def main():
args = get_macro_args()
if args.recommended_arch:
filename = args.recommended_arch
ctx = get_extension_context(args.context,
device_id=args.device_id,
type_config=args.type_config)
nn.set_default_context(ctx)
ext = nn.ext_utils.import_extension_module(args.context)
data_iterator = data_iterator_cifar10
tdata = data_iterator(args.batch_size, True)
vdata = data_iterator(args.batch_size, False)
mean_val_train, std_val_train, channel, img_height, img_width, num_class = get_data_stats(
tdata)
mean_val_valid, std_val_valid, _, _, _, _ = get_data_stats(vdata)
data_dict = {
"train_data": (tdata, mean_val_train, std_val_train),
"valid_data": (vdata, mean_val_valid, std_val_valid),
"basic_info": (channel, img_height, img_width, num_class)
}
check_arch = np.load(filename, allow_pickle=True)
print("Train the model whose architecture is:")
show_arch(check_arch)
val_acc = CNN_run(args,
check_arch.tolist(),
data_dict,
with_train=True,
after_search=True)
def main():
"""
Start architecture search and save the architecture found by the controller during the search.
"""
args = get_macro_args()
arguments_assertion(args)
ctx = get_extension_context(args.context,
device_id=args.device_id,
type_config=args.type_config)
nn.set_default_context(ctx)
ext = nn.ext_utils.import_extension_module(args.context)
if args.sampling_only:
sample_from_pretrained_controller(args)
return
data_iterator = data_iterator_cifar10
tdata = data_iterator(args.batch_size, True)
vdata = data_iterator(args.batch_size, False)
mean_val_train, std_val_train, channel, img_height, img_width, num_class = get_data_stats(
tdata)
mean_val_valid, std_val_valid, _, _, _, _ = get_data_stats(vdata)
data_dict = {
"train_data": (tdata, mean_val_train, std_val_train),
"valid_data": (vdata, mean_val_valid, std_val_valid),
"basic_info": (channel, img_height, img_width, num_class)
}
initializer = I.UniformInitializer((-0.1, 0.1))
# Prepare all the weights in advance
controller_weights_and_shape = {
'controller_lstm/0/lstm/affine/W':
(2 * args.lstm_size, 4, args.lstm_size),
'controller_lstm/0/lstm/affine/b': (4, args.lstm_size),
'controller_lstm/1/lstm/affine/W':
(2 * args.lstm_size, 4, args.lstm_size),
'controller_lstm/1/lstm/affine/b': (4, args.lstm_size),
'ops/affine/W': (args.lstm_size, args.num_ops),
'skip_affine_1/affine/W': (args.lstm_size, args.lstm_size),
'skip_affine_2/affine/W': (args.lstm_size, 1),
'skip_affine_3/affine/W': (args.lstm_size, args.lstm_size)
}
for w_name, w_shape in controller_weights_and_shape.items():
nn.parameter.get_parameter_or_create(w_name,
w_shape,
initializer=initializer,
need_grad=True)
# create dictionary of controller's weights
controller_weights_dict = {
w_name: nn.get_parameters()[w_name]
for w_name in controller_weights_and_shape.keys()
}
arch_change, best_arch = search_architecture(args, data_dict,
controller_weights_dict)
if args.select_strategy == "best":
print(
"saving the model which achieved the best validation accuracy as {}."
.format(args.recommended_arch))
check_arch = best_arch
else:
# Use the latest architecture. it's not necessarily the one with the best architecture.
print("saving the latest model recommended by the controller as {}.".
format(args.recommended_arch))
check_arch = arch_change[-1]
np.save(args.recommended_arch, np.array(check_arch))
print("The saved architecture is;")
show_arch(check_arch)
print("when you want to train the network from scratch,\n\
type 'python macro_retrain.py <OPTION> --recommended-arch {}'".format(
args.recommended_arch))
# save the controller's weights so that another architectures can be made.
all_params = nn.get_parameters(grad_only=False)
controller_weights = list(controller_weights_and_shape.keys()) + ["w_emb"]
for param_name in all_params.keys():
if param_name not in controller_weights:
nn.parameter.pop_parameter(param_name)
nn.save_parameters(
os.path.join(args.model_save_path, 'controller_params.h5'))
# If you want to train the model recommended by the controller from scratch
# right after architecture search, uncomment the lines below
# nn.clear_parameters()
# ext.clear_memory_cache() # clear all the Variables
# val_acc = CNN_run(args, check_arch, data_dict, with_train=True, after_search=True)
return