astro-emulators-toolkit 0.2.1__tar.gz → 0.2.2__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {astro_emulators_toolkit-0.2.1/src/astro_emulators_toolkit.egg-info → astro_emulators_toolkit-0.2.2}/PKG-INFO +73 -1
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/README.md +72 -0
- astro_emulators_toolkit-0.2.2/src/astro_emulators_toolkit/_version.py +1 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/card.py +5 -1
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/versions.py +12 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/experimental/models/__init__.py +20 -3
- astro_emulators_toolkit-0.2.2/src/astro_emulators_toolkit/experimental/models/mhe.py +302 -0
- astro_emulators_toolkit-0.2.2/src/astro_emulators_toolkit/experimental/models/mhe_config.py +242 -0
- astro_emulators_toolkit-0.2.2/src/astro_emulators_toolkit/experimental/models/mhe_layers.py +177 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2/src/astro_emulators_toolkit.egg-info}/PKG-INFO +73 -1
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit.egg-info/SOURCES.txt +3 -0
- astro_emulators_toolkit-0.2.1/src/astro_emulators_toolkit/_version.py +0 -1
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/LICENSE +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/pyproject.toml +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/setup.cfg +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/_typing.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/_vendor/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/_vendor/get_soap_jax.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/_vendor/soap_jax/LICENSE +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/_vendor/soap_jax/README.txt +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/_vendor/soap_jax/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/_vendor/soap_jax/soap.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/artifact.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/bundle.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/extras.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/hub.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/integrity.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/loaded_state.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/metadata.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/minmax_sidecars.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/provenance.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/publish.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/readme.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/release.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/bundle/safetensors_io.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/cli.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/config/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/config/io.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/config/migrations.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/config/parsing.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/config/schema.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/data/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/data/array_dataset.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/data/loader.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/data/mapped_dataset.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/data/npy_table.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/data/preprocess.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/data/protocols.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/data/subset.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/data/toy.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/data/toy_utils.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/emulator.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/emulator_runtime.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/experimental/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/experimental/models/explicit_wavelength_mlp.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/experimental/models/mlp_2d_regression.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/experimental/models/siren.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/experimental/runtime_adapters.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/experimental/tasks/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/experimental/tasks/classification.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/inference/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/inference/compose.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/io_trees.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/models/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/models/activations.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/models/cannon.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/models/canonical_wrappers.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/models/dtypes.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/models/mlp.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/models/runtime_adapters.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/models/transformer_payne.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/models/transformer_payne_batch.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/optimizers/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/optimizers/_soap.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/presets/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/presets/cannon.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/presets/mlp.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/presets/transformer_payne.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/py.typed +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/resolver.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/spec.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/tasks/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/tasks/common.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/tasks/regression.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/training/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/training/callbacks.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/training/checkpoint.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/training/fit_backends.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/training/paths.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/training/solvers.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/training/state.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/training/trainer.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/utils/__init__.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit/utils/tree.py +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit.egg-info/dependency_links.txt +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit.egg-info/entry_points.txt +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit.egg-info/requires.txt +0 -0
- {astro_emulators_toolkit-0.2.1 → astro_emulators_toolkit-0.2.2}/src/astro_emulators_toolkit.egg-info/top_level.txt +0 -0
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: astro-emulators-toolkit
|
|
3
|
-
Version: 0.2.
|
|
3
|
+
Version: 0.2.2
|
|
4
4
|
Summary: General-purpose astrophysical emulators: spectra, isochrones, and more.
|
|
5
5
|
Author: Tomasz Rozanski
|
|
6
6
|
License-Expression: MIT
|
|
@@ -833,6 +833,78 @@ cfg = load_config("runs/demo/config.yaml")
|
|
|
833
833
|
|
|
834
834
|
Prototype model families and tasks live under `experimental/...`. They are useful for research and development but are intentionally outside the stable public contract.
|
|
835
835
|
|
|
836
|
+
The wavelength-explicit MHE (multiresolution hash encoding) families are:
|
|
837
|
+
|
|
838
|
+
- `experimental/mhe_mlp`: a shared parameter MLP followed by a wavelength
|
|
839
|
+
decoder that receives learned hash features at every hidden layer.
|
|
840
|
+
- `experimental/mhe_transformer_payne`: parameter tokens and cross-attention,
|
|
841
|
+
with a learned hash encoding and projection for wavelength queries.
|
|
842
|
+
|
|
843
|
+
Both reuse the `transformer_payne` runtime adapter: the same canonical input/output
|
|
844
|
+
keys, channel metadata, custom nested leaf names, flux/intensity device batch
|
|
845
|
+
transforms, `fit`, `predict`, JAX composition, and portable bundle APIs. Parameters
|
|
846
|
+
have shape `(batch, parameter_dim)` and public wavelengths have shape
|
|
847
|
+
`(batch, n_wavelength)`. Outputs have shape `(batch, n_wavelength)` for one channel,
|
|
848
|
+
or `(batch, n_wavelength, channels)` otherwise. Initialize from an example batch
|
|
849
|
+
or `init_hints={"parameter_dim": ...}`.
