helmenov/s3vm_pines

A study on multiclass S3VM for Indian Pines

This Package provides some tools for a study.

• recategorize17to10_csv : csv map to recategorize original 17 categories to 10 categories
• train_test_split() : function to split annotated data into train and test (proportion with respect to the numbers of annotated data is defined prop_train)
• labeled_unlabeled_test_split() : function to split annotated data into labeled, unlabeled, and test (unlabeled proportion with respect to the numbers of train data is defined prop_train_l)
• colored_map() : plot land cover image colored by category

And, experiment examples under mybook directory.

Using

pip install git+https://www.github.com/helmenov/s3vm_pines

and, in your python code,

import s3vm_pines

functions

s3vm_pines.module.train_test_split()

Note: train dataset is sampled aggromeratively in the map space. it is NOT random sampled.

Args:

• p_train : (float) train data proportion with respect to the number of annotated data (train:test = p_train:(1-p_train)). default is 0.5, it means train:test = 1:1
• recategorize_rule : (str) csv file name which defines recategorize rule. default is recategorize17to10_csv. if None then it means "not to recategorize": using original 17 categories.
• gt_gic : (bool) read IndianPines package. default is True

Outputs: status, status_name (tuple)

• status : (list[int]) status_number's list for each data ID
• status_names : (list[str]) status_label's for each status_number. it must be ['backgound', 'test', 'train']

It clustering by each target labels, firstly. Then sampled train data from the broadest cluster in zigzag sampling.

s3vm_pines.module.labeled_unlabeled_sample()

sample labeled and unlabeled data from train data in the each proportion. labeled data sampled in coordinately continuous, and unlabeled data sampled in random.

Args:

• p_labeled : (float): 0 < p_labeled <= 0.5, proportion of the # of labeled with respect to the # of training data.
• p_unlabeled : (flaot): 0 < p_unlabeled <= 0.5, proportion of the # of unlabeled with respect to the # of training data.
• train_test_status : (NDArray): referenced train_test_status labels which train_test_split method yields.
• unlabeled_neighbor_labeled: (bool): If true, sample unlabeled data from neighboring labeld data. Else, sample them randomly (default: False)
• seed_labeled : (float) random seed for sampling start point
• seed_unlabeled : (float) random seed for random sampling

Outputs: status, status_names (tuple)

• status : (list[int]) status_number's list for each data ID
• status_names : (list[str]) status labels for each status_number. it must be ["background", "test", "traininig_rest", "labeled", "unlabeled"].

s3vm_pines.module.labeled_unlabeled_test_split()

Args:

• prop_train_l : (float) labeled data proportion with respect to the number of train data. (not to the number of annotated data). the proportion with respect to the number of annotated data is prop_train_l * prop_train
• status : (list[int]) status_number's list for each data ID (provided from train_test_split)
• unlabeled_type: (str) 'from_train' or 'from_spatial' or 'from_annot'
  • 'from_train': unlabeled data is sampled from train data
  • 'from_annot': unlabeled data is sampled from annotated data except for labeled data in random sampling
  • 'from_spatialneighbor': unlabeled data is sampled from annotated data except for labeled data in a condition: neighboring with labeled data and the coherency is over coh_threshold
• coh_thresold: (float) coherency infmum threshold (>0, <1). high coherency means similar. low coherency means independent.
• seed_l: random seed for defining start index for sampling labeled data from training data
• seed_u: random seed for random sampling unlabeled data from annotated data except for labeled data
• recategorize_rule: you should set same value with that of train_test_split
• gt_gic: you should set same value with that of train_set_split

Outputs: l_u_t_status, l_u_t_status_name (tuple)

• l_u_t_status : (list[int]) status_number's list for each data ID
• l_u_t_status_names : (list[str]) status_label's for each status_number. it must be ['background', 'test', 'unlabeled', 'labeled']

s3vm_pines.module.colored_map()

plot land-cover colored image with the legend of categories.

target data is colored heavily and rest annoted data is colored lightly in background.

Args:

• ax: (matplotlib.pyplot.figure object) axes ploted. you should prepare ax before you use this function
• target: (list[int]) category number's list for sample ID
• cordinates: (list[list[int],list[int]]) cordinates list for sample ID. 1st column is 'x'-axis, and 2nd column is 'y'-axis
• recategorize_rule : you should set same value with that of train_test_split
• gt_gic: you should set same value with that of train_test_split

Output: None