v1.10.0

Added

• Added n_jobs and chunksize arguments to MuminDataset, to enable customisation
  of these.

Changed

• Lowered the default value of chunksize from 50 to 10, which also lowers the memory
  requirements when processing articles and images, as fewer of these are kept in
  memory at a time.
• Now stores all images as uint8 NumPy arrays rather than int64, reducing memory
  usage of images significantly.

v1.9.0

Added

• Added checkpoint after rehydration. This means that if compilation fails for whatever
  reason after this point, the next compilation will resume after the rehydration
  process.
• Added some more unit tests.

Fixed

• Fixed bug on Windows where some tweet IDs were negative.
• Fixed another bug on Windows where the timeout decorator did not work, due to the use
  of signals, which are not available on Windows machines.
• Fixed bug on MacOS causing Python to crash during parallel extraction of articles and
  images.

Changed

• Refactored repository to use the more modern pyproject.toml with poetry.

v1.8.0

Changed

• Now allows instantiation of MuminDataset without having any Twitter bearer
  token, neither as an explicit argument nor as an environment variable, which
  is useful for pre-compiled datasets. If the dataset needs to be compiled then
  a RuntimeError will be raised when calling the compile method.

v1.7.0

Added

• Now allows setting twitter_bearer_token=None in the constructor of
  MuminDataset, which uses the environment variable TWITTER_API_KEY
  instead, which can be stored in a separate .env file. This is now the
  default value of twitter_bearer_token.

Changed

• Replaced DataFrame.append calls with pd.concat, as the former is
  deprecated and will be removed from pandas in the future.

v1.6.2

Fixed

• Now removes claims that are only connected to deleted tweets when calling
  to_dgl. This previously caused a bug that was due to a mismatch between
  nodes in the dataset (which includes deleted ones) and nodes in the DGL graph
  (which does not contain the deleted ones).

v1.6.1

Fixed

• Now correctly catches JSONDecodeError during rehydration.

v1.6.0

• Changed the download link from Git-LFS to the official data.bris data
  repository, with URI https://doi.org/10.5523/bris.23yv276we2mll25fjakkfim2ml.

v1.5.0

Changed

• Now using dicts rather than Series in to_dgl. This improved the wall time
  from 1.5 hours to 2 seconds!

Fixed

• There was a bug in the call to dgl.data.utils.load_graphs causing
  load_dgl_graph to fail. This is fixed now.

v1.4.1

Changed

• Now only saves dataset at the end of add_embeddings if any embeddings were
  added.

v1.4.0

Added

• The to_dgl method is now being parallelised, speeding export up
  significantly.
• Added convenience functions save_dgl_graph and load_dgl_graph, which
  stores the Boolean train/val/test masks as unsigned 8-bit integers and
  handles the conversion. Using the dgl-native save_graphs and
  load_graphs causes an error, as it cannot handle Boolean tensors. These two
  convenience functions can be loaded simply as
  from mumin import save_dgl_graph, load_dgl_graph.