|
|
850
|
+
|
|
851
|
+
Start from the existing flux or intensity preset and replace the model spec:
|
|
852
|
+
|
|
853
|
+
```python
|
|
854
|
+
from dataclasses import replace
|
|
855
|
+
from astro_emulators_toolkit import Emulator
|
|
856
|
+
from astro_emulators_toolkit.presets import transformer_payne_flux
|
|
857
|
+
|
|
858
|
+
cfg = transformer_payne_flux(profile="smoke")
|
|
859
|
+
cfg = replace(cfg, model=replace(
|
|
860
|
+
cfg.model,
|
|
861
|
+
name="experimental/mhe_transformer_payne",
|
|
862
|
+
params={**cfg.model.params,
|
|
863
|
+
"mhe": {"levels": 8, "log2_hashmap_size": 12,
|
|
864
|
+
"wavelength_range": [0.0, 1.0]}},
|
|
865
|
+
))
|
|
866
|
+
emu = Emulator.from_config(cfg)
|
|
867
|
+
# Use explicitly preprocessed data and emu.make_device_batch_transform(...)
|
|
868
|
+
# with emu.fit(...), just as for transformer_payne.
|
|
869
|
+
```
|
|
870
|
+
|
|
871
|
+
For `experimental/mhe_mlp`, replace the model name and parameter dictionary with,
|
|
872
|
+
for example, `{"hidden_sizes": [128, 128, 128], "parameter_hidden_sizes": [128, 128],
|
|
873
|
+
"min_period": 1e-4, "max_period": 2.0, "mhe": {"wavelength_range": [0.0, 1.0]}}`.
|
|
874
|
+
The transformer uses the usual `dim`, `dim_head`, `no_tokens`, `no_layers`, and
|
|
875
|
+
`dim_ff_multiplier`, plus optional `parameter_hidden_dim`, per-block
|
|
876
|
+
`ff_hidden_sizes`, and `head_hidden_dim`. Architecture parameters and the stable
|
|
877
|
+
models' optional `si`/`mup`/`mup_depth` LR policies are not interchangeable: use
|
|
878
|
+
`lr_scaling=None` or an explicit Optax optimizer for the MHE families.
|
|
879
|
+
|
|
880
|
+
MHE preprocessing is explicit. `mhe.wavelength_range` (default `[0, 1]`),
|
|
881
|
+
`min_period`, and `max_period` are in the wavelength coordinates actually supplied
|
|
882
|
+
to the model, whether physical, normalized, or logarithmic. Neither model takes a
|
|
883
|
+
logarithm or applies reference scaling automatically. Record transforms, units,
|
|
884
|
+
scaling, and intended-use bounds in the usual I/O metadata. Enable
|
|
885
|
+
`JAX_ENABLE_X64=1` and preserve float64 wavelengths even with float32 weights.
|
|
886
|
+
|
|
887
|
+
The nested `mhe` configuration also accepts `levels`, `features_per_level`,
|
|
888
|
+
`log2_hashmap_size`, and `table_init_scale`. Each level linearly interpolates
|
|
889
|
+
learned features with cell widths geometrically spaced from `max_period / 2` to
|
|
890
|
+
`min_period / 2`. Power-of-two tables are capped at `2 ** log2_hashmap_size`, with
|
|
891
|
+
wrapped indices. The range sets the grid origin and allocation; it does not clip
|
|
892
|
+
queries or define a safe extrapolation policy. Wavelength derivatives are
|
|
893
|
+
piecewise defined at grid boundaries. The transformer also supports
|
|
894
|
+
`projection_hidden_dim`, `projection_activation`, `projection_bias`, and optional
|
|
895
|
+
`normalize_output` inside `mhe`; normalization acts on encoded features, not
|
|
896
|
+
physical spectra, and defaults to off.
|
|
897
|
+
|
|
898
|
+
These families require toolkit 0.2.2 or newer, including when loading their bundles.
|
|
899
|
+
They train entirely within aemu without the original development project
|
|
900
|
+
or pretrained weights. Implementation comparisons and bundle round trips establish
|
|
901
|
+
numerical parity and portability; held-out residuals, boundaries, and realistic
|
|
902
|
+
inference tests remain separate scientific validation work.
|
|
903
|
+
|
|
904
|
+
Run [the MHE flux example](examples/experimental/08_train_mhe_flux.py) with
|
|
905
|
+
`--model mhe_mlp` or `--model mhe_transformer_payne` for a complete native workflow
|
|
906
|
+
with explicit input normalization and bundle export.
|
|
907
|
+
|
|
836
908
|
### Transformer-specific notes
|
|
837
909
|
|
|
838
910
|
The `transformer_payne` family **requires** `JAX_ENABLE_X64=1`: wavelengths are always handled in float64 up to the embedding cast, and model construction fails without it. Enable it before building or loading these models:
|
|
@@ -794,6 +794,78 @@ cfg = load_config("runs/demo/config.yaml")
|
|
|
794
794
|
|
|
795
795
|
Prototype model families and tasks live under `experimental/...`. They are useful for research and development but are intentionally outside the stable public contract.
