In this project, we retrain Inception's Final Layer for new classes in tensorflow framework.
For more details, please read 2 below articles:

• https://www.tensorflow.org/versions/master/how_tos/image_retraining/index.html
• https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#0

Here are steps to create training data model over 60,000 images into 4 classes: floor plans, map, inside, outside.

1. In project folder (Ex: house), creating training/ for classified images. Create subfolder accordingly:

   • house/
     • training/
       • floor_plans/
       • map/
       • inside/
       • outside/

2. Classify first 200 images manually, move images to folder above accordingly.

3. Retrain Inceptionv3 network in tensorflow as steps:
   3.1 Run docker container and map project folder into docker container:

    docker run -it -v ~/house:/house  gcr.io/tensorflow/tensorflow:latest-devel`
    cd /house`
   

   3.2 Training images:

    python tensorflow/examples/image_retraining/retrain.py 
        --bottleneck_dir=/house/bottlenecks \  
        --how_many_training_steps 4000 \  
        --model_dir=/house/inception \  
        --output_graph=/house/retrained_graph.pb \  
        --output_labels=/house/retrained_labels.txt \  
        --image_dir /house/training.
   

   Training output is a graph and folder data: retrained_graph.pb, retrained_labels.txt

   3.3 Classify next images by label_image.py

    python label_image.py --output_dir=prediction/ --image_path=unclassified-images/ --threshold=90  
   

   • image_path: a unclassified folder or image
   • output_dir: a folder to hold images after classifying
   • threshold: Percent threshold to decide an image is belong to a class

   After this commands, images will be moved to images folder above accordingly.

   3.4 Check again images folder above manually to correct if neccessary.

   Repeat steps from 3.2 to 3.4 for all images to create better training model

4. Use label_image.py script to predict images with training model data above.
   Finally, we can use training mode data above to predict images

    python label_image.py --output_dir=prediction/ --image_path=unclassified-images/ --threshold=90  
   

   • image_path: a unclassified folder or image (only support jpeg images now)
   • output_dir: a folder to hold images after classifying. If it is set, images in image_path will be removed
   • threshold: Percent threshold to decide an image is belong to a class

   Example:

    python label_image.py --image_path=tests/outside1.jpg
   

   Output:

   outside (score = 0.99965)
   map (score = 0.00016)
   floor plans (score = 0.00014)
   inside (score = 0.00006)

   With result above, outside probability = 99.96%

   Prediction Report:

   Place 939 images from internet into sample folder and classify them.

    python label_image.py --output_dir=prediction/ --image_path=sample/ --threshold=80
   

   Output:

   9 images are classified incorrectly
   72 images can not be classified
   858 images are classified correctly
   correct rate = 858/939 = 91.4%

   72 images can not be classified because:

   • some of them is not clear enough even checking by manual.
   • some of them is different from tranining set. So we need to add more training images to improve training model. Ex: house/apartment in Vietnam, ..

   Note: After classification with option --output_dir, classified images in --image_path will be removed. So we need to checkout sample folder if we want to predict with sample folder again.