Reference / Old checkpoints

Old checkpoints¶

This page lists how a checkpoint saved before a feature existed loads and behaves. load_checkpoint_config() fills in each missing setting, and load_checkpoint() rebuilds the model.

Missing settings¶

A config.yaml without one of these settings loads with the value in the second column. The model then predicts and evaluates as it did when it was saved.

Setting	Loads as	Training
`data.normalize_spectra`	`false`, with a warning; see below	Allowed
`data.precision_encoding`	`false`	Rejected
`model.label_aware_uncertainty`, `multi_head_model.label_aware_uncertainty`	`false`	Rejected
`multi_head_model.residual_stream`	`false`	Allowed

If one of these values is wrong for a checkpoint, add the setting to its config.yaml.

Spectrum normalization¶

A configuration without data.normalize_spectra loads as true when it comes from the dorothy_update branch and does not set data.use_dense_loading: true. A configuration comes from that branch when it has a multi_head_model.head_dropout entry or 20 predicted parameters. Every other configuration without the setting loads as false. Both cases print a warning with the value used.

Precision encoding¶

A checkpoint saved before the precision encoding reads \(\sigma\) in the error channel of its inputs, not \(\rho = \ln(1 + 1/\sigma)\).

Prediction and evaluation give the spectra \(\sigma\), as in its training.
Saliency derivatives with respect to the error channel are derivatives with respect to \(\sigma\).
dorothy train --resume rejects the checkpoint.

Label-aware uncertainty¶

A checkpoint saved before label_aware_uncertainty has the additive model.

	Additive model (old)	Label-aware model
Output layer	One linear layer predicts \(\mu\) and \(\ln s\)	A label decoder predicts \(\mu\), an error decoder \(\ln s\)
What \(s\) measures	The scatter beyond the reported label error \(\sigma_y\)	The scatter around the label source’s measurement, label noise included
Loss variance	\(\sigma_y^2 + s^2\)	\(s^2\)
z-score denominator in evaluation	\(\sqrt{\sigma_y^2 + s^2}\); `HeadEvaluation.z_label_errors` holds \(\sigma_y\)	\(s\)
`dorothy info`	`additive`	`label-aware`

Prediction reports \(s\), which excludes the label noise.
The default embedding of an old MLP (MLP.get_embeddings, AnomalyDetector) is its last hidden layer.
The checkpoint cannot be trained further: training rejects label_aware_uncertainty: false, and --resume needs the same layers.

Residual stream¶

A multi-head checkpoint saved before residual_stream has plain blocks. Its trunk has a 128-number output. The error decoder of each head concatenates the log label errors to its input. dorothy train --resume needs the same residual_stream value in the new configuration.

Calibration factors¶

A checkpoint without calibration.json reports the raw predicted scatter \(s\). dorothy info prints Calibration: none.

Head names¶

A multi-head model with one label source used to name its head default. Its weights load into a head named after the first label source. Select the head by that name: --label-source apogee, not --label-source default.

Several label sources¶

Checkpoint	Loads as	Evaluation
`model` with several label sources	One head, the first label source, the one it trained on	Works
`multi_head_model`, one survey and several label sources	One head, the first label source, as saved	`DataSplit.load` raises `ValueError`; retrain the model

The second kind trained only its first label source, and its split indexes a data layout that this version no longer builds.

Data split¶

File	Behaviour
`data_split.pkl` of a dense layout without `index_space: rows`	Its indices point into the filtered arrays of the old loaders. `DataSplit.load` converts them to rows of the loaded arrays with `TrainingData.legacy_filtered_rows`
`data_split.pkl` of sparse storage without `index_space: rows`	Its indices are already catalogue rows and load unchanged
No `data_split.pkl`	`dorothy evaluate` recreates the split from the configured `seed`

Other settings¶

Setting	Behaviour
`training.learning_rate`	Becomes `training.scheduler.max_lr` for the `none`, `reduce_on_plateau` and `cosine_annealing` schedules; ignored with a warning for the others
A setting that this version does not have	Ignored with a warning
`multi_head_model.head_dropout` for a head the model does not have	Ignored with a warning

Weight files¶

File	Behaviour
`final_model.pth` saved before it held the last epoch	Holds the best weights, like `best_model.pth`
A directory without `config.yaml`, such as an old epoch checkpoint	See Weights without a configuration

Weights without a configuration¶

load_checkpoint rebuilds a directory without config.yaml from the weights alone.

The model is an MLP with the additive output layer. Its layer sizes and normalization come from the weights, and its activation is the default, GELU.
Label-aware weights raise ValueError: they need the configuration.
The parameter names come from normalizer.pkl, or else are the first names of STELLAR_PARAMETERS.
The scatter floor is 0.01, and there are no calibration factors.
dorothy predict normalizes the spectra and does not apply the precision encoding.