def get_dict(name,
data_path,
nasbench_file,
num_samples,
num_operations,
val_acc_threshold=0.,
seed=1234,
**kwargs):
nasbench_file_abs = os.path.join(data_path, nasbench_file)
print(f'Loading nasbench101: {nasbench_file_abs}')
nasbench = api.NASBench(nasbench_file_abs,
num_samples=num_samples,
seed=seed)
archs = []
seqs = []
valid_accs = []
all_keys = list(nasbench.hash_iterator())
dataset_dicts = []
min_val_acc = float('inf')
max_val_acc = 0
min_data = max_data = None
for idx, key in enumerate(all_keys):
fixed_stat, computed_stat = nasbench.get_metrics_from_hash(key)
if len(fixed_stat['module_operations']) not in num_operations:
continue
arch = api.ModelSpec(matrix=fixed_stat['module_adjacency'],
ops=fixed_stat['module_operations'])
data = nasbench.query(arch)
if data['validation_accuracy'] < val_acc_threshold:
continue
if min_val_acc > data['validation_accuracy']:
min_val_acc = data['validation_accuracy']
min_data = data
if max_val_acc < data['validation_accuracy']:
max_val_acc = data['validation_accuracy']
max_data = data
data["id"] = idx
dataset_dicts.append(data)
meta_dict = {}
meta_dict['num_samples'] = len(dataset_dicts)
meta_dict['num_operations'] = num_operations
meta_dict['min_val_acc'] = min_val_acc
meta_dict['max_val_acc'] = max_val_acc
meta_dict['min_data'] = min_data
meta_dict['max_data'] = max_data
print(f'min_val_acc: {min_val_acc}, max_val_acc: {max_val_acc}')
if name not in MetadataCatalog.list():
MetadataCatalog.get(name).set(**meta_dict)
return dataset_dicts
def register_dataset(self):
# register the training dataset, only need to do this once
catalog_list = MetadataCatalog.list()
if 'wb_train' not in catalog_list:
register_coco_instances("wb_train", {}, "roboflow/train/_annotations.coco.json", "/roboflow/train")
if 'wb_val' not in catalog_list:
register_coco_instances("wb_val", {}, "roboflow/valid/_annotations.coco.json", "/roboflow/valid")
if 'wb_test' not in catalog_list:
register_coco_instances("wb_test", {}, "roboflow/test/_annotations.coco.json", "/roboflow/test")
cv2_imshow(masked_image)
else:
if save_out:
cv2.imwrite(os.path.join(image_outpath, filename), im_input)
else:
cv2_imshow(im_input)
check_annotations(image_infile, im_input)
print("\tSaving to ", os.path.join(image_outpath, filename))
return predictions, vis_output
"""
python model_segm.py --input
"""
if __name__ == "__main__":
print("Metadata List: ", MetadataCatalog.list())
input_folders = _PREDEFINED_SPLITS_GRC_MD["rdd2020_source"][
"rdd2020_train"]
try:
for folder in input_folders:
print("\n----------- ", folder, "------------\n")
#image_filepath = os.path.join(ROADDAMAGE_DATASET, "train/Japan/images", "Japan_000000.jpg") # single image
image_filepath = os.path.join(ROADDAMAGE_DATASET, folder,
"images") # in directory
image_outpath = os.path.join(image_filepath,
"../images_segm") # out directory
os.makedirs(image_outpath, exist_ok=True)
if os.path.isdir(image_filepath):
for id, imfile in enumerate(
sorted(glob.glob(os.path.join(
image_filepath, '*.jpg')))): # assuming jpg images
from detectron2.data import DatasetCatalog, MetadataCatalog
import copy
from .build import DATASET_MAPPER_REGISTRY
@DATASET_MAPPER_REGISTRY.register()
class NoneMapper(object):
def __init__(self, cfg, **kwargs):
pass
def __call__(self, dataset_dict):
dataset_dict = copy.deepcopy(
dataset_dict) # it will be modified by code below
return dataset_dict
DatasetCatalog.register('NoneDataset', (lambda: [{'test': 'test'}] * 10))
if 'NoneDataset' not in MetadataCatalog.list():
MetadataCatalog.get('NoneDataset').set(**{'num_samples': 10})
