|
|
|
|
|
import torch |
|
|
|
|
|
try: |
|
|
from deepspeed.profiling.flops_profiler import get_model_profile |
|
|
has_deepspeed_profiling = True |
|
|
except ImportError as e: |
|
|
has_deepspeed_profiling = False |
|
|
|
|
|
try: |
|
|
from fvcore.nn import FlopCountAnalysis, flop_count_str, flop_count_table |
|
|
from fvcore.nn import ActivationCountAnalysis |
|
|
has_fvcore_profiling = True |
|
|
except ImportError as e: |
|
|
FlopCountAnalysis = None |
|
|
ActivationCountAnalysis = None |
|
|
has_fvcore_profiling = False |
|
|
|
|
|
|
|
|
def profile_deepspeed(model, input_size=(3, 224, 224), input_dtype=torch.float32, |
|
|
batch_size=1, detailed=False): |
|
|
device, dtype = next(model.parameters()).device, next(model.parameters()).dtype |
|
|
flops, macs, params = get_model_profile( |
|
|
model=model, |
|
|
args=torch.zeros((batch_size,) + input_size, device=device, dtype=input_dtype), |
|
|
print_profile=detailed, |
|
|
detailed=detailed, |
|
|
warm_up=10, |
|
|
as_string=False, |
|
|
output_file=None, |
|
|
ignore_modules=None) |
|
|
return macs, 0 |
|
|
|
|
|
|
|
|
def profile_fvcore(model, input_size=(3, 224, 224), input_dtype=torch.float32, max_depth=4, |
|
|
batch_size=1, detailed=False, force_cpu=False): |
|
|
if force_cpu: |
|
|
model = model.to('cpu') |
|
|
device, dtype = next(model.parameters()).device, next(model.parameters()).dtype |
|
|
example_input = torch.zeros((batch_size,) + input_size, device=device, dtype=input_dtype) |
|
|
fca = FlopCountAnalysis(model, example_input) |
|
|
aca = ActivationCountAnalysis(model, example_input) |
|
|
if detailed: |
|
|
print(flop_count_table(fca, max_depth=max_depth)) |
|
|
return fca, fca.total(), aca, aca.total() |
|
|
|