def __init__(self, args, custom_args, platform_util):
super(ModelInitializer, self).__init__(args, custom_args, platform_util)
set_env_var("OMP_NUM_THREADS", "1")
if args.batch_size == -1:
args.batch_size = 1
if args.verbose:
print("Setting batch_size to 1 since it is not supplied.")
if args.batch_size == 1:
if args.verbose:
print("Running Wide_Deep model Inference in Latency mode")
else:
if args.verbose:
print("Running Wide_Deep model Inference in Throughput mode")
# Select script based on batch size
if args.batch_size == 1:
executable = os.path.join(args.mode, args.precision,
"wide_deep_inference_bs1_latency.py")
else:
executable = os.path.join(args.mode, args.precision,
"wide_deep_inference.py")
self.run_cmd = " OMP_NUM_THREADS=1" + \
" numactl --cpunodebind=0 --membind=0 " + \
" python " + executable + \
" --data_dir=" + self.args.data_location + \
" --model_dir=" + self.args.checkpoint + \
" --batch_size=" + str(self.args.batch_size)
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args, platform_util)
self.perf_script_path = os.path.join(
self.args.intelai_models, self.args.mode, self.args.precision,
self.RFCN_PERF_SCRIPT)
self.accuracy_script_path = os.path.join(
self.args.intelai_models, self.args.mode, self.args.precision,
self.RFCN_ACCURACY_SCRIPT)
# remove intelai models path, so that imports don't conflict
if "MOUNT_BENCHMARK" in os.environ and \
os.environ["MOUNT_BENCHMARK"] in sys.path:
sys.path.remove(os.environ["MOUNT_BENCHMARK"])
if self.args.intelai_models in sys.path:
sys.path.remove(self.args.intelai_models)
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
# set num_inter_threads and num_intra_threds
self.set_num_inter_intra_threads(num_inter_threads=self.args.num_inter_threads,
num_intra_threads=self.args.num_intra_threads)
omp_num_threads = self.args.num_intra_threads if self.args.num_cores == -1 else self.args.num_cores
set_env_var("OMP_NUM_THREADS", omp_num_threads)
self.parse_args()
def __init__(self, args, custom_args, platform_util):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.command = ""
command_prefix = "{} generate.py".format(self.python_exe)
# Set default KMP env vars, except for KMP_SETTINGS
self.set_kmp_vars(kmp_settings=None)
self.parse_custom_args()
# Set the num_inter_threads and num_intra_threads (override inter threads to 1)
self.set_num_inter_intra_threads(num_inter_threads=1)
set_env_var("OMP_NUM_THREADS", self.args.num_cores)
if self.args.socket_id != -1:
command_prefix = "numactl --physcpubind=0-{} --membind={} {}".\
format(str(int(self.args.num_cores) - 1), self.args.socket_id,
command_prefix)
else:
command_prefix = "numactl --physcpubind=0-{} -l {}".format(
str(int(self.args.num_cores) - 1), command_prefix)
checkpoint_path = os.path.join(self.args.checkpoint,
self.args.checkpoint_name)
# create command to run benchmarking
self.command = ("{} {} --num_inter_threads={} "
"--num_intra_threads={} --sample={}").format(
command_prefix, checkpoint_path,
str(self.args.num_inter_threads),
str(self.args.num_intra_threads),
str(self.args.sample))
def __init__(self, args, custom_args, platform_util):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.run_inference_sanity_checks(self.args, self.custom_args)
self.set_kmp_vars()
self.set_num_inter_intra_threads()
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
self.model_dir = os.path.join(self.args.intelai_models, self.args.mode,
self.args.precision)
# get benchmark command
benchmark_script = os.path.join(self.model_dir, "infer_detections.py")
# get command with numactl
self.run_cmd = self.get_numactl_command(self.args.socket_id)
self.run_cmd += "{0} {1}".format(self.python_exe, benchmark_script)
self.run_cmd += " --input-graph {0}".format(self.args.input_graph)
self.run_cmd += " --batch-size {0}".format(args.batch_size)
self.run_cmd += " --inter-op-parallelism-threads {0}".format(
self.args.num_inter_threads)
self.run_cmd += " --intra-op-parallelism-threads {0}".format(
self.args.num_intra_threads)
if self.args.accuracy_only:
self.run_cmd += " --accuracy-only "
self.run_cmd += " --data-location {0}".format(
self.args.data_location)
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.perf_script_path = os.path.join(self.args.intelai_models,
self.args.mode,
self.args.precision,
self.RFCN_PERF_SCRIPT)
self.accuracy_script_path = os.path.join(self.args.intelai_models,
