Minkowski Engine¶
Minkowski Engine is an auto-differentiation library for sparse tensors. It supports all standard neural network layers such as convolution, pooling, unpooling, and broadcasting operations for sparse tensors.
Minkowski Engine is available as an Apptainer container on Apocrita.
Usage¶
Minkowski Engine requires a suite of supporting tools to be installed, so for reproducibility, we provide all of the tools in a container along with Minkowski Engine.
To run the default version of Minkowski Engine, simply load the
minkowski_engine module:
module load minkowski_engine
Calling python after loading the minkowski_engine module will invoke the
installed version of Python inside the container. Additionally, this entry
point will automatically use any requested GPU cards.
Example jobs¶
CUDA support
CUDA has been installed inside the Minkowski Engine container so you do not need to load a CUDA module before running your analysis.
Serial job¶
Here is an example job running on 1 core and 5GB of memory:
#!/bin/bash
#$ -cwd
#$ -j y
#$ -pe smp 1
#$ -l h_rt=1:0:0
#$ -l h_vmem=5G
module load minkowski_engine
python train_network.py
GPU job¶
Here is an example job running on 8 cores and 1 GPU:
#!/bin/bash
#$ -cwd
#$ -j y
#$ -pe smp 12
#$ -l h_rt=240:0:0
#$ -l h_vmem=7.5G
#$ -l gpu=1
module load minkowski_engine
python train_network.py