• Cropping1D Support
• User specified stop layers (undocumented)
• Support for MultiHeadAttention layers

• Refinements to segmenting
• Segment starting with see-through layers followed by Conv1Ds with valid padding are kept in one segment
• Layers are duplicated with from_config and get_config
• Generalized pooling layers and added ability to add custom pooling layers

• Runs for batch size 1
• Multi-input layers that had the same input twice (e.g. Add()([x,x])) would not run, fixed this
• Support for newer versions of tensorflow which changed sub-models class from keras to tf.keras (in flatten_model)
• Stop layers were traversed redundantly

• Sequences for benchmarking in notebooks dir and a notebook to process the sequence
• Benchmarking notebooks
• Notebook to time Basset conv and fc separately
• Ability to specify custom mutations
• For each mutation, models only run on input sequences for which character is different from mutation. As a result, each batch usually has a different size. This is slow for the first few batches as it entails a one-time cost.
• Lots of documentation and a logo!

• Models updated:
  • Activation added to Basset
  • Num output for Basset and Factorized Basset
  • For BPNet, only one channel output and one counts instead of two

• FastISM object would keep intermediate outputs of a batch even after it was used, as a result it would occupy extra memory. Get rid of such objects now through a cleanup() function. This has stopped GPU Resource errors that popped up after running a few batches

• Support for multi-input models where alternate input does not merge with primary sequence input before a stop layer.
• Support for layers that dependend on exact order of inputs, e.g. Subtract and Concat.

• Support for recursively defined networks with 3 test cases
• This Changelog file.

• BPNet test cases atol changed to 1e-5 so they pass deterministically

• First PyPI release and tagged version
• Tested and working on non-recursively defined single-input, single and multi-output architectures
• Tested and working on arcitectures with skip connections

The format is based on Keep a Changelog.