SDSD
- Understand batch matrix multiplication, Keras
- Keras dot, tf dot
- tf.matmul, K.batxh_dot
- 2차원에서는 K.dot와 tf.matmul 동일
-
from keras import backend as K
a = K.ones((3,4))
b = K.ones((4,5))
c = K.dot(a, b)
print(c.shape) #(3,5)
mport tensorflow as tf
a = tf.ones((3,4))
b = tf.ones((4,5))
c = tf.matmul(a, b)
print(c.shape) #(3,5)
- 고차원이면 Note that this behavior is specific to Keras dot. It is a reproduction of Theano behavior.
-
from keras import backend as K
a = K.ones((2,3,4))
b = K.ones((7,4,5))
c = K.dot(a, b)
print(c.shape) #(2,3,7,5)
- Matrix multiplication when tensors are matrices
-
from keras import backend as K
a = K.ones((9, 8, 7, 4, 2))
b = K.ones((9, 8, 7, 2, 5))
c = K.batch_dot(a, b)
print(c.shape) #(9, 8, 7, 4, 5)
import tensorflow as tf
a = tf.ones((9, 8, 7, 4, 2))
b = tf.ones((9, 8, 7, 2, 5))
c = tf.matmul(a, b)
print(c.shape) #(9, 8, 7, 4, 5)
frfrom keras.layers.merge import dot
import tensorflow as tf
from keras.layers import Input, Dense, Masking, Dropout, LSTM, Bidirectional, Activation
ZX
u=tf.constant(
[[1,2,3],
[2,3,3]],dtype=float) #2,3 (B,F)
y=tf.constant(
[[[3,2,3] ,
[4,2,3]] ,
[[5,2,3] ,
[6,2,3]]],dtype=float) #2,2,3 (B,T,F)
#2,2,3
alpha1 = dot([y,u], axes=-1) # (2,2) dot(-1)(B,F)(B,T,F)=>(B,T)
alpha = Activation('softmax')(alpha1)
#(?,T)
z = dot([y,alpha], axes=1) # (2,3) dot(1)(B,F)(B,T,F)=>(B,T)
with tf.Session() as sess:
u_,y_,alpha1_,alpha_,z_=sess.run([u,y,alpha1,alpha,z])
print('u:{}\n,y:{}\n,alpha1:{}\n,alpha:{}\n,z:{}\n'.format(u_,y_,alpha1_,alpha_,z_))
print('u:{}\n,y:{}\n,alpha1:{}\n,alpha:{}\n,z:{}\n'.format(u_.shape,y_.shape,alpha1_.shape,alpha_.shape,z_.shape))
"""
u:(2,3) (B,F)
[[1. 2. 3.]
[2. 3. 3.]]
y:(2,2,3) (B,T,F)
[[[3. 2. 3.]
[4. 2. 3.]]
[[5. 2. 3.]
[6. 2. 3.]]]
alpha1:(2,2)=dot([u, y], axes=-1) (B,F)(B,T,F)=>(B,T)
[[16. 17.]
[25. 27.]]
alpha:(2,2)
[[0.26894143 0.7310586 ]
[0.11920291 0.880797 ]]
z:(2,3)=dot([alpha, y], axes=1) dot([(B,T) (B,T,F)],1)=>(B,F)
[[3.7310586 2. 3. ]
[5.880797 1.9999999 3. ]]
dot([(B,T) (B,T,F)],1)=>(B,F)
dot([(B,T) (B,T,F)],-1)=>(B,T)
"""
vu = tf.tensordot(v, u_omega, axes=1, name='vu') # (B,T) shape (B,T,A)*(A)=(B,T)
alphas = tf.nn.softmax(vu, name='alphas') # (B,T) shape (B,T)=(B,T)
v=tf.constant(
[[[111. 112. 113. 114.]
[121. 122. 123. 124.]
[131. 132. 133. 134.]]
[[211. 212. 213. 214.]
[221. 222. 223. 224.]
[231. 232. 233. 234.]]], dtype=tf.float32)
(2,3,4)
[[[111. 112. 113. 114.]
[121. 122. 123. 124.]
[131. 132. 133. 134.]]
[[211. 212. 213. 214.]
[221. 222. 223. 224.]
[231. 232. 233. 234.]]]
u=tf.constant( [1 1 1 1], dtype=float32)
(4,)
[1. 1. 1. 1.]
vu=tf.tensordot(v,u,axes=1)
(2, 3)
[[450 490 530]
[850 890 930]]
A.B=|A| |B| cos(th)=trans(A)B
동일한 각도 때 최대값, 90도이면 0, 반대방향이면 -최대값
(a1,