转载需注明出处:https://www.codelast.com/

OS:Ubuntu 14.04

在台式机上执行ELL的demo程序 cntkDemo.py 时,可能会遇到程序僵死的问题。
cntkDemo.py 这个程序会调用OpenCV,在一个GUI窗口中显示USB摄像头拍摄的实时视频流,而僵死的现象正是:执行到弹出GUI窗口显示摄像头拍摄的视频流的代码的时候,程序进入僵死状态,不能执行后续逻辑。此时,只能Ctrl+C终止掉程序。

我的Ubuntu 14.04是一台老爷机,性能非常差,我觉得这有可能程序僵死的原因之一?我试了几次都是这样,于是我打算换一个思路来跑这个demo,不再纠结于解决窗口僵死的问题。
文章来源:https://www.codelast.com/
先来看一下原版的 cntkDemo.py 部分代码:

while (True):
# Grab next frame
ret, frame = cap.read()
# Prepare the image to send to the model.
# This involves scaling to the required input dimension and re-ordering from BGR to RGB
data = helper.prepare_image_for_predictor(frame)
# Get the model to classify the image, by returning a list of probabilities for the classes it can detect
predictions = model.Predict(data)
# Get the (at most) top 5 predictions that meet our threshold. This is returned as a list of tuples,
# each with the text label and the prediction score.
top5 = helper.get_top_n(predictions, 5)
# Turn the top5 into a text string to display
text = "".join([str(element[0]) + "(" + str(int(100*element[1])) + "%)  " for element in top5])
# Draw the text on the frame
frameToShow = frame
helper.draw_label(frameToShow, text)
helper.draw_fps(frameToShow)
# Show the new frame
cv2.imshow('frame', frameToShow)
# Wait for Esc key
if cv2.waitKey(1) & 0xFF == 27:
break

这段代码的注释非常清晰,它的功能是:在一个无限循环中,不断地去抓取USB摄像头拍摄的一帧图像,然后用model预测其分类及概率,最后再把预测结果叠加显示在GUI窗口中,类似于下面这样:

ELL coffee mug

既然 cntkDemo.py 主要是为了测试model能不能正常跑,那么我在命令行以文字形式显示预测结果也是一样的啊,没有必要非得在GUI窗口中展示。
文章来源:https://www.codelast.com/
于是我把程序改成了下面这样(完整程序):

import sys
import os
import numpy as np
import cv2
import time
import findEll
import cntk_to_ell
import modelHelper as mh
def get_ell_predictor(modelConfig):
"""Imports a model and returns an ELL.Predictor."""
return cntk_to_ell.predictor_from_cntk_model(modelConfig.model_files[0])
def main():
if (not os.path.exists('VGG16_ImageNet_Caffe.model')):
print("Please download the 'VGG16_ImageNet_Caffe.model' file, see README.md")
sys.exit(1)
# ModelConfig for VGG16 model from CNTK Model Gallery
# Follow the instructions in README.md to download the model if you intend to use it.
helper = mh.ModelHelper("VGG16ImageNet", ["VGG16_ImageNet_Caffe.model"], "cntkVgg16ImageNetLabels.txt", scaleFactor=1.0)
# Import the model
model = get_ell_predictor(helper)
# Save the model
helper.save_ell_predictor_to_file(model, "vgg16ImageNet.map")
camera = 0
if (len(sys.argv) > 1):
camera = int(sys.argv[1]) 
# Start video capture device
cap = cv2.VideoCapture(camera)
while (True):
print('Read a frame from camera...')
ret, frame = cap.read()
# Prepare the image to send to the model.
# This involves scaling to the required input dimension and re-ordering from BGR to RGB
data = helper.prepare_image_for_predictor(frame)
# Get the model to classify the image, by returning a list of probabilities for the classes it can detect
predictions = model.Predict(data)
# Get the (at most) top 5 predictions that meet our threshold. This is returned as a list of tuples,
# each with the text label and the prediction score.
top5 = helper.get_top_n(predictions, 5)
# Turn the top5 into a text string to display
text = "".join([str(element[0]) + "(" + str(int(100*element[1])) + "%)  " for element in top5])
# Output the text on command line
print(text)
if __name__ == "__main__":
main()

其中最关键的是,不再用 cv2.imshow() 的方式来弹出GUI窗口,而是用 print(text) 的方式把结果打印到command line。
经过实测,在我的老爷机上这个程序就完全不会僵死了。此时,你把摄像头对准哪个物体,它拍摄的就是哪个物体的图像,model也就是对这个物体的图像进行分类预测。
文章来源:https://www.codelast.com/
下面是我的一次test的command line输出:

OpenBLAS : Your OS does not support AVX instructions. OpenBLAS is using Nehalem kernels as a fallback, which may give poorer performance.
Read a frame from camera, time 1
Frame 1 saved to disk
Read a frame from camera, time 2
Frame 2 saved to disk
Read a frame from camera, time 3
Frame 3 saved to disk
Read a frame from camera, time 4
Frame 4 saved to disk
Read a frame from camera, time 5
Frame 5 saved to disk
Loading...
Selected CPU as the process wide default device.
 
