search

rueberger / division_detection

Cell-division classifier
MIT License
3 stars 2 forks source link

readme

Division Detection

This repository provides the cell-division classifier used in the paper:

In toto imaging and reconstruction of post-implantation mouse development at the single-cell level
Katie McDole, Léo Guignard, Fernando Amat, Andrew Berger, Grégoire Malandain, Loïc A. Royer, Srinivas C. Turaga, Kristin Branson, and Philipp J. Keller
Cell, 2018
https://www.sciencedirect.com/science/article/pii/S0092867418312431

Installation

Use docker to run the classifier. Images are available at https://hub.docker.com/r/rueberger/division_detection/. Docker will pull the latest image at runtime.

Prereqs

Data volume

Note: in future releases a script for pulling the data and building the volume will be provided

The base images do not include any data. To use them, put the data you wish into a Docker volume and attach it at container runtime.

Please refer to online docker documentation on volume creation and usage.

Set the name of your volume to the environment variable DATA_VOL such as in the following:

export DATA_VOL=your_volume_name

It is used to mount the data volume at container runtime.

Building the container

The containers may be built for local use by executing

make cpu_image

and/or

make gpu_image

from the root repo directory.

Note pre-built images are available at the tag rueberger/division_detection:latest for the CPU image and rueberger/division_detection:latest_gpu for the GPU image.

Usage

Running the image

The following commands download the image if necessary, start them, and drop you into a bash shell. $DATA_VOL is mounted at /data.

CPU

docker run --name div_det -it --mount type=volume,source=$DATA_VOL,destination=/data rueberger/division_detection:latest

GPU

New nvidia-docker runtime

docker run --runtime=nvidia --name div_det -it --mount type=volume,source=$DATA_VOL,destination=/data rueberger/division_detection:latest_gpu

Old nvidia-docker

nvidia-docker run --name div_det -it --mount type=volume,source=$DATA_VOL,destination=/data rueberger/division_detection:latest_gpu

Running the included example

As script for pulling example data is included. Here is how to use it:

  1. Run the download script:

    source division_detection/scripts/fetch_data.sh

    This downloads the example data set into ~/data/division_detection/klb

  2. Run the container, mounting the data where expected:

    docker run --runtime=nvidia --name div_det -it --mount type=bind,source=~/data/division_detection,destination=/data rueberger/division_detection:latest_gpu python division_detection/scripts/predict.py /data --chunk_size 50 100 100

    Notes:

    • ~/ needs to expanded to the absolute path for source=
    • Notice that here we use a bind-mount instead of putting the data in a proper volume as discussed above=. Doesn't matter to the container.
    • This took 8500MB gpu memory
    • A single gpu is used by default. More may be used by appending --allowed_gpus 0 1 2 3, for instance, which would be used if 4 GPUs were available. The numbers following --allowed_gpus are the $CUDA_VISIBLE_DEVICES id of the GPUs

Generating predictions

A script is provided for running predictions at scripts/predict.py. For now, run it with python predict.py - usage instructions will be printed.

Later releases will automatically add a system wide alias for this.

The most basic usage is

python predict.py /data/path

The directory /data/path should contain only a directory named klb which contains the klb files you wish to process.

Training

This release officially supports prediction only. While all code used to train the model is included and has inline docstrings, it has not been tested and high level docs are not available at this time.

~Training will be supported in future releases. We aim to release a fully featured package for machine learning on spatiotemporal bioimagery in 2018.~

As of July 2020, I am no longer at Janelia and there are no plans to introduce support for training to this package. That being said, the training is relatively straightforward and should be easily reproducible by the experienced ML practitioner - I am happy to correspond via email.

Disclaimer

THIS SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.