Skip to content

README.md

Latest commit

 

History

History
87 lines (73 loc) · 3.03 KB

README.md

File metadata and controls

87 lines (73 loc) · 3.03 KB

Introduction

Dockerfile to build a MySQL container image which can be used for PAD machine learning activities.

Installation

Build the container,

docker build --tag pad_nn_training:1.0 .

Run the container. Note ports only required for external mysql access,

docker run -p 3306:3306 -dit  --name padnn1 pad_nn_training:1.0

View logs,

docker logs padnn1

you should see something like,

Installing database...
Starting MySQL server...
Waiting for database server to accept connections.
Creating debian-sys-maint user...
Creating database "pad"...
Granting access to database "pad" for user "pad_user"...
Loading SQL file "/working/pad-database.sql"...

Run bash inside container,

docker exec -it padnn1 /bin/bash

Remove the container when you are finished.

docker rm --force padnn1

Use

To access the container through a bash terminal run

docker exec -it padnn1 /bin/bash

The container loads the MSH training/test images to folder /working/msh_msh_tanzania_data/. The pad_caffenet_train.py script will train the NN using these images,

python3 pad_caffenet_train.py

The checkpointed weight files can be found in tmp/caffenet_pad_msh_tanzania.ckpt which can be loaded as follows (given that the checkpoint is model_checkpoint),

#create session
with tf.Session() as sess:
    new_saver = tf.train.import_meta_graph(model_checkpoint+'.meta')
    new_saver.restore(sess, model_checkpoint)
    #load in the saved weights
    saver.restore(sess, model_checkpoint)
    ...

For generating different image partitions the container loads all of the MSH images to /var/www/html/joomla/images/padimages/msh/processed, the images can be located in the SQL database to obtain full metadata.

All provided images are cropped to (71, 359, 71+636, 359+490) and resized to (227,227) for training. Full sized images sets are available at http://www.crc.nd.edu/~csweet1/padimages/, for example msh.tar.

To crop and resize full size images modify the resize_images_bundle.py script to point to the selected folders and run,

python3 resize_images_bundle.py

To predict the catagory for an image, given a set of trained weights,

python3 predict_graph.py -n msh_tanzania_3k_12.nnet -i Acetaminophen-12LanePADKenya2015-1-58861.processed.png

The nnet file contains,

DRUGS,Paracetamol Starch,Penicillin Procaine,Starch,Lactose,Amoxicillin,Cellulose,Vitamin C,Quinine,Benzyl Penicillin,Paracetamol,DI water,Blank PAD
LANES,
WEIGHTS,tf_checkpoints/caffenet_pad_msh_tanzania_blank.ckpt
TYPE,tensorflow
DESCRIPTION,MSH Tanzania data with blank and DI water cards: 12 drug NN, 3k images (Tensorflow version), no image brightness leveling.
TEST,12LanePADKenya2015

which describes the network parameters. The weights described above can be downloaded from http://www.crc.nd.edu/~csweet1/padimages/caffenet_pad_msh_tanzania_blank.tar.gz.

Note: This script works with rectified images, not the cropped and resized (227x227) images provided in the docker container.