def add_experiment_args(parser: ArgumentParser) -> None:
"""
Adds the arguments used by all the models.
:param parser: the parser instance
"""
parser.add_argument('--dataset',
type=str,
required=True,
choices=DATASET_NAMES,
help='Which dataset to perform experiments on.')
parser.add_argument('--model',
type=str,
required=True,
help='Model name.',
choices=get_all_models())
parser.add_argument('--device', type=str, default=None)
parser.add_argument('--lr',
type=float,
required=True,
help='Learning rate.')
parser.add_argument('--batch_size',
type=int,
required=True,
help='Batch size.')
parser.add_argument('--n_epochs',
type=int,
required=True,
help='The number of epochs for each task.')
def main():
lecun_fix()
parser = ArgumentParser(description='mammoth', allow_abbrev=False)
parser.add_argument('--model',
type=str,
required=True,
help='Model name.',
choices=get_all_models())
parser.add_argument('--load_best_args',
action='store_true',
help='Loads the best arguments for each method, '
'dataset and memory buffer.')
add_management_args(parser)
args = parser.parse_known_args()[0]
mod = importlib.import_module('models.' + args.model)
if args.load_best_args:
parser.add_argument('--dataset',
type=str,
required=True,
choices=DATASET_NAMES,
help='Which dataset to perform experiments on.')
if hasattr(mod, 'Buffer'):
parser.add_argument('--buffer_size',
type=int,
required=True,
help='The size of the memory buffer.')
args = parser.parse_args()
if args.model == 'joint':
best = best_args[args.dataset]['sgd']
else:
best = best_args[args.dataset][args.model]
if args.model == 'joint' and args.dataset == 'mnist-360':
args.model = 'joint_gcl'
if hasattr(args, 'buffer_size'):
best = best[args.buffer_size]
else:
best = best[-1]
for key, value in best.items():
setattr(args, key, value)
else:
get_parser = getattr(mod, 'get_parser')
parser = get_parser()
args = parser.parse_args()
if args.seed is not None:
set_random_seed(args.seed)
if args.model == 'mer': setattr(args, 'batch_size', 1)
dataset = get_dataset(args)
backbone = dataset.get_backbone()
loss = dataset.get_loss()
model = get_model(args, backbone, loss, dataset.get_transform())
if isinstance(dataset, ContinualDataset):
train(model, dataset, args)
else:
assert not hasattr(model, 'end_task') or model.NAME == 'joint_gcl'
ctrain(args)
def show_options():
from models import get_all_models
from samplers import get_all_multi_samplers, get_all_single_samplers
print("models")
print((',\n').join(get_all_models()))
print("single samplers")
print((',\n').join(get_all_single_samplers()))
print("multi samplers")
print((',\n').join(get_all_multi_samplers()))
def create_tables():
db.connect_db()
# creates tables if not exist
db.database.create_tables(get_all_models(), True)
def main():
if torch.cuda.device_count() > 1:
torch.set_num_threads(6 * torch.cuda.device_count())
else:
torch.set_num_threads(2)
parser = ArgumentParser(description='mammoth', allow_abbrev=False)
parser.add_argument('--model',
type=str,
required=True,
help='Model name.',
choices=get_all_models())
parser.add_argument('--load_best_args',
action='store_true',
help='Loads the best arguments for each method, '
'dataset and memory buffer.')
add_management_args(parser)
args = parser.parse_known_args()[0]
mod = importlib.import_module('models.' + args.model)
if args.load_best_args:
parser.add_argument('--dataset',
type=str,
required=True,
choices=DATASET_NAMES,
help='Which dataset to perform experiments on.')
if hasattr(mod, 'Buffer'):
parser.add_argument('--buffer_size',
type=int,
required=True,
help='The size of the memory buffer.')
args = parser.parse_args()
model = args.model
if model == 'joint':
model = 'sgd'
best = best_args[args.dataset][model]
if hasattr(args, 'buffer_size'):
best = best[args.buffer_size]
else:
best = best[-1]
for key, value in best.items():
setattr(args, key, value)
else:
get_parser = getattr(mod, 'get_parser')
parser = get_parser()
args = parser.parse_args()
if args.seed is not None:
set_random_seed(args.seed)
off_joint = False
if args.model == 'joint' and args.dataset == 'seq-core50':
args.dataset = 'seq-core50j'
args.model = 'sgd'
off_joint = True
dataset = get_dataset(args)
# continual learning
backbone = dataset.get_backbone()
loss = dataset.get_loss()
model = get_model(args, backbone, loss, dataset.get_transform())
if off_joint:
print('BEGIN JOINT TRAINING')
jtrain(model, dataset, args)
else:
print('BEGIN CONTINUAL TRAINING')
train(model, dataset, args)
def get_cfg(name):
module = sys.modules[__name__]
name = name.lower()
try:
return getattr(module, '{}_cfg'.format(name))
except Exception as e:
raise ValueError("Model config {} not found".format(name))
def build_model(name):
cfg = get_cfg(name)
cfg['name'] = name
cfg['pretrained'] = False
return model_builder.build(cfg)
@pytest.mark.parametrize("model", get_all_models())
def test_model(model):
model = build_model(model)
model.eval()
test_input = torch.rand(1, 3, 256, 128)
with torch.no_grad():
endpoints = model(test_input, model.endpoints)
print(model.create_endpoints())
print(model.dimensions)
# foward one image compare to dimensions
for key, dim in model.dimensions.items():
assert endpoints[key].shape[1:] == dim, key