当前位置：网站首页>Tensorflow 2 detailed explanation (TF ecosystem, installation, housekeeping, basic operation)

TensorFlow Born in 2011 year , As Google An internal and closed source project of the company , It was named at that time DisBelief.DisBelief It is a machine learning system with deep learning neural network . In the system 2015 year 11 On August th Apache 2.0 Released to the developer community under an open source license , And become TensorFlow.TensorFlow 1.0.0 Version released at 2017 year 2 month 11 Japan .TensorFlow 2.0.0 Alpha Version released at 2019 year 3 month 6 Japanese TensorFlow Developer Summit . During and after , A large number of versions have been released , And contains a lot of new features .

TensorFlow Its name comes from tensor (tensor）. Tensors are the generalization of vectors and matrices in higher dimensions . The rank of the tensor is the index number required to specify each element of the tensor . Scalar ( A number ) Is rank 1 Tensor , The vector is rank 1 Tensor , The matrix is rank 2 Tensor , The three-dimensional array is rank 3 Tensor . Tensors have data types and shapes ( All data items in the tensor must have the same type ）.

Four dimensional tensor ( That is, the rank is 4） An example of this is Images , for instance , These four dimensions can be batch (batch）、 Height (height)、 Width (width） and Color (color) passageway .

imagel = tf. zeros([15, 500, 300, 3])  #  batch 、 Height 、 Width 、 Color channel

Even though TensorFlow It can be used in general numerical calculation fields , Especially machine learning , But its main research and development field is usually deep neural network (Deep Neural Networks, DNN) Application . Its application fields mainly include speech recognition , Such as voice assistant, which is widely used now , Text based applications 、 Language translator 、 Image recognition 、 Alien search 、 Cancer detection and diagnosis 、 Time series applications, etc .

TensorFlow The ecological system

Dynamic graph mechanism

First introduced TensorFlow Dynamic graph mechanism in (Eager Execution). TensorFlow At first, we need to construct a computational graph composed of operations and tensors , Next , These operations and tensors must be in Google Calculation in the so-called session ( It's called declarative programming ). This is the writing TensorFlow A common method of program . However , From the form of research 1.5 Release start , Dynamic graph mechanism comes into effect , And from 1.7 The release is officially integrated into TensorFlow in . Its features include the immediate execution of operations , So that the tensor can be like NumPy Arrays are processed as well （ It is called imperative programming ）.

tf.data

tf.data It's a API, It allows users to move from simpler 、 Reusable parts build complex data input pipelines . The highest level of abstraction is Dataset, It contains elements of tensor nested structure , It also contains transformation plans that act on these elements . The following is about Dataset Common classes for ：

FixedLengthRecordDataset： Used to get data sets with fixed length records , The data set comes from one or more binary files .
TFRecordDataset： Indicates that it contains one or more TFRecord The record data set of the file .
TextLineDataset： Represents a dataset containing rows from one or more text files .
tf. data. Iterator： Express Dataset Iteration state .

Estimator

Estimator It's an advanced API, Allow users to build very simplified machine learning programs , And be responsible for training 、 assessment 、 Prediction and service export .TensorFlow. jis It's a group. API, Allow users to use the underlying JavaScript Linear algebra library or other high-level API Building and training models . With the help of this API, The model can be trained and run in the browser .

TensorFlow Lite

TensorFlow Lite yes TensorFlow Lightweight version of , Suitable for mobile and embedded equipment . It consists of a runtime interpreter and a set of utilities . The original intention is , Train the model on a more powerful machine , Then use the utility to convert the model into .tflite Format , The model can then be loaded into the selected device . In writing this book ,Tensor Flow Lite use C++ API To support Android and i0s, And it is applicable to Android Of Java Wrappers .

TensorFlow Hub

TensorFlow Hub Is a library , It aims to promote the release of reusable modules of machine learning models 、 Discover and use . ad locum , Module by TensorFlow Network structure and its weight composition . This module can be reused in different tasks through migration learning . The original intention is , Train models on large datasets , Then reuse the appropriate modules for different but related tasks . This method has many advantages , Not only can smaller datasets be used to train models , Generalization can also be improved , Significantly speed up training .

