鏈接：https://tensorflow.juejin.im/get_started/

學(xué)習(xí)它主要為了解決yunyang tensorflow-yolov-3GPU多線程調(diào)用的問題

文章目錄

• • • 坑1
    • • 啥是Eager Execution？
      • Eager Execution 有啥優(yōu)點(diǎn)？
      • 啥是張量？
      • 張量的基本創(chuàng)建與使用
      • 張量的屬性
      • NumPy array和TensorFlow張量之間最明顯的區(qū)別
      • TensorFlow張量和NumPy nararrays之間的轉(zhuǎn)換
    • 坑2
    • • GPU加速
      • • （1）設(shè)備名稱
        • （2）顯示設(shè)備配置
    • 坑3
    • • 數(shù)據(jù)集
      • • （1）創(chuàng)建源數(shù)據(jù)集
        • （2）應(yīng)用transformations
        • (3)迭代

坑1

開始學(xué)習(xí) TensorFlow 最簡單的方法是使用 Eager Execution，官方提供的教程為Colab notebook，打不開，參考其他的吧，比如這個(gè)：tensorflow之Eager execution基礎(chǔ)

從tensorflow之Eager execution基礎(chǔ)中，我了解到：

啥是Eager Execution？

「Eager Execution」，它是一個(gè)命令式、由運(yùn)行定義的接口，一旦從 Python 被調(diào)用，其操作立即被執(zhí)行。 這使得入門 TensorFlow 變的更簡單，也使研發(fā)更直觀。

Eager Execution 有啥優(yōu)點(diǎn)？

1、快速調(diào)試即刻的運(yùn)行錯(cuò)誤并通過 Python 工具進(jìn)行整合 2、借助易于使用的 Python 控制流支持動(dòng)態(tài)模型 3、為自定義和高階梯度提供強(qiáng)大支持 4、適用于幾乎所有可用的 TensorFlow 運(yùn)算

啥是張量？

張量是一個(gè)多維數(shù)組。與NumPy ndarray對象類似，Tensor對象具有數(shù)據(jù)類型和形狀。 此外，Tensors可以駐留在加速器（如GPU）內(nèi)存中。 TensorFlow提供了豐富的操作庫（tf.add，tf.matmul，tf.linalg.inv等）， 它們使用和生成Tensors。這些操作自動(dòng)轉(zhuǎn)換本機(jī)Python類型。

張量的基本創(chuàng)建與使用

# -*- coding: utf-8 -*- """ @File : 191206_test_Eager_execution.py @Time : 2019/12/6 11:11 @Author : Dontla @Email : sxana@qq.com @Software: PyCharm """ # 導(dǎo)入tensorflow import tensorflow as tf tf.enable_eager_execution()# 創(chuàng)建和使用張量 print(tf.add(1,2)) # tf.Tensor(3, shape=(), dtype=int32) print(tf.add([1, 2], [3, 4])) # tf.Tensor([4 6], shape=(2,), dtype=int32) print(tf.square(5)) # tf.Tensor(25, shape=(), dtype=int32) print(tf.reduce_sum([1, 2, 3])) # tf.Tensor(6, shape=(), dtype=int32) print(tf.encode_base64("hello world")) # tf.Tensor(b'aGVsbG8gd29ybGQ', shape=(), dtype=string) print(tf.square(2) + tf.square(3)) # tf.Tensor(13, shape=(), dtype=int32)x = tf.matmul([[1]], [[2, 3]]) print(x) # tf.Tensor([[2 3]], shape=(1, 2), dtype=int32) print(x.shape) # (1, 2) print(x.dtype) # <dtype: 'int32'>

張量的屬性

每個(gè)Tensor都有一個(gè)形狀和數(shù)據(jù)類型

x = tf.matmul([[1]], [[2, 3]]) print(x.shape) print(x.dtype)

NumPy array和TensorFlow張量之間最明顯的區(qū)別

• 張量可以由加速器內(nèi)存（如GPU，TPU）支持。
• 張量是不可改變的。

TensorFlow張量和NumPy nararrays之間的轉(zhuǎn)換

• TensorFlow操作自動(dòng)將NumPy ndarrays轉(zhuǎn)換為Tensors。
• NumPy操作自動(dòng)將Tensors轉(zhuǎn)換為NumPy ndarrays。

