v2.8.5

• Fix a bug where passing a config to train did not override output URL
• Raise a warning when local and remote sources clash
• Fix a bug where automated download does not work for cross validation

v2.8.2

• Enable fit_powerlaw to pull models from wandb directly.

v2.8.1

• Enable pooling of different token files
• Update documentation for associated scripts

v2.8.0

• Can now compare empirical token distributions across runs

v2.7.2

• Fix a bug where casing and sequence splitting did not occur correctly during tokenisation

v2.7.1

• Fix a bug where tokenise_bio did not load dependencies correctly
• Add sequence breaking functionality to tokenise_bio -b for splitting long seqs that may cause memory issues (eg chr1)
• Add sequence casing functionality to tokenise_bio -c for changing data input during tokeniser training to upper or lower case

v2.6.3

• Can now sweep on a subset of data by using the --partition_percent option

v2.5.0

• Fix a bug where train was not functioning correctly (class label encodings)

v2.4.4

• Fix a bug where csv files were truncated if the input sequence is too long
• Fix a bug where embeddings were not generated correctly in create_embedding_bio_sp.py

v2.4.3

• Add support for reverse complementing some non-standard nucleotides
• Fix bug in k-merisation process where a non-existing tokeniser file was parsed