MCMCChainsStorage.jl: Storing Your Chains on Disk

The MCMCChainsStorage.jl package provides options for storing your MCMCChains.jl chains on disk without using serialization. Serialization is not suitable for long-term storage; or for sharing your chains to colleagues with different operating systems, Julia versions, or even without Julia. MCMCChainsStorage.jl solves these problems.

Currently only storage in HDF5 file formats is supported, but other storage options may be added in the future.

MCMCChainsStorage.jl is in the general Julia registry. That means all you need to do to install it is to start Julia, activate your desired environment, enter the package management context (type ]), and issue the command

pkg> add MCMCChainsStorage

The MCMCChainsStorage package depends on the MCMCChains and the HDF5 packages. If you do not have these packages installed on your system, installing MCMCChainsStorage will install them automatically.

The packages provides methods for Base.read and Base.write that read an MCMCChains object from or write it to HDF5 storage:

using HDF5
using MCMCChains
using MCMCChainsStorage

# Construct a chain and write it out...
chain = Chains(randn(500, 2, 4), [:a, :b])
h5open("an_hdf5_file.h5", "w") do f
  write(f, chain)
end

# ...and we can get it back
chain = h5open("an_hdf5_file.h5", "r") do f
  read(f, Chains)
end

Reading and writing preserves the sections of the chain, so if you have metadata stored in, for example, the "internals" section, it will be written out and read back properly.

It is also possible to write a chain to a group in a larger HDF5 file:

h5open("another_hdf5_file.h5", "w") do f
  g = create_group(f, "a_chain")
  write(g, chain)
end

chain = h5open("another_hdf5_file.h5", "r") do f
  read(f["a_chain"], Chains)
end

The package provides one additional utility function: if your model returns a named tuple of generated quantities, then you can call

model = ... # Construct a Turing model
trace = Turing.sample(model, ...) # Construct a chain, of shape `(nsamp, nparams, nchain)`
full_trace = append_generated_quantities(trace, Turing.generated_quantities(model, trace))

to obtain an MCMCChains object that incorporates both the original samples and the generated quantities.

The chain is stored with one group for each section (parameters, internals, etc). Each "name" within the section is stored as a separate HDF5 data set, so arrays in the chain will be placed in data sets named "x[1]", "x[2]", etc. Compression is enabled by default; currently there is no way to change this default, but why would you want to? An advantage of this format is that generic tools like h5ls will produce a reasonable description of the chain; and it is straightforward to reconstruct the chain without too much code in any language that can interface with the HDF5 storage format.