Install Tensorflow

pip install --upgrade "tensorflow==1.7.*"

Fork the git repository

Clone the git repository

git clone https://github.com/Navan0/build-your-first-imageClassifier.git

cd build-your-first-imageClassifier 

Download the training images

download your datasets
    

Put the images into the classes

ls tf_files/data

yourclass1/
yourclass2/
yourclass3/
yourclass4/
yourclass5/
LICENSE.txt

(Re)training the network

In this exercise, we will retrain a MobileNet. MobileNet is a a small efficient convolutional neural network. "Convolutional" just means that the same calculations are performed at each location in the image.

Set those variables in your shell

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

Investigate the retraining script

python -m scripts.retrain -h

Run the training

python -m scripts.retrain \
  --bottleneck_dir=tf_files/bottlenecks \
  --how_many_training_steps=4000\
  --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

Classifying an image

python -m scripts.label_image \
    --graph=tf_files/retrained_graph.pb  \
    --image=tf_files/test/test_m.jpg

tflite_convert --help

IMAGE_SIZE=224 tflite_convert
--graph_def_file=tf_files/retrained_graph.pb
--output_file=tf_files/optimized_graph.lite
--input_format=TENSORFLOW_GRAPHDEF
--output_format=TFLITE
--input_shape=1,${IMAGE_SIZE},${IMAGE_SIZE},3
--input_array=input
--output_array=final_result
--inference_type=FLOAT
--input_data_type=FLOAT

cp tf_files/optimized_graph.lite android/tflite/app/src/main/assets/graph.lite cp tf_files/retrained_labels.txt android/tflite/app/src/main/assets/labels.txt

test

python -m scripts.label_image
--graph=/home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/retrained_graph.pb
--image=/home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/pizztest.jpg

optimize

python -m tensorflow.python.tools.optimize_for_inference
--input=/home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/retrained_graph.pb
--output=/home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/optimized_graph.pb
--input_names="input"
--output_names="final_result"

Verify the optimized model

python -m scripts.label_image
--graph=/home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/optimized_graph.pb
--image=/home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/pizztest.jpg

du -h /home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/optimized_graph.pb

gzip -c /home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/optimized_graph.pb > /home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/optimized_graph.pb.gz

Quantize an Image

python -m scripts.quantize_graph
--input=/home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/optimized_graph.pb
--output=/home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/rounded_graph.pb
--output_node_names=final_result
--mode=weights_rounded

gzip -c /home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/rounded_graph.pb > /home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/rounded_graph.pb.gz

gzip -l /home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/rounded_graph.pb.gz

cp /home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/rounded_graph.pb /home/navaneeth/work/tot/build-your-first-imageClassifier/android/tfmobile/assets/graph.pb

cp /home/navaneeth/work/tot/build-your-first-imageClassifier/tf_files/ /home/navaneeth/work/tot/build-your-first-imageClassifier/android/tfmobile/assets/labels.txt