Finished loading.
Pre-processing...
 
Will not process Dropout - skipping this layer as irrelevant.
 
Will not process Dropout - skipping this layer as irrelevant.
 
Will not process Combine - skipping this layer as irrelevant.
Convolution :  226x226x3  ->  224x224x64 | padding  1
ReLU :  224x224x64  ->  226x226x64 | padding  0
Convolution :  226x226x64  ->  224x224x64 | padding  1
ReLU :  224x224x64  ->  224x224x64 | padding  0
Pooling :  224x224x64  ->  114x114x64 | padding  0
Convolution :  114x114x64  ->  112x112x128 | padding  1
ReLU :  112x112x128  ->  114x114x128 | padding  0
Convolution :  114x114x128  ->  112x112x128 | padding  1
ReLU :  112x112x128  ->  112x112x128 | padding  0
Pooling :  112x112x128  ->  58x58x128 | padding  0
Convolution :  58x58x128  ->  56x56x256 | padding  1
ReLU :  56x56x256  ->  58x58x256 | padding  0
Convolution :  58x58x256  ->  56x56x256 | padding  1
ReLU :  56x56x256  ->  58x58x256 | padding  0
Convolution :  58x58x256  ->  56x56x256 | padding  1
ReLU :  56x56x256  ->  56x56x256 | padding  0
Pooling :  56x56x256  ->  30x30x256 | padding  0
Convolution :  30x30x256  ->  28x28x512 | padding  1
ReLU :  28x28x512  ->  30x30x512 | padding  0
Convolution :  30x30x512  ->  28x28x512 | padding  1
ReLU :  28x28x512  ->  30x30x512 | padding  0
Convolution :  30x30x512  ->  28x28x512 | padding  1
ReLU :  28x28x512  ->  28x28x512 | padding  0
Pooling :  28x28x512  ->  16x16x512 | padding  0
Convolution :  16x16x512  ->  14x14x512 | padding  1
ReLU :  14x14x512  ->  16x16x512 | padding  0
Convolution :  16x16x512  ->  14x14x512 | padding  1
ReLU :  14x14x512  ->  16x16x512 | padding  0
Convolution :  16x16x512  ->  14x14x512 | padding  1
ReLU :  14x14x512  ->  14x14x512 | padding  0
Pooling :  14x14x512  ->  7x7x512 | padding  0
linear :  7x7x512  ->  1x1x4096 | padding  0
ReLU :  1x1x4096  ->  1x1x4096 | padding  0
linear :  1x1x4096  ->  1x1x4096 | padding  0
ReLU :  1x1x4096  ->  1x1x4096 | padding  0
linear :  1x1x4096  ->  1x1x1000 | padding  0
Softmax :  1x1x1000  ->  1x1x1000 | padding  0
 
Finished pre-processing.
 
