Running Inception_V3 using On-device AI

 Introduction

 The Inception v3 Imagenet classification model is trained to classify images with 1000 labels.

Inception-V3 is a convolutional neural network architecture designed for image classification and recognition tasks.


Tensorflow Setup :

To set up TensorFlow, perform the following procedure:

1.  Download android-ndk-, and set the ANDROID_NDK_ROOT environment variable. 

  •  $export ANDROID_NDK_ROOT=<path to Android Ndk root>

2. Go to the SNPE SDK root directory, and run the following script to set up the SNPE SDK environment (the TensorFlow home directory must be provided):

  •  $export TENSORFLOW_DIR=/usr2//.local/bin/ $source bin/envsetup.sh -t $TENSORFLOW_DIR 

 where $TENSORFLOW_DIR is the path to the TensorFlow installation.

The script sets up the following environment variables:

  •  SNPE_ROOT: root directory of the SNPE SDK installation
  •  TENSORFLOW_HOME: root directory of the TensorFlow installation provided 

 The script also updates:

  •  PATH, LD_LIBRARY_PATH, and PYTHONPATH.

Running Inception_V3 Model 

Convert the TensorFlow Inception_v3 model to DLC 

1. Download the Inception_v3 model and assets from https://storage.googleapis.com/download.tensorflow.org/models/inception_v3_2016_ 08_28_frozen.pb.tar.gz into the tmp folder.

  •  $cd $SNPE_ROOT $mkdir tmp $wget -N ./tmp “https://storage.googleapis.com/download.tensorflow.org/models/inception _v3_2016_08_28_frozen.pb.tar.gz”

2. Prepare input data for the Inception_v3 model. 

  •  The SNPE requires that the input image be in a format that is different from Caffe or TensorFlow networks. 
  • In the SNPE, the image must be presented as a tensor of shape (height × width × channel), where channel is the fastest-changing dimension.

3. Create the folder structure like this  in the examples/Models/inception_V3 folder










4. Extract the downloaded inception_v3xxx.tar file.
  •  $cd $SNPE_ROOT/models/inception_v3/ $mkdir -p tensorflow_dlc_data/cropped $tar -xzf ./tmp/inception_v3_2016_08_28_frozen.pb.tar.gz -C models/inception_v3/tensorflow

 == Copy List of trained data set to data folder for classification==
  •  $cp tensorflow/imagenet_slim_labels.txt data/
5. Copy the test image data set from $SNPE_ROOT/models/alexnet/data/. 
  • $cp $SNPE_ROOT/models/alexnet/data/*.jpg  data/ 

A sample test image data set is provided by the SNPE SDK in the 
$SNPE_ROOT/models/alexnet/data/ directory. 

 Input data preparation is not related to AlexNet. 

6. The inception_v3 model accepts 299 × 299 raw images as input 
The create_inception_v3_raws.py script CAB file is used to convert the .jpeg files 
into 299 × 299 raw files. 

By default, this script converts to 299 × 299. Use -S <size> to generate the other 
resolutions. 

Create SNPE inception_v3 raw data (convert JPEG to raw): 

  • $python $SNPE_ROOT 
  • /models/inception_v3/scripts/create_inceptionv3_raws.py -i data -d 
  • data/cropped/ -s 299 
The size of raw images must be equal to the model input layer dimensions. Here, the 
size of raw images 299 × 299 ×  3 (three RGB channels) is equal to the TensorFlow 
model input layer dimensions.
 
7. Create the raw file list test files. 

Create image list data files: 

  • $python $SNPE_ROOT /models/inception_v3/scripts/create_file_list.py -i 
  • data/cropped  -o data/cropped/raw_list.txt -e *.raw 
  • $python $SNPE_ROOT/ /models/inception_v3/scripts/ create_file_list.py -i 
  • data/cropped  -o data/target_raw_list.txt  -e *.raw -r 
 Convert the Inception_v3 model

After preparing the assets  convert the .pb format to the DLC format using the snpe-tensorflow-to-dlc tool.
 Convert the .pb file to DLC format:
  •  $snpe-tensorflow-to-dlc --graph tensorflow/inception_v3_2016_08_28_frozen.pb --input_dim input "1,299,299,3" --out_node InceptionV3/Predictions/Reshape_1 --dlc inception_v3.dlc

Run the Inception_v3 model
   
   Run on the host :

Go to the base location for the model, and run snpe-net-run. 
  •  $cd $SNPE_ROOT/models/inception_v3 $snpe-net-run --container dlc/inception_v3_quantized.dlc --input_list data/cropped/raw_list.txt 
A Python script to interpret the classification results is provided and can be used as follows:
  •  $python $SNPE_ROOT/models/inception_v3/scripts/show_inceptionv3_classifications.py -i data/cropped/raw_list.txt \ -o output/ \ -l data/imagenet_slim_labels.txt
The output should show classification results for all images, like the following:



Run on an Android target :

Prerequisite: Push SNPE libraries and the prebuilt snpe-net-run executable to /data/local/tmp/snpeexample on the Android target.

