// the find
cerndb/dist-keras
Distributed Deep Learning, with a focus on distributed training, using Keras and Apache Spark.
dist-keras is a 2016-era library for distributed deep learning that wraps Keras and Apache Spark, implementing several async gradient descent algorithms (DOWNPOUR, EASGD, ADAG) via a parameter server architecture. It was built at CERN for training on Hadoop clusters and is aimed at researchers who want to experiment with distributed optimizer variants without rebuilding the plumbing from scratch.
Implements several non-trivial distributed optimizers (ADAG, EASGD, DOWNPOUR) that aren't bundled into standard frameworks, with actual paper references for each. The parameter server separation in the code is clean — workers, trainers, and parameter servers are distinct modules, not a tangled mess. Includes worked Jupyter notebook examples with real datasets (MNIST, ATLAS Higgs) rather than toy hello-world scripts. The ADAG optimizer reportedly outperforms vanilla DOWNPOUR on statistical convergence, and there's experimental data in the blog posts to back that claim.
Dead since July 2018 — the Python 3 compatibility issue is listed as a known bug and a TODO simultaneously, which tells you everything about the maintenance state. It targets TensorFlow 1.x and the pre-tf.keras era; none of this works with TensorFlow 2 or PyTorch, which is where every serious distributed training workload has lived for years. Horovod, PyTorch DDP, and Ray Train have made this approach obsolete: they're actively maintained, handle model parallelism, and don't require you to run an Apache Spark cluster just to distribute gradient updates. The Punchcard job server referenced in the docs appears to be an internal CERN tool with no public code, so the remote deployment feature is essentially undocumented vaporware for external users.