|
|
796
796
|
|
|
797
|
+
The wavelength-explicit MHE (multiresolution hash encoding) families are:
|
|
798
|
+
|
|
799
|
+
- `experimental/mhe_mlp`: a shared parameter MLP followed by a wavelength
|
|
800
|
+
decoder that receives learned hash features at every hidden layer.
|
|
801
|
+
- `experimental/mhe_transformer_payne`: parameter tokens and cross-attention,
|
|
802
|
+
with a learned hash encoding and projection for wavelength queries.
|
|
803
|
+
|
|
804
|
+
Both reuse the `transformer_payne` runtime adapter: the same canonical input/output
|
|
805
|
+
keys, channel metadata, custom nested leaf names, flux/intensity device batch
|
|
806
|
+
transforms, `fit`, `predict`, JAX composition, and portable bundle APIs. Parameters
|
|
807
|
+
have shape `(batch, parameter_dim)` and public wavelengths have shape
|
|
808
|
+
`(batch, n_wavelength)`. Outputs have shape `(batch, n_wavelength)` for one channel,
|
|
809
|
+
or `(batch, n_wavelength, channels)` otherwise. Initialize from an example batch
|
|
810
|
+
or `init_hints={"parameter_dim": ...}`.
|
|
811
|
+
|
|
812
|
+
Start from the existing flux or intensity preset and replace the model spec:
|
|
813
|
+
|
|
814
|
+
```python
|
|
815
|
+
from dataclasses import replace
|
|
816
|
+
from astro_emulators_toolkit import Emulator
|
|
817
|
+
from astro_emulators_toolkit.presets import transformer_payne_flux
|
|
818
|
+
|
|
819
|
+
cfg = transformer_payne_flux(profile="smoke")
|
|
820
|
+
cfg = replace(cfg, model=replace(
|
|
821
|
+
cfg.model,
|
|
822
|
+
name="experimental/mhe_transformer_payne",
|
|
823
|
+
params={**cfg.model.params,
|
|
824
|
+
"mhe": {"levels": 8, "log2_hashmap_size": 12,
|
|
825
|
+
"wavelength_range": [0.0, 1.0]}},
|
|
826
|
+
))
|
|
827
|
+
emu = Emulator.from_config(cfg)
|
|
828
|
+
# Use explicitly preprocessed data and emu.make_device_batch_transform(...)
|
|
829
|
+
# with emu.fit(...), just as for transformer_payne.
|
|
830
|
+
```
|
|
831
|
+
|
|
832
|
+
For `experimental/mhe_mlp`, replace the model name and parameter dictionary with,
|
|
833
|
+
for example, `{"hidden_sizes": [128, 128, 128], "parameter_hidden_sizes": [128, 128],
|
|
834
|
+
"min_period": 1e-4, "max_period": 2.0, "mhe": {"wavelength_range": [0.0, 1.0]}}`.
|
|
835
|
+
The transformer uses the usual `dim`, `dim_head`, `no_tokens`, `no_layers`, and
|
|
836
|
+
`dim_ff_multiplier`, plus optional `parameter_hidden_dim`, per-block
|
|
837
|
+
`ff_hidden_sizes`, and `head_hidden_dim`. Architecture parameters and the stable
|
|
838
|
+
models' optional `si`/`mup`/`mup_depth` LR policies are not interchangeable: use
|
|
839
|
+
`lr_scaling=None` or an explicit Optax optimizer for the MHE families.
|
|
840
|
+
|
|
841
|
+
MHE preprocessing is explicit. `mhe.wavelength_range` (default `[0, 1]`),
|
|
842
|
+
`min_period`, and `max_period` are in the wavelength coordinates actually supplied
|
|
843
|
+
to the model, whether physical, normalized, or logarithmic. Neither model takes a
|
|
844
|
+
logarithm or applies reference scaling automatically. Record transforms, units,
|
|
845
|
+
scaling, and intended-use bounds in the usual I/O metadata. Enable
|
|
846
|
+
`JAX_ENABLE_X64=1` and preserve float64 wavelengths even with float32 weights.
|
|
847
|
+
|
|
848
|
+
The nested `mhe` configuration also accepts `levels`, `features_per_level`,
|
|
849
|
+
`log2_hashmap_size`, and `table_init_scale`. Each level linearly interpolates
|
|
850
|
+
learned features with cell widths geometrically spaced from `max_period / 2` to
|
|
851
|
+
`min_period / 2`. Power-of-two tables are capped at `2 ** log2_hashmap_size`, with
|
|
852
|
+
wrapped indices. The range sets the grid origin and allocation; it does not clip
|
|
853
|
+
queries or define a safe extrapolation policy. Wavelength derivatives are
|
|
854
|
+
piecewise defined at grid boundaries. The transformer also supports
|
|
855
|
+
`projection_hidden_dim`, `projection_activation`, `projection_bias`, and optional
|
|
856
|
+
`normalize_output` inside `mhe`; normalization acts on encoded features, not
|
|
857
|
+
physical spectra, and defaults to off.