  •  $export SNPE_TARGET_ARCH= aarch64-android-clang5.0 
  •  $export SNPE_TARGET_STL=libgnustl_shared.so $adb shell "mkdir -p /data/local/tmp/snpeexample/$SNPE_TARGET_ARCH/bin" 
  •  $adb shell "mkdir -p /data/local/tmp/snpeexample/
  •  $SNPE_TARGET_ARCH/lib"
  •  $adb shell "mkdir -p /data/local/tmp/snpeexample/dsp/lib" 
  •  $adb push
  •  $SNPE_ROOT/lib/
  •  $SNPE_TARGET_ARCH/* /data/local/tmp/snpeexample/
  •  $SNPE_TARGET_ARCH/lib $adb push $SNPE_ROOT/lib/dsp/* \ \ /data/local/tmp/snpeexample/dsp/lib $adb push 
  •  $SNPE_ROOT/bin/$SNPE_TARGET_ARCH/snpe-net-run \ /data/local/tmp/snpeexample/$SNPE_TARGET_ARCH/bin  
To execute the Inception_v3 classification model on an Android target, perform the following procedure:
     1. Prepare assets in the Android device.
  •  $cd $SNPE_ROOT/models/inception_v3 
  •  $mkdir data/rawfiles && cp data/cropped/*.raw data/rawfiles/ 
  •  $adb shell "mkdir -p /data/local/tmp/inception_v3" 
  •  $adb push data/rawfiles /data/local/tmp/inception_v3/cropped
  •  $adb push data/target_raw_list.txt /data/local/tmp/inception_v3
  •  $adb push dlc/inception_v3_quantized.dlc /data/local/tmp/inception_v3 $rm -rf data/rawfiles

      2. Execute the Inception_v3 model on the Android device using CPU runtime.

 $adb shell 
 #export SNPE_TARGET_ARCH=aarch64-android-clang5.0 
 #export
 LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/data/local/tmp/snpeexample/
$SNPE_TARGE T_ARCH/lib 
 #export PATH=$PATH:/data/local/tmp/snpeexample/$SNPE_TARGET_ARCH/bin
 #snpe-net-run -h
 #cd /data/local/tmp/inception_v3
 #snpe-net-run --container inception_v3_quantized.dlc --input_list target_raw_list.txt
 #exit The executable creates the results folder: /data/local/tmp/inception_v3/output.

    3. To pull the output, run the following:
  •  $adb pull /data/local/tmp/inception_v3/output output_android

 Run the following Python script to check the classification results: 

  •  $python scripts/show_inceptionv3_classifications.py \ -i data/target_raw_list.txt \ -o output_android \ -l data/imagenet_slim_labels.txt 

 The output should show classification results for all images, like the following:



Running Benchmarks for Inception_V3

Pre-requisite :
  • your_model.dlc 
  •   A text file listing all the input data. For an example, see: $SNPE_ROOT/models/alexnet/data/image_list.txt. 
  •   All the input data listed in the text file. For an example, see $SNPE_ROOT/models/alexnet/data/cropped.
$cd $SNPE_ROOT/benchmark  
$cp alexnet_sample.json inception_v3.json 
$vi inception_v3.json 
// modify the inception_v3.json parameters according to inception_v3 
{
  "Name":"inceptionV3", 
  "HostRootPath": "inception_v3",
  "HostResultsDir":"inception_v3/results",
  "DevicePath":"/data/local/tmp/snpeexample", 
  "Devices":["2d07d8f9"], 
  "HostName": "localhost", 
  "Runs":5,

"Model": { 
      "Name": "Inception_v3", 
      "Dlc": 
  "$SNPE_ROOT/models/inception_v3/dlc/inception_v3_quantized.dlc",
                 "InputList":
 "$SNPE_ROOT/models/inception_v3/data/target_raw_list.txt",
                     "Data": [
       "$SNPE_ROOT/models/inception_v3/data/cropped"
              ]
              }, 
     "Runtimes":["GPU", "CPU", "DSP"], 
     "Measurements": ["timing"] 
 }

 Run the benchmark for Inception_v3 

run the following command to run on the only device connected to your computer
  • $cd $SNPE_ROOT/benchmarks  
  • $python snpe_bench.py -c inception_v3.json  
View the results

A latest_results link that points to the most recent run is created.

 # In inception_v3.json, "HostResultDir" is set to "inception_v3/results"
  •  cd $SNPE_ROOT/benchmarks/inception_v3/results 
 # Notice the time stamped directories and the "latest_results" link. 
  •  cd $SNPE_ROOT/benchmarks/inception_v3/results/latest_results 
 # Notice the .csv file, open this file in a csv viewer (Excel, LibreOffice Calc) 


Comments

Post a Comment

Popular posts from this blog

Examples of running Machine Learning Model on Device using Qualcomm AI HUB

Getting started  Installation  We recommend using Miniconda to manage your python versions and environments. Step 1: Python environment Install  miniconda  on your machine. Windows : When the installation finishes, open Anaconda Prompt from the Start menu. macOS/Linux : When the installation finished, open a new shell window. Set up an environment for Qualcomm® AI Hub: conda create python = 3 .8 -n qai_hub conda activate qai_hub Warning If any of these steps fail with  SSL:   CERTIFICATE_VERIFY_FAILED  errors, you have an SSL interception and traffic inspection tool installed. Ask your IT department for instructions on how to set up certificates for Python  pip  and the Python  Requests  library. They should provide you with a new certificate file. You will also need to add this certificate to Miniconda prior to creating your environment. This can be done with  conda   config   --set   ssl_verify   <path ...
Read more

Medical Report Analyzer - Progess

Image
Content : You get to know the approach what we have followed for this Medical report Analyzer project. there are different steps involved in this process and get to know  which libraries are used with code snippets. Extracting the Text From the Input PDF step 1 : We have extracted text from the pdf using the Python library  Pytesseract  installing the usefull libraries using pip command :- ! sudo apt-get install tesseract-ocr ! pip install pytesseract pip install PyMuPDF step 2 : now code to extract text from pdf you will get the output text as : Now there are two ways to Extract the specific field details from the extracted field :  Using the HardCode Method where you will Write code with many if-else condition to handle' different cases for different format of Medical Report you have to parse the text using regular expressions to extract the relevant field details like Name , Age , Gender, DOB , etc.       2. You Can use LLMs to extract the req...
Read more