mirrors
/
pytorch-image-models
mirror of https://github.com/huggingface/pytorch-image-models.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Add activation count to fvcore based profiling in benchmark.py

pull/964/head
Ross Wightman 2021-10-24 15:30:12 -07:00
parent 51f488b7f5
commit 5882e62ada
1 changed files with 13 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
benchmark.py
View File

@ -45,7 +45,7 @@ except ImportError as e:
has_deepspeed_profiling = False has_deepspeed_profiling = False
try: try:
from fvcore.nn import FlopCountAnalysis, flop_count_str from fvcore.nn import FlopCountAnalysis, flop_count_str, ActivationCountAnalysis
has_fvcore_profiling = True has_fvcore_profiling = True
except ImportError as e: except ImportError as e:
FlopCountAnalysis = None FlopCountAnalysis = None
@ -167,16 +167,18 @@ def profile_deepspeed(model, input_size=(3, 224, 224), batch_size=1, detailed=Fa
as_string=False, # print raw numbers (e.g. 1000) or as human-readable strings (e.g. 1k) as_string=False, # print raw numbers (e.g. 1000) or as human-readable strings (e.g. 1k)
output_file=None, # path to the output file. If None, the profiler prints to stdout. output_file=None, # path to the output file. If None, the profiler prints to stdout.
ignore_modules=None) # the list of modules to ignore in the profiling ignore_modules=None) # the list of modules to ignore in the profiling
return macs return macs, 0 # no activation count in DS
def profile_fvcore(model, input_size=(3, 224, 224), batch_size=1, detailed=False): def profile_fvcore(model, input_size=(3, 224, 224), batch_size=1, detailed=False):
device, dtype = next(model.parameters()).device, next(model.parameters()).dtype device, dtype = next(model.parameters()).device, next(model.parameters()).dtype
fca = FlopCountAnalysis(model, torch.ones((batch_size,) + input_size, device=device, dtype=dtype)) example_input = torch.ones((batch_size,) + input_size, device=device, dtype=dtype)
fca = FlopCountAnalysis(model, example_input)
aca = ActivationCountAnalysis(model, example_input)
if detailed: if detailed:
fcs = flop_count_str(fca) fcs = flop_count_str(fca)
print(fcs) print(fcs)
return fca.total() return fca.total(), aca.total()
class BenchmarkRunner: class BenchmarkRunner:
@ -275,11 +277,12 @@ class InferenceBenchmarkRunner(BenchmarkRunner):
) )
if has_deepspeed_profiling: if has_deepspeed_profiling:
macs = profile_deepspeed(self.model, self.input_size) macs, _ = profile_deepspeed(self.model, self.input_size)
results['gmacs'] = round(macs / 1e9, 2) results['gmacs'] = round(macs / 1e9, 2)
elif has_fvcore_profiling: elif has_fvcore_profiling:
macs = profile_fvcore(self.model, self.input_size) macs, activations = profile_fvcore(self.model, self.input_size)
results['gmacs'] = round(macs / 1e9, 2) results['gmacs'] = round(macs / 1e9, 2)
results['macts'] = round(activations / 1e6, 2)
_logger.info( _logger.info(
f"Inference benchmark of {self.model_name} done. " f"Inference benchmark of {self.model_name} done. "
@ -427,13 +430,15 @@ class ProfileRunner(BenchmarkRunner):
f'input size {self.input_size} and batch size {self.batch_size}.') f'input size {self.input_size} and batch size {self.batch_size}.')
macs = 0 macs = 0
activations = 0
if self.profiler == 'deepspeed': if self.profiler == 'deepspeed':
macs = profile_deepspeed(self.model, self.input_size, batch_size=self.batch_size, detailed=True) macs, _ = profile_deepspeed(self.model, self.input_size, batch_size=self.batch_size, detailed=True)
elif self.profiler == 'fvcore': elif self.profiler == 'fvcore':
macs = profile_fvcore(self.model, self.input_size, batch_size=self.batch_size, detailed=True) macs, activations = profile_fvcore(self.model, self.input_size, batch_size=self.batch_size, detailed=True)
results = dict( results = dict(
gmacs=round(macs / 1e9, 2), gmacs=round(macs / 1e9, 2),
macts=round(activations / 1e6, 2),
batch_size=self.batch_size, batch_size=self.batch_size,
img_size=self.input_size[-1], img_size=self.input_size[-1],
param_count=round(self.param_count / 1e6, 2), param_count=round(self.param_count / 1e6, 2),