def main():
args = parse_args()
if args.dtype == 'float32':
args.dtype = 'float'
# report some available info
if args.device == 'gpu':
assert args.num_gpus > 0, "Number of GPUs must be specified in GPU mode"
print("__caffe2.cuda_version__=%s" % (json.dumps(workspace.GetCUDAVersion())))
print("__caffe2.cudnn_version__=%s" % (json.dumps(workspace.GetCuDNNVersion())))
try:
opts = vars(args)
opts['phase'] = 'inference' if args.forward_only else 'training'
model_title, times = benchmark(opts)
except Exception as err:
#TODO: this is not happenning, program terminates earlier.
# For now, do not rely on __results.status__=...
times = np.zeros(0)
model_title = 'Unk'
print ("Critical error while running benchmarks (%s). See stacktrace below." % (str(err)))
traceback.print_exc(file=sys.stdout)
if len(times) > 0:
mean_time = np.mean(times) # seconds
# Compute mean throughput
num_local_devices = 1 if args.device == 'cpu' else args.num_gpus #Number of compute devices per node
num_devices = num_local_devices * args.num_workers #Global number of devices
replica_batch = args.batch_size #Input is a replica batch
mean_throughput = num_devices * replica_batch / mean_time #images / sec
#
print("__results.time__=%s" % (json.dumps(1000.0 * mean_time)))
print("__results.throughput__=%s" % (json.dumps(int(mean_throughput))))
print("__exp.model_title__=%s" % (json.dumps(model_title)))
print("__results.time_data__=%s" % (json.dumps((1000.0*times).tolist())))
else:
print("__results.status__=%s" % (json.dumps("failure")))
def get_nvidia_info():
return (
get_nvidia_smi_output(),
workspace.GetCUDAVersion(),
workspace.GetCuDNNVersion(),
)
parser.add_argument(
'--enable_tensor_core',
action='store_true',
help=
'Enable volta\'s tensor ops (requires CUDA >= 9, cuDNN >= 7 and NVIDIA Volta GPU)'
)
args = parser.parse_args()
if args.dtype == 'float32':
args.dtype = 'float'
# report some available info
if args.device == 'gpu':
assert args.num_gpus > 0, "Number of GPUs must be specified in GPU mode"
print("__caffe2.cuda_version__=%s" %
(json.dumps(workspace.GetCUDAVersion())))
print("__caffe2.cudnn_version__=%s" %
(json.dumps(workspace.GetCuDNNVersion())))
try:
opts = vars(args)
opts['phase'] = 'inference' if args.forward_only else 'training'
model_title, times = benchmark(opts)
except Exception as err:
#TODO: this is not happenning, program terminates earlier.
# For now, do not rely on __results.status__=...
times = np.zeros(0)
model_title = 'Unk'
print(
"Critical error while running benchmarks (%s). See stacktrace below."
% (str(err)))
parser.add_argument('--enable_tensor_core', action='store_true', help='Enable volta\'s tensor ops (requires CUDA >= 9, cuDNN >= 7 and NVIDIA Volta GPU)')
parser.add_argument('--num_decode_threads', type=int, required=False, default=1, help='Number of image decode threads. For high throughput models such as AlexNetOWT set to 6-8 for 4 Voltas.')
parser.add_argument('--float16_compute', nargs='?', const=True, default=False, type=str2bool, help='If true, use FP16 SGD optimizer else use multi-precision SGD optimizer')
# These parameters affect the ModelHelper behaviour and are now applied for GPU benchmarks
parser.add_argument('--cudnn_workspace_limit_mb', type=int, required=False, default=64, help='CuDNN workspace limit in MBs')
parser.add_argument('--use_cudnn', nargs='?', const=True, default=True, type=str2bool, help='Use NVIDIA cuDNN library.')
parser.add_argument('--cudnn_exhaustive_search', nargs='?', const=True, default=True, type=str2bool, help='Benchmark inference (if true) else benchmark training.')
args = parser.parse_args()
if args.dtype == 'float32':
args.dtype = 'float'
# report some available info
if args.device == 'gpu':
assert args.num_gpus > 0, "Number of GPUs must be specified in GPU mode"
print("__caffe2.cuda_version__=%s" % (json.dumps(workspace.GetCUDAVersion())))
print("__caffe2.cudnn_version__=%s" % (json.dumps(workspace.GetCuDNNVersion())))
try:
opts = vars(args)
opts['phase'] = 'inference' if args.forward_only else 'training'
model_title, times = benchmark(opts)
except Exception as err:
#TODO: this is not happenning, program terminates earlier.
# For now, do not rely on __results.status__=...
times = np.zeros(0)
model_title = 'Unk'
print ("Critical error while running benchmarks (%s). See stacktrace below." % (str(err)))
traceback.print_exc(file=sys.stdout)
if len(times) > 0: