ModelIOConfigABC¶

class ModelIOConfigABC(input_resolutions, patch_input_shape, stride_shape=None, output_resolutions=<factory>)[source]¶

Defines a data class for holding a deep learning model’s I/O information.

Enforcing such that following attributes must always be defined by the subclass.

Parameters:

  • input_resolutions (list(dict)) – Resolution of each input head of model inference, must be in the same order as target model.forward().

  • patch_input_shape (numpy.ndarray, list(int), tuple(int, int)) – Shape of the largest input in (height, width).

  • stride_shape (numpy.ndarray, list(int), tuple(int)) – Stride in (x, y) direction for patch extraction.

  • output_resolutions (list(dict)) – Resolution of each output head from model inference, must be in the same order as target model.infer_batch().

input_resolutions¶

Resolution of each input head of model inference, must be in the same order as target model.forward().

Type:

list(dict)

patch_input_shape¶

Shape of the largest input in (height, width).

Type:

numpy.ndarray, list(int), tuple(int, int)

stride_shape¶

Stride in (x, y) direction for patch extraction.

Type:

numpy.ndarray, list(int), tuple(int, int)

output_resolutions¶

Resolution of each output head from model inference, must be in the same order as target model.infer_batch().

Type:

list(dict)

highest_input_resolution¶

Highest resolution to process the image based on input and output resolutions. This helps to read the image at the optimal resolution and improves performance.

Type:

dict

Examples

>>> # Defining io for a base network and converting to baseline.
>>> ioconfig = ModelIOConfigABC(
...     input_resolutions=[{"units": "mpp", "resolution": 0.5}],
...     output_resolutions=[{"units": "mpp", "resolution": 1.0}],
...     patch_input_shape=(224, 224),
...     stride_shape=(224, 224),
... )
>>> ioconfig = ioconfig.to_baseline()

Methods

scale_to_highest

Get the scaling factor from input resolutions.

to_baseline

Returns a new config object converted to baseline form.

Attributes

stride_shape

input_resolutions

patch_input_shape

output_resolutions

static scale_to_highest(resolutions, units)[source]¶

Get the scaling factor from input resolutions.

This will convert resolutions to a scaling factor with respect to the highest resolution found in the input resolutions list. If a model requires images at multiple resolutions. This helps to read the image a single resolution. The image will be read at the highest required resolution and will be scaled for low resolution requirements using interpolation.

Parameters:

  • resolutions (list(dict(Units, Resolution))) – A list of resolutions where one is defined as {‘resolution’: value, ‘unit’: value}

  • units (Units) – Resolution units.

Returns:

A 1D array of scaling factors having the same length as resolutions.

Return type:

numpy.ndarray

Examples

>>> # Defining io for a base network and converting to baseline.
>>> ioconfig = ModelIOConfigABC(
...     input_resolutions=[
...                {"units": "mpp", "resolution": 0.25},
...                {"units": "mpp", "resolution": 0.5},
...                ],
...     output_resolutions=[{"units": "mpp", "resolution": 1.0}],
...     patch_input_shape=(224, 224),
...     stride_shape=(224, 224),
... )
>>> ioconfig = ioconfig.scale_to_highest()
... array([1. , 0.5])  # output
>>>
>>> # Defining io for a base network and converting to baseline.
>>> ioconfig = ModelIOConfigABC(
...     input_resolutions=[
...                {"units": "mpp", "resolution": 0.5},
...                {"units": "mpp", "resolution": 0.25},
...                ],
...     output_resolutions=[{"units": "mpp", "resolution": 1.0}],
...     patch_input_shape=(224, 224),
...     stride_shape=(224, 224),
... )
>>> ioconfig = ioconfig.scale_to_highest()
... array([0.5 , 1.])  # output
to_baseline()[source]¶

Returns a new config object converted to baseline form.

This will return a new ModelIOConfigABC where resolutions have been converted to baseline format with the highest possible resolution found in both input and output as reference.

Parameters:

self (ModelIOConfigABC)

Return type:

ModelIOConfigABC