Brassica napus image analysis

Segmentation of images of Brassica napus plants in the field using a deep-learning network based on densenet:

• "Densely Connected Convolutional Networks" https://arxiv.org/pdf/1608.06993.pdf

Code Author: Sam De Meyer

Installation / Upgrading

Installation

Warning: When installing the GPU enabled pytorch version, over a gigabyte of disk space might be used. If your device doesn't have a GPU, installing the CPU only version will save a lot of disk space.

Note: Below are instructions to install the hooloovoo package as a regular package and this package in development mode. The advantage of development mode is that all the files of the repository are accessible and easy to edit. You can install both packages in development mode using pip install -e.

Conda (recommended)

1. Create / use a conda python 3 environment with at least pip version 19:

   conda create -n $MYENV python=3.7
   source activate $MYENV

   Note: pip can upgrade itself in the current env:

   pip install --upgrade pip

2. Install the correct pytorch version for your system, see the instructions here: https://pytorch.org/get-started/locally/.

3. Install all conda requirements:

   conda install cytoolz future matplotlib numpy pandas pillow pyyaml scikit-image scipy tensorboard tensorflow

4. Install this package with pip:

   # The hooloovoo dependency
   pip install 'git+https://github.com/MMichaelVdV/hooloovoo.git' or pip install git+https://github.com/MMichaelVdV/hooloovoo.git (Windows)
   
   # This repo
   git clone 'git+https://github.com/MMichaelVdV/Brassica_segmentation.git' or git clone https://github.com/MMichaelVdV/Brassica_segmentation.git (Windows)
   cd Brassica_segmentation
   pip install -e .

   After installing a package with pip in conda (if not using -e), package files can be found here:

   $HOME/.conda/envs/$MYENV/lib/python3.$X/site-packages/$SOME_PACKAGE
   

   Where $X is the python version of $MYENV.

Pip

1. (optional) Create a venv for pip. You will at need at least pip version 19.

   Note: pip can upgrade itself in the current env:

   pip install --upgrade pip

2. Install the correct pytorch version for your system, see the instructions here: https://pytorch.org/get-started/locally/.

3. Install this package with pip:

   # The hooloovoo dependency
   pip install 'git+https://github.com/MMichaelVdV/hooloovoo.git'
   
   # This repo
   git clone 'git+https://github.com/MMichaelVdV/Brassica_segmentation.git'
   cd Brassica_segmentation
   pip install -e .

Upgrading

pip install --upgrade 'git+https://github.com/MMichaelVdV/hooloovoo.git'

cd Brassica_segmentation
git pull
pip install -e .

Usage

Once installed, try the following:

cnn_segmentation -h

If you get a help message, the installation worked.

The cnn_segmentation command takes an application and a settings file as argument. Example settings files can be found here:

• https://github.com/MMichaelVdV/Brassica_segmentation/blob/main/cnn_segmentation/applications/brassomics/settings/settings_devel.yaml

Settings file can be either in yaml or json format.

Running on SGE:

Below is a template of a qsub script to run the program on an SGE cluster. This assumes you have installed the package in a conda env called $MYENV.

#!/bin/bash
#$ -l h_vmem=24G
#$ -l h_stack=24M

module load anaconda
source activate $MYENV
Brassica_segmentation $APP $SETTINGS_FILE

Where $SETTINGS_FILE contains the path to the .json or .yaml settings file, and $APP is brassomics.

Tensorboard

If you are training a model, you can follow the training progress remotely by starting a tensorboard server. Inside the python environment, issue the following command:

tensorboard --logdir=$LOGDIR/tensorboard

Where $LOGDIR is the logging path declared in the settings file.

Warning: Tensorboard is not real-time, usually it lags a few minutes behind. As such it is possible that not all metrics are visible right after starting training.

Note: Tensorboard starts scanning the most recent files first, as such, to see the most recent results faster, restart tensorboard and refresh the page.