def load(model_dir):
"""
Loads a network from its model directory
:param model_dir: The directory
:return: The loaded RUN-CSP Network
"""
with open(os.path.join(model_dir, "parameters.json"), 'r') as f:
parameters = json.load(f)
state_size = parameters['state_size']
language = Constraint_Language.load(os.path.join(model_dir, 'language.json'))
network = RUN_CSP(model_dir, language, state_size)
return network
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
'-l',
'--language_config_path',
type=str,
help='The path to a json file that specifies the constraint language')
parser.add_argument(
'-m',
'--model_dir',
type=str,
help=
'Path to the model directory where the trained RUN-CSP instance will be stored'
)
parser.add_argument('-v',
'--n_variables',
type=int,
default=100,
help='Number of variables in each training instance.')
parser.add_argument(
'--c_min',
type=int,
default=100,
help='Minimum number of clauses in each training instance.')
parser.add_argument(
'--c_max',
type=int,
default=600,
help='Maximum number of clauses in each training instance.')
parser.add_argument('-i',
'--n_instances',
type=int,
default=4000,
help='Number of instances for training.')
parser.add_argument(
'-t',
'--t_max',
type=int,
default=30,
help=
'Number of iterations t_max for which RUN-CSP runs on each instance')
parser.add_argument('-s',
'--state_size',
type=int,
default=128,
help='Size of the variable states in RUN-CSP')
parser.add_argument('-b',
'--batch_size',
type=int,
default=10,
help='Batch size used during training')
parser.add_argument('-e',
'--epochs',
type=int,
default=25,
help='Number of training epochs')
args = parser.parse_args()
print(f'Loading constraint language from {args.language_config_path}')
language = Constraint_Language.load(args.language_config_path)
# create RUN_CSP instance for given constraint language
network = RUN_CSP(args.model_dir, language, args.state_size)
print(f'Generating {args.n_instances} training instances')
train_instances = [
CSP_Instance.generate_random(args.n_variables,
np.random.randint(args.c_min, args.c_max),
language)
for _ in tqdm(range(args.n_instances))
]
# combine instances into batches
train_batches = CSP_Instance.batch_instances(train_instances,
args.batch_size)
# train and store the network
train(network, train_batches, args.t_max, args.epochs)