Quantcast
Channel: CodeSection,代码区,Linux操作系统:Ubuntu_Centos_Debian - CodeSec
Viewing all articles
Browse latest Browse all 11063

加快TensorFlow在树莓派上的执行速度――模型预热

$
0
0

转载请注明出处: http://www.codelast.com/

本文软硬件环境:

树莓派:3代 Model B V1.2,内存1GB

OS:Arch linux ARM

在 上一篇文章 中,我写了在树莓派上用TensorFlow做的一个深度学习(图像识别)实验,但正如文中所说,50秒执行一次预测的实用性为0。因此,有必要采取一些措施来加快TensorFlow的执行速度,其中一个可行的方法就是“预热”(warm-up),把TensorFlow移植到树莓派上的作者Sam Abrahams已经比较详细地在GitHub上列出了 性能测试的结果 。依照作者的描述,我也测试了一下,看看那悲催的50秒时间能减少到多少秒。

『1』 什么是预热(warm-up)

首先,本文还是对TensorFlow的python图像分类程序 classify_image.py 来描述的。

预热就是指在真正执行一次预测之前,先执行若干次 Session.run() 方法,从而达到加快一次预测的执行速度的目的。

文章来源: http://www.codelast.com/

『2』 代码修改

代码改起来其实很简单。为了能衡量程序运行时间,需要使用Python的time模块,因此在一开始需要import:

import time

然后对 run_inference_on_image 方法做一些修改,如下:

def run_inference_on_image(image): """Runs inference on an image. Args: image: Image file name. Returns: Nothing """ if not tf.gfile.Exists(image): tf.logging.fatal('File does not exist %s', image) image_data = tf.gfile.FastGFile(image, 'rb').read() # the image used to warm-up TensorFlow model warm_up_image_data = tf.gfile.FastGFile('/root/tensorflow-related/test-images/ubike.jpg', 'rb').read() # Creates graph from saved GraphDef. create_graph() with tf.Session() as sess: # Some useful tensors: # 'softmax:0': A tensor containing the normalized prediction across # 1000 labels. # 'pool_3:0': A tensor containing the next-to-last layer containing 2048 # float description of the image. # 'DecodeJpeg/contents:0': A tensor containing a string providing JPEG # encoding of the image. # Runs the softmax tensor by feeding the image_data as input to the graph. softmax_tensor = sess.graph.get_tensor_by_name('softmax:0') print("Warm-up start") for i in range(10): print("Warm-up for time {}".format(i)) predictions = sess.run(softmax_tensor, {'DecodeJpeg/contents:0': warm_up_image_data}) print("Warm-up finished") # record the start time of the actual prediction start_time = time.time() predictions = sess.run(softmax_tensor, {'DecodeJpeg/contents:0': image_data}) predictions = np.squeeze(predictions) # Creates node ID --> English string lookup. node_lookup = NodeLookup() top_k = predictions.argsort()[-FLAGS.num_top_predictions:][::-1] for node_id in top_k: human_string = node_lookup.id_to_string(node_id) score = predictions[node_id] print('%s (score = %.5f)' % (human_string, score)) print("Prediction used time:{} S".format(time.time() - start_time))

其中,我们自己添加的代码有如下几部分:

# the image used to warm-up TensorFlow model warm_up_image_data = tf.gfile.FastGFile('/root/tensorflow-related/test-images/ubike.jpg', 'rb').read()

这里使用了另外一张图片来预热模型(和真正预测时使用的不是同一张图片),为了简单写死了路径。

文章来源: http://www.codelast.com/

print("Warm-up start") for i in range(10): print("Warm-up for time {}".format(i)) predictions = sess.run(softmax_tensor, {'DecodeJpeg/contents:0': warm_up_image_data}) print("Warm-up finished")

这里循环10次来预热模型。

文章来源: http://www.codelast.com/

# record the start time of the actual prediction start_time = time.time() # (中间省略) print("Prediction used time:{} S".format(time.time() - start_time))

这里打印出了真正预测一张图片的执行时间(秒数),这个时间就是我们真正需要关心的,看它能减少到多少秒。

文章来源: http://www.codelast.com/

『3』 测试结果

执行和上一篇文章一样的命令,输出如下:

/usr/lib/python3.5/site-packages/tensorflow/python/ops/array_ops.py:1750: VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future

result_shape.insert(dim, 1)

Warm-up start

Warm-up for time 0

W tensorflow/core/framework/op_def_util.cc:332] Op BatchNormWithGlobalNormalization is deprecated. It will cease to work in GraphDef version 9. Use tf.nn.batch_normalization().

Warm-up for time 1

Warm-up for time 2

Warm-up for time 3

Warm-up for time 4

Warm-up for time 5

Warm-up for time 6

Warm-up for time 7

Warm-up for time 8

Warm-up for time 9

Warm-up finished

mountain bike, all-terrain bike, off-roader (score = 0.56671)

tricycle, trike, velocipede (score = 0.12035)

bicycle-built-for-two, tandem bicycle, tandem (score = 0.08768)

lawn mower, mower (score = 0.00651)

alp (score = 0.00387)

Prediction used time: 4.141446590423584 Seconds

文章来源: http://www.codelast.com/

可见:在10次预热之后,一次预测消耗的时间是 4.14 秒,虽然4秒多还是没有达到我们心目中的理想速度,但这已经比之前的50秒强太多了。

此外,从测试结果我们可以体会到的是:预热(Session.run())的头几次特别慢,后面就快起来了,所以,预热次数太少是不行的。


Viewing all articles
Browse latest Browse all 11063

Trending Articles