def save_class_scores_hdf(path, bin_id, scores, roi_ids, classes):
np_scores = np.array(scores)
#print('length compare: len(scores)==len(inputs) -> {}=={} -> {}'.format(len(scores),len(roi_ids),len(scores)==len(roi_ids)))
#print('class compare: len(score[0])==len(classes) -> {}=={} -> {}'.format(len(scores[0]),len(classes),len(scores[0])==len(classes)))
assert np_scores.shape[0] == len(roi_ids), 'wrong number of ROI numbers'
assert np_scores.shape[1] == len(classes), 'wrong number of class labels'
roi_numbers = [ifcb.Pid(roi_id).target for roi_id in roi_ids]
try:
with h5.File(path, 'w') as f:
ds = f.create_dataset('scores',
data=np_scores,
compression='gzip',
dtype='float16')
ds.attrs['bin_id'] = bin_id
ds.attrs['class_labels'] = [l.encode('ascii') for l in classes]
f.create_dataset('roi_numbers',
data=roi_numbers,
compression='gzip',
dtype='int16')
except RuntimeError as e:
# RuntimeError: Unable to create attribute (object header message is too large)
# see: https://github.com/h5py/h5py/issues/1053#issuecomment-525363860
print(' ', type(e), e, 'SWITCHING TO CSV OUTPUT')
path = path.replace('.h5', '.csv').replace('.hdf', '.csv')
save_class_scores_csv(path, bin_id, scores, roi_ids, classes)
def save_class_scores_mat(path, bin_id, scores, roi_ids, classes):
from scipy.io import savemat
np_scores = np.array(scores)
roi_numbers = [ifcb.Pid(roi_id).target for roi_id in roi_ids]
d = {
'version': 2,
'roinum': roi_numbers,
'TBscores': np_scores.astype('f4'),
'class2useTB': np.asarray(classes + ['unclassified'], dtype='object')
}
savemat(path, d, do_compression=True)
def save_run_results(input_images,
output_scores,
class_labels,
timestamp,
outdir,
outfile,
model_id=None,
input_obj=None):
output_classranks = np.max(output_scores, axis=1)
output_classes = np.argmax(output_scores, axis=1)
assert output_scores.shape[0] == len(
output_classes), 'wrong number inputs-to-outputs'
assert output_scores.shape[1] == len(
class_labels), 'wrong number of class labels'
results = dict(version='v3',
model_id=model_id,
timestamp=timestamp,
class_labels=class_labels,
input_images=input_images,
output_classes=output_classes,
output_scores=output_scores)
outfile = os.path.join(outdir, outfile)
if isinstance(input_obj, ifcb.Pid):
bin_obj = input_obj
results['bin_id'] = bin_obj.pid
results['roi_numbers'] = [ifcb.Pid(img).target for img in input_images]
outfile_dict = dict(BIN_ID=bin_obj.pid,
INPUT_SUBDIRS=bin_obj.namespace,
BIN_YEAR=bin_obj.year,
BIN_DATE=bin_obj.yearday)
outfile = outfile.format(**outfile_dict).replace(2 * os.sep, os.sep)
os.makedirs(os.path.dirname(outfile), exist_ok=True)
_save_run_results(outfile, results)
else: # ImageDataset
if '{INPUT_SUBDIRS}' in outfile:
dir_groups = {}
input_src = input_obj if os.path.isdir(input_obj) else ''
for img_path, img_classidx, img_scores in zip(
input_images, output_classes, output_scores):
parent_dir = os.path.dirname(img_path.replace(input_src,
'')) #+os.sep
if parent_dir not in dir_groups:
dir_groups[parent_dir] = {
k: v if k not in [
'input_images', 'output_classes', 'output_scores'
] else []
for k, v in results.items()
}
dir_groups[parent_dir]['input_images'].append(
os.path.basename(img_path))
dir_groups[parent_dir]['output_classes'].append(img_classidx)
dir_groups[parent_dir]['output_scores'].append(img_scores)
for parent_dir, sub_results in dir_groups.items():
sub_outfile = outfile.format(INPUT_SUBDIRS=parent_dir)
os.makedirs(os.path.dirname(sub_outfile), exist_ok=True)
sub_results['output_classes'] = np.asarray(
sub_results['output_classes'],
dtype=results['output_classes'].dtype)
sub_results['output_scores'] = np.asarray(
sub_results['output_scores'],
dtype=results['output_scores'].dtype)
_save_run_results(sub_outfile, sub_results)
else: #easy
os.makedirs(os.path.dirname(outfile), exist_ok=True)
_save_run_results(outfile, results)
def target_id(self, target_number):
return ifcb.Pid(self.pid).with_target(target_number)