Broadcasting
- expand（扩展数据）
- without copying data（不复制数据）
- tf.broadcast_to
Key idea
- Insert 1 dim ahead if needed
- Expand dims with size 1 to same size
- example:
[4,16,16,32]
[32]
[4,16,16,32]
[1,1,1,32]
[4,16,16,32]
[4,16,16,32]
How to understand?
When it has no axis
- Create a new concepy
- [classes, students, scores] + [scores]
When it has dim of size 1
- Treat it shared by all
- [classes,students,scores] + [students,1]
Broadcasting可以理解成把维度分成大维度和小维度，小维度较为具体，大维度更加抽象。也就是小维度针对某个示例，然后让这个示例通用语大维度。
Why broadcasting？
for real demanding
- [classes, students, scores]
- Add bias for every student: +5 score
- [4,32,8] + [4,32,8]
- [4,32,8] + [5.0]
memory consumption
- [4,32,8] -> 1024
- bias = [8]: [5.0,5.0,5.0,...] -> 8
Broadcastable?
Match from Last dim!
- if current dim=1, expand to same
- if either has no dim, insert one dim and expand to same
- otherwise, Not Broadcastable
[4,32,14,14]
[1,32,1,1] -> [4,32,14,14] √
[14,14] -> [1,1,14,14] -> [4,32,14,14] √
[2,32,14,14] ×
[3] √
[32,32,1] √
[4,1,1,1] √
import tensorflow as tf
x = tf.random.normal([4,32,32,3])
x.shape
TensorShape([4, 32, 32, 3])
(x+tf.random.normal([3])).shape
TensorShape([4, 32, 32, 3])
(x+tf.random.normal([32,32,1])).shape
TensorShape([4, 32, 32, 3])
(x+tf.random.normal([4,1,1,1])).shape
TensorShape([4, 32, 32, 3])
try:
(x+tf.random.normal([1,4,1,1])).shape
except Exception as e:
print(e)
Incompatible shapes: [4,32,32,3] vs. [1,4,1,1] [Op:Add] name: add/
(x+tf.random.normal([4,1,1,1])).shape
TensorShape([4, 32, 32, 3])
b = tf.broadcast_to(tf.random.normal([4,1,1,1]),[4,32,32,3])
b.shape
TensorShape([4, 32, 32, 3])
Broadcast VS Tile
a = tf.ones([3,4])
a.shape
TensorShape([3, 4])
a1 = tf.broadcast_to(a,[2,3,4])
a1.shape
TensorShape([2, 3, 4])
a2 = tf.expand_dims(a,axis=0) # 0前插入一维
a2.shape
TensorShape([1, 3, 4])
a2 = tf.tile(a2,[2,1,1]) # 复制一维2次，复制二、三维1次
a2.shape
TensorShape([2, 3, 4])