Command-Line Interface

The glab command provides a high-level interface to run experiments from the command line. The command-line interface to the GLAB API can be run as follows.

$ glab [ARGUMENTS]

Note

The location of the glab script is controlled by the pip installer. On Linux, this location is $HOME/.local/bin. If glab can not be found, try adding this location to your system path.

As an alternative, this script can always be run as:

$ python -m glimpse.glab.cli [ARGUMENTS]

Arguments

Arguments to this command are described below, where they are grouped depending on their functionality.

Corpus arguments

-C DIR, --corpus-subdir=DIR
	Use DIR as a corpus sub-directory (use -C repeatedly to specify all sub-directories in corpus)
-c DIR, --corpus=DIR
	Use DIR as corpus directory.
-b, --balance	Choose equal number of images per class

Extraction arguments

-A, --save-all	Save activity for all layers, rather than just the layers from which features are extracted.
-N, --no-activity
	Do not compute activity model activity for each image (implies no classifier). This can be used to learn prototypes without immediately evaluating them.
-n NUM, --num-prototypes=NUM
	Generate NUM S2 prototypes (default: 10)
-O FILE, --options=FILE
	Read model options from FILE

Prototype learning arguments

--base-weight=w
	Add a weight of w for all training patches (default: 0.0)
--high=VALUE	Use VALUE as the high end of uniform distribution for random prototypes (default: 1.0)
--low=VALUE	Use VALUE as the low end of uniform distribution for random prototypes (default: 0.0)
--masks=DIR	Set mask directory for object_mask_wkmeans to DIR (default: ‘’)
-P FILE, --prototypes=FILE
	Read S2 prototypes from FILE (overrides -p)
-p ALG, --prototype-algorithm=ALG
	Use prototype learning algorithm ALG (one of: histogram, ica, imprint, kmeans, kmedoids, meta_feature_wkmeans, nearest_kmeans, nmf, normal, object_mask_wkmeans, pca, shuffle, sparse_pca, uniform)
--regr-samples=NUM
	Sample NUM patches when training regression model for meta_feature_wkmeans (default: 0)
--samples=NUM	Sample NUM training patches for S2 prototype learning (default: 0)

Evaluation arguments

-E, --evaluate	Train and test a classifier
-f NUM, --num-folds=NUM
	Use NUM folds for cross-validation (default: 10)
-H, --hist-features
	Use histograms (accumulated over space and scale) for each feature band (requires spatial features, such as C1)
-l LAYER, --layer=LAYER
	Extract image features from model layer LAYER (default: ‘C2’)
-L ALG, --learner=ALG
	Use learning algorithm ALG to use for classification. ALG can be a Python expression, or one of ‘svm’ or ‘logreg’) (default: ‘svm’)
-S FUNC, --score-function=FUNC
	Use scoring function FUNC for classifier evaluation (one of: accuracy or auc) (default: ‘accuracy’)
-T SIZE, --train-size=SIZE
	Set the size of the training set to SIZE (number of instances or fraction of total)
-x, --cross-validate
	Compute test accuracy via (10x10-way)cross- validation instead of fixed training/testing split

Other arguments

--command=cmd	Execute cmd after running the experiment (but before results are saved
-h, --help	Print this help and exit
-i FILE, --input=FILE
	Read initial experiment data from FILE.
-o FILE, --output=FILE
	Store results to FILE
-t TYPE, --pool-type=TYPE
	Set the worker pool type to TYPE (one of: s, singlecore, m, multicore, c, cluster)
-v, --verbose	Enable verbose logging

Examples

Evaluate with a logistic regression classifier using 100 C2 features based on imprinted prototypes.

$ glab -v -n 100 -p imprint -c cats_and_dogs --learner=logreg -E
INFO:root:Reading class sub-directories from: cats_and_dogs
INFO:root:Reading images from class directories: ['cats_and_dogs/cat', 'cats_and_dogs/dog']
INFO:root:Using pool: MulticorePool
INFO:root:Learning 100 prototypes at 1 sizes from 4 images by imprinting
Time: 0:00:01   |##################################################|   Speed:  3.10  unit/s
INFO:root:Learning prototypes took 1.294s
INFO:root:Computing C2 activation maps for 8 images
Time: 0:00:02   |##################################################|   Speed:  3.32  unit/s
INFO:root:Computing activation maps took 2.409s
INFO:root:Evaluating classifier on fixed train/test split on 8 images using 100 features from layer(s): C2
INFO:root:Training on 4 images took 0.002s
INFO:root:Classifier is Pipeline(learner=LogisticRegression([...OUTPUT REMOVED...]))
INFO:root:Classifier accuracy on training set is 1.000000
INFO:root:Scoring on training set (4 images) took 0.001s
INFO:root:Scoring on testing set (4 images) took 0.000s
INFO:root:Classifier accuracy on test set is 0.250000