|
|
858
|
+
|
|
859
|
+
These families require toolkit 0.2.2 or newer, including when loading their bundles.
|
|
860
|
+
They train entirely within aemu without the original development project
|
|
861
|
+
or pretrained weights. Implementation comparisons and bundle round trips establish
|
|
862
|
+
numerical parity and portability; held-out residuals, boundaries, and realistic
|
|
863
|
+
inference tests remain separate scientific validation work.
|
|
864
|
+
|
|
865
|
+
Run [the MHE flux example](examples/experimental/08_train_mhe_flux.py) with
|
|
866
|
+
`--model mhe_mlp` or `--model mhe_transformer_payne` for a complete native workflow
|
|
867
|
+
with explicit input normalization and bundle export.
|
|
868
|
+
|
|
797
869
|
### Transformer-specific notes
|
|
798
870
|
|
|
799
871
|
The `transformer_payne` family **requires** `JAX_ENABLE_X64=1`: wavelengths are always handled in float64 up to the embedding cast, and model construction fails without it. Enable it before building or loading these models:
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
__version__ = "0.2.2"
|
|
@@ -34,7 +34,11 @@ GENERATED_END = "<!-- aemu:generated:end -->"
|
|
|
34
34
|
TODO_MARKER = "<!-- aemu:todo"
|
|
35
35
|
|
|
36
36
|
# Families whose wavelength handling needs JAX float64.
|
|
37
|
-
_X64_FAMILY_PREFIXES = (
|
|
37
|
+
_X64_FAMILY_PREFIXES = (
|
|
38
|
+
"transformer_payne",
|
|
39
|
+
"experimental_mhe_mlp_",
|
|
40
|
+
"experimental_mhe_transformer_payne_",
|
|
41
|
+
)
|
|
38
42
|
|
|
39
43
|
_AUTHOR_SECTIONS = (
|
|
40
44
|
(
|
|
@@ -114,6 +114,13 @@ FIRST_LOADING_TOOLKIT: dict[str, dict[Any, str]] = {
|
|
|
114
114
|
}
|
|
115
115
|
|
|
116
116
|
|
|
117
|
+
# New model families can require a newer runtime without changing bundle stamps.
|
|
118
|
+
_MODEL_FIRST_LOADING_TOOLKIT = {
|
|
119
|
+
"experimental_mhe_mlp_v1": "0.2.2",
|
|
120
|
+
"experimental_mhe_transformer_payne_v1": "0.2.2",
|
|
121
|
+
}
|
|
122
|
+
|
|
123
|
+
|
|
117
124
|
def _version_key(version: str) -> tuple[int, ...]:
|
|
118
125
|
return tuple(int(part) for part in version.split("."))
|
|
119
126
|
|
|
@@ -137,6 +144,11 @@ def minimum_loading_toolkit_version(metadata: dict[str, Any]) -> str | None:
|
|
|
137
144
|
if stamp not in table:
|
|
138
145
|
return None
|
|
139
146
|
needed.append(table[stamp])
|
|
147
|
+
family_minimum = _MODEL_FIRST_LOADING_TOOLKIT.get(
|
|
148
|
+
str(metadata.get("model_family_id") or "")
|
|
149
|
+
)
|
|
150
|
+
if family_minimum is not None:
|
|
151
|
+
needed.append(family_minimum)
|
|
140
152
|
provenance = metadata.get("provenance") or {}
|
|
141
153
|
if wrote_v2_meaning(provenance.get("toolkit_version")):
|
|
142
154
|
# 0.2.0 pre-releases stamped config schema v1 but wrote keys that
|
|
@@ -3,16 +3,33 @@ from __future__ import annotations
|
|
|
3
3
|
from typing import Any
|
|
4
4
|
|
|
5
5
|
from ...models import ModelRegistryEntry
|
|
6
|
-
from .
