Variopinta 0.7.0

Variopinta 0.7.0

• Add LongestMaxSize for aspect-preserving resize with exact half-up dimension
  rounding, optional antialiasing, and shared image/mask geometry. Compose it
  with PadIfNeeded for square output without stretching.
• Accelerate constant padding for RGB and grayscale images and label masks,
  preserving fill values and output pixels.
• Renew benchmark evidence, record participant-specific resize rounding, and
  fix fingerprint coverage for tensor-output benchmarks.

Variopinta 0.6.0

Variopinta 0.6.0

• Rename target configuration to input_spec and output_specs, require names
  for public targets and output ports, and give explicit targets a default
  ReturnArray(name="array") output.
• Make secondary pipeline, target, output, I/O, and transform configuration
  keyword-only. Rename PNG compression to compression_level.
• Replace sampled scalar shorthands with explicit *_range tuples, clarify
  affine and crop units, and add explicit pixel/fraction units to
  CoarseDropout hole sizes.
• Construct compiled pipelines and result containers through pipeline methods
  only. Pickles remain same-release artifacts; pickles from releases with the
  earlier field names are not supported.

Migration mappings for this breaking pre-1.0 release:

Earlier spelling	Current spelling
Image(carrier=..., outputs=..., name=...)	Image(input_spec=..., name=..., output_specs=...)
Encode(..., compression=n)	Encode(..., name=..., compression_level=n)
RandomResizedCrop(..., scale=..., ratio=...)	RandomResizedCrop(..., area_range=..., aspect_ratio_range=...)
Affine(degrees=d, translate=..., scale=s, shear=...)	Affine(degrees_range=(-d, d), translate_max_fraction=..., scale_range=(s, s), shear_x_range=..., shear_y_range=...)
GaussianNoise(mean=m, std=s)	GaussianNoise(mean_range=(m, m), std_range=(s, s))
Sharpen(alpha=a, lightness=s)	Sharpen(blend_weight_range=(a, a), strength_range=(s, s))

Convert every other sampled scalar to an explicit two-endpoint range. For old
ColorJitter scalar amounts, use the already-canonical factor endpoints:
brightness=0.2 becomes brightness_range=(0.8, 1.2). Add
hole_height_unit="pixels" or hole_width_unit="pixels" when migrating old
integer dropout ranges; omitted units mean fractions.

Variopinta 0.5.0

Variopinta 0.5.0

• Support native grayscale uint8 images in HW and HWC1 layouts across the
  transform catalog, preserving array rank and returning owned, contiguous output.
• Accept scalar and one-element normalization statistics and image fills that
  broadcast to grayscale or RGB. Active three-element values require RGB;
  grayscale normalization requires compatible values such as mean=0.5, std=0.5.
• Add Image(decode_mode="gray") for encoded-buffer and path inputs, grayscale
  JPEG/PNG outputs, and contiguous single-channel CHW tensor output.
• Upgrade explain() to schema 5 with per-target channel alternatives and
  channel-aware execution reporting. Consumers of schema 4 must update.
• Add a Fashion-MNIST notebook with grayscale augmentation previews and PyTorch
  classifier training on CPU or CUDA.
• Add an Oxford-IIIT Pet notebook with shared image/mask augmentation, keyed
  replay, and segmentation training with PyTorch Lightning.

Variopinta 0.4.4

Variopinta 0.4.4

• Support standard-library pickle for Pipeline and CompiledPipeline,
  preserving execution mode, random sequence position, and shared target/output
  port identities while rebuilding independent native execution state.
• Allow Datasets to own compiled pipelines directly with spawn and forkserver
  workers, without custom serialization or lazy compilation.
• Add an executed Imagenette notebook with augmentation examples, keyed replay,
  and ten epochs of Lightning GPU training using four persistent workers per loader.

Variopinta 0.4.3

Variopinta 0.4.3

• Infer NumPy array and PipelineResult return types for implicit and explicit
  calls to both pipeline executors, preserving typed output-port lookup.
• Accept sequences of differently typed output ports in static type checking.

Variopinta 0.4.2

Variopinta 0.4.2

• Accept a single Image or Mask directly in Pipeline(targets=...), with
  the same bindings and result structure as a one-element sequence.

Variopinta 0.4.1

Variopinta 0.4.1

• Accept a single output port directly in Image and Mask target signatures,
  while preserving tuple and sequence declarations for multiple outputs.
• Add focused examples for returning, encoding, and writing multiple outputs
  from one transformed target.

Variopinta 0.4.0

Variopinta 0.4.0

• Replace pipeline-global I/O and special-case masks with immutable Image and
  Mask targets. Each target selects an Array, Encoded, or Path carrier
  and fans its final raster out to named ReturnArray, ReturnTensor, Encode,
  and Write ports.
• Replace positional explicit calls and arity-collapsing tuples with keyword
  target bindings plus immutable PipelineResult/TargetResult values.
• Remove the Return output and ToTorch transform. ReturnTensor now owns
  the optional Torch adapter and the CHW terminal layout.
• Rename the public reference executor to Pipeline; .compile() returns
  CompiledPipeline with the same static target signature.
• Move semantic border fill to each Mask port, keep mask rasterization nearest
  and lossless, and make standalone I/O generic. PNG unchanged decoding
  preserves grayscale samples and palette indices from static 1/2/4/8-bit PNG.
• Make target writes atomic per file after all sources are acquired and every
  encoded result is prepared. Multiple destinations are not one transaction.

Variopinta 0.3.1

Variopinta 0.3.1

• Accelerate the dominant single-image CPU paths for padding, sharpening,
  affine and generic remapping, Gaussian noise, hue jitter, and standalone
  HWC-to-CHW conversion. GaussianNoise now uses the pinned rand_distr
  ZIGNOR stream, so keyed noise pixels differ from earlier releases while
  later transform sampling remains unchanged.

Variopinta 0.3.0

Variopinta 0.3.0

• Add immutable array, encoded-buffer, and path pipeline inputs plus returned,
  encoded-buffer, and path outputs, with native decode/augment/encode routes,
  size limits, GIL-aware execution, and source/sink introspection.
• Add max_encoded_bytes to decode_image and read_image.