6/deep-cat-or-dog

use deep learning to answer the question, "Cat or Dog?"

deep-cat-or-dog

use deep learning (tensorflow) to answer the question, "Cat or Dog?"

using the classifier

try it out with example images that weren't in the training set:

python label_image.py example/cat1.jpg
cat (score = 0.99855)
dog (score = 0.00145)

python label_image.py example/dog1.jpg
dog (score = 0.98780)
cat (score = 0.01220)

training the model yourself (optional)

• download train.zip from https://www.kaggle.com/c/dogs-vs-cats/data
• unzip that into /train/ folder in this directory
• follow instructions from https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#1

On 3. Retrieving the images, rather than downloading the flowers dataset, run ruby sort-dataset.rb to sort the cat/dog dataset into the tensorflow retrain.py's expected folder structure.

Also, replace the following command:

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

with:

docker run -it -v /path/to/deep-cat-or-dog:/tf_files  gcr.io/tensorflow/tensorflow:latest-devel

Finally, on 4. (Re)training Inception, replace

--image_dir /tf_files/flower_photos

with:

--image_dir /tf_files/tf_train

optimizing the model (optional)

This step makes classification faster but potentially lower accuracy:

# In Docker:
python /tensorflow/tensorflow/python/tools/optimize_for_inference.py \
  --input=/tf_files/retrained_graph.pb \
  --output=/tf_files/optimized_graph.pb \
  --input_names='DecodeJpeg/contents:0' \
  --output_names=final_result

python /tensorflow/tensorflow/tools/quantization/quantize_graph.py \
  --input=/tf_files/optimized_graph.pb
  --output=/tf_files/quantized_graph.pb
  --output_node_names=final_result
  --mode=weights_rounded

Finally, use bazel to run memory mapping (why?):

# In Docker:
cd /tensorflow
bazel build tensorflow/contrib/util:convert_graphdef_memmapped_format
bazel-bin/tensorflow/contrib/util/convert_graphdef_memmapped_format \
  --in_graph=/tf_files/quantized_graph.pb \
  --out_graph=/tf_files/memmapped_graph.pb

Note: you may have to append --local_resources 256,2.0,1.0 to the bazel build command to get it to work within Docker.