|
|
7
|
-
from .mlp_2d_regression import MLP2DRegression, MLP2DRegressionConfig
|
|
8
|
-
from .siren import Siren, SirenConfig
|
|
6
|
+
from ...models.runtime_adapters import TransformerPayneRuntimeAdapter
|
|
9
7
|
from ..runtime_adapters import (
|
|
10
8
|
ExplicitWavelengthMLPRuntimeAdapter,
|
|
11
9
|
MLP2DRegressionRuntimeAdapter,
|
|
12
10
|
SirenRuntimeAdapter,
|
|
13
11
|
)
|
|
12
|
+
from .explicit_wavelength_mlp import ExplicitWavelengthMLP, ExplicitWavelengthMLPConfig
|
|
13
|
+
from .mhe import MHEMLP, MHETransformerPayne
|
|
14
|
+
from .mhe_config import MHEMLPConfig, MHETransformerPayneConfig
|
|
15
|
+
from .mlp_2d_regression import MLP2DRegression, MLP2DRegressionConfig
|
|
16
|
+
from .siren import Siren, SirenConfig
|
|
14
17
|
|
|
15
18
|
_EXPERIMENTAL_MODEL_REGISTRY: dict[str, ModelRegistryEntry] = {
|
|
19
|
+
"mhe_mlp": ModelRegistryEntry(
|
|
20
|
+
public_name="experimental/mhe_mlp",
|
|
21
|
+
family_id="experimental_mhe_mlp_v1",
|
|
22
|
+
config_cls=MHEMLPConfig,
|
|
23
|
+
model_cls=MHEMLP,
|
|
24
|
+
runtime=TransformerPayneRuntimeAdapter(),
|
|
25
|
+
),
|
|
26
|
+
"mhe_transformer_payne": ModelRegistryEntry(
|
|
27
|
+
public_name="experimental/mhe_transformer_payne",
|
|
28
|
+
family_id="experimental_mhe_transformer_payne_v1",
|
|
29
|
+
config_cls=MHETransformerPayneConfig,
|
|
30
|
+
model_cls=MHETransformerPayne,
|
|
31
|
+
runtime=TransformerPayneRuntimeAdapter(),
|
|
32
|
+
),
|
|
16
33
|
"explicit_wavelength_mlp": ModelRegistryEntry(
|
|
17
34
|
public_name="experimental/explicit_wavelength_mlp",
|
|
18
35
|
family_id="experimental_explicit_wavelength_mlp_v1",
|
|
@@ -0,0 +1,302 @@
|
|
|
1
|
+
"""MHE networks with a shared parameter trunk and independent wavelength queries.
|
|
2
|
+
|
|
3
|
+
Adapted from exojax_emulator.mhe_models. Trained parameter names and numerical
|
|
4
|
+
operations are retained for reference comparisons.
|
|
5
|
+
"""
|
|
6
|
+
|
|
7
|
+
from __future__ import annotations
|
|
8
|
+
|
|
9
|
+
from typing import NamedTuple
|
|
10
|
+
import jax
|
|
11
|
+
import jax.numpy as jnp
|
|
12
|
+
from flax import nnx
|
|
13
|
+
|
|
14
|
+
from .mhe_config import MHETransformerPayneConfig, MHEMLPConfig, positive_int
|
|
15
|
+
from .mhe_layers import (
|
|
16
|
+
MHEHashEncoding,
|
|
17
|
+
MHEWavelengthEncoding,
|
|
18
|
+
activation,
|
|
19
|
+
linear,
|
|
20
|
+
rms_norm,
|
|
21
|
+
)
|
|
22
|
+
|
|
23
|
+
|
|
24
|
+
def _validate_dims(in_dim, out_dim):
|
|
25
|
+
positive_int("in_dim", in_dim)
|
|
26
|
+
if out_dim != 1:
|
|
27
|
+
raise ValueError(
|
|
28
|
+
"MHE models require output init size 1; wavelength is an explicit input axis."
|
|
29
|
+
)
|
|
30
|
+
|
|
31
|
+
|
|
32
|
+
def _labels(labels, in_dim, dtype):
|
|
33
|
+
labels = jnp.asarray(labels, dtype=dtype)
|
|
34
|
+
if labels.ndim != 2 or labels.shape[1] != in_dim:
|
|
35
|
+
raise ValueError(f"parameters must have shape (batch, {in_dim})")
|
|
36
|
+
return labels
|
|
37
|
+
|
|
38
|
+
|
|
39
|
+
def _waves(waves, batch, *, one=False):
|
|
40
|
+
# Match TransformerPayne's float64 coordinate path. Callers must preserve
|
|
41
|
+
# precision before this point; promoting an already rounded array cannot restore it.
|
|
42
|
+
waves = jnp.asarray(waves, dtype=jnp.float64)
|
|
43
|
+
if one:
|
|
44
|
+
if waves.ndim == 0:
|
|
45
|
+
waves = jnp.broadcast_to(waves, (batch,))
|
|
46
|
+
if waves.shape != (batch,):
|
|
47
|
+
raise ValueError(
|
|
48
|
+
"predict_one expects a scalar or one wavelength per atmosphere"
|
|
49
|
+
)
|
|
50
|
+
else:
|
|
51
|
+
if waves.ndim == 1:
|
|
52
|
+
waves = jnp.broadcast_to(waves, (batch, waves.shape[0]))
|
|
53
|
+
if waves.ndim != 2 or waves.shape[0] != batch:
|
|
54
|
+
raise ValueError("queries must have shape (queries,) or (batch, queries)")
|
|
55
|
+
return waves
|
|
56
|
+
|
|
57
|
+
|
|
58
|
+
class TransformerContext(NamedTuple):
|
|
59
|
+
"""Differentiable per-block K/V without a wavelength axis."""
