这是一份练习代码,但感觉有点意思。
效果
import
numpy as np
class
Dot(
object
):
def
forward(
self
, X, W):
self
.X
=
X
self
.W
=
W
return
np.dot(X, W)
def
backward(
self
, dout):
return
np.dot(dout,
self
.W.T), np.dot(
self
.X.T, dout)
class
Add(
object
):
def
forward(
self
, D, B):
return
D
+
B
def
backward(
self
, dout):
return
np.
sum
(dout, axis
=
0
)
class
Affine(
object
):
def
__init__(
self
):
self
.dot
=
Dot()
self
.add
=
Add()
def
forward(
self
, X, W, B):
T
=
self
.dot.forward(X, W)
return
self
.add.forward(T, B)
def
backward(
self
, dout):
dB
=
self
.add.backward(dout)
dX, dW
=
self
.dot.backward(dout)
return
dX, dW, dB
class
Sigmoid(
object
):
def
forward(
self
, A):
self
.out
=
1
/
(
1
+
np.exp(
-
A))
return
self
.out
def
backward(
self
, dout):
return
dout
*
(
1.0
-
self
.out)
*
self
.out
class
Relu(
object
):
def
forward(
self
, A):
self
.mask
=
(A<
=
0
)
A[
self
.mask]
=
0
return
A
def
backward(
self
, dout):
dout[
self
.mask]
=
0
return
dout
class
SoftmaxWithLoss(
object
):
def
softmax(
self
, A):
A
=
A
-
A.
max
(axis
=
1
, keepdims
=
True
)
T
=
np.exp(A)
return
T
/
np.
sum
(T, axis
=
1
, keepdims
=
True
)
def
crossEntropyError(
self
, Z, Label):
delta
=
0.000000001
return
-
np.
sum
(np.log(Z
+
delta)
*
Label)
/
Z.shape[
0
]
def
forward(
self
, A, Label):
self
.Z
=
self
.softmax(A)
self
.Label
=
Label
return
self
.Z,
self
.crossEntropyError(
self
.Z, Label)
def
backward(
self
, dout
=
1
):
return
(
self
.Z
-
self
.Label)
*
dout
/
self
.Z.shape[
0
]
class
MSELoss(
object
):
def
forward(
self
, A, Label):
self
.A
=
A
self
.Label
=
Label
return
np.
sum
((A
-
Label)
*
(A
-
Label))
/
A.shape[
0
]
def
backward(
self
, dout
=
1
):
return
2
*
(
self
.A
-
self
.Label)
*
dout
/
self
.A.shape[
0
]
class
MyNN(
object
):
def
__init__(
self
):
self
.lr
=
0.01
self
.MAX_NUM
=
1000000
def
buildThreeLayerNet(
self
, input_dim, output_dim):
N0
=
input_dim
N1
=
10
N2
=
10
N3
=
output_dim
self
.W1
=
np.random.randn(N0, N1)
self
.B1
=
np.random.randn(N1)
self
.W2
=
np.random.randn(N1, N2)
self
.B2
=
np.random.randn(N2)
self
.W3
=
np.random.randn(N2, N3)
self
.B3
=
np.random.randn(N3)
self
.affine1
=
Affine()
self
.activation1
=
Sigmoid()
self
.affine2
=
Affine()
self
.activation2
=
Sigmoid()
self
.affine3
=
Affine()
self
.mseloss
=
MSELoss()
def
getMaxAbsValue(
self
, X):
a
=
np.
max
(X)
b
=
np.
min
(X)
if
b<
0
:
b
=
-
b
ret
=
max
(a,b)
if
ret
=
=
0
:
ret
=
0.0000001
return
ret
def
learn(
self
, X, Y):
input_dim
=
len
(X[
0
])
output_dim
=
len
(Y[
0
])
self
.MAX_X
=
self
.getMaxAbsValue(X)
self
.MAX_Y
=
self
.getMaxAbsValue(Y)
X
=
np.array(X)
/
self
.MAX_X
Y
=
np.array(Y)
/
self
.MAX_Y
self
.buildThreeLayerNet(input_dim, output_dim)
for
i
in
range
(
0
,
self
.MAX_NUM):
A1
=
self
.affine1.forward(X,
self
.W1,
self
.B1)
Z1
=
self
.activation1.forward(A1)
A2
=
self
.affine2.forward(Z1,
self
.W2,
self
.B2)
Z2
=
self
.activation2.forward(A2)
A3
=
self
.affine3.forward(Z2,
self
.W3,
self
.B3)
l
=
self
.mseloss.forward(A3, Y)
if
i
%
10000
=
=
0
:
print
(l)
if
l<
0.0001
:
break
dA3
=
self
.mseloss.backward(dout
=
1
)
dZ2, dW3, dB3
=
self
.affine3.backward(dA3)
dA2
=
self
.activation2.backward(dZ2)
dZ1, dW2, dB2
=
self
.affine2.backward(dA2)
dA1
=
self
.activation1.backward(dZ1)
dX, dW1, dB1
=
self
.affine1.backward(dA1)
self
.W1
-
=
self
.lr
*
dW1
self
.B1
-
=
self
.lr
*
dB1
self
.W2
-
=
self
.lr
*
dW2
self
.B2
-
=
self
.lr
*
dB2
self
