Awesome

Transfer Learning Tensorflow

Edit: The following contains the code for freezing the first k layers and retraining the last (n-k) layers: Retraining multiple layers

June Python Pune meetup slides

I followed Tensorflow's tutorial on retraining the final layer of Inception model and tested the results on the flowers dataset(mentioned in the tutorial) as well as on a custom dataset(cats vs dogs).

This tutorial documents the same process along with the problems that I faced while doing so and the links to the solutions.

The definition, from Wikipedia:

Transfer learning or inductive transfer is a research problem in machine learning that focuses on storing knowledge gained while solving one problem and applying it to a different but related problem.

Links:

Setup:

• Install bazel:

  echo "deb [arch=amd64] http://storage.googleapis.com/bazel-apt stable jdk1.8" | sudo tee/etc/apt/sources.list.d/bazel.list
  curl https://bazel.build/bazel-release.pub.gpg | sudo apt-key add -
  sudo apt-get update && sudo apt-get install bazel
  

• Clone tensorflow:

  git clone https://github.com/tensorflow/tensorflow 
  

• Change into tensorflow directory:

  ./configure
  

Dataset 1: Flower dataset

5 categories (600-800 images for each )

• daisy
• sunflowers
• dandelion
• tulips
• roses

Download the dataset

  cd ~
  curl -O http://download.tensorflow.org/example_images/flower_photos.tgz
  tar xzf flower_photos.tgz

Retraining:

  bazel build tensorflow/examples/image_retraining:retrain

Retrain on the flower dataset

  bazel-bin/tensorflow/examples/image_retraining/retrain --image_dir ~/flower_photos 

Final test accuracy = 89.1% (N=384)

Visualize the retraining

tensorboard --logdir /tmp/retrain_logs

Trained model is stored as /tmp/output_graph.pb It is reused for further testing.

/tmp/output_labels.txt contains the labels given for training, i.e., the folder names.

Testing on an image 21652746_cc379e0eea_m.jpg:

bazel-bin/tensorflow/examples/label_image/label_image  --graph=/tmp/output_graph.pb --labels=/tmp/output_labels.txt  --output_layer=final_result  --image=$HOME/Downloads/flower_photos/daisy/21652746_cc379e0eea_m.jpg --input_layer=Mul

Result:

2017-06-05 13:05:24.554667: I tensorflow/examples/label_image/main.cc:251] daisy (2): 0.998173

2017-06-05 13:05:24.554708: I tensorflow/examples/label_image/main.cc:251] sunflowers (3): 0.00125897

2017-06-05 13:05:24.554717: I tensorflow/examples/label_image/main.cc:251] dandelion (4): 0.000368108

2017-06-05 13:05:24.554725: I tensorflow/examples/label_image/main.cc:251] tulips (0): 0.000137791

2017-06-05 13:05:24.554735: I tensorflow/examples/label_image/main.cc:251] roses (1): 6.20492e-05

The flower was correctly detected.

Dataset 2: Cats Dogs dataset

I downloaded 218 images of cats and dogs, each and created a folder 'animals', containg 'cats' and 'dogs' folders, in the home directory.

Retraining on the animals folder:

bazel-bin/tensorflow/examples/image_retraining/retrain --image_dir ~/animals

Final test accuracy = 100.0% (N=36)

Let's test on an image of a dog and a cat:

bazel-bin/tensorflow/examples/label_image/label_image  --graph=/tmp/output_graph.pb --labels=/tmp/output_labels.txt  --output_layer=final_result  --image=$HOME/Pictures/132.jpg --input_layer=Mul

Result: 2017-06-05 13:25:35.234769: I tensorflow/examples/label_image/main.cc:251] dogs (1): 0.998014

2017-06-05 13:25:35.234809: I tensorflow/examples/label_image/main.cc:251] cats (0): 0.00198587

bazel-bin/tensorflow/examples/label_image/label_image  --graph=/tmp/output_graph.pb --labels=/tmp/output_labels.txt  --output_layer=final_result  --image=$HOME/Pictures/117.jpg --input_layer=Mul

Result: 2017-06-05 13:26:53.620682: I tensorflow/examples/label_image/main.cc:251] cats (0): 0.99999

2017-06-05 13:26:53.620725: I tensorflow/examples/label_image/main.cc:251] dogs (1): 1.00321e-05

Errors encountered

• E tensorflow/examples/label_image/main.cc:350] Running model failed: Not found: FeedInputs: unable to find feed output input

  Solution: https://github.com/tensorflow/serving/issues/295 @davidsmandrade

• InvalidArgumentError (see above for traceback): NodeDef mentions attr 'dct_method' not in Op image:uint8; attr=channels ...

  Solution: Tensorflow version issues. Graph was created using another version and being tested using another version.

• ImportError: cannot import name pywrap_tensorflow

  Solution: https://stackoverflow.com/a/35963479

License

MIT