dtype = torch.float
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
print("device used: ", str(device))
args = Arguments()
args.in_channels = 3
# args.conv_type = "FFT2D"
args.conv_type = ConvType.STANDARD2D
args.compress_rate = None
args.preserve_energy = None
args.is_debug = False
args.next_power2 = True
args.compress_type = CompressType.STANDARD
args.tensor_type = TensorType.FLOAT32
args.num_classes = 10
args.min_batch_size = 16
args.test_batch_size = 16
batch_size = 16
inputs = torch.randn(batch_size,
args.in_channels,
32,
32,
dtype=dtype,
device=device)
model = resnet18(args=args)
model.to(device)
def test_forward_pass_resnet18(self):
"""
total time for (ConvType.STANDARD2D-ConvExecType.SERIAL): 6.813918352127075
total time for (ConvType.FFT2D-ConvExecType.CUDA): 53.35197567939758
total time for (ConvType.FFT2D-ConvExecType.SGEMM): 55.51149845123291
total time for (ConvType.STANDARD2D-ConvExecType.SERIAL): 6.736859083175659
total time for (ConvType.FFT2D-ConvExecType.CUDA): 53.84979581832886
total time for (ConvType.FFT2D-ConvExecType.SGEMM): 56.26755166053772
global init time: 0.24471688270568848
global pad time: 4.250756025314331
(r)fft time: 8.754997730255127
conjugate time: 3.734828233718872
correlation time: 25.324009656906128
restore time (de-compress/concat output): 0.021800994873046875
i(r)fft time: 8.525353193283081
total time for (ConvType.FFT2D-ConvExecType.SGEMM): 56.27733850479126
GPU mem: 2903
global init time: 0.2371835708618164
global pad time: 4.492943286895752
(r)fft time: 9.08437442779541
conjugate time: 3.8394811153411865
correlation time: 25.043412446975708
restore time (de-compress/concat output): 0.021334409713745117
i(r)fft time: 5.491833925247192
total time for (ConvType.FFT2D-ConvExecType.CUDA): 53.804604053497314
GPU mem: 2679
"""
if not torch.cuda.is_available():
print("CUDA device is not available.")
return
device = torch.device("cuda")
print("\ndevice used: ", str(device))
C = 3
# dtype = torch.float
# random mini batch imitating cifar-10
# N, H, W = 128, 32, 32
# inputs = torch.randn(N, C, H, W, dtype=dtype, device=device,
# requires_grad=True)
args = Arguments()
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
args.sample_count_limit = 10000
args.min_batch_size = 32
args.dataset_name = "cifar10"
args.test_batch_size = args.min_batch_size
args.network_type = NetworkType.ResNet18
from cnns.nnlib.datasets.cifar import get_cifar
train_loader, test_loader, _, _ = get_cifar(
args=args, dataset_name=args.dataset_name)
repetition = 1
args.in_channels = C
args.compress_rate = None
args.preserve_energy = 100
args.is_debug = True
args.next_power2 = True
args.compress_type = CompressType.STANDARD
args.tensor_type = TensorType.FLOAT32
args.num_classes = 10
args.test_batch_size = args.min_batch_size
args.in_channels = C
args.dtype = torch.float32
conv_exec_types = [ # (ConvType.STANDARD2D, ConvExecType.SERIAL),
# (ConvType.FFT2D, ConvExecType.CUDA),
(ConvType.FFT2D, ConvExecType.SGEMM),
# (ConvType.FFT2D, ConvExecType.CUDA_SHARED_LOG),
# (ConvType.FFT2D, ConvExecType.CUDA_DEEP),
# (ConvType.FFT2D, ConvExecType.SERIAL),
# (ConvType.FFT2D, ConvExecType.BATCH),
]
for conv_type, conv_exec_type in conv_exec_types:
args.conv_type = conv_type
args.conv_exec_type = conv_exec_type
model = resnet18(args=args)
model.to(device)
model.eval()
start_eval = time.time()
for _ in range(repetition):
for inputs, _ in train_loader:
inputs = inputs.to(device)
outputs_standard = model(inputs)
standard_time = time.time() - start_eval
print(f"total time for ({conv_type}-{conv_exec_type}):"
f" {standard_time}")