Real Estate Image Classifer using Tensorflow

I followed this O'Reilly tutorial to train an image classifier for real estate photos using Tensorflow.

The first version of the model is based on Inception trained with ImageNet photos. I've retrained the final layer with a set of real estate images that have nine different labels (bathroom, bedroom, closet, dining room, etc). The overal accuracy of the first version is 84.6%.

Follow instructions in this tutorial and set up real estate photos in tf_files directory (with categories already labeled) as instructed.

Load Docker Quickstart Terminal app and use iTerm as terminal.

$ docker run -it -v $HOME/tf_files:/tf_files b.gcr.io/tensorflow/tensorflow:latest-devel

Updating the code…(this is apain - have to do it every time). Pulling the code requires a default email address, which you can set to anything, since we’re not planning on pushing any changes back.

$ git config --global user.email "you@example.com"
$ git config --global user.name "Your Name"

Now you should be able to pull the latest source code.

$ cd /tensorflow/
$ git pull origin master

error...

$ git commit -a
$ git pull origin master

THIS IS VERY IMPORTANT: You should now have fully up-to-date code. We want to sync it to a version we know works, though, so we’ll run this command:

$ git checkout 6d46c0b370836698a3195a6d73398f15fa44bcb2

Now build the code:

$ cd /tensorflow/
$ bazel build -c opt --copt=-mavx tensorflow/examples/image_retraining:retrain

Running the code (or some version of below with different parameters):

$ bazel-bin/tensorflow/examples/image_retraining/retrain \
--bottleneck_dir=/tf_files/home_bottlenecks \
--model_dir=/tf_files/inception \
--output_graph=/tf_files/home_graph.pb \
--output_labels=/tf_files/home_labels.txt \
--image_dir /tf_files/home_photos

The model is saved as a protobuf file (in this case home_graph.pb) and training label names are saved as a text file (home_labels.txt). Here is a version of the training model with different learning rate and batch size:

$ bazel-bin/tensorflow/examples/image_retraining/retrain \
--bottleneck_dir=/tf_files/home_bottlenecks_3 \
--model_dir=/tf_files/inception \
--output_graph=/tf_files/home_graph_4.pb \
--output_labels=/tf_files/home_labels_4.txt \
--image_dir /tf_files/home_photos \
--learning_rate 0.001 \
--training_batch_size 200

1. Using other topologies: e.g. VGGNet, MXNet (which are better for localized images)
2. Using other base training datasets: e.g. MIT Places dataset (better for location, architecture, room images)
3. Tweaking model hyperparameters (e.g. learning rate, training batch size), input distortions (to make training more robust), size of training/test set (control for overfitting).