tensorflow

A quick how-to on tensorflow based on Siraj Raval tutorial, "Build a TensorFlow Image Classifier in 5 Min"

Install Docker

1. For Mac user:
   • Download and install this, https://www.docker.com/docker-mac
   • Once you have the above installed, download and install the following, https://www.docker.com/products/docker-toolbox
2. Once you have it installed, make sure to run 'Docker Quickstart Terminal'
3. Let's prepare some other things, before we install tensorflow image
   • Open a new terminal
   • Create a tensorflow/tf_files under your home directory; type the following
     • mkdir ~/tensorflow/tf_files
   • This directory will hold all directories of images that we will be used for training
   • There should be at least 2 directories under ~/tensorflow/tf_files
   • The directory structure should be as follows:

    ~/tensorflow/tf_files
      +-- darth_vader
         +-- dv1.jpg
         +-- dv2.jpg
         +-- dv3.jpg
         +-- ...
         +-- dv{n}.jpg
      +-- darth_maul
         +-- dm1.jpg
         +-- dm2.jpg
         +-- dm3.jpg
         +-- ...
         +-- dm{n}.jpg
      +-- kitten
         +-- ...
     

4. To make easier to download all required files, make sure to install this chrome extension, Fatkun Batch Download Image
5. Now, let's download all required images to be saved in those directories above (darth_vader, darth_maul). You can use the following links to download darth vader and darth maul images:
   • Darth Vader
     • Once you open the above link opened, click Fatkun Batch Download Image icon
     • A modal window will be opened, click on This Tab
     • Click on More Options and make sure to select Rename based on pic_{NO001}.{EXT} option
     • Once you have it downloaded, make sure to move these files to ~/tensorflow/tf_files
   • Darth Maul
     • Do the same steps as the [Dart Vader] above
6. Now, let's install tensoflow image on the docker that you opened through 'Docker Quickstart Terminal'. (See step #2). Type the following in your 'Docker Quickstart Terminal'
   • docker run -it -v ~/tensorflow/tf_files/:/star_wars/ gcr.io/tensorflow/tensorflow:latest-devel
   • Remember: Do not close or exit this tensorflow docker container session
     What the above command does is:
     • It maps your host's ~/tensorflow/tf_files directory to /star_wars/ directory in tensorflow docker container
     • You will be logged in to tensorflow docker container
7. Now, you are in tensorflow docker container, do the following:
   • cd /tensorflow
   • git pull
8. Now, it's time to train the model (darth_vader and darth_maul): Make sure you are in tensorflow docker container, and in /tensorflow/ directory, and type the following:

python tensorflow/examples/image_retraining/retrain.py \
--bottleneck_dir=/tf_files/bottlenecks \
--how_many_training_steps 500 \
--model_dir=/tf_files/inception \
--output_graph=/tf_files/retrained_graph.pb \
--output_labels=/tf_files/retrained_labels.txt \
--image_dir /star_wars

If everything is setup properly, you will see similar output as follows:

>> Downloading inception-2015-12-05.tgz 100.0%
Successfully downloaded inception-2015-12-05.tgz 88931400 bytes.
Looking for images in 'darth_maul'
Looking for images in 'darth_vader'
Looking for images in 'kitten'
No files found
2017-08-08 00:49:13.566091: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-08-08 00:49:13.566173: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-08-08 00:49:13.566235: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
Creating bottleneck at /tf_files/bottlenecks/darth_maul/pic_003.jpg.txt
2017-08-08 00:49:14.095826: W tensorflow/core/framework/op_def_util.cc:332] Op BatchNormWithGlobalNormalization is deprecated. It will cease to work in GraphDef version 9. Use tf.nn.batch_normalization().
Creating bottleneck at /tf_files/bottlenecks/darth_maul/pic_004.jpg.txt
Creating bottleneck at /tf_files/bottlenecks/darth_maul/pic_006.jpg.txt
...
...
...
2017-08-08 00:51:07.991467: Step 480: Validation accuracy = 100.0% (N=100)
2017-08-08 00:51:08.811397: Step 490: Train accuracy = 100.0%
2017-08-08 00:51:08.811650: Step 490: Cross entropy = 0.020678
2017-08-08 00:51:08.896350: Step 490: Validation accuracy = 100.0% (N=100)
2017-08-08 00:51:09.644563: Step 499: Train accuracy = 100.0%
2017-08-08 00:51:09.644881: Step 499: Cross entropy = 0.019680
2017-08-08 00:51:09.738435: Step 499: Validation accuracy = 100.0% (N=100)
Final test accuracy = 100.0% (N=7)
Converted 2 variables to const ops.

9. Now, we want to write a script that uses trained classifier to detect if a given image contains Darth Vader. Open a new terminal windows, and do the following: (Make sure you are not in tensorflow docker container)
   • cd ~/tensorflow/tf_files
   • Create a new file. You can name it whatever you want. Source code
10. Before we can test our sample data/images, let's download any Darth Vader and Darth Maul images:
    • Let's use this Dart Vader image, and save it under ~/tensorflow/tf_files, named it dv1.jpeg
    • Let's use this Dart Maul image, and save it under ~/tensorflow/tf_files, named it nondv1.jpeg
11. Now, go back to the other terminal that tensorflow docker container session, and do the following:
    • python /star_wars/tf_classify.py /star_wars/nondv1.jpeg The output will be similiar as follows:

2017-08-08 01:14:25.582835: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-08-08 01:14:25.583106: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-08-08 01:14:25.583156: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-08-08 01:14:26.187402: W tensorflow/core/framework/op_def_util.cc:332] Op BatchNormWithGlobalNormalization is deprecated. It will cease to work in GraphDef version 9. Use tf.nn.batch_normalization().
darth maul (score = 0.78427)
darth vader (score = 0.21573)

From the above, our application detects nondv1.jpeg is actually Darth Maul image with score 0.78427 (78%)

12. Still in the tensorflow docker container session, do the following:
    • python /star_wars/tf_classify.py /star_wars/dv1.jpeg The output will be similar as follows:

2017-08-08 01:14:39.293931: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-08-08 01:14:39.294094: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-08-08 01:14:39.294141: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-08-08 01:14:39.910169: W tensorflow/core/framework/op_def_util.cc:332] Op BatchNormWithGlobalNormalization is deprecated. It will cease to work in GraphDef version 9. Use tf.nn.batch_normalization().
darth vader (score = 0.89057)
darth maul (score = 0.10943)

From the above, our application detects dv1.jpeg is actually Darth Vader image with score 0.89057 (89%)