap_flip_K = np.swapaxes(flip_K, 0, 1)
# 增加高度和宽度0填充
ppadding_next_dz = np.lib.pad(padding_next_dz,
((0, 0), (0, 0), (k1 - 1, k1 - 1),
(k2 - 1, k2 - 1)),
'constant',
constant_values=0)
dz = conv_forward(ppadding_next_dz.astype(np.float64),
swap_flip_K.astype(np.float64),
np.zeros((C, ), dtype=np.float64))
# 求卷积和的梯度dK
swap_z = np.swapaxes(z, 0, 1) # 变为(C,N,H,W)与
dK = conv_forward(swap_z.astype(np.float64),
padding_next_dz.astype(np.float64),
np.zeros((D, ), dtype=np.float64))
# 偏置的梯度
db = np.sum(np.sum(np.sum(next_dz, axis=-1), axis=-1),
axis=0) # 在高度、宽度上相加；批量大小上相加
# 把padding减掉
dz = _remove_padding(
dz,
padding) # dz[:, :, padding[0]:-padding[0], padding[1]:-padding[1]]
return dK / N, db / N, dz
