def params_prauc_tables(h5d_fl1, h5d_fl2, curvetype, kfolds, params_od, mix, strata, trec):
"""Area Under the Curve(AUC) paired with table of parameters for PR curve.
# # # Make proper Definition here # # #
"""
# Selecting whether the resaults should be retured in binary(i.e. Trueth-Table)...
# ...or multi-class value form.
if curvetype == 'multiclass':
binary = True
else:
binary = False
# Beginning AUC-Params table building.
res_lst = list()
# Loading data in a convenient form.
for params_lst, params_path in zip(
param_comb.ParamGridIter(params_od, 'list'),
param_comb.ParamGridIter(params_od, 'path')):
# Defining list for AUC values storage. For this loop.
auc_values = list()
if params_lst[0] > params_lst[1]:
if mix:
pred_scores, expd_y, pred_y = rfse_multiclass_multimeasure_res(
h5d_fl1, h5d_fl2, kfolds, params_path, binary=binary, strata=strata
)
else:
pred_scores, expd_y, pred_y = multiclass_res(
h5d_fl1, kfolds, params_path, binary=binary, strata=strata
)
# NOTE: Crossckecking and replacing the class-tags of the experiment to virtual...
# ...class tags refering to the index of the np.unique(expd_y) vector in order...
# ...to ease the calculations of the curves.
tags2idx_ref = np.unique(expd_y)
i_fix = 0
if tags2idx_ref[0] > 0:
i_fix = 1
for i, tg in enumerate(tags2idx_ref):
expd_y[np.where(expd_y == tg)] = i + i_fix
pred_y[np.where(pred_y == tg)] = i + i_fix
# Selecting the case and calculating the precision recall curves.
if curvetype == 'multiclass':
# NOTE: Option 'is_truth_tbl' is critical to be selected correctly depending...
# ...on the input.
prec, recl, t = mx.pr_curve(
expd_y, pred_scores, full_curve=True, is_truth_tbl=True
)
# Interpolated at 11-Recall-Levels.
prec, recl = mx.reclev11_max(prec, recl, trec=trec)
elif curvetype == 'multiclass_macro':
# NOTE: Option 'unknow_class' is critical to be selected correctly depending...
# ...on the input.
prec, recl, t = mx.pr_curve_macro(
expd_y, pred_y, pred_scores, full_curve=True,
)
# Interpolated at 11-Recall-Levels.
prec, recl = mx.reclev11_max(prec, recl, trec=trec)
elif curvetype == 'onevsall':
# Finding unique genres.
gnr_tgs = np.unique(expd_y)
# Precsion and Recall scores lists of the PR curve per genre.
prec_lst = list()
recl_lst = list()
# Calculating AUC per genre tag.
for gnr in gnr_tgs:
if mix:
pred_scores, expd_y, pred_y = onevsall_multimeasure_res(
h5d_fl1, h5d_fl2, gnr, kfolds, params_path
)
else:
pred_scores, expd_y, pred_y = onevsall_res(
h5d_fl1, gnr, kfolds, params_path
)
# NOTE: Option 'is_truth_tbl' is critical to be selected correctly depending...
# ...on the input.
prec_val, recl_val, t = mx.pr_curve(
expd_y, pred_scores, full_curve=True, is_truth_tbl=False
)
# Interpolated at 11-Recall-Levels.
prec_val, recl_val = mx.reclev11_max(prec_val, recl_val, trec=trec)
# Keeping Precsion and Recall scores of the PR curve per genre.
prec_lst.append(prec)
recl_lst.append(recl)
# Calculating the PR Averaged Macro Curves values for 1-vs-All case.
prec = np.mean(np.vstack(prec_lst), axis=0)
recl = np.mean(np.vstack(recl_lst), axis=0)
else:
raise Exception('Invalide curvetype argument value.')