|
|
60
|
+
|
|
61
|
+
keys_values: tuple
|
|
62
|
+
|
|
63
|
+
|
|
64
|
+
class CrossAttention(nnx.Module):
|
|
65
|
+
def __init__(self, cfg, *, rngs):
|
|
66
|
+
self.heads, self.head_dim, self.alpha_att = (
|
|
67
|
+
(cfg.dim // cfg.dim_head),
|
|
68
|
+
cfg.dim_head,
|
|
69
|
+
cfg.alpha_att,
|
|
70
|
+
)
|
|
71
|
+
options = dict(
|
|
72
|
+
rngs=rngs,
|
|
73
|
+
bias=cfg.bias_attention,
|
|
74
|
+
dtype=cfg.dtype,
|
|
75
|
+
initializer=cfg.kernel_init,
|
|
76
|
+
sigma=cfg.sigma,
|
|
77
|
+
)
|
|
78
|
+
self.q = linear(cfg.dim, cfg.dim, zero=cfg.init_att_q == "zero", **options)
|
|
79
|
+
self.k = linear(cfg.dim, cfg.dim, **options)
|
|
80
|
+
self.v = linear(cfg.dim, cfg.dim, **options)
|
|
81
|
+
self.out = linear(cfg.dim, cfg.dim, zero=cfg.init_att_o == "zero", **options)
|
|
82
|
+
|
|
83
|
+
def prepare_context(self, tokens):
|
|
84
|
+
shape = (*tokens.shape[:-1], self.heads, self.head_dim)
|
|
85
|
+
return self.k(tokens).reshape(shape), self.v(tokens).reshape(shape)
|
|
86
|
+
|
|
87
|
+
def __call__(self, queries, context):
|
|
88
|
+
k, v = context
|
|
89
|
+
q = self.q(queries).reshape(*queries.shape[:-1], self.heads, self.head_dim)
|
|
90
|
+
scores = jnp.einsum("bqhd,bthd->bhqt", q, k) / self.head_dim
|
|
91
|
+
if self.alpha_att != 1.0:
|
|
92
|
+
scores = scores * self.alpha_att
|
|
93
|
+
values = jnp.einsum("bhqt,bthd->bqhd", jax.nn.softmax(scores, axis=-1), v)
|
|
94
|
+
return self.out(values.reshape(queries.shape))
|
|
95
|
+
|
|
96
|
+
|
|
97
|
+
class Block(nnx.Module):
|
|
98
|
+
def __init__(self, cfg, width, *, rngs):
|
|
99
|
+
self.attention = CrossAttention(cfg, rngs=rngs)
|
|
100
|
+
options = dict(
|
|
101
|
+
rngs=rngs,
|
|
102
|
+
bias=cfg.bias("feed_forward"),
|
|
103
|
+
dtype=cfg.dtype,
|
|
104
|
+
initializer=cfg.kernel_init,
|
|
105
|
+
sigma=cfg.sigma,
|
|
106
|
+
)
|
|
107
|
+
self.ff_in = linear(cfg.dim, width, **options)
|
|
108
|
+
self.ff_out = linear(width, cfg.dim, zero=cfg.ff_init == "zero", **options)
|
|
109
|
+
self.activation, self.epsilon, self.residual_scale = (
|
|
110
|
+
cfg.activation,
|
|
111
|
+
cfg.rms_epsilon,
|
|
112
|
+
cfg.residual_scale,
|
|
113
|
+
)
|
|
114
|
+
|
|
115
|
+
def __call__(self, h, context):
|
|
116
|
+
update = self.attention(rms_norm(h, self.epsilon), context)
|
|
117
|
+
h = h + (update if self.residual_scale == 1.0 else self.residual_scale * update)
|
|
118
|
+
update = self.ff_out(
|
|
119
|
+
activation(self.activation, self.ff_in(rms_norm(h, self.epsilon)))
|
|
120
|
+
)
|
|
121
|
+
return h + (
|
|
122
|
+
update if self.residual_scale == 1.0 else self.residual_scale * update
|
|
123
|
+
)
|
|
124
|
+
|
|
125
|
+
|
|
126
|
+
class MHETransformerPayne(nnx.Module):
|
|
127
|
+
"""Shared atmospheric tokens/KV, with independent wavelength predictions.
|
|
128
|
+
|
|
129
|
+
Inputs: labels (B,D), float64 wavelength queries (Q,) or (B,Q).
|
|
130
|
+
Outputs: (B,Q) or (B,Q,C). Rebuild prepared context if labels or weights change.