通過在Tensors上調(diào)用.numpy（）方法，可以將張量顯式轉(zhuǎn)換為NumPy ndarrays。這些轉(zhuǎn)換通常很容易，因?yàn)槿绻赡?#xff0c;數(shù)組和Tensor共享底層內(nèi)存表示。但是，共享底層表示并不總是可行的，因?yàn)門ensor可能托管在GPU內(nèi)存中，而NumPy陣列總是由主機(jī)內(nèi)存支持，因此轉(zhuǎn)換將涉及從GPU到主機(jī)內(nèi)存的復(fù)制。

import tensorflow as tf import numpy as np tf.enable_eager_execution()ndarray = np.ones([3, 3]) print(ndarray) # [[1. 1. 1.] # [1. 1. 1.] # [1. 1. 1.]]print("TensorFlow operations convert numpy arrays to Tensors automatically") tensor = tf.multiply(ndarray, 42) print(tensor) # tf.Tensor( # [[42. 42. 42.] # [42. 42. 42.] # [42. 42. 42.]], shape=(3, 3), dtype=float64)print("And NumPy operations convert Tensors to numpy arrays automatically") print(np.add(tensor, 1)) # [[43. 43. 43.] # [43. 43. 43.] # [43. 43. 43.]]print("The .numpy() method explicitly converts a Tensor to a numpy array") print(tensor.numpy()) # [[42. 42. 42.] # [42. 42. 42.] # [42. 42. 42.]]

坑2

GPU加速

通過使用GPU進(jìn)行計(jì)算，可以加速許多TensorFlow操作。在沒有任何注釋的情況下，TensorFlow會(huì)自動(dòng)決定是使用GPU還是CPU進(jìn)行操作（如有必要，還可以復(fù)制CPU和GPU內(nèi)存之間的張量）。由操作產(chǎn)生的張量通常由執(zhí)行操作的設(shè)備的存儲(chǔ)器支持。例如：

# -*- coding: utf-8 -*- """ @File : 191208_test_Eager_execution_once_cls.py @Time : 2019/12/8 12:25 @Author : Dontla @Email : sxana@qq.com @Software: PyCharm """ import tensorflow as tftf.enable_eager_execution()x = tf.random_uniform([3, 3])print("Is there a GPU available: ") print(tf.test.is_gpu_available()) # Trueprint("Is the Tensor on GPU #0: "), print(x.device) # /job:localhost/replica:0/task:0/device:GPU:0 print(x.device.endswith('GPU:0')) # True

（1）設(shè)備名稱

Tensor.device屬性提供托管Tensor內(nèi)容的設(shè)備的完全限定字符串名稱。此名稱對一組詳細(xì)信息進(jìn)行編碼，例如，正在執(zhí)行此程序的主機(jī)的網(wǎng)絡(luò)地址的標(biāo)識符以及該主機(jī)中的設(shè)備。這是分布式執(zhí)行TensorFlow程序所必需的，但我們暫時(shí)不會(huì)這樣做。如果張量位于主機(jī)上的第N個(gè)張量上，則字符串將以GPU：結(jié)尾。

（2）顯示設(shè)備配置

TensorFlow中的術(shù)語“placement"指的是如何為執(zhí)行設(shè)備分配（放置）各個(gè)操作。如上所述，當(dāng)沒有提供明確的指導(dǎo)時(shí)，TensorFlow會(huì)自動(dòng)決定執(zhí)行操作的設(shè)備，并在需要時(shí)將Tensors復(fù)制到該設(shè)備。但是，可以使用tf.device上下文管理器將TensorFlow操作顯式放置在特定設(shè)備上。例如：