self.args.mode,
self.args.precision,
self.RFCN_ACCURACY_SCRIPT)
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
# set num_inter_threads and num_intra_threds
self.set_num_inter_intra_threads(
num_inter_threads=self.args.num_inter_threads,
num_intra_threads=self.args.num_intra_threads)
omp_num_threads = self.args.num_intra_threads if self.args.num_cores == -1 else self.args.num_cores
set_env_var("OMP_NUM_THREADS", omp_num_threads)
self.parse_args()
self.validate_args()
def run_benchmark(self):
command_prefix = self.get_command_prefix(self.args.socket_id) + \
self.python_exe + " " + self.benchmark_script
# set num_inter_threads and num_intra_threads
self.set_num_inter_intra_threads()
if self.args.socket_id == -1:
if self.args.num_cores == 1:
command_prefix = "taskset -c 0 " + \
command_prefix
self.args.num_intra_threads = 1
else:
command_prefix = "taskset -c 0-" + \
str(self.args.num_cores - 1) + \
" " + command_prefix
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
config_file_path = os.path.join(self.args.checkpoint,
self.args.config_file)
run_cmd = command_prefix + \
" --inter_op " + str(self.args.num_inter_threads) + \
" --intra_op " + str(self.args.num_intra_threads) + \
" --omp " + str(self.args.num_intra_threads) + \
" --pipeline_config_path " + config_file_path + \
" --checkpoint_dir " + str(self.args.checkpoint) + \
" --eval_dir " + self.research_dir + \
"/object_detection/models/rfcn/eval " + \
" --logtostderr " + \
" --blocktime=0 " + \
" --run_once=True"
self.run_command(run_cmd)
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.set_num_inter_intra_threads()
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
set_env_var("KMP_HW_SUBSET",
"{}c,1T".format(self.args.num_intra_threads))
script_path = os.path.join(self.args.intelai_models, self.args.mode,
self.args.precision, "brats", "predict.py")
# add numactl prefix to the command
self.command_prefix = self.get_command_prefix(self.args.socket_id) + \
"python " + script_path
# add additional args to the command
self.command_prefix += \
" --inter {} --intra {} --nw 1 --nb 5 --bs {}".\
format(self.args.num_inter_threads, self.args.num_intra_threads,
self.args.batch_size)
def __init__(self, args, custom_args=[], platform_util=None):
self.args = args
self.custom_args = custom_args
self.platform_util = platform_util
# Set KMP env vars, if they haven't already been set
self.set_kmp_vars()
set_env_var("KMP_HW_SUBSET", "1T")
if self.args.accuracy_only:
print("Accuracy testing for DRAW inference is not supported yet.")
sys.exit(1)
# Set the num_inter_threads and num_intra_threads
self.set_default_inter_intra_threads(self.platform_util)
# Create the command prefix with numactl and executing the script
script_path = os.path.join(self.args.intelai_models, self.args.mode,
self.args.precision, "draw_inf.py")
self.command_prefix = self.get_numactl_command(args.socket_id) + \
" python {} ".format(script_path)
# Add additional args to the command
self.command_prefix += "--cp {} --num_inter_threads {} " \
"--num_intra_threads {} --bs {} --dl {} " \
"--nw 100 --nb 200".\
format(self.args.checkpoint, self.args.num_inter_threads,
self.args.num_intra_threads, self.args.batch_size,
self.args.data_location)
def __init__(self, args, custom_args, platform_util):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.accuracy_script_path = os.path.join(self.args.intelai_models,
self.args.mode,
self.args.precision,
self.accuracy_script)
self.perf_script_path = os.path.join(self.args.intelai_models,
self.args.mode,
self.args.precision,
self.benchmark_script)
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
# set num_inter_threads and num_intra_threds
self.set_num_inter_intra_threads(
num_inter_threads=self.args.num_inter_threads,
num_intra_threads=self.args.num_intra_threads)
self.omp_num_threads = self.args.num_intra_threads if self.args.num_cores == -1 else self.args.num_cores
set_env_var("OMP_NUM_THREADS", self.omp_num_threads)
self.parse_custom_args()
self.run_inference_sanity_checks(self.args, self.custom_args)
self.research_dir = os.path.join(self.args.model_source_dir,
"research")
def __init__(self, args, custom_args, platform_util):
super(ModelInitializer, self).__init__(args, custom_args, platform_util)
self.command = ""
command_prefix = "{} generate.py".format(self.python_exe)
