- Agnostic to centroid calculation method.
- Propagates all frame example metadata to individual instance examples.
- Tests for coordinate transformations correctness, especially for
  tricky bounding box and cropping calculations.

- Low level confidence map generation functions
- Multi-instance confidence map generator
- Multi-instance centroid confidence map generator
- Instance-cropped center instance confidence map generator
- Instance-cropped all instance confidence map generator
- Tests

- Useful for downstream processing that affects shape.

- "Full" to indicate it refers to the input to the instance cropper,
  which may have been modified relative to the "raw" image, e.g., due to
  padding/scaling.
- Rename attribute `drop_full_image` to `keep_full_image`

- The general expectation is that points in individual examples are
  already on the image grid scale.
- The "scale" key can be used to track scaling operations, including
  aspect ratio changes. This information is sufficient to map points on
  example images/confidence maps to the raw images.
- Commented scaling behavior across pipeline modules.
- Changed behavior of `MultiConfidenceMaps` to compute confidence maps
  on the current image grid under the assumption that the "instances"
  key is at the same scale as the "image".

- Documented the normalization modes expected when using different
  pretrained models from tf.keras.applications.
- Still needs tests and a transformer block class.

- Track issue: aleju/imgaug#537

- Add test.

- Implements edge confidence maps; equivalent of part confidence maps
  where scalar values are scaled by the distance of the sampling grid
  points to the line segment formed by edges.
- Implement part affinity fields as simple unit vectors weighted by the
  edge confidence maps for masking.

- Transformers that operate over multiple elements of the data:
    - Shuffler
    - Batcher (this one is tricky with variable length elements)
    - Prefetcher

- BottomUpPipeline: for building standard multi-confmap + PAF pipelines
- TopDownPipeline: for building standard instance cropped pipelines

- float32 precision sometimes results in bounding box sizes that are
  #.9997, which can result in an incorrect crop, even though the
  centered bounding boxes are created with exact dimensions.

- Loads elements of the pipeline at creation time rather than during
  execution. This is a workaround to weird thread safety issues with
  using Labels as a provider during training.

- These resulted random hangs during training that could not be
  interrupted via SIGINT or otherwise.
- Solution was to remove TF ops from inside py_functions in the data
  pipeline.
- Interesting side effect is a big performance speedup.
- See: tensorflow/tensorflow#32454

- Replaces the `ds_index` argument to the `make_dataset` method
- Allows for flexible specification of subsets of the data to be used
  without having to create new source data objects (e.g., sleap.Labels).
  This makes it easier to read subsets of videos or to split the labels
  for train/val/test sets.
- Examples generated by `LabelsReader` now include an "example_ind" key
  to indicate the index of the example within the original data.

- This contains simple attr attribute classes with all user-facing
  configuration options related to training.
- This supersedes the previous TrainingJob class and should contain all
  serializable job-related parameters.
- Functional modules should be able to be instantiated from these, but
  can be decoupled in the actual implementation.

- This is necessary to build partial pipelines without providers.

- Cleaned up nn.architectures to remove architectures that need major
  updating before they can be ported, removed unused utility functions.
- We'll port in the deleted architectures from other branches when we're
  ready, but in the meanwhile they can't be used without major rewrites.
- Exception: hrnet module which is not pretrained and will require minor
  adjustments to fit new format.
- Replace old nn.model module with nn.model_config
- Removed PreprocessingConfig from nn.model since that's covered
  elsewhere in nn.config.DataConfig.
- Removed preprocessing-related args from legacy classmethod builders in
  ModelConfig.

- Useful for increasing receptive field without adding more blocks.
- Defaults still match the canonical architecture and tests.

- Config sections now sit in their own submodules, but the syntax is the
  same since all config classes are imported in nn/config/__init__.py.
- Initial implementation of ModelConfig. Totally decoupled from previous
  Model/ModelConfig implementations.
- High level architecture parameters are now explicitly specified for
  each architecture and "oneof"-style options will just be specified by
  having different attributes which validators can check for mutual
  exclusivity. This should make it easier to serialize into more
  readable json as opposed to the old system of specifying the class
  name and doing some voodoo to construct the correct one.
- This decoupling comes at the cost of reduced flexibility in model
  customization (though this can overcome with "generic" model configs
  like EncoderDecoder), as well as requiring manual specification of
  parameters for each architecture.

- This achieves validation to enforce the behavior desired for the
  mutually-exclusive attribute classes.
- Classes decorated with `attr.s` and `oneof` will raise an error if
  more than one attribute is set to not None when instantiated.

…integration

…integration

- classmethods for instantiation from config classes
- Moved old modules to *_legacy.py until TrainingJob conversion is done
- Several tweaks to configs

- This allows for (un)structuring instances without needing to write
  any extra code for instances nested in larger attrs classes.

- Removed part_metrics since this will always be enabled.

- Created standalone OHKM loss computation function for easier testing.
- Added heuristics for ratio-based dynamic hard keypoint detection.
- Fix default OHKM ratio config (should be >1)

- KerasModelPredictor supports multi-input/output and runs extremely
  fast on GPU (but the device can be set at the class-level).
- Refactored global peak finding to return coordinates in points format
  and apply thresholding before returning.
- GlobalPeakFinder transformer for generic peak finding on any confmaps
  key with stride adjustment.

- Should probably get moved somewhere else when nn is cleaned up

- Moved old training module to *_legacy.py for now
- Added several generic builder functions. These will likely be replaced
  by the TrainingJob/Trainer classes, but writing them out now to get a
  sense of where we need coupling.

- Previously the first stem block started with a max pooling directly on
  the input layer

- Should probably look into this issue more closely since it breaks
  dynamic shaped inputs (could be an issue for multi-instance
  inference potentially?)

- Implements several versions of each pipeline:
  - base_pipeline: Fetches input data with no optimization. Useful for
      composing with other pipelines and for quick inspection.
  - training_pipeline: Generates examples with all data structures
      needed for training. Uses optimizations to minimize I/O latency
      across iterations like preloading/prefetching. Applies data
      augmentation and shuffling.
  - viz_pipeline: Produces minimal examples necessary for generating
      visualizations of predictions during training. Necessarily, this
      implements basic inference, but does not have some optimizations
      that a full inference pipeline might benefit from.
- Lots of repetition but keeping it as-is for now for flexibility

…integration

…integration

If field type is optional_int (optional_double), then we'll show
a widget with a spin box (for int or double) and a checkbox to
disable the spin box. Box is checked iff field value is None.

…integration

If options are set then will show as a list (i.e., QComboBox),
otherwise will be a text field (i.e., QLineEdit).

- support for "text" field type
- for optional field types, return specified string for val when None
- add set_field_enabled method
- emit valueChanged when fields in StackBuilderWidget changed

…integration