Wocalei，我在pycharm上跑半天沒跑起來，最后放到ipython上才跑起來的！！！

C:\Users\Dontla>ipython Python 3.6.8 (tags/v3.6.8:3c6b436a57, Dec 24 2018, 00:16:47) [MSC v.1916 64 bit (AMD64)] Type 'copyright', 'credits' or 'license' for more information IPython 7.9.0 -- An enhanced Interactive Python. Type '?' for help.In [1]: import tensorflow as tf...:...:...: def time_matmul(x):...: %timeit tf.matmul(x, x)...:...:...: # Force execution on CPU...: print("On CPU:")...: with tf.device("CPU:0"):...: x = tf.random_uniform([1000, 1000])...: assert x.device.endswith("CPU:0")...: time_matmul(x)...:...: # Force execution on GPU #0 if available...: if tf.test.is_gpu_available():...: with tf.device("GPU:0"): # Or GPU:1 for the 2nd GPU, GPU:2 for the 3rd etc....: x = tf.random_uniform([1000, 1000])...: assert x.device.endswith("GPU:0")...: time_matmul(x)...: On CPU: 608 μs ± 40.2 μs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 2019-12-09 11:33:50.419224: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 2019-12-09 11:33:51.356242: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1411] Found device 0 with properties: name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.6575 pciBusID: 0000:0e:00.0 totalMemory: 11.00GiB freeMemory: 9.10GiB 2019-12-09 11:33:51.533039: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1411] Found device 1 with properties: name: GeForce GT 710 major: 3 minor: 5 memoryClockRate(GHz): 0.954 pciBusID: 0000:05:00.0 totalMemory: 2.00GiB freeMemory: 1.67GiB 2019-12-09 11:33:51.546476: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1461] Ignoring visible gpu device (device: 1, name: GeForce GT 710, pci bus id: 0000:05:00.0, compute capability: 3.5) with Cuda compute capability 3.5. The minimum required Cuda capability is 3.7. 2019-12-09 11:33:51.567213: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1490] Adding visible gpu devices: 0 2019-12-09 11:33:53.149081: I tensorflow/core/common_runtime/gpu/gpu_device.cc:971] Device interconnect StreamExecutor with strength 1 edge matrix: 2019-12-09 11:33:53.164773: I tensorflow/core/common_runtime/gpu/gpu_device.cc:977] 0 1 2019-12-09 11:33:53.172301: I tensorflow/core/common_runtime/gpu/gpu_device.cc:990] 0: N N 2019-12-09 11:33:53.182373: I tensorflow/core/common_runtime/gpu/gpu_device.cc:990] 1: N N 2019-12-09 11:33:53.187598: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1103] Created TensorFlow device (/device:GPU:0 with 8789 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:0e:00.0, compute capability: 6.1) 620 μs ± 48.5 μs per loop (mean ± std. dev. of 7 runs, 1000 loops each)In [2]:

不過，在我搞了半天之后，終于能使程序再pycharm上運(yùn)行了：

# -*- coding: utf-8 -*- """ @File : 191208_test_Eager_execution_once_cls.py @Time : 2019/12/8 12:25 @Author : Dontla @Email : sxana@qq.com @Software: PyCharm """from timeit import timeit import tensorflow as tfdef func(x):tf.matmul(x, x)def time_matmul(x):print(timeit('func', 'from __main__ import func', number=1))# Force execution on CPU print("On CPU:") with tf.device("CPU:0"):x = tf.random_uniform([1000, 1000])assert x.device.endswith("CPU:0")time_matmul(x)# Force execution on GPU #0 if available if tf.test.is_gpu_available():with tf.device("GPU:0"): # Or GPU:1 for the 2nd GPU, GPU:2 for the 3rd etc.x = tf.random_uniform([1000, 1000])assert x.device.endswith("GPU:0")time_matmul(x)

