Installing

I recommend installing Anaconda to manage dependencies in isolated environments. Once you have downloaded the latest version of Anaconda for your OS, you can use the provided environment.yml file to automatically install the latest compatible packages for running the scripts in this repo.

Simply run:

conda env create -f environment.yml

(downloads packages)

source activate tf101gpu-py35

There are three different files for training:

• train_ph_iou.py - train by manually feeding batches of images through tensorflow placeholders mechanism, uses intersection-over-union loss approximation from http://www.cs.umanitoba.ca/~ywang/papers/isvc16.pdf - only supports binary classification
• train_ph_xent.py - train by manually feeding batches of images through tensorflow placeholders mechanism, uses classic cross-entropy loss and supports arbitrary number of classes through --out argument.
• train_q.py - train by tensorflow queue prefetching mechanism, used for normal RGB images. Supports both loss types.

If using train_q.py, you must create tfrecords first:

training - should result in a 140.9MB *.tfrecords file

python preprocessing/rgb_to_byte_tfrecords.py --in_path <path/to/>ciona17_farm1_training1/ --out_path <path/to/>tfrecord/ --rec ciona17-rgb-train

validation - should result in a 67.1MB *.tfrecords file

python preprocessing/rgb_to_byte_tfrecords.py --in_path <path/to/>ciona17_farm1_validatio
n/ --out_path <path/to/>tfrecord/ --rec ciona17-rgb-valid

Then, usage of training scripts:

usage: train_*.py [-h] [--train_dir TRAIN_DIR] [--restore RESTORE]
                       [--gpu GPU] [--out OUT] [--bs BS] [--lr LR]
                       [--model MODEL] [--loss LOSS] [--fmt FMT] [--plot PLOT]
                       [--root_path ROOT_PATH] [--debug] [--email]
                       save

The annotated dataset used with this code can be found at: https://dataverse.scholarsportal.info/dataset.xhtml?persistentId=doi:10.5683/SP/NTUOK9

To use the --email functionality, create a .env file in the root of this repo with the following format:

EMAIL:<user>@gmail.com
PASSWORD:<password>

Note that only gmail is supported for now. If you are a MLRG member ask me for permission to use the [email protected] account so that you don't have to insecure your personal gmail.