01-TensorFlow basis
What is Tensorflow
Google's open-source software library
- Take data flow diagram, a numerical calculation
- Support for multiple platforms - GPU, CPU, mobile device
- Initially used for deep learning, becoming more and more common
Tensorflow data structure
Data flow diagram #
Input-output relationship between the nodes, the transport line tensor: Line. Tensor: tensor - refers to data
- Node: operation (op): The special arithmetic operation node, all operations are an op, data processing
- Just use tensorflow of API functions are defined in OP
- Node is assigned to run on a variety of computing devices
- graph: the overall program structure of FIG.
- Is essentially a memory location allocated, there is a default map, all tensor op and memory addresses are the same.
- FIG different memory addresses are not the same, the process of calculating the interfering
- session: session: FIG computation program (running only a map, can be specified in the session to run FIG graph = g)
- Structure diagram of
- Allocation of computing resources
- Control over resources (variables, queues, thread)
The characteristics Tensorflow
- High degree of flexibility, ease of calling the function, you can also write your own package
- True portability, you can simply run on different devices
- Product research and combination
- Automatic differentiated, mainly for the reverse propagation calculations
- Multi-language support, C ++, Java, JS, R
- Performance optimization
The front end system Tensorflow
- FIG mechanism defined procedure: front-end system
- Backend system: calculation map structure
Tensorflow version changes
Tensorflow1.0 Key Features
- XLA: Accelerate Linear Algebra
- Enhance the training speed 58 times
- You can be run on the mobile device
- It cited higher-level API
- tf.layers/ tf.metrics / tf.losses/ tf.keras
- Tensorflow debugger
- Support docker mirroring, introduced tensorflow serving services
Tensorflow 2.0 Key Features
- Tf.keras and eager mode using a simple model building
- Robust cross-platform deployment model
- Strong research experiments
- 清除了不推荐使用和重复的API
Tensorflow2.0 简化模型开发流程
- 使用tf.data加载数据
- 使用tf.keras 构建模型,也可以使用premade estimator 验证模型
- 使用tensorflow hub进行迁移学习
- 注: 迁移学习 - 使用一个前人预先训练好的,应用在其他领域的网络作为模型训练的起点,站在前人基础上更进一步,不必重新发明轮子。
- 使用eager mode 进行运行和调试
- 使用分发策略进行分布式训练
- 导出到SavedModel
- 使用Tensorflow Serve, Tensorflow Lite, Tensorflow.js
Tensorflow 强大的跨平台能力
- Tensorflow 服务
- 直接通过HTTP/ TEST 或 GTPC/协议缓冲区
- Tensorflow Lite - Android, iOS 和嵌入式
- Tensorflow.js - Javascript 部署
- 其他语言
Tensorflow vs. Pytorch
入门时间(易用性)
- Tensorflow 1.*
- 静态图 ,构建完之后不可以更改, 效率高
- 额外概念, 会话,变量,占位符
- 写样本代码
- Tensorflow 2.0
- 动态图, 构建完之后可以更改, 效率不高,调试容易
- Eager mode 直接集成在python中
- Pytorch
- 动态图
- numpy扩展,集成在python
"""
不同方式求解 1 + 1/2 + 1/2^2 + 1/2^3 + ...... + 1/2^50
"""
# 1. tensorflow 1.*求解
import tensorflow as tf
print(tf.__version__)
x = tf.Variable(0.)
y = tf.Variable(1.)
add_op = x.assign(x + y)
div_op = y.assign(y / 2)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for iteration in range(50):
sess.run(add_op)
sess.run(div_op)
print(x.eval())
# 2. pytorch 求解
import torch
print(torch.__version__)
x = torch.Tensor([0.])
y = torch.Tensor([1.])
for iteration in range(50):
x = x + y
y = y / 2
print(x)
# 3. tensorflow 2.0 求解
import tensorflow as tf
print(tf.__version__)
x = tf.constant(0.)
y = tf.constant(1.)
for iteration in range(50):
x = x + y
y = y / 2
print(x.numpy())
# 4. 纯python求解
x = 0
y = 1
for iteration in range(50):
x = x + y
y = y / 2
print(x) # 精度有点不一样
图创建和调试
- Tensorflow 1.*
- 静态图,难以调试, 需要使用tfdbg
- Tensorflow 2.0 与 pytorch
- 动态图,python自带的调试工具
全面性
- python缺少少量的功能,使用频次很低
- 沿维翻转张量 (np.flip, np.flipud, np.fliplr)
- 检查无穷与非数值张量(np.is_nan, np.is_inf)
- 快速傅里叶变换 (np.fft)
序列化和部署
- Tensorflow 支持更加广泛,多语言,跨平台
- pytorch 支持比较简单