Configuration¶

Overview¶

AnomalyMatch uses DotMap for dot-notation configuration access. The default configuration is defined in get_default_cfg.

Loading Configuration¶

import anomaly_match as am

cfg = am.get_default_cfg()

The bundled-test-data defaults differ

Until you set cfg.data_dir, it points at the test data shipped with the repository, and get_default_cfg then overrides num_train_iter to 32 and image_size to [64, 64] so start_ui() works out of the box. The defaults in the tables below are what you get once data_dir is your own.

Paths¶

Key	Description
`cfg.save_dir`	Path to store the trained model output
`cfg.data_dir`	Location of data to search, should be a folder of `.zarr`, images or Cutana-compatible catalogues (`.csv , *.parquet`)
`cfg.label_file`	CSV mapping annotated images to labels
`cfg.metadata_file`	Optional CSV containing metadata for images (automatically merged with labelled data)
`cfg.prediction_search_dir`	Path where data to be predicted is stored

Training Parameters¶

Key	Default	Description
`cfg.batch_size`	16	Training batch size
`cfg.num_train_iter`	200	Number of training iterations
`cfg.lr`	0.0075	Learning rate
`cfg.weight_decay`	7.5e-4	L2 regularization parameter
`cfg.opt`	`"SGD"`	Optimizer type
`cfg.momentum`	0.9	SGD momentum
`cfg.bn_momentum`	`1.0 - ema_m`	Batch normalization momentum
`cfg.eval_batch_size`	500	Batch size for evaluation
`cfg.num_eval_iter`	-1	Evaluation frequency (-1 = no evaluation)
`cfg.pretrained`	`True`	Use pretrained backbone
`cfg.net`	`"efficientnet-lite0"`	Backbone network architecture
`cfg.N_to_load`	1000	Number of unlabeled images loaded into the training dataset at once
`cfg.num_workers`	4	Parallel workers for data loading
`cfg.seed`	42	Seed for the unlabelled-pool sample and model init
`cfg.gpu`	0	GPU index to train and predict on
`cfg.test_ratio`	0.0	Proportion for evaluation (> 0 shows AUROC/AUPRC curves)
`cfg.log_level`	`"INFO"`	Controls verbosity of training/session logs

FixMatch Parameters¶

Key	Default	Description
`cfg.ema_m`	0.99	Exponential moving average momentum
`cfg.hard_label`	`True`	Use hard labels for unlabelled data
`cfg.temperature`	0.5	Temperature for softmax
`cfg.ulb_loss_ratio`	1.0	Weight of the unlabeled loss
`cfg.p_cutoff`	0.95	Confidence threshold for pseudo-labeling
`cfg.uratio`	5	Ratio of unlabeled to labeled data in each batch

Prediction Parameters¶

Key	Default	Description
`cfg.N_batch_prediction`	`None`	Batch size for prediction; auto-estimated from GPU memory when `None`
`cfg.prediction_search_dir`	`None`	Directory containing data to predict
`cfg.prediction_db_dir`	`None`	Pin `predictions.db` to a directory; by default it relocates to local scratch when the session directory is networked
`cfg.compile_model`	`False`	`torch.compile` the eval model; worth enabling only for backbones larger than `efficientnet-lite0`

Normalisation¶

Normalisation can be selected in the UI via a dropdown, or set in code:

cfg.normalisation.normalisation_method = am.NormalisationMethod.ZSCALE

The training setup screen remembers your last choice

Starting a training run records that screen's normalisation settings in ~/.config/anomalymatch/ui_state.json, and they become the defaults the next time it opens — taking precedence over the values set here, so your last run's settings survive a kernel restart. Delete that file (or its normalisation_settings entry) to go back to the values from your notebook. The channel-combination matrix is only restored when the selected source exposes the same input channels it was recorded against.

Available methods:

Method	Description
`CONVERSION_ONLY`	No normalisation
`LOG`	Logarithmic normalisation
`ZSCALE`	Linear normalisation based on zscale min and max
`ASINH`	Asinh normalisation with configurable scale and percentile clipping

Normalisation Config¶

Key	Default	Description
`cfg.normalisation.image_size`	`[150, 150]`	Target image size (below 96x96 not recommended)
`cfg.normalisation.n_output_channels`	3	Number of output channels (auto-inferred from `fits_extension`)
`cfg.normalisation.output_dtype`	`np.uint8`	Output data type
`cfg.normalisation.normalisation_method`	`CONVERSION_ONLY`	Stretch applied while loading; also selectable in the UI dropdown
`cfg.normalisation.fits_extension`	`None`	FITS extension(s) to load: an index, a name, or a list of either — see Prediction
`cfg.normalisation.channel_combination`	`None`	`(n_output_channels, n_extensions)` matrix mixing extensions into channels; identity when `None`. An all-zero row blanks that output channel — see Blanking an output channel
`cfg.normalisation.interpolation_order`	1 (bi-linear)	0-5, matching skimage resize orders
`cfg.normalisation.norm_asinh_scale`	`[0.7, …]`	Per-channel asinh scale (`ASINH` only)
`cfg.normalisation.norm_asinh_clip`	`[99.8, …]`	Per-channel percentile clip (`ASINH` only)
`cfg.normalisation.apply_flux_conversion`	`True`	Convert pixel values to flux density in Jansky using the AB zeropoint (`MAGZERO`) from FITS headers

For the full stretch semantics see the normalisation notes in the repository.

Weights in `channel_combination`¶

Each row of channel_combination is one output channel and each column one input band. For image folders and Zarr stores the columns are the source's own channels, so a [[1/3, 1/3, 1/3]] matrix turns RGB into a single greyscale channel. The training preview and the scoring gallery show the result, decoded with your normalisation settings rather than as raw pixels. An all-zero row blanks that output channel — see Blanking an Output Channel.

Keep row weights non-negative and summing to at most 1. A row summing to more than 1 is rescaled for you, with a warning — otherwise bright pixels would exceed the output range and be silently clipped.

Negative weights need fits_extension unset

With fits_extension unset (Cutana catalogues, Zarr arrays, image folders), AnomalyMatch applies the matrix itself: negative weights are allowed but warned about, because the combined value can fall below 0 and clip to 0 on unsigned output. With fits_extension set — including Cutana catalogues that name their bands — fitsbolt applies the matrix and rejects negative weights, so validation fails with channel_combination must not have negative values.

Normalisation Consistency¶

Normalisation settings are saved in the model checkpoint during training. During prediction, they are loaded automatically — there is no need to re-specify them. Both training and Cutana streaming use fitsbolt for normalisation, guaranteeing identical results.