# Saving the AUC value and extending parameters list with AUC(s).
try:
params_lst.extend([mx.auc(recl, prec)])
except:
print "Warning:", params_path, "PR AUC is for these params has set to 0.0"
params_lst.extend([0.0])
# Appending the parameters list together with their respective AUC(s).
res_lst.append(params_lst)
# Stacking and returning the data collected in a 2D array. Last column contain the AUC for...
# ...every parameters values possible combination.
return np.vstack(res_lst)
def PRConf_table(h5d_fl1, h5d_fl2, kfolds, params_path, mix, strata, prereccon=0):
"""Precision Recall Tables and Contigency tables from H5D files.
### Make proper Definition here ###
"""
# Beginning Contingency table building
if mix:
rfse_data = rfse_multiclass_multimeasure_res(
h5d_fl1, h5d_fl2, kfolds, params_path, binary=False, strata=strata
)
else:
rfse_data = multiclass_res(
h5d_fl1, kfolds, params_path, binary=False, strata=strata
)
# 3rd element contain predicted y values list.
pred_y = rfse_data[2]
# 2rd element contain predicted y values list.
exp_y = rfse_data[1]
# Getting the expected classes.
exp_cls_tags_set = np.unique(exp_y)
# Calculating contigency table.
conf_mtrx = mx.seq_contingency_table(
exp_y, pred_y, exp_cls_tags_set=exp_cls_tags_set, arr_type=np.int32
)
if prereccon in [0, 1]:
# Calculating precision recall scores.
# Getting the number of samples per class. Zero tag is inlcuded.
smpls_per_cls = np.bincount(np.array(exp_y, dtype=np.int))
# Keeping from 1 to end array in case the expected class tags start with above zero values.
if smpls_per_cls[0] == 0 and exp_cls_tags_set[0] > 0:
smpls_per_cls = smpls_per_cls[1::]
elif smpls_per_cls[0] > 0 and exp_cls_tags_set[0] == 0:
pass # same as --> smpls_per_cls = smpls_per_cls
# Anythig else should rase an Exception.
else:
raise Exception("Samples count in zero bin is different to the expected class tag cnt!")
# Calculating Precision per class.
precisions = [
dg / float(pred_docs)
for dg, pred_docs in zip(np.diag(conf_mtrx), np.sum(conf_mtrx, axis=1))
if pred_docs > 0
]
# Calculating Recall per class.
recalls = [
dg / float(splpc)
for dg, splpc in zip(np.diag(conf_mtrx), smpls_per_cls)
if splpc > 0
]
# This funciton works only for the mx.contingency_table() output.
# pr_tbl = mx.precision_recall_scores(conf_mtrx)
pr_tbl = [precisions, recalls]
if prereccon in [0, 2]:
col_sums = conf_mtrx.sum(axis=0)
conf_mtrx = np.vstack((conf_mtrx, col_sums))
# conf_percent = np.divide(conf_mtrx, np.bincount(expected_y)) * 100
if prereccon == 0:
# Returning...
return (pr_tbl, conf_mtrx)
elif prereccon == 1:
# Returning...
return pr_tbl
elif prereccon == 2:
# Returning...
return conf_mtrx
else:
raise Exception("Returning mode 'prereccon' variable invalid value. Valid values {0,1,2}.")
params_path = plist2ppath(comb_val[4], ensbl=comb_val[0])
pred_scores, expd_y, pred_y = rfse_multiclass_multimeasure_res(
# h5d_fl1, h5d_fl2, kfolds, params_path, comb_val[4][2],
# genre_tag=None, binary=True, strata=None
h5d_fl1, h5d_fl2, kfolds, params_path, binary=False, strata=None
# binary=False <- for Micro
)
else:
# Building the parapmeters path
params_path = plist2ppath(comb_val[4], ensbl=comb_val[0])
pred_scores, expd_y, pred_y = multiclass_res(
h5d_fl1, kfolds, params_path, binary=False, strata=None
)
# Closing the h5d files.
if comb_val[3] == 'Comb':
h5d_fl1.close()
h5d_fl2.close()
else:
h5d_fl1.close()
# Creating the Actual PRC.
# y, x, t = pr_curve(expd_y, pred_scores, full_curve=True, is_truth_tbl=True)
# Creating the Actual MACRO PRC.
y, x, t = pr_curve_macro(
expd_y, pred_y, pred_scores, full_curve=True
