First complete version of the BGNN image segmentation workflow managed using snakemake.
The image segmentation workflow is managed using snakemake, a user-friendly python workflow manager. Snakemake uses a syntax based on python and can use python code in the definition of the workflow.
The segmentation workflow consists of the following steps each defined by a "rule". The output of each rule is stored to a specific folder that is produced by the workflow:
-
Download the fish Images from URLs within the input CSV file. The CSV file must have a column named
paththat contains the URL of the image. The CSV file must have a column namedoriginal_file_namethat will be used to create a unique filename for the images downloaded.- Images are saved to the folder Images.
-
Extract Metadata information using Detectron2 (deep learning segmentation). The code developed by Drexel and the script used can be found here.
- This code extracts the ruler scale (pixels/cm).
- This code identified and extracts a bounding box around the fish in the image.
- Information about the ruler scale and bounding box size are saved as .json files in the folder Metadata.
- The output filename is based on the image filename adding the filename suffix "_mask" and changing the extension.
- The mask is saved in the folder Mask.
-
Create Cropped images of the fish using the bounding box from Metadata. The code is under Crop_image_main.py. * We increased the bounding box around the fish by 2.5% on each side (5% total) for the crop to prevent truncation of the file. * The output filename is based on the image filename adding the filename suffix "_cropped". * The cropped image is in the folder Cropped.
-
Segmented traits using code developed by Maruf and reorganized by Thibault here.
- This code produces an image of a segmented fish (color coding following example in "Stage 1" below).
- The input cropped image must be resized to 800x320 pixels.
- We then resize the output segmented image (800x320) to the size of the cropped image (which is the size of the bounding box plus 5% increase).
- This ensures that the image is at the same scale as the ruler when the ruler scale was extracted in Metadata.
- The output filename is based on the image filename adding the filename suffix "_segmented".
- The segmented image is saved in the folder Segmented.
The steps are represented in the following workflow diagram.
-
Snakemake package: installed and managed by conda or mamba, which is similar to conda but optimized for snakemake. To access snakemake, use the bioconda channel or mamba conda channel.
-
Snakefile
-
Scripts for
-
Containers
- these are pulled from the respective GitHub repositories
-
Data
- Images : store the ouput from the Download step. Images downloaded from Tulane server
- Metadata : store the output from generate_metadata.py code developed by Drexel team. One file ".json" per image
- Cropped : store the ouput from Crop image.
- Segmented : store the ouput from Segment trait using code developed by M. Maruf (Virginia Tech)
Some of the containers are created using GitHub action, some other are created using singularity remote builder. We are currently transitioning all the containers to github action.
There are 3 containers of interest:
- Metadata_generator :
singularity pull --arch amd64 library://thibaulttabarin/bgnn/metadata_generator:v2 Usage : gen_metadata.py {image.jpg} {metadata.json} {mask.jpg} - Crop_image :
docker pull ghcr.io/hdr-bgnn/crop_image/crop_image:0.0.2 Usage : Crop_image_main.py {image.jpg} {metadata.json} {Cropped.jpg} - Segment_trait:
docker://ghcr.io/hdr-bgnn/bgnn-trait-segmentation:0.0.4 Usage : segmentation_main.py {Cropped.png} {Segmented.png}
This is the best way to start.
Requirement: have an acount at OSC. It is easy to register for an accout.
1- Log onto a login node... Be gentle with them, they don't like to work too hard!
ssh <username>@pitzer.osc.edu
2- Clone the repository, do it only the first time.
git clone [email protected]:hdr-bgnn/BGNN_Snakemake.git
3- Only the first.
module load miniconda3
4- Create an environment named snakemake with module snakemake
conda create -n snakemake -c bioconda -c conda-forge snakemake -y
5- Activate the environment, so now you have access to the package snakemake
source activate snakemake
6- you should see environment named "snakmake" if not check here for more info
conda info -e
9- Request an interactive session on a computing node. Replace <PROJECT_NAME> by the appropriete value.
sinteractive -N 1 -n 4 -t 00:10:00 -A <PROJECT_NAME> -J test -p debug squeue -u $USER
10- Again! yes since you are on different node (computing node that you have requested)
module load miniconda3
11- # Again! Same as before
source activate snakemake
12- Test snakeme. The first time, it may take sometime to download the container. In this simple version, snakemake will call Snakefile, all the results will be dump in the directory where you are (containing Snakefile)
snakemake --cores 4 --use-singularity --config list=List/list_lepomis_INHS.csv #
13- To exit the computing node
exit
14- To check the result. You shloud see folders Images/ Metadata/ Cropped/ Segmented/... populated with multiple fish_file on some sort.
```ls ~/BGNN_Snakemake```
To submit a job to OSC I use the script SLURM_Snake, it will call the Snakefile from the same directory.
Usage, connect to the login node
ssh <username>@pitzer
Clone this BGNN_Snakemake repo (if necessary) and cd into the repo.
git clone [email protected]:hdr-bgnn/BGNN_Snakemake.git
cd BGNN_Snakemake
The SLURM_snake script requires a snakemake conda environment.
If you have followed the Using interactive instructions this environment will already exist.
If not, run the following to load miniconda, create the environment, and unload miniconda.
module load miniconda3/4.10.3-py37
conda create -n snakemake -c bioconda -c conda-forge snakemake -y
module purge
The SLURM_snake script has the following positional arguments:
- a data directory to hold all data for a single run - required
- a CSV file that contains details about the image files to process - required
- the number of jobs for Snakemake to run at once - optional defaults to 4
The SLURM_Snake script should be run with arguments like so:
sbatch SLURM_Snake <data_directory> <path_to_csv> [number_of_jobs]
For example if you want to store the data files at the relative path of data/list_test and processs List/list_test.csv run the following:
sbatch SLURM_Snake data/list_test List/list_test.csv
To run the example with up to 8 parallel jobs run the command like so:
sbatch SLURM_Snake data/list_test List/list_test.csv 8
To check the process
squeue -u $USER
That's it!
Comment: this script will create a slurm-job_ID.out log file.
The data_directory will contain the following directory structure (plus some log and cache file):
Images/
Cropped/
Metadata/
Mask/
Segmented/
The SLURM_Snake script configures Snakemake to submit separate sbatch jobs for each step run.
Details about individual steps can be seen by running the sacct command.
For example:
sacct --format JobID,JobName%40,State,Elapsed,ReqMem,MaxRSS
Keep in mind that sacct defaults to showing jobs from the current day. See sacct docs for options to specify a different time range.
The pipeline can be run using docker to provide the required snakemake and singularity environment.
After installing docker clone this repo and cd into the BGNN_Snakemake directory.
To process the images in List/list_test.csv using the pipeline run (on macOS or Linux):
docker run --privileged -it -v $(pwd):/src -w /src snakemake/snakemake:v7.12.1 \
snakemake -c4 --use-singularity --config list=List/list_test.csv
For Windows use PowerShell replacing $(pwd) with ${PWD} in the command above.