結(jié)果：

D:\20191031_tensorflow_yolov3\python\python.exe D:/20191031_tensorflow_yolov3/needed/test_tensorflow-gpu/191208_test_Eager_execution_once_cls.py On CPU: 2019-12-10 09:56:02.601521: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 6.209648925078723e-07 2019-12-10 09:56:03.389699: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1411] Found device 0 with properties: name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.6575 pciBusID: 0000:0e:00.0 totalMemory: 11.00GiB freeMemory: 9.10GiB 2019-12-10 09:56:03.537904: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1411] Found device 1 with properties: name: GeForce GT 710 major: 3 minor: 5 memoryClockRate(GHz): 0.954 pciBusID: 0000:05:00.0 totalMemory: 2.00GiB freeMemory: 1.67GiB 2019-12-10 09:56:03.538533: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1461] Ignoring visible gpu device (device: 1, name: GeForce GT 710, pci bus id: 0000:05:00.0, compute capability: 3.5) with Cuda compute capability 3.5. The minimum required Cuda capability is 3.7. 2019-12-10 09:56:03.539099: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1490] Adding visible gpu devices: 0 2019-12-10 09:56:05.750966: I tensorflow/core/common_runtime/gpu/gpu_device.cc:971] Device interconnect StreamExecutor with strength 1 edge matrix: 2019-12-10 09:56:05.751268: I tensorflow/core/common_runtime/gpu/gpu_device.cc:977] 0 1 2019-12-10 09:56:05.751452: I tensorflow/core/common_runtime/gpu/gpu_device.cc:990] 0: N N 2019-12-10 09:56:05.751652: I tensorflow/core/common_runtime/gpu/gpu_device.cc:990] 1: N N 2019-12-10 09:56:05.752012: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1103] Created TensorFlow device (/device:GPU:0 with 8789 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:0e:00.0, compute capability: 6.1) 9.314473379262722e-07Process finished with exit code 0

以上我是重新定義了一個(gè)func()函數(shù)了，但并未抓住問題的本質(zhì)，以下才是：

# -*- coding: utf-8 -*- """ @File : 191208_test_Eager_execution_once_cls.py @Time : 2019/12/8 12:25 @Author : Dontla @Email : sxana@qq.com @Software: PyCharm """from timeit import timeit import tensorflow as tfdef time_matmul(x):print(timeit('tf.matmul(x, x)', 'from __main__ import tf,x', number=1))# Force execution on CPU print("On CPU:") with tf.device("CPU:0"):x = tf.random_uniform([1000, 1000])assert x.device.endswith("CPU:0")time_matmul(x)# Force execution on GPU #0 if available if tf.test.is_gpu_available():with tf.device("GPU:0"): # Or GPU:1 for the 2nd GPU, GPU:2 for the 3rd etc.x = tf.random_uniform([1000, 1000])assert x.device.endswith("GPU:0")time_matmul(x)

結(jié)果：

D:\20191031_tensorflow_yolov3\python\python.exe D:/20191031_tensorflow_yolov3/needed/test_tensorflow-gpu/191208_test_Eager_execution_once_cls.py On CPU: 2019-12-10 10:45:10.788751: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 0.0006731259434785336 2019-12-10 10:45:11.520722: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1411] Found device 0 with properties: name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.6575 pciBusID: 0000:0e:00.0 totalMemory: 11.00GiB freeMemory: 9.10GiB 2019-12-10 10:45:11.647470: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1411] Found device 1 with properties: name: GeForce GT 710 major: 3 minor: 5 memoryClockRate(GHz): 0.954 pciBusID: 0000:05:00.0 totalMemory: 2.00GiB freeMemory: 1.67GiB 2019-12-10 10:45:11.648053: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1461] Ignoring visible gpu device (device: 1, name: GeForce GT 710, pci bus id: 0000:05:00.0, compute capability: 3.5) with Cuda compute capability 3.5. The minimum required Cuda capability is 3.7. 2019-12-10 10:45:11.648577: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1490] Adding visible gpu devices: 0 2019-12-10 10:45:13.033777: I tensorflow/core/common_runtime/gpu/gpu_device.cc:971] Device interconnect StreamExecutor with strength 1 edge matrix: 2019-12-10 10:45:13.034072: I tensorflow/core/common_runtime/gpu/gpu_device.cc:977] 0 1 2019-12-10 10:45:13.034252: I tensorflow/core/common_runtime/gpu/gpu_device.cc:990] 0: N N 2019-12-10 10:45:13.034427: I tensorflow/core/common_runtime/gpu/gpu_device.cc:990] 1: N N 2019-12-10 10:45:13.034762: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1103] Created TensorFlow device (/device:GPU:0 with 8789 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:0e:00.0, compute capability: 6.1) 0.0015999160455466566Process finished with exit code 0