TensorFlow Extended

TensorFlow Extended(TFX) It's based on TensorFlow General machine learning platform . The open source libraries released so far include TensorFlow Transform、Tensor FlowModel Analysis and TensorFlow Serving.

tf.keras It's a use. Python Advanced neural network API, As Tensor Flow Interface with other tensor tools .tf.keras Support rapid prototyping , It is user-friendly, modular and extensible . It supports convolution and cyclic Networks , And you can CPU and GPU Up operation .Keras Is in TensorFlow 2 The first choice for development in API.

TensorBoard

TensorBoard It is a set of visualization tools , Support for TensorFlow Understanding of procedures 、 Debugging and optimization , It is compatible with both dynamic graph and computational graph execution environments . During model training , have access to TensorBoard Carry out various visual operations .

TensorFlow install

（Linux） Use conda Configuration compatibility TensorFlow and PyTorch Of env Environmental Science _ It's time for Haohao to cook -CSDN Blog _linux Get into tensorflow Environmental Science If you want to use it at the same time Tensorflow and pytorch, It needs to be installed in the same environment . After many attempts, I configured a compatible environment , In this environment , Use TensorFlow The model and pytorch The model of can be deployed on the unified system architecture https://blog.csdn.net/qq_52213943/article/details/124617468

TensorFlow Basic operation

Import TensorFlow And look at

import tensorflow as tf
print("TensorFlow version: {}".format(tf.__version__))
print("Eager execution is: {}".format(tf.executing_eagerly()))
print("Keras version: {}".format(tf.keras.__version__))

Running results ：

Dynamic graph check GPU

var = tf.Variable([3, 3])

if tf.test.is_gpu_available():
    print('Running on GPU')
    print('GPU #0?')
    print(var.device.endswith('GPU:0'))
else:
    print('CPU')

Running results ：

Statement Eager Variable

t0 = 24 
# python Variable 

t1 = tf.Variable(42) 
# 0 D tensor 

t2 = tf.Variable([ [ [0., 1., 2.], [3., 4., 5.] ], [ [6., 7., 8.], [9., 10., 11.] ] ])
#  Three dimensional tensor

Running results ：

Statement TensorFlow Constant

mol = tf.constant(42)
mol

Running results ：

establish tensor tensor

The shape of the tensor can be accessed through attributes

t2 = tf.Variable([ [ [0., 1., 2.], [3., 4., 5.] ], [ [6., 7., 8.], [9., 10., 11.] ] ])
# tensor variable
print(t2.shape)

Output shape results ：

Tensor shape remodeling

r1 = tf.reshape(t2,[2,6]) # 2  That's ok  6  Column 
r2 = tf.reshape(t2,[1,12]) # 1  That's ok  12  Column 

# The value cannot be changed

Running results ：

Tensor and python Installation and replacement of variables

print(t2.numpy())
# perhaps 
print(t2[1, 0, 2].numpy())

Calculate the tensor size

s = tf.size(input=t2).numpy()

View tensor data types

t3.dtype

# Output 
tf.float32

Tensor basic operation rules

Tensor operations can be performed by overloading operators +、-、*、/ Realization

t2*t2

t4 = t2*4

Calculate the square difference of tensor

x = [1,3,5,7,11]
y = 5
s = tf.math.squared_difference(x, y)

Calculate average

tf.reduce_mean(input_tensor=numbers) 
#( 4. + 5. + 7. + 3.)/4 = 4.75

Random initialization tensor

# tf.random.normal() Output the tensor of a given shape   Normal distribution filling 
tf.random.normal(shape = (3,2), mean=10, stddev=2, dtype=tf.float32, seed=None,  name=None)

# tf.random.uniform() Output the tensor of a given shape   The range of maximum and minimum values of the upper and lower boundaries is filled 
tf.random.uniform(shape = (2,4),  minval=0, maxval=None, dtype=tf.float32, seed=None,  name=None)

Save the recovery tensor using checkpoints

Later, , This method can save Zhang Liang

variable = tf.Variable([[1,3,5,7],[11,13,17,19]])
checkpoint= tf.train.Checkpoint(var=variable)
save_path = checkpoint.save('./vars') # Exist under this path 
variable.assign([[0,0,0,0],[0,0,0,0]])
variable

Welcome to thumb up 、 Collection 、 Comment area exchange , Reprint mark the source of .

-----------------------------

原网站

版权声明
本文为[Haohao, it's time to cook]所创，转载请带上原文链接，感谢
https://yzsam.com/2022/201/202207190442570575.html