Learning about tensorflow (II)

type ：

list,np.array,tf.Tensor, among np.array And Tensor be similar , But there are problems ： Not very good GPU bonus , No automatic derivation

scalar：1.1   dim=0

vec： [1.1] dim =1

matrix:[[1.1,2.2]] dim = 2

tensor:rank>2   , Generally speaking >0 Just go , occasionally scalar Also belong to tensor.

  Use gpu And use cpu：

with tf.device('gpu'):   #s Use gpu
    b=tf.range(4)
print(b.device)
c=b.cpu()
print(c.device)

print(c.numpy)#numpy cpu
print(c.shape)
print(c.ndim)  # dimension 
print(tf.rank(c))

  see c Information about ：

Judge whether it is tensor type ：

# Judge whether it is tensor
a = tf.constant([1.])
b = np.array(4)

s=isinstance(a,tf.Tensor)
print(s)
s1=isinstance(b,tf.Tensor)
print(s1)
print(a.dtype)  # The specific type 

Conversion of types ：

# Type conversion 
a = np.array(5)
print(a.dtype)

aa = tf.convert_to_tensor(a,dtype=tf.int32)# Turn into tensor     dtype  Specify the conversion type 

# Data type conversion generally uses cast
aaa = tf.cast(aa,dtype=tf.double)
# Boolean and plastic can also be transformed 

# Turn into Varible, It has the property of derivation 
w = tf.Variable(a)
w.name
w.trainable  # Gradient information is required 

# int(tensor)

  Initialization problem ：

# Initialize to 0
tf.zeros([])
# Understood as a shape

a = tf.zeros([2,3,3])
tf.zeros_like(a)
#tf.zero（a.shape）
#shape = (2,3,3)
#tf.zeros a.shape Functions are similar 

# Initialize to 1
tf.ones()

# Initialize to 9（ Customizable ）
tf.fill([2,2],9)

# Random initialization 
tf.random.normal([2,2],mean=1,stddev=1)
#mean  mean value   stddev  variance 

# Don't specify     Normal is 0 Of 

tf.random.truncated_normal()  # truncation     The gradient of 0-- The gradient disappears     Better performance 

zeros-- Initialize to 0,ones-- Initialize to 1,fill-- Custom initialization ,random.normal Random initialization

  Distribution problem ：

# Uniform distribution 
tf.random.uniform([2,2],minval=0,maxval=1)

# Break up randomly 
idx = tf.range(10)
idx = tf.random.shuffle(idx)

a = tf.random.normal([10,784])
b = tf.random.uniform([10],maxval=10,dtype=tf.int32)

a = tf.gather(a,idx)   # same idx
b = tf.gather(b,idx)

  This is a simplified method of summing ：

loss  = tf.keras.losses.mes()   #(y-out) And 

Dim = 3 It's more about dealing with natural language
Dim = 4  More pictures

DIm = 5 Multitasking