Feature Extractor radtorch.extractor

FeatureExtractor

Feature Extractor performs feature extraction from images using a pytorch model pretrained on ImageNet. Features can be accessed using below attributes after running the feature extraction process through the .run() method.

Parameters:

Name	Type	Description	Default
model_arch	str	Model architecture to be used for feature extraction.	required
dataset	ImageDataset	ImageDataset to be used for training.	required
subset	str	the subset op the dataset to extract features from. Default: 'train'. Options: 'train', 'valid', 'test'.	'train'
device	str	Device to be used for training. Default: 'auto' which automtically detects GPU presence and uses it for feature extraction. Options: 'auto', 'cuda', 'cpu'.	'auto'

Attributes:

Name	Type	Description
loader	pytorch dataloader object	Training pytorch DataLoader.
table	pandas dataframe	the table of images to be used for feature extraction.
model	pytorch neural network	Instance of the pytorch model to be used for feature extraction.
features	pandas dataframe	table of extracted features.
feature_table	pandas dataframe	table of extracted features and corresponding uid for each image instance.
feature_names	list	names of the extracted features.

Examples:

import radtorch
import albumentations as A

ds = radtorch.data.ImageDataset('data/PROTOTYPE/FILE/', transforms={'train': A.Compose([A.Resize(64,64)])})

ext = radtorch.extractor.FeatureExtractor('vgg16', dataset=ds)
ext.run()

Methods

hybrid_table(self, sklearn_ready=False, label_id=True)

Use this method to create pandas dataframes of features and labels that can be used directly into training using scikit-learn.

Parameters:

Name	Type	Description	Default
sklearn_ready	bool	True returns a tuple of extracted features dataframe and labels dataframe. False returns table of features, uid, path, label and label_id.	False
label_id	bool	True returns the label ids (integer) instead of the label string.	True

num_features(self)

Returns the expected number of features to be extracted.

run(self)

Runs the feature extraction process