Skip to content

v0.1.9

Latest

Choose a tag to compare

@github-actions github-actions released this 10 Nov 15:34
ba022cd

In this release: The ADMIRE model-based beamformer, the UltraPIPS perceptual metric for B-mode images, and improved streaming infrastructure for training on large ultrasound datasets through Hugging Face!

New Features

Changed

  • Streaming hf:// datasets no longer runs into the Hugging Face rate limit, so you can train on many batches with many workers without downloading the data. The API is unchanged; two new environment variables, ZEA_MAX_CONCURRENT_REQUESTS and ZEA_FIND_H5_SHAPES_WORKERS, tune it (see Environment variables) (#671 by @tristan-deep)
  • Conversion scripts for EchoNet-Dynamic, CAMUS and PICMUS accept the .zip file itself. EchoNet-Dynamic reproduces the published split by default, and its polar conversion is about 15 times faster (#676 by @wesselvannierop)
  • (Breaking) In compute_pfield, alpha and percentile also shape the field when norm=False, the unused n_el argument is removed, and probes other than linear and phased arrays raise NotImplementedError (#657 by @wesselvannierop)
  • (Breaking) zea.func.vmap validates axes like jax.vmap, so a list no longer matches a tuple. out_axes=None only works on the native jax.vmap, and the _use_torch_vmap argument is removed (#692 by @wesselvannierop)
  • Parameters.copy() and set_transmits are much faster (#656 by @wesselvannierop)
  • (Breaking) The check on cartesian beamforming grids uses per-axis limits from the image's k-space support instead of λ/2, see How fine should the beamforming grid be?. Auto-sized grids get one more pixel per axis (#673 by @wesselvannierop)
  • The default probe_bandwidth_percent is 100 instead of 200, which also narrows the band used by compute_pfield (#644 by @tristan-deep)

Fixed

  • A streamed hf:// read that still fails after its retries, such as a dropped connection, no longer ends the epoch: the Dataloader reopens the file and reads the sample again (#690 by @tristan-deep)
  • zea.func.vmap passes arguments at the right positions with multiple None arguments. It also keeps all outputs, supports nested and None outputs, supports negative axes when batching, and raises an error for arguments of different lengths (#691 by @wesselvannierop)
  • The pressure field uses the one-way transmit spectrum, averages over frequencies instead of summing, and takes the element positions from probe_geometry (#657 by @wesselvannierop)
  • t_peak estimated from waveforms_two_way uses the envelope peak; values change where the carrier peak differs from it (#656 by @wesselvannierop)
  • Conversion of Verasonics BS100BW and BS50BW data samples I and Q at the same time instants. Data converted with earlier versions, such as some Open-H-RF data, may be affected (#672 by @vincentvdschaft)
  • zea.ops.Normalize clips its output instead of its input, so results stay within output_range (#681 by @wesselvannierop)
  • waveforms_one_way and waveforms_two_way follow selected_transmits, like the other per-transmit parameters (#683 by @wesselvannierop)
  • Operations and pipelines can be deep-copied and pickled, also when not registered with Keras, and copies keep their JIT settings (#698 by @tristan-deep)
  • The shared GIF palette keeps exact colors, so pure black stays (0, 0, 0) (#680 by @wesselvannierop)
  • Deriving xlims on a GPU with the torch backend (#661 by @tristan-deep)
  • Jitted ops on the torch backend in the Docker images, which lacked a C++ compiler (#658 by @wesselvannierop)
  • zea app and load_model_from_hf download to the zea cache instead of the Hugging Face cache (#689 by @Louisvh)
  • @cache_output functions that call a recursive function can now be cached. Additionally, the cache key now also includes other @cache_output functions it calls, so editing those invalidates its cache (#703 by @wesselvannierop)

Full Changelog: v0.1.8...v0.1.9