giffordlab/deepsea-docker repository overview

⁠deepsea-docker

This contains a Docker image for training a new DeepSEA model (v0.93⁠). If you are looking for a docker that predicts with the published DeepSEA model on 919 epigenomics features without re-training, check out our another repo⁠.

⁠Building the image

We have set up an automated build here⁠. Pushes should automatically trigger builds on the Docker hub. If you want to build a local image, you can run:

docker build -t giffordlab/deepsea-docker .

⁠Getting the image

You can get the (public) automated build image from any machine, including EC2 ones, by running:

docker pull giffordlab/deepsea-docker

Prerequisites are Docker and having the NVIDIA 346.46 driver installed (the default for CUDA 7.0; see the upstream image documentation).

⁠Using the image

Run the image with:

docker run -it --rm \
--device /dev/nvidia0 \
--device /dev/nvidia1 \
--device /dev/nvidia2 \
--device /dev/nvidia-uvm \
--device /dev/nvidiactl \
giffordlab/deepsea-docker

This will launch the training script (main.lua) with a subsampled input file that's part of the image (train.mat). Their full training input file is around 4 GB and therefore is not on Github.

Check the main.lua source for command-line options. Right now, the only way to change the input is to use docker run -v to replace /root/deepsea/input with a physical directory containing files train.mat and valid.mat (formats TBD).

To capture the output, run commands interactively or mount /root/deepsea/results to a physical directory (you can use the -save option in main.lua for more flexibility).

⁠License

The original code is from the Troyanskaya lab, see here⁠.