Changelog of the Inference YAML Configuration

Version 8

This version removes the Sampling section from the configuration file.

The Sampling section was used to specify the splitting and recombination strategy for the input image, in cases where the input image are too large to be processed in a single pass. This is now handled by dedicated operations that can be specified more flexibly in the PreProcessing and PostProcessing sections respectively.

More specifically, the SplitIntoPatchesOperation is used to split the input image into patches, while the RecombinePatches operation is used to recombine the patches into a single prediction image. When the input image needs to be padded to a multiple of the DimensionDivisor, the PadToNextMultiple operation is used to pad the input image. When the model prediction needs to be unpadded, the Inverse operation is used to remove the padding from the model prediction.

Warning

The new mechanism uses a more efficient implementation of the splitting and recombination logic, which can lead to slightly different final predictions. As such, upgrading a configuration file to version 8 is a breaking change for existing models.

In order to prevent unwanted regressions, the yaml configuration file for existing models is not automatically updated to version 8. If you want to let the yaml configuration file be updated automatically to version 8, you need to manually set the flag UpgradeToMLModelV2: true (default is false).

Upgrading to version 8 will automatically remove any specified Sampling section and convert it to the corresponding pre- and post-processing operations.

In particular, the following logic is applied:

If Sampling.MaxSizeSubdivision was set, the SplitIntoPatchesOperation operation is appended to the PreProcessing section and the RecombinePatches operation is inserted as first operation in the PostProcessing section.
If Sampling.MaxSizeSubdivision was not set and Sampling.DimensionDivisor was set, a PadToNextMultiple operation is added to the PreProcessing section.
If DimensionDivisor was set together with MaxSizeSubdivision, it is checked that MaxSizeSubdivision is a multiple of the DimensionDivisor, but no extra padding is added to the PreProcessing.
Sampling.PaddingMode is used to configure any potential padding of the input image.
Sampling.RecombineWeighted and Sampling.RecombineDevice are used to configure the recombination of the patches.
Any other parameter of the Sampling section not mentioned above is ignored.

Example 1

The following configuration example with splitting and recombination:

Version: 7
PreProcessing:
  ...
PostProcessing:
  ...
Sampling:
- DimensionDivisor: 32
- MaxSizeSubdivision: 96
- PixelsOverlap: 16
- RecombineWeighted: true
- PaddingMode: Mirror

is converted to:

Version: 8
PreProcessing:
  ... # operations from previous version
  - SplitIntoPatchesOperation:
      patch_size: [96, 96, 96]
      patch_step_size: 1.0
      device: GPUIfOpenGL # or ForceCPU

PostProcessing:
  - RecombinePatches:
      device: GPUIfOpenGL # or ForceCPU
      mode: Weighted
  ... # operations from previous version

Example 2

The following configuration example with padding and unpadding:

Version: 7
PreProcessing:
  ...
PostProcessing:
  ...

Sampling:
- DimensionDivisor: 32
- MaxSizeSubdivision: -1
- PixelsOverlap: 16
- PaddingMode: Mirror
- SkipUnpadding: false

is converted to:

Version: 8
PreProcessing:
  ... # operations from previous version
  - PadToNextMultiple: # new operation for padding the input image to a multiple of the DimensionDivisor
      dimension_divisor: 32
      padding_mode: Mirror
      allow_dimension_change: false
      record_identifier: pad_to_next_dim_ev7to8_1 # identifier of the padding operation, used to identify which operation instance to inverse in the PostProcessing

PostProcessing:
  - Inverse: # new operation for removing the padding from the model prediction
      target_identifier: pad_to_next_dim_ev7to8_1 # identifier of the target operation to be inverted
  ... # operations from previous version

Note

To update an old configuration file to the latest version, you can load it by instantiating a ModelConfiguration object and then use the ModelConfiguration::save() function to save it to disk. The save function will write a configuration file with the latest version.

If you need to bulk upgrade a large number of configurations, you can use the convenience function imfusion.machinelearning.update_model_configuration in the imfusion-sdk python package, or use it directly from the command line with python -m imfusion.machinelearning.update_model_configuration --input <input/path.yaml> --output <output/path.yaml>.

Version 7

Sampling.EnsureInputImageMultipleOfDivisor is added to control the behavior of padding the input image to a multiple of the DimensionDivisor. This behavior was default in versions 6 and less, but from version 7 will default to false.

Version 6

All engine related parameters are now listed in a subsection under the keyword Engine. The engine name is specified by the field Name.

Version 5

Sampling.SkipUnpadding now defaults to true.
Sampling.MirrorPadding is deprecated, use PaddingMode: Mirror instead. Other possible padding modes are Zero or Clamp.

Version 4

We no longer automatically insert a MakeFloat operation as first step of the PreProcessing sequence.

Version 3

Parameter Type now strictly refers to the type of model, and can take the values NeuralNetwork, RandomForest, and NeuralNetworkLegacy. Previously, type was sometimes used as PredictionType. This is still supported, but will be discontinued at some point.
Model configuration in the format .configtxt are no longer supported.