def compute_raw_weight(model: nn.Module,
input_size: typing.Tuple[int, int],
cuda=False):
"""
compute d_flops_in and d_flops_out for every convolutional layers in the model
Note: this method needs to do forward pass, which is time-consuming
"""
model._flops_weight_computed = True
# register hooks
hook_handles = []
for submodule in model.modules():
if isinstance(submodule, nn.Conv2d):
hook_handles.append(
submodule.register_forward_hook(_conv_raw_weight_hook))
elif isinstance(submodule, nn.Linear):
hook_handles.append(
submodule.register_forward_hook(_linear_raw_weight_hook))
# do forward pass to compute the input spatial size for each layer
random_input = torch.rand(8, 3, *input_size)
if cuda:
random_input = random_input.cuda()
model = model.cuda()
model(random_input)
# remove hooks
for h in hook_handles:
h.remove()
def compute_conv_flops_weight(model: nn.Module,
building_block,
input_size: typing.Tuple[int, int] = (32, 32),
cuda=False) -> typing.List[typing.Tuple[int]]:
"""
compute the conv_flops_weight for the model
:param building_block: the basic building block for CNN.
Use BasicBlock for ResNet-56. Use VGGBlock for CIFAR VGG.
:param input_size: the input spatial size of the network
"""
if not model.flops_weight_computed or model.input_size is None or model.input_size != input_size:
# initialization
# set flag
model._flops_weight_computed = True
model.input_size = input_size
# update when
# 1. the flops is never be computed
# 2. the input size is changed
compute_raw_weight(model, input_size,
cuda=cuda) # compute d_flops_in and d_flops_out
# the weight is raw weight (without scaling!)
# now compute the weight min and max for rescaling
conv_flops_weight_raw: typing.List[typing.Tuple[int, int]] = []
for submodule in model.modules():
if isinstance(submodule, building_block):
submodule: building_block
block_raw_weight = submodule.get_conv_flops_weight(
update=True, scaling=False)
conv_flops_weight_raw.append(block_raw_weight)
# scale weight to [0, 1]
# set weight_max and weight_min for each building blocks
weights = []
for block in conv_flops_weight_raw:
for i in block:
weights.append(i)
max_weights = max(weights)
min_weights = min(weights)
# compute the min_weights and max_weights for rescaling
# for all blocks, the raw_weight_min and the raw_weight_max are same
for submodule in model.modules():
if isinstance(submodule, building_block):
submodule: building_block
submodule.raw_weight_min = min_weights
submodule.raw_weight_max = max_weights
conv_flops_weight: typing.List[typing.Tuple[int]] = []
for submodule in model.modules():
if isinstance(submodule, building_block):
submodule: building_block
conv_flops_weight.append(submodule.conv_flops_weight)
return conv_flops_weight
