def add_default_args(parser: ArgumentParser):
parser.add_argument('dataset_name',
choices=list(DATASETS.keys()),
type=str,
help='dataset name')
# Network
parser.add_argument('--backbone',
type=str,
choices=BACKBONES,
default=Backbone.RESNET_50,
help='Backbone network type.')
parser.add_argument('--pretrained',
type=str2bool,
default=True,
help='No pretraining.')
# Others
parser.add_argument('--set_determinism', type=str2bool, default=False)
parser.add_argument('--test_batch', default=32, type=int)
parser.add_argument('--img_test_batch', default=512, type=int)
parser.add_argument('--verbose',
type=str2bool,
default=True,
help='Debug mode')
parser.add_argument('-j', '--workers', default=4, type=int)
parser.add_argument('--p', type=int, default=18, help='')
parser.add_argument('--k', type=int, default=4, help='')
parser.add_argument('--num_test_images', type=int, default=8)
return parser
def main(_argv):
datasets = DATASETS.keys()
if len(FLAGS.dataset) > 0:
datasets = FLAGS.dataset
items = SPLIT_OPTIONS
if len(FLAGS.train_skip) > 0:
items = {
train_skip: SPLIT_OPTIONS[train_skip] for train_skip in FLAGS.train_skip
}
methods = METHODS.values()
if len(FLAGS.method) > 0:
methods = [METHODS[method] for method in FLAGS.method]
for model_version in FLAGS.model_version:
for dataset in datasets:
for label, skip in items.items():
weights_dir = os.path.join(
model_folder,
f"{model_version}-{dataset}-{label}.pth",
)
if (
FLAGS.weights is not None
and len(FLAGS.train_skip) == 0
and len(FLAGS.dataset) == 0
):
weights_dir = FLAGS.weights
for method in methods:
Path(
os.path.join(
out_folder,
dataset,
f"{model_version}-{label}",
method,
)
).mkdir(parents=True, exist_ok=True)
step = 1
if dataset == DATASETS["food101"]:
step = 5
if dataset == DATASETS["stanford-dogs"]:
step = 2
if dataset == DATASETS["plant-data"]:
step = 2
measure_model(
model_version,
dataset,
os.path.join(
out_folder,
dataset,
f"{model_version}-{label}",
method,
),
weights_dir,
device,
method=method,
step=step,
)
def main(_argv):
for model_version in FLAGS.model_version:
for dataset in DATASETS.keys():
for label, skip in SPLIT_OPTIONS.items():
print(
f"Training {model_version} model with {label} of data ({dataset})"
)
if model_version in AVAILABLE_MODELS[:1]:
trained_model = train_resnet(dataset,
model_version,
data_dir=data_dir,
skip=skip)
if model_version == AVAILABLE_MODELS[2]:
trained_model = train_efficientnet(dataset,
model_version,
data_dir=data_dir,
skip=skip)
if model_version == AVAILABLE_MODELS[3]:
trained_model = train_densenet(dataset,
model_version,
data_dir=data_dir,
skip=skip)
print(
f"Saving model to '{os.path.join(model_folder, f'{model_version}-{dataset}-{label}.pth')}'"
)
torch.save(
trained_model.state_dict(),
os.path.join(model_folder,
f"{model_version}-{dataset}-{label}.pth"),
)
print(
f"Testing {model_version} model trained on {label} of data ({dataset})"
)
weights_dir = os.path.join(
model_folder, f"{model_version}-{dataset}-{label}.pth")
test_model(model_version, dataset, out_folder, weights_dir,
device, label)
def main(_argv):
datasets = DATASETS.keys()
if len(FLAGS.dataset) > 0:
datasets = FLAGS.dataset
items = SPLIT_OPTIONS
if len(FLAGS.train_skip) > 0:
items = {
train_skip: SPLIT_OPTIONS[train_skip]
for train_skip in FLAGS.train_skip
}
methods = METHODS.values()
if len(FLAGS.method) > 0:
methods = [METHODS[method] for method in FLAGS.method]
ids = None
if FLAGS.image_ids is not None:
ids = [int(x) - 1 for x in FLAGS.image_ids.split(',')]
for model_version in FLAGS.model_version:
for dataset in datasets:
for label, skip in items.items():
weights_dir = os.path.join(
model_folder,
f"{model_version}-{dataset}-{label}.pth",
)
if (FLAGS.weights is not None and len(FLAGS.train_skip) == 0
and len(FLAGS.dataset) == 0):
weights_dir = FLAGS.weights
for method in methods:
Path(
os.path.join(
out_folder,
dataset,
f"{model_version}-{label}",
method,
)).mkdir(parents=True, exist_ok=True)
step = 20
if dataset == DATASETS["food101"]:
step = 1300
if dataset == DATASETS["stanford-dogs"]:
step = 250
if dataset == DATASETS["plant-data"]:
step = 250
measure_filter_model(
model_version,
dataset,
os.path.join(
out_folder,
dataset,
f"{model_version}-{label}",
method,
),
weights_dir,
device,
method=method,
step=step,
use_infidelity=FLAGS.use_infidelity,
use_sensitivity=FLAGS.use_sensitivity,
render=FLAGS.render_results,
ids=ids,
)
import warnings
warnings.filterwarnings("ignore")
FLAGS = flags.FLAGS
flags.DEFINE_multi_enum(
"model_version",
["resnet18"],
AVAILABLE_MODELS,
f"Model version {AVAILABLE_MODELS}",
)
flags.DEFINE_multi_enum(
"dataset",
[],
list(DATASETS.keys()),
f"(optional) Dataset name, one of available datasets: {list(DATASETS.keys())}",
)
flags.DEFINE_multi_enum(
"train_skip",
[],
list(SPLIT_OPTIONS.keys()),
f"(optional) version of the train dataset size: {list(SPLIT_OPTIONS.keys())}",
)
flags.DEFINE_multi_enum(
"method",
[],
list(METHODS.keys()),
f"(optional) select on of available methods: {list(METHODS.keys())}",
)
flags.DEFINE_string(