坑3

數(shù)據(jù)集

演示如何使用tf.data.Dataset API構(gòu)建pipelines以將數(shù)據(jù)提供給模型。它涵蓋：

• 創(chuàng)建數(shù)據(jù)集。
• 在啟用了eager execution的情況下對數(shù)據(jù)集進(jìn)行迭代。
  我們建議使用數(shù)據(jù)集API從簡單，可重復(fù)使用的部分構(gòu)建高性能，復(fù)雜的pipelines，這些部分將為模型的訓(xùn)練或評估循環(huán)提供支持。

如果您熟悉TensorFlow圖，則在啟用eager execution時(shí)，構(gòu)建數(shù)據(jù)集對象的API保持完全相同，但迭代數(shù)據(jù)集元素的過程稍微簡單一些。您可以對tf.data.Dataset對象使用Python迭代，而不需要顯式創(chuàng)建tf.data.Iterator對象。因此，在啟用eager執(zhí)行時(shí)，TensorFlow指南中對迭代器的討論無關(guān)緊要。

（1）創(chuàng)建源數(shù)據(jù)集

使用其中一個(gè)工廠函數(shù)（如Dataset.from_tensors，Dataset.from_tensor_slices）或使用從TextLineDataset或TFRecordDataset等文件讀取的對象創(chuàng)建源數(shù)據(jù)集。

# -*- coding: utf-8 -*- """ @File : 191208_test_Eager_execution_once_cls.py @Time : 2019/12/8 12:25 @Author : Dontla @Email : sxana@qq.com @Software: PyCharm """import tensorflow as tftf.enable_eager_execution()ds_tensors = tf.data.Dataset.from_tensor_slices([1, 2, 3, 4, 5, 6]) # print(type(ds_tensors)) # <class 'tensorflow.python.data.ops.dataset_ops.TensorSliceDataset'> # print(ds_tensors) # TensorSliceDataset shapes: (), types: tf.int32># Create a CSV file import tempfile# tempfile.mkstemp()返回os.open返回的文件描述符和文件名 _, filename = tempfile.mkstemp(dir='./') # print(_) # 3 # print(filename) # D:\20191031_tensorflow_yolov3\needed\test_tensorflow-gpu\tmpjrwgapybwith open(filename, 'w') as f:f.write("""Line 1 Line 2 Line 3""")ds_file = tf.data.TextLineDataset(filename)with open(filename, 'r') as f:print(f.read()) # Line 1 # Line 2 # Line 3

（2）應(yīng)用transformations

使用map，batch，shuffle等轉(zhuǎn)換函數(shù)將轉(zhuǎn)換應(yīng)用于數(shù)據(jù)集的記錄。（啥意思？？）

ds_tensors = ds_tensors.map(tf.square).shuffle(2).batch(2)ds_file = ds_file.batch(2)

(3)迭代

啟用eager執(zhí)行時(shí)，Dataset對象支持迭代。如果您熟悉TensorFlow圖中數(shù)據(jù)集的使用，請注意不需要調(diào)用Dataset.make_one_shot_iterator（）或get_next（）。

print('Elements of ds_tensors:') for x in ds_tensors:print(x) # Elements of ds_tensors: # tf.Tensor([4 9], shape=(2,), dtype=int32) # tf.Tensor([16 25], shape=(2,), dtype=int32) # tf.Tensor([ 1 36], shape=(2,), dtype=int32)print('\nElements in ds_file:') for x in ds_file:print(x) # Elements in ds_file: # tf.Tensor([b'Line 1 ' b'Line 2'], shape=(2,), dtype=string) # tf.Tensor([b'Line 3'], shape=(1,), dtype=string)

引用文章：tensorflow之Eager execution基礎(chǔ)

總結(jié)

以上是生活随笔為你收集整理的tensorflow教程 学习笔记 之 Eager execution 急切执行的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔推薦給好友。