Tags

### Changes

- subclass layer for frequency reduction in 1DConv model for serialization. Fixes serialization of 1D Conv models.
- save data of all trials as csv in hpsearch

### Changes

- Buffer size back to 1000. OOM errors on euler with 20000.

### Changes

- __Breaking__: rename module hp_search to hpsearch
- Potential fix for training routine that lead to incomplete epochs because not enough batches could be extracted.
- increase Shuffle buffer size to 20000
- add initializers as parameters to model architectures.
  - two new parameter in orcai_parameter["model"]
    - "conv_initializer": default "glorot_uniform" (the keras default), "he_normal" might be better suited for layers with ReLU activations
    - "lstm_initializer": default "glorot_uniform" (the keras default), which should be fine for LSTMs. Included for completeness.
    - if values are not in orcai_parameters.json, the default values are used.
- __Breaking__: removed maskedAUC reporting. Class is still available but unused.
  - parameter orcai_parameter["monitor"] can only be "val_MBA"
- minor refactoring

### Changes

- fix ResNet1DConv for keras 3

### Changes

- __Breaking__: switch back to keras 3. keras_tuner is not compatible with tf_keras.

### Changes

- fix path passed to to callbacks

### Changes

- __Breaking__: remove `--save_best_model` flag from `hpsearch` command
- update pipeline.
- minor prettifications.

## [0.19.0] - 2025-04-28

### Changes

- __Breaking__: filter duplicate snippets. These duplicates stem from randomly sampling.
- Minor reformatting of code and help

### Changes

- __Breaking__: downgrade keras to keras 2 (tf-keras)
- __Breaking__: move JSONEncoderExt to seperate module to optimize loading
- __Breaking__: new parameter in orcai_parameter: "ReduceLROnPlateau_min_learning_rate": 1e-06
  - add to orcai_parameters.json to unbreak.

### Notes

- keras 2 trains better models. Reason unknown.
- tried to revert to Sebs DataLoader function, inheriting from keras.utils.Sequence. No difference detected except worse performance.

### Changes

- cleanup imports
- refactor error-prone sorting of annotations
- __Breaking__: new parameter in orcai_parameter: "EarlyStopping_patience": 10, "ReduceLROnPlateau_patience": 3, "ReduceLROnPlateau_factor": 0.5
  - add to orcai_parameters.json to unbreak.

### Changes

- __Breaking__: in train, load model instead of weights when restarting
  - the argument `load_weights` is now `load_model`
  - the flag option in the cli `--load_weights` is now `--load_model`
- __Breaking__: implement class weights.
  - orcai_parameters.json now has a model.call_weights key.
    - to unbreak add `"call_weights": null` to the model section of orcai_parameters.json
  - Implemented "three" methods for calculating call weights:
    - `"balanced"` is the same heuristic as is used in sklearn (= total / (n_calls \* count)),
    - `"max"` is 1/count \* total,
    - `"uniform"` is all ones and equal to None.
    - Use `null` (in Json, `None` in python) to disable class weights.
- __Breaking__: switch to class based metrics
- __Breaking__: implement AUC ROC metric
- __Breaking__: making the choice of metric to monitor for callbacks an option in orcai_parameters.json
  - to unbreak add `"monitor": "val_MBA"` to the model section of orcai_parameters.json
- make ReduceLROnPlateau callback patience == model_parameters['patience'] // __3__
- define metrics in architectures.py
- new arg parameter to overwrite default orcai parameter on project init

### Changes

- make ReduceLROnPlateau callback patience == model_parameters['patience'] // 2

### Changes

- save best hyperparameters
- flag to save model when running hp_search

### Changes

- __Breaking__ save all_snippets.csv.gz in tvt_data not recording data
- enable changing of initial learning rate with orcai_parameters.json
- switch formatter and format imports

### Changes

- fix: set backend explicitly

### Changes

- set backend explicitly
- use class for loss function and set from_logits to False (should be equivalent)
- use __version__ attribute for version number

### Changes

- update zarr to zarr>=3,<4. Is backwards compatible in terms of loading files

### Changes

- fix for zarr 2.8.15: fix numcodecs to < 0.16 (until zarr is updated or we switch to zarr 3.0)

### Changes

- fix dataset compression option in cli
- fallback dataset_shape
- fix keras imports

### Changes

- add option for dataset compression to create_tvt_data, train, test and hp_search
- save test results to subdir by default
- improve saved tables in test