.W3
-
=
self
.lr
*
dW3
self
.B3
-
=
self
.lr
*
dB3
def
predict(
self
, X):
X
=
np.array(X)
/
self
.MAX_X
A1
=
self
.affine1.forward(X,
self
.W1,
self
.B1)
Z1
=
self
.activation1.forward(A1)
A2
=
self
.affine2.forward(Z1,
self
.W2,
self
.B2)
Z2
=
self
.activation2.forward(A2)
A3
=
self
.affine3.forward(Z2,
self
.W3,
self
.B3)
return
A3
*
self
.MAX_Y
class
FitFx(
object
):
def
fx1(
self
, x):
return
x
-
5
def
fx2(
self
, x):
return
(x
-
5
)
*
*
2
def
fx3(
self
, x):
return
(x
-
5
)
*
*
3
def
fx4(
self
, x):
return
(x
-
5
)
*
*
4
def
fx_xor(
self
, x1, x2):
return
x1^x2
def
fit11(
self
, fx):
print
(fx)
X
=
[[i]
for
i
in
range
(
0
,
11
)]
print
(X)
Y
=
[[fx(a[
0
])]
for
a
in
X]
print
(Y)
nn
=
MyNN()
nn.learn(X,Y)
print
(nn.predict(X))
def
fit21(
self
, fx):
print
(fx)
X
=
[[d1,d2]
for
d1,d2
in
[(
0
,
0
), (
0
,
1
), (
1
,
0
), (
1
,
1
)]]
print
(X)
Y
=
[[fx(
*
a)]
for
a
in
X]
print
(Y)
nn
=
MyNN()
nn.learn(X,Y)
print
(nn.predict(X))
def
main():
m
=
FitFx()
m.fit11(m.fx2)
if
"__main__"
=
=
__name__:
main()
import
numpy as np
class
Dot(
object
):
def
forward(
self
, X, W):
self
.X
=
X
self
.W
=
W
return
np.dot(X, W)
def
backward(
self
, dout):
return
np.dot(dout,
self
.W.T), np.dot(
self
.X.T, dout)
class
Add(
object
):
def
forward(
self
, D, B):
return
D
+
B
def
backward(
self
, dout):
return
np.
sum
(dout, axis
=
0
)
class
Affine(
object
):
def
__init__(
self
):
self
.dot
=
Dot()
self
.add
=
Add()
def
forward(
self
, X, W, B):
T
=
self
.dot.forward(X, W)
return
self
.add.forward(T, B)
def
backward(
self
, dout):
dB
=
self
.add.backward(dout)
dX, dW
=
self
.dot.backward(dout)
return
dX, dW, dB
class
Sigmoid(
object
):
def
forward(
self
, A):
self
.out
=
1
/
(
1
+
np.exp(
-
A))
return
self
.out
def
backward(
self
, dout):
return
dout
*
(
1.0
-
self
.out)
*
self
.out
class
Relu(
object
):
def
forward(
self
, A):
self
.mask
=
(A<
=
0
)
A[
self
.mask]
=
0
return
A
def
backward(
self
, dout):
dout[
self
.mask]
=
0
return
dout
class
SoftmaxWithLoss(
object
):
def
softmax(
self
, A):
A
=
A
-
A.
max
(axis
=
1
, keepdims
=
True
)
T
=
np.exp(A)
return
T
/
np.
sum
(T, axis
=
1
, keepdims
=
True
)
def
crossEntropyError(
self
, Z, Label):
delta
=
0.000000001
return
-
np.
sum
(np.log(Z
+
delta)
*
Label)
/
Z.shape[
0
]
def
forward(
self
, A, Label):
self
.Z
=
self
.softmax(A)
self
.Label
=
Label
return
self
.Z,
self
.crossEntropyError(
self
.Z, Label)
def
backward(
self
, dout
=
1
):
return
(
self
.Z
-
self
.Label)
*
dout
/
self
.Z.shape[
0
]
class
MSELoss(
object
):
def
forward(
self
, A, Label):
self
.A
=
A
self
.Label
=
Label
return
np.
sum
((A
-
Label)
*
(A
-
Label))
/
A.shape[
0
]
def
backward(
self
, dout
=
1
):
return
2
*
(
self
.A
-
self
.Label)
*
dout
/
self
.A.shape[
0
]
class
MyNN(
object
):
def
__init__(
self
):
self
.lr
=
0.01
self
.MAX_NUM
=
1000000
def
buildThreeLayerNet(
self
, input_dim, output_dim):
N0
=
input_dim
N1
=
10
N2
=
10
N3
=
output_dim
self
.W1
=
np.random.randn(N0, N1)
self
.B1
=
np.random.randn(N1)
self
.W2
=
np.random.randn(N1, N2)
self
.B2
=
np.random.randn(N2)
self
.W3
=
np.random.randn(N2, N3)
self
.B3
=
np.random.randn(N3)
self
.affine1
=
Affine()
self
.activation1
=
Sigmoid()
self
.affine2
=
Affine()
self
.activation2
=
Sigmoid()
self
.affine3
=
Affine()
self
.mseloss
=
MSELoss()
def
getMaxAbsValue(
self
, X):
a
=
np.
max
(X)
b
=
np.
min
(X)
[培训]内核驱动高级班,冲击BAT一流互联网大厂工作,每周日13:00-18:00直播授课