Skip to content

Latest commit

 

History

History
105 lines (81 loc) · 4.61 KB

README.md

File metadata and controls

105 lines (81 loc) · 4.61 KB
Raw

Docker Containerization of the Multiview Registration Plugin (https://github.com/preibischLab/multiview-reconstruction).

This is the repository for the docker image located at:

Description

This docker image contains FIJI and the Multiview Reconstruction plugin developped by Stephan Preibisch (@StephanPreibisch). It has been built for CPU and GPU computing (Nvidia only) and works using the SnakeMake automated pipeline built by Stephan Preibisch (@StephanPreibisch) and Christopher Schmied (@schmiedc) (https://github.com/mpicbg-scicomp/snakemake-workflows).

The CUDA Library has been built under Ubuntu 16.04 for CUDA 9 compatibility and is integrated in the image, along with all dependencies.

Usage

This docker image uses a mounted volume and launches N jobs as a parameter.

Install Docker and Nvidia-Docker

Follow the installation steps to install docker on your platform: https://docs.docker.com/engine/installation/

GPU computing requirements

If you want to perform GPU computing for the deconvolution step you need to install the CUDA driver on your machine (refer the the Nvidia CUDA install page: https://developer.nvidia.com/cuda-downloads) Once this is working you need to install nvidia-docker as explained here: https://github.com/NVIDIA/nvidia-docker

Prepare the data

Launch the computation CPU

There are 3 flags that needs to be set, 2 volumes and 1 environnment variable:

  • FULL_PATH_TO_YOUR_DATA: data volume
  • FULL_PATH_TO_A_LOG_FOLDER: log volume
  • NUMBER_OF_CONCURRENT_JOBS: number of concurent jobs
sudo docker run \
-v FULL_PATH_TO_YOUR_DATA:/data/ \
-v FULL_PATH_TO_A_LOG_FOLDER:/opt/multiview/snakemake-workflows/spim_registration/timelapse/.snakemake/log/ \
-e 'NUM_JOBS=NUMBER_OF_CONCURRENT_JOBS' \
xqua/multiview_reconstruction

example:

sudo docker run \
-v /scratch/flyEmbryo:/data/ \
-v /var/log/snakemake:/opt/multiview/snakemake-workflows/spim_registration/timelapse/.snakemake/log/ \
-e 'NUM_JOBS=8' \
xqua/multiview_reconstruction

GPU enabled

You need to replace the docker command with the nvidia-docker tool as such:

sudo nvidia-docker run \
-v FULL_PATH_TO_YOUR_DATA:/data/ \
-v FULL_PATH_TO_A_LOG_FOLDER:/opt/multiview/snakemake-workflows/spim_registration/timelapse/.snakemake/log/ \
-e 'NUM_JOBS=NUMBER_OF_CONCURRENT_JOBS' \
xqua/multiview_reconstruction

Important notes

What if it crashes ?

The beauty of Docker is that as long as you mount the same volume, it will keep the computation state (thanks to snakemake). Thus, by relaunching the same command it will continue where it has left. If some timepoints crash in the process, just relaunch the docker run and it will restart where it failed.

Using your own config.yaml

If you are using your own config.yaml there are 3 lines that you must change for as shown below it to work with the docker container:

# current working Fiji
fiji-app: "/opt/multiview/2015-06-30_Fiji.app/ImageJ-linux64",
# bean shell scripts and Snakefile
bsh_directory: "/opt/multiview/snakemake-workflows/spim_registration/timelapse/",
# Directory that contains the cuda libraries
directory_cuda: "/opt/multiview/CUDA/lib",

Deconvolution missing file error:

If you see this error trying to launch a deconvolution:

MissingInputException in line 377 of /opt/multiview/snakemake-workflows/spim_registration/timelapse/Snakefile:
Missing input files for rule deconvolution:
dataset_one_merge.xml_timelapse_transform

You need to either activate the external_transfrom to downsample 2x (or more) your dataset:

  external_trafo_switch: "_external_trafo",

Or add an empty file in your data repository: touch /PATH/TO/DATA/dataset_one_merge.xml_timelapse_transform

Deconvolution GPU Error CUFFT_ALLOC_FAILED

If you get a CUFFT_ALLOC_FAILED error, it means your GPU is running out of memory. The probable reason for this is that the block size you defined for the deconvolution is too big to be handle by your card. It is easily fixed by changing the block_size parameter in the config.yaml: Valid values (that I know off) are:

  • "in 64x64x64 blocks"
  • "in 128x128x128 blocks"
  • "in 256x256x256 blocks"
  • "in 512x512x512 blocks"
compute: '"in 256x256x256 blocks"',