def convert_and_test(network_type, train_weights):
builder = ACNetBuilder(base_config=None, deploy=False)
general_test(network_type=network_type,
weights=train_weights,
builder=builder)
deploy_weights = train_weights.replace('.hdf5', '_deploy.hdf5')
convert_acnet_weights(train_weights,
deploy_weights=deploy_weights,
eps=1e-5)
builder.switch_to_deploy()
general_test(network_type=network_type,
weights=deploy_weights,
builder=builder)
def acnet_cfqkbnc():
try_arg = start_exp()
network_type = 'cfqkbnc'
dataset_name = 'cifar10'
log_dir = 'acnet_exps/{}_{}_train'.format(network_type, try_arg)
save_weights = 'acnet_exps/{}_{}_savedweights.pth'.format(
network_type, try_arg)
weight_decay_strength = 1e-4
batch_size = 64
lrs = parse_usual_lr_schedule(try_arg)
if 'bias' in try_arg:
weight_decay_bias = weight_decay_strength
else:
weight_decay_bias = 0
if 'warmup' in try_arg:
warmup_factor = 0
else:
warmup_factor = 1
config = get_baseconfig_by_epoch(
network_type=network_type,
dataset_name=dataset_name,
dataset_subset='train',
global_batch_size=batch_size,
num_node=1,
weight_decay=weight_decay_strength,
optimizer_type='sgd',
momentum=0.9,
max_epochs=lrs.max_epochs,
base_lr=lrs.base_lr,
lr_epoch_boundaries=lrs.lr_epoch_boundaries,
lr_decay_factor=lrs.lr_decay_factor,
warmup_epochs=5,
warmup_method='linear',
warmup_factor=warmup_factor,
ckpt_iter_period=20000,
tb_iter_period=100,
output_dir=log_dir,
tb_dir=log_dir,
save_weights=save_weights,
val_epoch_period=2,
linear_final_lr=lrs.linear_final_lr,
weight_decay_bias=weight_decay_bias)
if 'normal' in try_arg:
builder = None
elif 'acnet' in try_arg:
from acnet.acnet_builder import ACNetBuilder
builder = ACNetBuilder(base_config=config, deploy=False)
else:
assert False
ding_train(config,
show_variables=True,
convbuilder=builder,
use_nesterov='nest' in try_arg)
log_dir = 'acnet_exps/{}_{}_train'.format(network_type, block_type)
weight_decay_bias = weight_decay_strength
config = get_baseconfig_by_epoch(network_type=network_type,
dataset_name=get_dataset_name_by_model_name(network_type), dataset_subset='train',
global_batch_size=batch_size, num_node=1,
weight_decay=weight_decay_strength, optimizer_type='sgd', momentum=0.9,
max_epochs=lrs.max_epochs, base_lr=lrs.base_lr, lr_epoch_boundaries=lrs.lr_epoch_boundaries, cosine_minimum=lrs.cosine_minimum,
lr_decay_factor=lrs.lr_decay_factor,
warmup_epochs=0, warmup_method='linear', warmup_factor=0,
ckpt_iter_period=40000, tb_iter_period=100, output_dir=log_dir,
tb_dir=log_dir, save_weights=None, val_epoch_period=5, linear_final_lr=lrs.linear_final_lr,
weight_decay_bias=weight_decay_bias, deps=None)
if block_type == 'acb':
builder = ACNetBuilder(base_config=config, deploy=False, gamma_init=gamma_init)
else:
builder = ConvBuilder(base_config=config)
target_weights = os.path.join(log_dir, 'finish.hdf5')
if not os.path.exists(target_weights):
train_main(local_rank=start_arg.local_rank, cfg=config, convbuilder=builder,
show_variables=True, auto_continue=auto_continue)
if block_type == 'acb' and start_arg.local_rank == 0:
convert_acnet_weights(target_weights, target_weights.replace('.hdf5', '_deploy.hdf5'), eps=1e-5)
deploy_builder = ACNetBuilder(base_config=config, deploy=True)
general_test(network_type=network_type, weights=target_weights.replace('.hdf5', '_deploy.hdf5'),
builder=deploy_builder)
deps = wrn_origin_deps_flattened(2, 8)
if batch_size is None:
batch_size = TEST_BATCH_SIZE
test_config = get_baseconfig_for_test(network_type=network_type,
dataset_subset='val',
global_batch_size=batch_size,
init_weights=init_weights,
deps=deps,
dataset_name=dataset_name)
return ding_test(cfg=test_config,
net=net,
show_variables=True,
init_hdf5=init_hdf5,
convbuilder=builder,
weights_dict=weights_dict)
if __name__ == '__main__':
import sys
network_type = 'resnet50'
weights = sys.argv[1]
dataset_name = 'imagenet_standard'
from acnet.acnet_builder import ACNetBuilder
builder = ACNetBuilder(base_config=None, deploy=False, gamma_init=1 / 3)
general_test(network_type=network_type,
weights=weights,
builder=builder,
dataset_name=dataset_name)
def ac_resnet50(pretrained=False, progress=True, **kwargs):
from acnet.acnet_builder import ACNetBuilder
#base_config is not None when using SE
builder = ACNetBuilder(base_config=None, deploy=False)
return ResNet(builder, Bottleneck, [3,4,6,3], num_classes=1000)