Cellseg-models-pytorch

Features

- Add pytorch lightning in-built `auto_lr_finder` option to `SegmentationExperiment`

<a id='changelog-0.1.18'></a>

Features

- Add Multi-scale-convolutional-attention (MSCA) module (SegNexT).
- Add TokenMixer & MetaFormer modules.

<a id='changelog-0.1.17'></a>

Features

- Add transformer modules
- Add exact, slice, and memory efficient (xformers) self attention computations
- Add transformers modules to `Decoder` modules
- Add common transformer mlp activation functions: star-relu, geglu, approximate-gelu.
- Add Linformer self-attention mechanism.
- Add support for model intialization from yaml-file in `MultiTaskUnet`.
- Add a new cross-attention long-skip module. Works with `long_skip='cross-attn'`

Refactor

- Added more verbose error messages for the abstract wrapper-modules in `modules.base_modules`
- Added more verbose error catching for xformers.ops.memory_efficient_attention.

<a id='changelog-0.1.16'></a>

Dependencies

- Bump old versions of numpy & scipy

<a id='changelog-0.1.15'></a>

Features

- Use the inferer class as input to segmentation benchmarker class

<a id='changelog-0.1.14'></a>

Performance

- Throw away some unnecessary parts of the cellpose post-proc pipeline that just brought overhead and did nothing.

Refactor

- Refactor the whole cellpose post-processing pipeline for readability.

- Refactored multiprocessing code to be reusable and moved it under `utils`.

Features

- Add exact euler integration (on CPU) for cellpose post-processing.

- added more pathos.Pool options for parallel processing. Added `ThreadPool`, `ProcessPool` & `SerialPool`
- add all the mapping methods for each Pool obj. I.e. `amap`, `imap`, `uimap` and `map`

Tests

- added tests for the multiprocessing tools.
<a id='changelog-0.1.13'></a>