|
|
131
|
+
"""
|
|
132
|
+
|
|
133
|
+
def __init__(
|
|
134
|
+
self,
|
|
135
|
+
*,
|
|
136
|
+
in_dim: int,
|
|
137
|
+
out_dim: int,
|
|
138
|
+
cfg: MHETransformerPayneConfig,
|
|
139
|
+
rngs: nnx.Rngs,
|
|
140
|
+
):
|
|
141
|
+
_validate_dims(in_dim, out_dim)
|
|
142
|
+
self.in_dim = in_dim
|
|
143
|
+
self.config = cfg
|
|
144
|
+
self.dim, self.parameter_tokens = cfg.dim, cfg.no_tokens
|
|
145
|
+
options = dict(
|
|
146
|
+
rngs=rngs, dtype=cfg.dtype, initializer=cfg.kernel_init, sigma=cfg.sigma
|
|
147
|
+
)
|
|
148
|
+
width = cfg.parameter_hidden_dim or cfg.dim
|
|
149
|
+
self.label_in = linear(
|
|
150
|
+
in_dim, width, bias=cfg.bias("parameter_embedding"), **options
|
|
151
|
+
)
|
|
152
|
+
self.label_out = linear(
|
|
153
|
+
width,
|
|
154
|
+
cfg.no_tokens * cfg.dim,
|
|
155
|
+
bias=cfg.bias("parameter_embedding"),
|
|
156
|
+
zero=cfg.emb_init == "zero",
|
|
157
|
+
**options,
|
|
158
|
+
)
|
|
159
|
+
widths = (
|
|
160
|
+
cfg.ff_hidden_sizes or (cfg.dim * cfg.dim_ff_multiplier,) * cfg.no_layers
|
|
161
|
+
)
|
|
162
|
+
self.blocks = nnx.List([Block(cfg, width, rngs=rngs) for width in widths])
|
|
163
|
+
width = cfg.head_hidden_dim or cfg.dim
|
|
164
|
+
self.head_in = linear(cfg.dim, width, bias=cfg.bias("output_head"), **options)
|
|
165
|
+
self.head_out = linear(
|
|
166
|
+
width,
|
|
167
|
+
cfg.channels,
|
|
168
|
+
bias=cfg.bias("output_head"),
|
|
169
|
+
zero=cfg.head_init == "zero",
|
|
170
|
+
**options,
|
|
171
|
+
)
|
|
172
|
+
self.wavelength_encoder = MHEWavelengthEncoding(
|
|
173
|
+
cfg.dim,
|
|
174
|
+
rngs=rngs,
|
|
175
|
+
min_period=cfg.min_period,
|
|
176
|
+
max_period=cfg.max_period,
|
|
177
|
+
cfg=cfg.mhe,
|
|
178
|
+
dtype=cfg.dtype,
|
|
179
|
+
)
|
|
180
|
+
|
|
181
|
+
def encode_wavelength(self, wavelengths):
|
|
182
|
+
return self.wavelength_encoder(wavelengths)
|
|
183
|
+
|
|
184
|
+
def prepare_context(self, labels):
|
|
185
|
+
labels = _labels(labels, self.in_dim, self.config.dtype)
|
|
186
|
+
tokens = self.label_out(
|
|
187
|
+
activation(self.config.activation, self.label_in(labels))
|
|
188
|
+
)
|
|
189
|
+
tokens = tokens.reshape(labels.shape[0], self.parameter_tokens, self.dim)
|
|
190
|
+
if self.config.alpha_emb != 1.0:
|
|
191
|
+
tokens = self.config.alpha_emb * tokens
|
|
192
|
+
tokens = rms_norm(tokens, self.config.rms_epsilon)
|
|
193
|
+
return TransformerContext(
|
|
194
|
+
tuple(block.attention.prepare_context(tokens) for block in self.blocks)
|
|
195
|
+
)
|
|
196
|
+
|
|
197
|
+
def predict_one(self, context, wavelengths):
|
|
198
|
+
"""One wavelength per atmosphere, with the trunk already evaluated."""
|
|
199
|
+
if len(context.keys_values) != len(self.blocks):
|
|
200
|
+
raise ValueError("Context must contain K/V for every transformer block")
|
|
201
|
+
wave = _waves(wavelengths, context.keys_values[0][0].shape[0], one=True)
|
|
202
|
+
h = self.encode_wavelength(wave[:, None])
|
|
203
|
+
for block, kv in zip(self.blocks, context.keys_values):
|
|
204
|
+
h = block(h, kv)
|
|
205
|
+
h = self.head_out(
|
|
206
|
+
activation(
|
|
207
|
+
self.config.activation,
|
|
208
|
+
self.head_in(rms_norm(h, self.config.rms_epsilon)),
|
|
209
|
+
)
|
|
210
|
+
)
|
|
211
|
+
h = activation(self.config.output_activation, h)[:, 0, :]
|
|
212
|
+
return h[:, 0] if self.config.channels == 1 else h
|
|
213
|
+
|
|
214
|
+
def predict_queries(self, context, wavelengths):
|
|
215
|
+
if len(context.keys_values) != len(self.blocks):
|
|
216
|
+
raise ValueError("Context must contain K/V for every transformer block")
|
|
217
|
+
waves = _waves(wavelengths, context.keys_values[0][0].shape[0])
|
|
218
|
+
return jax.vmap(
|
|
219
|
+
lambda wave: self.predict_one(context, wave), in_axes=1, out_axes=1
|
|
220
|
+
)(waves)
|
|
221
|
+
|
|
222
|
+
def __call__(self, x, *, train: bool = False, rngs: nnx.Rngs | None = None):
|
|
223
|
+
del train, rngs
|
|
224
|
+
labels, wavelengths = x
|
|
225
|
+
return self.predict_queries(self.prepare_context(labels), wavelengths)
|
|
226
|
+
|
|
227
|
+
|
|
228
|
+
class MHEMLP(nnx.Module):
|
|
229
|
+
"""Shared parameter MLP; MHE features enter each vmapped query layer.