Constructing equivalent ELL layers from CNTK...
Converting layer  conv1_1: Convolution(data: Tensor[3,224,224]) -> Tensor[64,224,224]
Converting layer  relu1_1: ReLU(conv1_1: Tensor[64,224,224]) -> Tensor[64,224,224]
Converting layer  conv1_2: Convolution(relu1_1: Tensor[64,224,224]) -> Tensor[64,224,224]
Converting layer  relu1_2: ReLU(conv1_2: Tensor[64,224,224]) -> Tensor[64,224,224]
Converting layer  pool1: Pooling(relu1_2: Tensor[64,224,224]) -> Tensor[64,112,112]
Converting layer  conv2_1: Convolution(pool1: Tensor[64,112,112]) -> Tensor[128,112,112]
Converting layer  relu2_1: ReLU(conv2_1: Tensor[128,112,112]) -> Tensor[128,112,112]
Converting layer  conv2_2: Convolution(relu2_1: Tensor[128,112,112]) -> Tensor[128,112,112]
Converting layer  relu2_2: ReLU(conv2_2: Tensor[128,112,112]) -> Tensor[128,112,112]
Converting layer  pool2: Pooling(relu2_2: Tensor[128,112,112]) -> Tensor[128,56,56]
Converting layer  conv3_1: Convolution(pool2: Tensor[128,56,56]) -> Tensor[256,56,56]
Converting layer  relu3_1: ReLU(conv3_1: Tensor[256,56,56]) -> Tensor[256,56,56]
Converting layer  conv3_2: Convolution(relu3_1: Tensor[256,56,56]) -> Tensor[256,56,56]
Converting layer  relu3_2: ReLU(conv3_2: Tensor[256,56,56]) -> Tensor[256,56,56]
Converting layer  conv3_3: Convolution(relu3_2: Tensor[256,56,56]) -> Tensor[256,56,56]
Converting layer  relu3_3: ReLU(conv3_3: Tensor[256,56,56]) -> Tensor[256,56,56]
Converting layer  pool3: Pooling(relu3_3: Tensor[256,56,56]) -> Tensor[256,28,28]
Converting layer  conv4_1: Convolution(pool3: Tensor[256,28,28]) -> Tensor[512,28,28]
Converting layer  relu4_1: ReLU(conv4_1: Tensor[512,28,28]) -> Tensor[512,28,28]
Converting layer  conv4_2: Convolution(relu4_1: Tensor[512,28,28]) -> Tensor[512,28,28]
Converting layer  relu4_2: ReLU(conv4_2: Tensor[512,28,28]) -> Tensor[512,28,28]
Converting layer  conv4_3: Convolution(relu4_2: Tensor[512,28,28]) -> Tensor[512,28,28]
Converting layer  relu4_3: ReLU(conv4_3: Tensor[512,28,28]) -> Tensor[512,28,28]
Converting layer  pool4: Pooling(relu4_3: Tensor[512,28,28]) -> Tensor[512,14,14]
Converting layer  conv5_1: Convolution(pool4: Tensor[512,14,14]) -> Tensor[512,14,14]
Converting layer  relu5_1: ReLU(conv5_1: Tensor[512,14,14]) -> Tensor[512,14,14]
Converting layer  conv5_2: Convolution(relu5_1: Tensor[512,14,14]) -> Tensor[512,14,14]
Converting layer  relu5_2: ReLU(conv5_2: Tensor[512,14,14]) -> Tensor[512,14,14]
Converting layer  conv5_3: Convolution(relu5_2: Tensor[512,14,14]) -> Tensor[512,14,14]
Converting layer  relu5_3: ReLU(conv5_3: Tensor[512,14,14]) -> Tensor[512,14,14]
Converting layer  pool5: Pooling(relu5_3: Tensor[512,14,14]) -> Tensor[512,7,7]
Converting layer  fc6: linear(pool5: Tensor[512,7,7]) -> Tensor[4096]
Converting layer  relu6: ReLU(fc6: Tensor[4096]) -> Tensor[4096]
Converting layer  fc7: linear(drop6: Tensor[4096]) -> Tensor[4096]
Converting layer  relu7: ReLU(fc7: Tensor[4096]) -> Tensor[4096]
Converting layer  fc8: linear(drop7: Tensor[4096]) -> Tensor[1000]
Converting layer  prob: Softmax(fc8: Tensor[1000]) -> Tensor[1000]
 
...Finished constructing ELL layers.
lighter, light, igniter, ignitor(28%)  
lighter, light, igniter, ignitor(28%)  
 
lighter, light, igniter, ignitor(32%)  
lighter, light, igniter, ignitor(30%)
......
有人可能会说这样做不直观,根本无法肯定摄像头当前正在拍摄的是什么东西。如果你非要看图片的话,倒是有一个折中的办法,就是用 cv2.imwrite('/home/codelast/current.jpg', frame) 把抓取的一帧图像保存到磁盘上,然后自己去打开文件看吧。
文章来源:https://www.codelast.com/
最后不得不感叹一下,我的台式机真的是太老了,一开头打印出的那一句“OpenBLAS : Your OS does not support AVX instructions. OpenBLAS is using Nehalem kernels as a fallback, which may give poorer performance” 就已经给我打了预防针,事实证明我的台式机跑这个demo真的很慢,没有个5分钟以上是根本没可能到开始预测的步骤的。
[原创] 执行ELL的demo程序cntkDemo.py时程序僵死的问题
Tagged on:             

发表评论

电子邮件地址不会被公开。 必填项已用*标注