Overview

This repo contains code for the "TensorFlow for poets 2" codelab.

Original Repo: Github repo: tensorflow-for-poets-2

This repo contains a simplified and trimmed down version of tensorflow's android image classification example in the android/ directory.

The scripts directory contains helpers for the codelab.

Instructions

Create a conda environment with Python 3.6 and Tensorflow 1.3 on it. Also install Jupyter (optional) in case we'd like to use Jupyter Notebook for quick and dirty experiments. Then activate that conda environment:

source activate
conda create --name py36-tf13 python=3.6 tensorflow=1.3 jupyter
source activate py36-tf13

Navigate to the root directory (i.e. this project directory).

Export environmental variables:

export IMAGE_SIZE=224
export ARCHITECTURE="mobilenet_0.50_${IMAGE_SIZE}"

Start tensorboard:

tensorboard --logdir tf_files/training_summaries --host=localhost &

Access tensorboard at http://localhost:6006/

Run training (change how_many_training_steps as you like):

python -m scripts.retrain \
  --bottleneck_dir=tf_files/bottlenecks \
  --how_many_training_steps=500 \
  --model_dir=tf_files/models/ \
  --summaries_dir=tf_files/training_summaries/"${ARCHITECTURE}" \
  --output_graph=tf_files/retrained_graph.pb \
  --output_labels=tf_files/retrained_labels.txt \
  --architecture="${ARCHITECTURE}" \
  --image_dir=tf_files/flower_photos

Note the creation of the retrained model:

tf_files/retrained_graph.pb, which contains a version of the selected network with a final layer retrained on your categories.
tf_files/retrained_labels.txt, which is a text file containing labels.

Inspect the label_image.py script:

python -m  scripts.label_image -h

Use the Retrained model to do image classification:

python -m scripts.label_image \
    --graph=tf_files/retrained_graph.pb \
    --image=tf_files/flower_photos/daisy/21652746_cc379e0eea_m.jpg

Should get something like this:

daisy 0.961026
dandelion 0.0227085
sunflowers 0.0160663
roses 0.000198295
tulips 1.20007e-06

Try another one:

python -m scripts.label_image \
    --graph=tf_files/retrained_graph.pb  \
    --image=tf_files/flower_photos/roses/2414954629_3708a1a04d.jpg

Should get something like this:

roses 0.994535
tulips 0.00540921
dandelion 4.29636e-05
sunflowers 9.80302e-06
daisy 3.06666e-06

Try a different learning rate (0.5), and create a new summary (see summaries_dir)

python -m scripts.retrain \
  --bottleneck_dir=tf_files/bottlenecks \
  --how_many_training_steps=500 \
  --model_dir=tf_files/models/ \
  --summaries_dir=tf_files/training_summaries/"${ARCHITECTURE}_LR_0.5" \
  --output_graph=tf_files/retrained_graph.pb \
  --output_labels=tf_files/retrained_labels.txt \
  --architecture="${ARCHITECTURE}" \
  --image_dir=tf_files/flower_photos \
  --learning_rate=0.5 \

Note that in Tensorboard, our original model (with default learning rate 0.01) trains much better. This new model (with learning rate 0.5) has persistent high entropy loss.

Atlas7 / tensorflow-for-poets-2

Overview

Instructions

Handy references

About

Languages