Dataset Tooling

The tooling for datasets is collected in the module hybrid_learning.datasets. For details have a look at the API Reference.

Base dataset handles

BaseDataset	Abstract base class for tuple datasets with storage location.
ActivationDatasetWrapper	Wrapper for image datasets that will generate and yield activation maps.

Custom dataset handles

coco	Standard and concept datasets derived from MS COCO dataset.
fasseg	Dataset handles for FASSEG-like datasets.
broden	Handles for Broden-like datasets.

Caching handles

Cache handles will allow to insert and read objects into/from a cache. For details see Dataset Caching.

Cache	Caching base handle.
CacheCascade	Combine several caches by trying to load from first to last.
CacheDict	Cache the values of dicts using different caches.
CacheTuple	Cache the values of tuples using different caches.
DictCache	Cache objects in a (multiprocessing capable) dictionary in memory.
FileCache	Base class to cache objects as files under a cache folder.
JPGCache	Cache for JPEG images using PIL.
NPYCache	File cache that uses numpy saving and loading mechanism to cache torch.Tensor objects.
NPZCache	File cache that uses numpy compressed saving and loading mechanism to cache torch.Tensor objects.
NoCache	Dummy cache that has no effect.
PTCache	File cache that uses torch saving and loading mechanism.
TensorDictCache	In-memory cache specifically for torch tensors.

Transformations

Transformations can be used to modify data tuples or values.

Transformations for tuples

ApplyMask	Given a 3-tuple of input, target, mask apply the mask to the first two.
FlattenTuple	Return the value of a sequence at selected index.
Identity	Simple identity transformation for example for defaults.
OnAll	Apply a given transformation to all tuple items.
OnBothSides	Apply a given transformation to both input and target of a tuple in parallel.
OnIndex	Apply a given transformation to tuple items at given indices.
OnInput	Apply a given one-value transformation only to the input of a two-tuple.
OnTarget	Apply a given one-value transformation only to the target of a two-tuple.
PackMask	Given a 3-tuple of prediction, target, mask tensors merge mask and target and return a two-tuple.
ReduceTuple	Transform an unpacked input tuple and then reduce it to one value using a reduction func.
SameSize	Given a tuple of input and target image, resize the target to the size of the input.
SubsetTuple	Return a tuple only containing the elements at given indices of input tuple.
UnfoldTuple	Convenience trafo that takes an iterable and returns it as tuple.
UnpackMask	Undo a PackMask operation on the target.

Transformations for dicts

DictTransform	Basic transformation for dicts.
DropAnn	Drop the annotation with given key from the annotations dict.
FlattenDict	Return the value of the annotations dict at selected key.
OnValues	Perform a transformation on all values of a dict.
RestrictDict	Restrict the annotation dictionary to the annotation items with featuring one of the selected keys.
SameSizeTensorValues	Up- or down-scale the tensor mask values of a dictionary to all have the same size.

Transformations for (tensor) images

AsBatch	Ensure that the given transformation is fed with a batch of inputs.
BatchWiseImageTransform	Wrap a transformation operating on a batch of masks to also work on single masks.
Binarize	Simple class for binarizing tensors into high and low class values.
BinarizeByQuantile	Set all but the given highest number of pixels / q-th quantile in an image to zero, rest to 1.
ConvOpWrapper	Base wrapper class to turn convolutional batch operations into single mask operations.
ImageTransform	Transformations that can be applied to images.
IntersectDecode	IoU encode a single mask.
IntersectEncode	Intersection encode a single mask.
IoUEncode	IoU encode a single mask.
NoGrad	Disable requires_grad for the given tensors.
PadAndResize	Transformation that pads an image to a given ratio and then resizes it to fixed size.
RecursiveLambda	Generic lambda transformation that applies the given function with the standard ImageTransform recursion.
Resize	Simple resize.
Threshold	Threshold tensors and set new values below and/or above the threshold.
ToActMap	Evaluate a given image by a torch model on the correct device.
ToBBoxes	Treat pixels of given mask as scores of constant-size bounding boxes, and return a mask with the non-max-suppressed bounding boxes.
ToFixedDims	Squeeze or unsqueeze a tensor to obtain specified number of dimensions.
ToTensor	Turn objects into tensors or move tensors to given device or dtype.
WithThresh	Wrap a batch transformation with binarizing (and unsqueezing) before and after.

Intersection and intersection over union encoders

BatchBoxBloat	Bloat single pixels to full boxes, always choosing the maximum box to be up front.
BatchConvOp	Base class for encoder that use convolution operations.
BatchIntersectDecode2D	Given batch of IoU encoded masks, estimates the original segmentation mask.
BatchIntersectEncode2D	Apply intersection encoding to batch of input masks of shape (batch, 1, height, width).
BatchIoUEncode2D	Apply intersection over union encoding to an input batch.
BatchPeakDetection	Keep only peak points, i.e. ones that take the maximum value within a window around them.
BatchWindowOp	Base class for encoder that use windowing operations.

Visualization and Utility Functions

From hybrid_learning.datasets.data_visualization

apply_mask(img, mask[, color, alpha])	Apply monochrome (possibly non-binary) mask to image of same size with alpha value.
apply_masks(img, masks[, colors, alphas])
compare_masks(*masks[, colors])	Merge several monochrome masks in different colors into the same image.
mean_proportion_pos_px(masks)	From the given samples, calculate the mean of the proportion of positive pixels per mask.
neg_pixel_prop(data[, max_num_samples, ...])	Collect the mean proportion of negative pixels in the binary segmentation mask data.
to_img(tens)	Transform a (cuda) tensor to a PIL.Image.Image.
to_monochrome_img(img_t)	torch.Tensor to monochrome PIL.Image.Image in 'L' (=8-bit) mode.
visualize_classification_data(dataset[, ...])	Visualize a dataset yielding tuples of the form (input, target_class_identifier).
visualize_mask_transforms(titled_datasets[, ...])	Visualize difference between original and the transformed masks of datasets.
visualize_segmentation_data(dataset[, ...])	Visualize a dataset yielding tuples of the form (input, target_mask).

From hybrid_learning.datasets.base

add_gaussian_peak(mask_np, centroid, ...[, ...])	Add a peak to the heatmap mask_np as a non-normalized gaussian at centroid.
cross_validation_splits(train_val_data, ...)	Split dataset it into num_splits and collect tuples of (rest, split).