def general_test(network_type, weights, builder=None, net=None, dataset_name=None, weights_dict=None,
batch_size=None):
if weights is None or weights == 'None':
init_weights = None
init_hdf5 = None
elif weights.endswith('.hdf5'):
init_weights = None
init_hdf5 = weights
else:
init_weights = weights
init_hdf5 = None
if init_hdf5 is not None:
deps = extract_deps_from_weights_file(init_hdf5)
else:
deps = None
if deps is None and ('wrnc16' in network_type or 'wrnh16' in network_type):
from constants import wrn_origin_deps_flattened
deps = wrn_origin_deps_flattened(2, 8)
if network_type == 'sres50':
from constants import RESNET50_ORIGIN_DEPS_FLATTENED
from rr.resrep_scripts import calculate_resnet_50_flops
flops_ratio = calculate_resnet_50_flops(deps) / calculate_resnet_50_flops(RESNET50_ORIGIN_DEPS_FLATTENED)
extra_msg = 'flops_r={:.4f}'.format(flops_ratio)
else:
extra_msg = None
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,
extra_msg=extra_msg, weights_dict=weights_dict)
def general_test(network_type, weights, builder=None):
if weights.endswith('.hdf5'):
init_weights = None
init_hdf5 = weights
else:
init_weights = weights
init_hdf5 = None
deps = extract_deps_from_weights_file(weights)
if deps is None and ('wrnc16' in network_type or 'wrnh16' in network_type):
from constants import wrn_origin_deps_flattened
deps = wrn_origin_deps_flattened(2, 8)
test_config = get_baseconfig_for_test(network_type=network_type, dataset_subset='val', global_batch_size=TEST_BATCH_SIZE,
init_weights=init_weights, deps=deps)
ding_test(cfg=test_config, show_variables=True, init_hdf5=init_hdf5, convbuilder=builder)
def general_test(network_type,
weights,
builder=None,
net=None,
dataset_name=None,
weights_dict=None,
batch_size=None):
if weights is None or weights == 'None':
init_weights = None
init_hdf5 = None
elif weights.endswith('.hdf5'):
init_weights = None
init_hdf5 = weights
else:
init_weights = weights
init_hdf5 = None
if init_hdf5 is not None:
deps = extract_deps_from_weights_file(init_hdf5)
else:
deps = None
if deps is None and ('wrnc16' in network_type or 'wrnh16' in network_type):
from constants import wrn_origin_deps_flattened
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)
iters_per_half=iters_per_half,
thresh=thresh)
aofp_train_main(local_rank=start_arg.local_rank,
target_layers=target_layers,
succ_strategy=succ_strategy,
warmup_iterations=warmup_iterations,
aofp_batches_per_half=iters_per_half,
cfg=aofp_config,
show_variables=True,
convbuilder=aofp_builder,
init_hdf5=init_hdf5,
flops_func=flops_func,
remain_flops_ratio=flops_remain_target)
# finetune
pruned_deps = extract_deps_from_weights_file(pruned_path)
finetune_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=finetune_lrs.max_epochs,
base_lr=finetune_lrs.base_lr,
lr_epoch_boundaries=finetune_lrs.lr_epoch_boundaries,
cosine_minimum=finetune_lrs.cosine_minimum,
lr_decay_factor=finetune_lrs.lr_decay_factor,
warmup_epochs=0,