Tensorflow Image Classifier

This project covers tensorflow image classification through an evolution model. It makes use of a docker instance to handle the training and guessing.

Requirements

• docker with wsl2

Usage

Provider a dataset with training images placed in category directories inside a data directory into the root directory of this project to act as a classifier. the classifier could be star wars characters, e.g.:

 [project-directory]/star-wars-characters/
 [project-directory]/star-wars-characters/data
 [project-directory]/star-wars-characters/data/car
 [project-directory]/star-wars-characters/data/moto
 [project-directory]/star-wars-characters/data/bus

build train/guess environment

./build.sh

The script will build a docker image called zhaw/image-classification:1.0.0. this will be used in further steps to train a model and make guesses for unclassified images.

Train

Just type

 ./train.sh [project-directory]/classifier-directory /path/to/training/images

After the train.sh was executed the classifier directory contains a trained model

Guess

single guesses can be done as follows. parameter 1 is the path to the training data (which contains a trained model now), the second parameter it he path to the unclassified file.

 ./guess.sh [project-directory]/project_directory /unknown/image.jpg

guessing an entire directory can be done as following:

./guessDir.sh [project-directory]/classifier [project-directory]/srcDir [project-directory]/destDir

Example of result

# ./guess.sh /synced/tensor-lib/moto-classifier/ /synced/imagesToTest/moto21.jpg
daisy (score = 0.99898)
sunflower (score = 0.00072)
dandelion (score = 0.00024)
tulip (score = 0.00006)
rose (score = 0.00000)

Use an absolute file path for classifier and images because the script dos not support relative path (volume mounting)

Credits

Credit goes to Xblaster for a lot of inspiration.