arnoldkuo/tf

Tensorflow

Tensorflow Installation (on Windows)

• Install Microsoft Visual Studio Community 2019 v16.0.4

• Install Python 3.7 for Windows

• Install nVidia CUDA 10.1 & CuDNN 7.6

• Install Tensorflow

  • pip install tensorflow-2.0.0-cp37-cp37m-win_amd64.whl
    Tensorflow 2.0.0 cpu for Windows10
    Tensorflow 2.0.0 cpu for Windows7
    

  • pip install tensorflow_gpu-2.0.0-cp37-cp37m-win_amd64.whl
    Tensorflow 2.0.0 cpu for Windows10
    Tensorflow 2.0.0 cpu for Windows7
    

  • pip install tensorflow-1.14.0-cp37-cp37m-win_amd64.whl
    Tensorflow 1.14.0 cpu for Windows10
    
    Tensorflow 1.14.0 cpu for Windows7
    

  • pip install tensorflow_gpu-1.14.0-cp37-cp37m-win_amd64.whl
    Tensorflow 1.14.0 gpu for Windows10
    
    Tensorflow 1.14.0 gpu for Windows7
    

  • tensorflow-optimized-wheels for linux
    

• Download Examples git clone https://github.com/rkuo2000/tf

• Go to directory cd tf

Tensorflow Sample Code

Intro - Tensorflow Session & Tensorboard

• jupyter notebook to run intro.ipynb
• run tensorboard : tensorboard --logdir=./
• use Chrome to open http://localhost:6006

Basic Machine Learning

• one neuron network : python3 easy_net.py
• ten neurons network: python3 hidden_net.py

MNIST : Handwritten Number Recognition

• plot data : python3 mnist_plotdata.py
• DNN : python3 mnist.py
• CNN : python3 mnist_cnn.py
• load model to predict :
  python3 mnist_cnn_test.py (test data)
  python3 mnist_cnn_image.py (image file)
  python3 mnist_cnn_webcam.py (camera)
  

Fashion-MNIST : Fashion Wearing Recongition

• CNN : python3 fashionmnist_cnn.py

Emotion Detection : Facial Expression Recognition

• Download the FER-2013 dataset from here and unzip it under data folder.
• change directory name from data/data to data/fer2013
• To train the model, run python3 emotion_detection.py --mode train
• To detect facial expression, run python3 emotion_detection.py --mode detect

Object Detection

• Download Google Images
  python3 download_google_images.py bread
  
• Download a pretrained model
  Download a pretrained model from TensorFlow Model Zoo
  cd ~/models/research/object_detection
tar zxvf ~/Downloads/ssd_mobilenet_v2_coco_2018_03_29.tar.gz
  
• Run Object_Detection
  cp ~/tf/Object_detection_*.py .
python3 Object_detection_image.py
python3 Object_detection_webcam.py
  
• Export Frozen_Inference_Graph
  cp ~/tf/export_inference_graph.sh .
./export_inference_graph.sh training model.ckpt-????
  
• Convert TF model to TFLite (for Android App)
  cp ~/tf/tflite_*.sh . (copy shell files)
  ./tflite_export.sh (convert from model.ckpt to tflite_graph.pb)
  ./tflite_convert_pb.sh (convert tflite_graph.pb to model.tflite)
  ./tflite_convert_pb_quant.sh (convert tflite_graph.pb to model_quant.tflite)
  

Transfer Learning (on Windows)

• Transfer Learning using Keras Mobilenet V2
  cd ~/tf
python3 download_google_images.py "blue tit" (download dataset)
  python3 download_google_images.py crow (download dataset)
  python3 transfer_learning_mobilenetv2.py (transfer learning)
  python3 transfer_learning_image.py (load model and test image file)
  python3 transfer_learning_webcam.py (load model and input from webcam )
  
• Convert Keras model to TFLite (for Android App)
  ./tflite_convert_h5.sh (convert tl_mobilenetv2.h5 to tl_mobilenetv2.tflite)
  ./tflite_convert_h5_quant.sh (convert tl_mobilenetv2.h5 to tl_mobilenetv2_quant.tflite)
  

Edge TPU (USB Accelerator on RPi3B)

• Convert _quant.tflite to _quant_edgetpu.tflite
  
• upload tl_mobilenetv2_quant.tflite to EdgeTPU online compiler
  
• download _quant_edgetput.tflite and copy to RPi3
  
• On RPi3B
  cd ~
git clone https://github.com/rkuo2000/tf (clone sample codes)
  cd ~/tf
vi model/bird_labels.txt (create label file)
  
• To test the model :
  python3 edgetpu_classify_webcam.py --model model/tl_mobilenetv2_quant.tflite --label model/bird_labels.txt
python3 edgetpu_classify_image.py --model model/tl_mobilenetv2_quant.tflite --label model/bird_labels.txt