RootNav 2.0

This is the RootNav 2.0 Code repository. This README and the repository code are being improved daily as we prepare for publication, please check back for new features and documentation soon!

17th Feb 2020 - Improvements coming to training code

• We are now working on some improvements to the training code, which had a few bits of python 2 and uneeded dependencies in etc. We'll also be doing a lot of code refactoring to make it easier to work with and edit for your needs. Once done we'll release this code, but in the meantime the most recent updates will be in the training branch.

Full updates may be found in the change log.

Upcoming Features

• Extended documentation to make clear how the tool is used

Longer term features

• Additional trained models. If you have datasets you'd like to see work with the tool and aren't covered by our examples, please get in touch and we can collaborate!

Installing RootNav 2.0

You will first need to download the code, either as a zip above, or by cloning the git repository (recommended):

git clone https://github.com/robail-yasrab/RootNav-2.0.git

Next, install the required dependencies. If you're using pip, then the following will work in Linux:

cd RootNav-2.0
pip install -r requirements.txt

Library support in other Operating Systems is more complex, and we recommend using Anaconda. You may find Anaconda is also simplest in Linux as well. First install anaconda, then create a new environment using the included yml dependencies file.

conda env create -f environment.yml
conda activate rootnav2

This will download and set up all the required libraries, providing you with a Python 3.6 installation. For those with access to Python 2.6 only, you will find a compatible version in the py2 branch .

Note: If you have Segmentation fault:11 errors on a Mac computer, this is likely caused by an issue with the opencv conda installation. We are working on removing the offending function, you will find a fix in the dev branch

Using the tool

The majority of users will want to run RootNav 2.0 on new images, in which case all the code you need is in the inference folder.

Inference

Running the tool using pre-trained models may be done within the inference folder. Our publication releases three trained models on three datasets:

• Wheat grown on blue paper
• Rapeseed grown on blue paper
• Arabidopsis grown on gel plates

The trained models are not included in the repository, instead they are downloaded automatically the first time they are required. You can list the currently available models with:

cd inference
python rootnav.py --list

Currently, the models available are wheat_bluepaper, osr_bluepaper and arabidopsis_plate. To run RootNav 2.0 on a directory of images, use the following command:

python rootnav.py --model wheat_bluepaper input_directory output_directory

RootNav will read any images within specified input directory, and output images and RSML to the output directory.

Analysing Root Systems

RootNav 2.0 doesn't measure root systems itself, it outputs architectures to RSML files, which may be analysed using a suitable tool. We have adapted the original viewer to perform this function, which you can find on github [RootNav Viewer]. The benefit of separating the extraction of root systems and measurement is that additional measurements may be taken at a later date without requiring re-analysis of the original images. This makes analysing and re-analysing many root systems very straightforward. The viewer can also be extended using plugin code to capture new measurements if desired, however most common phenotypes are already built into the tool.

Training

Training code may be found in the training folder. Instructions on training models are given below. If you would like to collaborate on the development of new models for RootNav 2.0, please contact us.

Dataset Directory Format

Please follow the below given directory formate for the training process. For dataset preparation (RSML to Annotations), please refer to our [RSML_to_Annotations] script.

- Roots
  -- train  (training dataset images)
  -- trainannot (RGB ground truth images)
  -- val (validation dataset images)
  -- valannot (RGB validation ground truth images)
  -- RSML (RSML annotation)

Usage

Setup config file

# Model Configuration
model:
    arch: <name> [options: 'Hourglass, Nested_HG:<value>

# Data Configuration
data:
    dataset: <name> [options: 'wheat_roots', OSR_roots] 
    train_split: <split_to_train_on>
    val_split: <spit_to_validate_on>
    img_rows: 512
    img_cols: 1024
    path: <path/to/dataset>


# Training Configuration
training:
    n_workers: 64
    train_iters: 35000
    batch_size: 16
    val_interval: 500
    print_interval: 25
    loss:
        name: <loss_type> [options: 'cross_entropy, bootstrapped_cross_entropy, multi_scale_crossentropy']
        <loss_keyarg1>:<value>

    # Optmizer Configuration
    optimizer:
        name: <optimizer_name> [options: 'sgd, adam, adamax, asgd, adadelta, adagrad, rmsprop']
        lr: 1.0e-3
        weight_decay: 0.00005
        momentum: 0
        
    # Augmentations Configuration
    augmentations:
        rotate: d                                    #[rotate -d to d degrees]
        hflip: p                                     #[flip horizontally with chance p]


    # LR Schedule Configuration
    lr_schedule:
        name: <schedule_type> [options: 'constant_lr, poly_lr, multi_step, cosine_annealing, exp_lr']
        <scheduler_keyarg1>:<value>

    # Transfer Leaning ... path to pre-trained model  
    # Resume from checkpoint  
    resume: <path_to_checkpoint>

Model Training:

python train.py [-h] [--config [CONFIG]] 

--config                Configuration file to use

# example: python train.py --config ./configs/rootnav2_Nested_HG.yml 

Contact

Publication details will appear here in due course. For enquiries please contact michael.pound@nottingham.ac.uk.