def classify_rf(
classify_rf_params,
train_rf_result,
sigproc_result,
sim_params,
progress=None,
):
if sigproc_result.n_channels != sim_params.n_channels:
raise ValueError(
f"Sigproc n_channels ({sigproc_result.n_channels}) does not match "
f"classifier n_channels ({sim_params.n_channels})")
if sigproc_result.n_cycles != sim_params.n_cycles:
raise ValueError(
f"Sigproc n_cycles ({sigproc_result.n_cycles}) does not match "
f"classifier n_cycles ({sim_params.n_cycles})")
pred_pep_iz, scores, all_class_scores = train_rf_result.classifier.classify(
utils.mat_flatter(sigproc_result.signal_radmat()), progress)
return ClassifyRFResult(
params=classify_rf_params,
pred_pep_iz=pred_pep_iz,
scores=scores,
all_class_scores=all_class_scores,
peps_pr=peps_pr,
)
def _sig_in_range(sigproc_v2):
"""
Returns a mask indicating which signals are in range. "In range" means
able to be represented as np.float32 for the moment. The radmat is float64,
but some downstream ops (e.g. classify_rf) want to represent as float32.
Or we could just truncate these signals? But these signals really are
probably bad.
"""
finfo = np.finfo(np.float32)
max_allowed = finfo.max
min_allowed = finfo.min
radmat = utils.mat_flatter(sigproc_v2.sig())
peak_max = np.max(radmat, axis=1)
peak_min = np.min(radmat, axis=1)
in_range_mask = (peak_min > min_allowed) & (peak_max < max_allowed)
return in_range_mask
def _create_flann(dt_mat):
pyflann.set_distance_type("euclidean")
flann = pyflann.FLANN()
flann.build_index(utils.mat_flatter(dt_mat), algorithm="kdtree_simple")
return flann
def flat_test_radmat(self):
assert self.test_radmat.ndim == 3
return utils.mat_flatter(self.test_radmat)
def flat_if_requested(self, mat, flat_chcy=False):
if flat_chcy:
return utils.mat_flatter(mat)
return mat
