emaballarin / RDDL

Rapid Deployment Deep Learning

RDDLeaner

Rapid Deployment Deep Learning, made leaner!
A Conda-based, templatable, extendable, R&D-focused infrastructure scaffold aiming at fast iterability and cross-unix robustness.

---

Some relevant specifications:

• The entire thing is Python-based.
• PyTorch and JAX (plus some respective extensions) are the supported deep learning frameworks. Some inter-operability is also guaranteed.
• The hardware stack of reference is Intel CPU and Nvidia GPU.
• A UNIX (or UNIX-like at least) system is a prerequisite.

---

Essential how-to:

1. Ensure that you have not an already-existing Anaconda environment called RDDL (or, for our purposes, whichever name you want to call the environment to-be-created). Otherwise, it will be overwritten without further notice!

2. Ensure that you are running bash or zsh as shell, since they are the only supported (no fish, sorry... although I use it as my default shell in most of the cases!), and verify that the conda (or mamba, if you want to use it instead) command is available and working correctly.

3. Clone the GitHub repository via git clone https://github.com/emaballarin/RDDL.git.

4. Open the init.sh and portablecuda.sh script files, and check if the environment variables, paths and versions are OK for your system. Please, note that this set of scripts is nothing more than a (useful, though) tool for personal use, shared in the hope it might be useful to someone else too. It is not a full-fledged, stable product. Some rough edges may still be present. In this case, however, minimal modifications should be enough!

5. Let the script prepare and install everything, via ./init.sh . Note that an active Internet connection is still required for everything to work as expected.

---

Experimental fish shell support

After having followed the Essential how-to right above, in order to properly use this specific conda setup from the fish shell, without sacrificing otherwise-working conda integration, you need to:

1. Ensure that you are starting from a clean (i.e. non-conda-integrated fish setup), undoing any step previously followed in this direction (note: any functionality will be restored by following such procedure!).

2. Add the line source "$HOME/anaconda3/etc/fish/conf.d/conda.fish" to one (any valid) fish script sourced on shell startup.

3. Install the bass plugin for fish.

4. Add the aliases alias conda_activate="bass source activate" and alias conda_deactivate="bass source deactivate" to one (any valid) fish script sourced on shell startup.

5. Use such aliases to activate (at least) the RDDL environment. Note that the usual conda command (from fish) currently fails to properly source bash-compatible scripts located in $CONDA_PREFIX/etc/conda/activate.d/ and $CONDA_PREFIX/etc/conda/deactivate.d/ which RDDL needs!