📊 features module¶

Factory for image feature extraction.

class pupyl.embeddings.features.Characteristics(value, names=None, *, module=None, qualname=None, type=None, start=1, boundary=None)¶

Describes high level characteristics of complex feature extractors.

Notes

The currently supported characteristics are:

MINIMUMWEIGHT_FAST_SMALL_PRECISION # MobileNet

LIGHTWEIGHT_FAST_SMALL_PRECISION # ResNet50V2

LIGHTWEIGHT_FAST_SHORT_PRECISION # ResNet101V2

LIGHTWEIGHT_QUICK_SHORT_PRECISION # DenseNet169

MEDIUMWEIGHT_QUICK_GOOD_PRECISION # DenseNet201

MIDDLEWEIGHT_QUICK_GOOD_PRECISION # InceptionV3

MIDDLEWEIGHT_SLOW_GOOD_PRECISION # Xception

HEAVYWEIGHT_SLOW_GOOD_PRECISION # EfficientNetV2M

HEAVYWEIGHT_SLOW_HUGE_PRECISION # EfficientNetV2L

static by_name(name)¶

Returns a characteristic by its name.

Parameters:: name (str) – String representation of enumerator
Raises:: UnknownCharacteristicsName: – If name is unknown.
Returns:: The corresponding characteristic.
Return type:: enum

static by_value(value)¶

Returns a characteristic by its value.

Parameters:: value (int) – Integer representing a characteristic
Raises:: UnknownCharacteristicsValue: – If value is unknown.
Returns:: The corresponding characteristic.
Return type:: enum

class pupyl.embeddings.features.Extractors(characteristics, extreme_mode=True)¶

Pretrained CNNs for embedding generation.

__enter__()¶: Opens Extractors context.

__exit__(exc_type, exc_val, exc_tb)¶: Close Extractors context.

__init__(characteristics, extreme_mode=True)¶

Creates embedding extractors.

Parameters:

characteristics (Enum) – Describing the intended characteristics to transform images into its underlying embeddings.
extreme_mode (bool) – Should the extreme mode (faster execution but not gentle with memory) be enabled or disabled?

_infer_network()¶

Translates a characteristic to a network architecture.

Raises:: UnknownCharacteristics: – If the characteristics passed through are unknown.
Returns:: With preprocessors and complete CNN architecture.
Return type:: tuple

acceleration_discovery()¶

Performs a hardware processing acceleration discovery.

Most NVIDIA® GPUs supported (through CUDA), which results on faster embeddings extraction.

extract(uri)¶

Converts image uri to its embeddings.

Parameters:: uri (str) – Location of the image to be converted to a embedding.
Returns:: 1D tensor with extracted features.
Return type:: numpy.ndarray

extract_save(path, uri)¶

Extracts features from an image, saving to disk after all.

Parameters:

path (str) – Where to store the tensor.
uri (str) – Location of the image to be converted to a embedding.

static load(path)¶

Loads up a tensor referencing index.

Parameters:: path (str) – Where to load from the tensor.
Returns:: Tensor loaded back again.
Return type:: numpy.ndarray

property output_shape¶

Getter for property output_shape.

Describes the output shape of extracted features.

Returns:: The output shape for the picked characteristics.
Return type:: int

preprocessor(uri)¶

Image preprocessing methods, suitable for posterior network ingestion. It may include image resizing, normalization, among others.

Parameters:: uri (str) – Location of the image to be preprocessed.
Returns:: Containing the processed image
Return type:: numpy.ndarray

static save(path, tensor)¶

Writes down a tensor referencing index.

Parameters:

path (str) – Where to store the tensor.
tensor (Tensor) – The tensor itself, to be saved.