This image classifier is based on an example from Google Tensorflow. It includes information on using tensorflow serving to do inference in production. The reason for packaging it up in it's own repo separate from pdf_trio is because it is useful stand-alone.
Tensorflow-hub is being used so that pre-trained models curated there can be obtained easily.
Clone the this repo, and set up the python env. (We are sssuming conda envs are in use here, adapt as needed.)
git clone https://github.com/tralfamadude/pdf_image_classifier
cd pdf_image_classifier
conda create --name tf_hub python=3.7
conda activate tf_hub
pip install -r requirements.txt
By editing this training script, it is possible to try different pre-trained image models,
alter the learning rate, and number of steps. Note that different pre-trained models like Inception and Mobilenet can have
slightly different images sizes. The images are resized, if necessary, by tf_hub_image_classifier.py
Arrange the your training .jpg files to be in 2 directories like:
image_training/
image_training/research/
image_training/research/foo.jpg
image_training/research/bar.jpg
image_training/other/
image_training/other/boo.jpg
image_training/other/baz.jpg
The directory is specified below to the train.sh.
conda activate tf_hub
./train.sh training_data_dir
Where the training_data_dir has 2 subdirectories: one for each class, which contain the prepared images.
When training is done, the name of the directory containing the SavedModel is emitted to stdout.
The SavedModel format is needed for tensorflow serving. As an example, here is an ls -R of a saved model:
1569277681
./1569277681:
saved_model.pb variables
./1569277681/variables:
variables.data-00000-of-00001 variables.index
We see a top directory 1569277681 which is a Unix epoch that can conveniently stand in for a version number,
a protocol buffer file save_model.pb which contains the serialized graph, and a deeper subdirectory
named variables which contains a large file of parameters variables.data-00000-of-00001 and an index
file.
Saved models are put in a directory so that multiple versions can coexist in the filesystem for tensorflow serving so that is is possible to make production transitions to new versions without downtime. The highest numeric version is served by default.
A withheld set of jpg files can be tested on the command line with this script to verify that the training went well. The target directory should contain *.jpg files:
./test_infer_validation.sh target_dir
output is a bit messy, but look for lines starting with your label names, for example:
research 0.6590455 validation_data/research//f94399f45fab3ef1e95cf7959a5580aa4b3c1d80.jpg
other 0.3409545 validation_data/research//f94399f45fab3ef1e95cf7959a5580aa4b3c1d80.jpg
The model load time and the inference time is printed too. Take heart that with tensorflow serving, the model load only happens once, instead of before each inference as here, and inference time for an image is going to be faster. This measurement should be considered an upper bound.