Version 0.3.1

Bugfixes

• regularizers: Fix dtype detection in RegularizeLayers
• web: Fix detection of Slurm job state through squeue

Version 0.3.0

Changes since v0.2 (2025-09-05).

Major features

• PyTorch backend (#31, #45). A fourth array backend alongside numpy, cupy and jax,
  with autodiff support for the gradient engine and Apple MPS device support.
• Arbitrary MTF support (#58). Real- and reciprocal-space convolution machinery
  (prepare_convolve2d/prepare_convolve2d_recip, DCT-based 'reflect' boundaries,
  separable-filter fast paths), a Filter algebra (products, sums, scalar multiplication),
  EMPAD preset MTFs, an apply_mtf preprocessing hook, and MTF in the forward model of
  both the gradient and conventional engines.
• Web UI overhaul (#54). Topic pub/sub websocket protocol (web/pubsub.py,
  web/views.py), tabbed/dockable dashboard with drag & drop and persisted layouts,
  migration from wasm-array to @hexane360/plotlib, new widgets (object amplitude &
  phase, sum & stack; probe modes in real and reciprocal space; scan positions),
  heartbeat/reconnect handling, version information, and an error page.
• New CLI structure (#46). phaser/cli/ with lazily-loaded subcommands and declared
  optional dependencies, plus two interactive tools: process_empad (EMPAD XML metadata
  fixup) and calc_drift (linear scan drift measurement), backed by a persisted YAML
  Config (utils/config.py).
• Arbitrary aberration support for the initial probe (#43), validated against Kirkland.
• New data loaders: Gatan/DM4 (#32, requires rsciio) and Nion, including scan rotation
  extraction from Gatan metadata.
• HDF5 state format v0.2 (#55). Reads remain backwards
  compatible with v0.1.
• Granular loss reporting (#35). progress is now one series per loss term (detector
  error plus each named regularizer), threaded through both engines, the observer and the
  web UI.
• Synthetic/constrained position gradients: positions_affine and positions_line allow constraining position updates.

Major bugfixes

• Websocket compression disabled via a monkeypatch on the pinned hypercorn, fixing server
  hangs and long response times (#54).
• Fixed slow compilation on the conventional engines, plus general conventional-engine fixes
  and gradient-engine speedups.
• Fixed underflow NaNs on torch/MPS, the adam solver's nu dtype, and remove_linear_ramp
  on torch (#45).
• Fixed torch.func.grad with non-tensor aux types, and recip_grid breaking tracing on torch.
• Compact grouping algorithm rewritten: recursive bisection seeding instead of kmeans, with
  vectorized refinement. Substantially better with a large number of groups.
• Exceptions are now passed to Observer.close(), and job failures propagate to the web UI.
• Scan affine transform is applied pre-rotation.
• Position updates are no longer scaled by grouping.
• Fixed drop_nan_patterns with tilt (#32).

Minor features

• nonneg_object_phase constraint, with an example plan (#42).
• clamp_object_amplitude accepts an optional lower bound (#38).
• Observer reporting for position & tilt updates, elapsed time and UTC time, and a detailed
  error breakdown.
• Py4DSTEM layout added to the default HDF5 search path.
• offset post-load hook.
• Better convolve1d on jax and torch, and better xp.pad support on torch (#52).
• Warning when the ground truth is too small for reliable cross-correlation in
  align_object_to_ground_truth (#41).

Packaging

• Version is now dynamic, read from phaser.__version__; the entry point moved to
  phaser.cli:cli.
• Dropped support for numpy 1.x (now numpy>=2.0,<2.7).
• New torch and cupy13 extras; jax widened to <0.11; py-pane bumped to 0.11.7.
• New dependencies: frozendict, optree, platformdirs, lxml.
• hypercorn pinned exactly in web feature.

Version 0.2

phaserEM version 0.2