|
|
230
|
+
|
|
231
|
+
GELU is the default; alternative activations are explicit ablations.
|
|
232
|
+
Context and query shape conventions match MHETransformerPayne.
|
|
233
|
+
"""
|
|
234
|
+
|
|
235
|
+
def __init__(self, *, in_dim: int, out_dim: int, cfg: MHEMLPConfig, rngs: nnx.Rngs):
|
|
236
|
+
_validate_dims(in_dim, out_dim)
|
|
237
|
+
self.in_dim = in_dim
|
|
238
|
+
self.config = cfg
|
|
239
|
+
assert cfg.mhe is not None and cfg.parameter_hidden_sizes is not None
|
|
240
|
+
self.wavelength_encoder = MHEHashEncoding(
|
|
241
|
+
rngs=rngs,
|
|
242
|
+
min_period=cfg.min_period,
|
|
243
|
+
max_period=cfg.max_period,
|
|
244
|
+
cfg=cfg.mhe,
|
|
245
|
+
dtype=cfg.dtype,
|
|
246
|
+
)
|
|
247
|
+
features = cfg.mhe.levels * cfg.mhe.features_per_level
|
|
248
|
+
options = dict(
|
|
249
|
+
rngs=rngs, dtype=cfg.dtype, initializer=cfg.kernel_init, sigma=cfg.sigma
|
|
250
|
+
)
|
|
251
|
+
sizes = (in_dim, *cfg.parameter_hidden_sizes)
|
|
252
|
+
self.parameter_layers = nnx.List(
|
|
253
|
+
[
|
|
254
|
+
linear(a, b, bias=cfg.bias_parameter_embedding, **options)
|
|
255
|
+
for a, b in zip(sizes[:-1], sizes[1:])
|
|
256
|
+
]
|
|
257
|
+
)
|
|
258
|
+
sizes = (cfg.parameter_hidden_sizes[-1], *cfg.hidden_sizes)
|
|
259
|
+
self.layers = nnx.List(
|
|
260
|
+
[
|
|
261
|
+
linear(a + features, b, bias=cfg.bias_decoder, **options)
|
|
262
|
+
for a, b in zip(sizes[:-1], sizes[1:])
|
|
263
|
+
]
|
|
264
|
+
+ [
|
|
265
|
+
linear(
|
|
266
|
+
cfg.hidden_sizes[-1],
|
|
267
|
+
cfg.channels,
|
|
268
|
+
bias=cfg.bias_output_head,
|
|
269
|
+
**options,
|
|
270
|
+
)
|
|
271
|
+
]
|
|
272
|
+
)
|
|
273
|
+
|
|
274
|
+
def encode_parameters(self, labels):
|
|
275
|
+
h = _labels(labels, self.in_dim, self.config.dtype)
|
|
276
|
+
for layer in self.parameter_layers:
|
|
277
|
+
h = activation(self.config.activation, layer(h))
|
|
278
|
+
return h
|
|
279
|
+
|
|
280
|
+
def prepare_context(self, labels):
|
|
281
|
+
return self.encode_parameters(labels)
|
|
282
|
+
|
|
283
|
+
def predict_one(self, context, wavelengths):
|
|
284
|
+
wave = _waves(wavelengths, context.shape[0], one=True)
|
|
285
|
+
features, h = self.wavelength_encoder(wave[:, None]), context[:, None, :]
|
|
286
|
+
for layer in self.layers[:-1]:
|
|
287
|
+
h = activation(
|
|
288
|
+
self.config.activation, layer(jnp.concatenate((h, features), axis=-1))
|
|
289
|
+
)
|
|
290
|
+
h = activation(self.config.output_activation, self.layers[-1](h))[:, 0, :]
|
|
291
|
+
return h[:, 0] if self.config.channels == 1 else h
|
|
292
|
+
|
|
293
|
+
def predict_queries(self, context, wavelengths):
|
|
294
|
+
waves = _waves(wavelengths, context.shape[0])
|
|
295
|
+
return jax.vmap(
|
|
296
|
+
lambda wave: self.predict_one(context, wave), in_axes=1, out_axes=1
|
|
297
|
+
)(waves)
|
|
298
|
+
|
|
299
|
+
def __call__(self, x, *, train: bool = False, rngs: nnx.Rngs | None = None):
|
|
300
|
+
del train, rngs
|
|
301
|
+
labels, wavelengths = x
|
|
302
|
+
return self.predict_queries(self.prepare_context(labels), wavelengths)
|