Philipp Eller

Issues running in fp16

4

Even though the fp16 unit tests run successfully, I cannot get clustering to work on my sample in fp16 mode. When is use my data `.astype(np.float32)` the clustering output looks...

When using the bridge sampler on samples generated with non-integer weights, it crashes (stack trace below). It is in particular the `nsamples=Int(sum(second_batch.weight))` here: https://github.com/bat/BAT.jl/blob/e4fcc532e5eb1ad192374d062ae50328d4ddfa21/src/integration/bridge_sampling_integration.jl#L134 I suggest for instance to use...

Sampling with `nchains=1` results in an AssertionError: ``` JULIA: AssertionError: length(chains) == nchains Stacktrace: [1] mcmc_init!(rng::Random123.Philox4x{UInt64, 10}, algorithm::BAT.MetropolisHastings{BAT.MvTDistProposal, BAT.RepetitionWeighting{Int64}, BAT.AdaptiveMHTuning}, density::BAT.PosteriorDensity{Float64, Float64, BAT.ReshapedDensity{BAT.TransformedDensity{BAT.DensityWithShape{BAT.LFDensity{PyObject}, ScalarShape{Real}}, BAT.DistributionTransform{BAT.StandardUvNormal{Float64}, Distributions.Uniform{Float64}, ScalarShape{Real}, ScalarShape{Real}}, BAT.TDNoCorr, ScalarShape{Real}},...

For splitting samples in two (and also for sub-sets in the covariance estimates), we should split samples in chucks of correlation lengths

Replacement of empty subsets

1

In https://github.com/bat/BAT.jl/blob/master/src/integration/ahmi/harmonic_mean_integration.jl#L453 integral estimates that are `nan` or `inf` (presumably when the hyper-rectangles are empty) are replaced by the estimate from the full sample. This should not be the case

adding stage for external osc calc

1

This is a new PISA stage, that allows one to use an external oscillation probability calculation. We use this in some (non IceCube) work we do. But still would be...

Osc next2

5

Hey All, This will be major work to make this PR, but this is absolutely critical! Please, everyone who worked on this branch start looking into this.

Currently there is (unsafe) ad hoc caching implemented for the KDE stage (PR #689). We should make this safer, and on a more global level. Ideas: I think cleanest would...

pretty print

1

Change to python 3 print function instead of carriage return