Releases: rs-station/careless
v0.2.7 Update CCpred
This version
- updates ccpred to add the
--overall
flag which pools results from multiple predictions files - updates behavior to skip normalization for metadata columns with zero variance
v0.2.6 Add stats submodule
This version just adds a new stats submodule with several command line entry points defined in setup.py
. Specifically, this version adds
careless.cchalf
for computing CC1/2 from*_xval_#.mtz
filescareless.ccanom
for computing CCanom from*_xval_#.mtz
filescareless.ccpred
for computing CCpred from (potentially multiple)*_predictions_#.mtz
filescareless.rsplit
for computing Rsplit from*_xval_#.mtz
filescareless.completeness
for computing completeness from mergedmtz
files
New Home
careless
is now part of the rs-station org!
v0.2.4 Fix CUDA on Some Systems
The previous version of careless
would pull the latest release of TensorFlow which was often a release candidate. This sometimes breaks compatibility with CUDA on making it impossible to train on GPU. This release restricts TF to major versions which are more stable.
v0.2.3 Make Graph Execution the Default Again
This release reverts an earlier change which made the slower eager execution the default during training and inadvertently flipped the semantics of the --run-eagerly
flag. There are no new features or CLI changes to note. Users will notice much faster training however.
v0.2.2 Fix numerical instability
This version fixes the numerical instability issue introduced in v0.2.1 (#61). This version adds a new command line parameter --epsilon
which controls the minimum width of variational distributions. The default value of 1e-7
should be fine for most any data.
v0.2.1
New Feature: Saving and loading weights in v0.2.1
- Output neural network weights to {output_prefix}_scale
- Output structure factor distribution weights {output_prefix}_structure_factor
- Load previous structure factors with
--structure-factor-file=
- Load previous neural network weights with
--scale-file=
- Freeze structure factors with
--freeze-structure-factors
- Freeze neural network scales with
--freeze-scales
Minor changes
- Switch all bijectors to
Exp
rather thanSoftplus
- Print a version number at run time
- Report the log likelihood of test sets
- Initialize the variational distribution to match the width of the prior
- Add CLI flag to control this initial width
- Add a tensorflow implementation of the fourth moment of truncated normals
Bug fixes
- Fixed a bug in metric reporting which lead to incorrect KL Divergences in the history
- Fix a bug whereby the second moment of the predicted intensities was inaccurate
v0.2.0
This version is a complete rewrite relative to the previous release. There is very little that this release doesn't touch. It is not safe to assume the CLI arguments are the same. Many features from the previous version (like reference priors) have been removed. If you rely on these for your work, please
- continue to use the previous version
- file an issue
I will make every effort to reintroduce the feature.
v0.1.7 PyPI release
This is the latest version of the original careless
implementation. This release is being made for debugging purposes.
v0.1.5 Original careless implementation
This is the stable version of the original careless
implementation with the original command line interface.
All subsequent releases are going to be based on a completely new API and will initially support a drastically reduced feature set. This rewrite is being conducted to
- support more advanced crossvalidation paradigms
- make it easier to support add multiple GPU support