Add 1D time-series support and automatic input projection to TransformerEmbedding

[satwiksps]

Support scalar time-series and simplify config handling in TransformerEmbedding

This PR improves the usability of TransformerEmbedding by adding native support for scalar (1D) time-series inputs and removing the requirement to manually specify a full configuration dictionary.

Problem

• Described in Support scalar time-series and simplify config handling in TransformerEmbedding #1696:
• TransformerEmbedding did not support scalar time-series shaped (batch, seq_len) or (batch, seq_len, 1).
• Users were required to manually construct a complete config dictionary, which was unintuitive and prone to errors.
• The model raised shape errors early and did not automatically project simple inputs.

What this PR does

• Adds automatic projection from scalar inputs to the required feature_space_dim via a lazily-initialized input_proj layer.
• Extends the forward pass to handle:
  • (batch, seq_len)
  • (batch, seq_len, 1)
  • (batch, seq_len, D) (existing behavior)
• Ensures compatibility with the attention mechanism (e.g., head dimension constraints).
• Adds new test test_transformer_embedding_scalar_timeseries to verify:
  • Correct handling of 1D inputs
  • Proper projection behavior
  • Successful integration into embedding API
• Updates docstrings to reflect new behavior.

Why this is valuable

• Makes the transformer embedding consistent with other embedding nets (RNN, CNN, LRU) that already accept scalar time-series.
• Eliminates user friction by making TransformerEmbedding usable out of the box.
• Enables simplified tutorials, workflows, and example code for time-series inference.

Testing

• New test added in embedding_net_test.py.
• All transformer-specific tests pass.
• Pre-commit (ruff, formatting, linting) passed successfully.

Checklist

• Code changes follow project style.
• Added targeted tests.
• Updated docstrings
• Pre-commit hooks passed.
• No breaking changes introduced.

Closes #1696

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@2c216c2). Learn more about missing BASE report.
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1703   +/-   ##
=======================================
  Coverage        ?   84.67%           
=======================================
  Files           ?      137           
  Lines           ?    11493           
  Branches        ?        0           
=======================================
  Hits            ?     9732           
  Misses          ?     1761           
  Partials        ?        0           

Flag	Coverage Δ
unittests	84.67% <100.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
sbi/neural_nets/embedding_nets/transformer.py	93.40% <100.00%> (ø)

[janfb]

Thanks @satwiksps , good first draft!

I suggest to make all the config options actual arguments to __init__ to give full control to the user and have explicit type hints and defaults.

We could even think about using a dataclass TransformerConfig, but this would create additional overhead for the user, having to instantiate the dataclass config object externall, not sure. what's your take here?

[janfb]

Looks good!

Added more suggestion for making it look even better in terms of documentation. Would be great if you could add this as well 🙏

[janfb]

Suggested change

	Basic configuration parameters:
	Args:

[janfb]

if arg docstrings break the line please add a small indent to set them apart from other arguments.

[janfb]

please add indent in line below.

[janfb]

please add indent in line below.

[janfb]

please add indent in line below.

[janfb]

Would you be for adding a short explanatory class docstring here?

e.g., for an SBI user working with time series or images but not so familiar with transformers, give a concise overview how they can use this class? e.g., what means "vit" (for images), what means "is_causal" (for time series). etc. not a tutorial, just a brief high-level explanation. Maybe even with a short code Example block.

When we add this docstring here on the top class level then it will show up nicely in the Sphinx Documentation, e.g., like with the EnsemblePosterior here: https://sbi.readthedocs.io/en/latest/reference/_autosummary/sbi.inference.EnsemblePosterior.html#sbi.inference.EnsemblePosterior