-
shape, ndim
-
reshape
-
expand_dims/squeeze
-
transpose
-
broadcast_to
import tensorflow as tf
view1-4不同的数据理解方式，不改变原有数据的content，并且数据全部使用。
-
view1: [b, 28, 28]
图片包含行列信息
-
view2: [b, 28*28]
图片的数据不变，理解方式改变， 没有了行列信息，常用于全连接层的输入
-
view3: [b, 2, 14*28]
图片的上下两部分
-
view4: [b, 28, 28, 1]
增加一个通道，channel概念，仅仅增加view1的理解方式，并未修改content
1. reshape
a = tf.random.normal([4, 28, 28, 3]) # [batch, row, col, channel] 行列理解方式
a.shape, a.ndim
(TensorShape([4, 28, 28, 3]), 4)
tf.reshape(a, [4, 28*28, 3]).shape # [batch, pixel, channel] 像素理解方式（抹掉行列信息）
TensorShape([4, 784, 3])
tf.reshape(a, [4, -1, 3]).shape
TensorShape([4, 784, 3])
tf.reshape(a, [4, 28*28*3]).shape # [batch, data_point] 数据点方式（抹掉行列和通道信息）
TensorShape([4, 2352])
tf.reshape(a, [4, -1]).shape
TensorShape([4, 2352])
reshape is flexible
2. tf.transpose
改变content
b = tf.random.normal([4, 28, 28, 3])
b.shape
TensorShape([4, 28, 28, 3])
tf.transpose(b).shape # [b, h, w, c]
TensorShape([3, 28, 28, 4])
tf.transpose(b, perm=[0, 2, 1, 3]).shape # [b, w, h, c]
TensorShape([4, 3, 1, 2])
tf.transpose(b, perm=[0, 3, 2, 1]).shape # [b, c, w, h]
TensorShape([4, 3, 28, 28])
3. tf.expand_dims
e.g.
-
a: [班级，学生， 成绩]
[4, 35, 8] 4个班级，35个学生，8门课成绩
-
增加学校维度dim=axis
[1 ,4, 35, 8] + [1, 4, 35, 8]
axis增加方式：正数在轴前，负数在轴后
c = tf.random.normal([4, 35, 8])
tf.expand_dims(c, axis=0).shape
TensorShape([1, 4, 35, 8])
tf.expand_dims(c, axis=3).shape
TensorShape([4, 35, 8, 1])
tf.expand_dims(c, axis=-3).shape # axis=0 vs -3
TensorShape([4, 1, 35, 8])
4. tf.squeeze
只能squeeze shape=1的维度
tf.squeeze(tf.zeros([1, 2, 1, 1, 5])).shape
TensorShape([2, 5])
d = tf.zeros([1, 2, 1, 3])
tf.squeeze(d, axis=0).shape, tf.squeeze(d, axis=2).shape, tf.squeeze(d, axis=-4).shape, tf.squeeze(d, axis=-2).shape
(TensorShape([2, 1, 3]),
TensorShape([1, 2, 3]),
TensorShape([2, 1, 3]),
TensorShape([1, 2, 3]))
Read more: