def main():
parser = argparse.ArgumentParser(description='profile torchvision models')
parser.add_argument('-d',
'--device',
type=str,
default="all",
help="device ids")
parser.add_argument('-v',
'--verbose',
action='store_true',
help="verbose msg")
parser.add_argument('-t',
'--timing',
action='store_true',
help="enable timing")
args = parser.parse_args()
device, devlist = get_device(args.device)
rasp.set_config({
'profile': {
'batch_size': 1,
'num_batches': 5,
'warmup_batches': 5,
'timing_max_depth': -1,
'compute_max_depth': -1,
'verbose': args.verbose,
},
})
fields = [
'name', 'type', 'in_shape', 'out_shape', 'params', 'lat', 'net_lat',
'lat[%]', 'flops', 'mem_r', 'mem_w', 'mem_rw', 'dev_mem_alloc',
'dev_mem_delta'
]
C = 16
w = 4
l = 8
model = RectNet(C, w, l).to(device)
bsize = 128
fm_size = 224
inputs = torch.randn(bsize, C, fm_size, fm_size).to(device)
stats = rasp.profile_compute_once(model, inputs=inputs)
if args.timing: stats = rasp.profile_timing_once(model, inputs=inputs)
summary, _ = rasp.summary_tape(stats, report_fields=fields)
print(summary)
summary, _ = rasp.summary_all(stats, report_fields=fields)
print(summary)
summary, _ = rasp.summary_node(stats, report_fields=fields)
print(summary)
rasp.profile_off(model)
def main():
parser = argparse.ArgumentParser(description='profile torchvision models')
parser.add_argument('-d',
'--device',
type=str,
default='all',
help='device ids')
parser.add_argument('-v',
'--verbose',
action='store_true',
help='verbose msg')
parser.add_argument('-t',
'--timing',
action='store_true',
help='enable timing')
parser.add_argument('-i',
'--input',
type=str,
default='(1, 3, 224, 224)',
help='input shape')
args = parser.parse_args()
device, devlist = get_device(args.device)
rasp.set_config({
'profile': {
'batch_size': 1,
'num_batches': 50,
'warmup_batches': 5,
'timing_max_depth': -1,
'compute_max_depth': -1,
'verbose': args.verbose,
},
})
print('%s | %s | %s | %s | %s' %
('Model', 'Params', 'FLOPs', 'latency', 'FLOPS'))
print('---|---|---|---|---')
fields = [
'name', 'type', 'in_shape', 'out_shape', 'params', 'lat', 'net_lat',
'lat[%]', 'flops', 'FLOPS', 'mem_rw', 'dev_mem_alloc', 'dev_mem_delta'
]
input_shape = tuple(eval(args.input))
inputs = torch.randn(input_shape, device=device)
for i, name in enumerate(model_names):
model = models.__dict__[name]().to(device=device)
summary, df = rasp.stat(model,
inputs=inputs,
device=device,
report_fields=fields,
timing=args.timing,
print_stat=False,
returns='sum, data')
if args.verbose:
print(summary)
total_f = df.tail(1)
rasp.profile_off(model)
flops, params = total_f.flops[0], total_f.params[0]
if args.timing:
latency = total_f.lat[0]
FLOPS = total_f.FLOPS[0]
else:
latency = 0
FLOPS = 0
print('%s | %s | %s | %s | %s' %
(name, rasp.round_value(params), rasp.round_value(flops),
rasp.round_value(latency), rasp.round_value(FLOPS)))
def main():
parser = argparse.ArgumentParser(description='profile torchvision models')
parser.add_argument('-d',
'--device',
type=str,
default="all",
help="device ids")
parser.add_argument('-v',
'--verbose',
action='store_true',
help="verbose msg")
parser.add_argument('-t',
'--timing',
action='store_true',
help="enable timing")
parser.add_argument('-i',
'--input',
type=str,
default='(8, 64, 224, 224)',
help='input shape')
args = parser.parse_args()
device, devlist = get_device(args.device)
rasp.set_config({
'profile': {
'batch_size': 1,
'num_batches': 100,
'warmup_batches': 10,
'timing_max_depth': -1,
'compute_max_depth': -1,
'verbose': args.verbose,
},
})
print("%s | %s | %s | %s" % ("Model", "Params", "FLOPs", "FLOPS"))
print("---|---|---|---")
fields = [
'name', 'type', 'in_shape', 'out_shape', 'params', 'lat', 'net_lat',
'lat[%]', 'flops', 'FLOPS', 'mem_r', 'mem_w', 'mem_rw',
'dev_mem_alloc', 'dev_mem_delta'
]
input_shape = tuple(eval(args.input))
chn_in = input_shape[1]
inputs = torch.randn(input_shape, device=device)
for i, name in enumerate(ops):
model = ops[name](chn_in, stride=1, affine=True)
summary, df = rasp.stat(model,
inputs=inputs,
device=device,
report_fields=fields,
timing=args.timing,
print_stat=False,
returns='sum, data')
if args.verbose:
print(summary)
rasp.profile_off(model)
total_f = df.tail(1)
total_flops, total_params, total_FLOPS = total_f.flops[
0], total_f.params[0], total_f.FLOPS[0]
print("%s | %s | %s | %s" %
(name, rasp.round_value(total_params),
rasp.round_value(total_flops), rasp.round_value(total_FLOPS)))