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
self.parse_custom_args()
# Set the num_inter_threads and num_intra_threads (override inter threads to 1)
self.set_num_inter_intra_threads(num_inter_threads=1)
set_env_var("OMP_NUM_THREADS", self.args.num_cores)
if self.args.socket_id != -1:
command_prefix = "numactl --physcpubind=0-{} --membind={} {}".\
format(str(int(self.args.num_cores) - 1), self.args.socket_id,
command_prefix)
else:
command_prefix = "numactl --physcpubind=0-{} -l {}".format(
str(int(self.args.num_cores) - 1), command_prefix)
checkpoint_path = os.path.join(self.args.checkpoint,
self.args.checkpoint_name)
# create command to run benchmarking
self.command = ("{} {} --num_inter_threads={} "
"--num_intra_threads={} --sample={}").format(
command_prefix, checkpoint_path, str(self.args.num_inter_threads),
str(self.args.num_intra_threads), str(self.args.sample))
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args, platform_util)
# Set the num_inter_threads and num_intra_threads
# if user did not provide then default value based on platform will be set
self.set_num_inter_intra_threads(self.args.num_inter_threads,
self.args.num_intra_threads)
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
benchmark_script = os.path.join(self.args.intelai_models, self.args.mode,
"infer_detections.py")
self.command_prefix = self.get_command_prefix(self.args.socket_id) \
+ "{} {}".format(self.python_exe, benchmark_script)
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
self.command_prefix += " -g {0}".format(self.args.input_graph)
self.command_prefix += " -i 1000"
self.command_prefix += " -w 200"
self.command_prefix += " -a {0}".format(self.args.num_intra_threads)
self.command_prefix += " -e {0}".format(self.args.num_inter_threads)
if self.args.data_location:
self.command_prefix += " -d {0}".format(self.args.data_location)
if self.args.accuracy_only:
self.command_prefix += " -r"
assert self.args.data_location, "accuracy must provide the data."
else:
# Did not support multi-batch accuracy check.
self.command_prefix += " -b {0}".format(self.args.batch_size)
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
# set num_inter_threads and num_intra_threads
self.set_num_inter_intra_threads()
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
benchmark_script = os.path.join(self.args.intelai_models, args.mode,
args.precision, "inference_bench.py")
self.benchmark_command = self.get_command_prefix(args.socket_id) + \
self.python_exe + " " + benchmark_script
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
self.cifar10_dir = os.path.join(args.model_source_dir, "research",
"gan", "cifar")
self.benchmark_command = self.benchmark_command + \
" -ckpt " + str(self.args.checkpoint) + \
" -dl " + str(self.args.data_location) + \
" --num_inter_threads " + str(self.args.num_inter_threads) + \
" --num_intra_threads " + str(self.args.num_intra_threads) + \
" -nw 100 -nb 500" + \
" --bs " + str(self.args.batch_size) + \
" --kmp_blocktime " + os.environ["KMP_BLOCKTIME"] + \
" --kmp_settings " + os.environ["KMP_SETTINGS"]
def __init__(self, args, custom_args, platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.benchmark_command = "" # use default batch size if -1
self.results_file_path = ""
if self.args.batch_size == -1:
self.args.batch_size = 1024
# get number of cores if there is no core info passed in.
if self.args.num_cores == -1:
self.args.num_cores = self.platform_util.num_cores_per_socket()
num_of_parallel_batches = self.platform_util.num_cores_per_socket()
kmp_blocktime = "0"
# Set KMP env vars, if they haven't already been set
self.set_kmp_vars(
kmp_settings="1",
kmp_blocktime=kmp_blocktime,
kmp_affinity="noverbose,warnings,respect,granularity=core,none")
# set num_inter_threads and num_intra_threads
self.set_num_inter_intra_threads(num_inter_threads=self.args.num_cores,
num_intra_threads=1)
benchmark_script = os.path.join(self.args.intelai_models,
self.args.mode, "inference.py")
self.benchmark_command = self.get_numactl_command(self.args.socket_id)\
+ "python " + benchmark_script
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
self.benchmark_command += " --input-graph=" + \
self.args.input_graph + \
" --inter-op-parallelism-threads=" + \
str(self.args.num_inter_threads) + \
" --intra-op-parallelism-threads=" + \
str(self.args.num_intra_threads) + \
" --omp-num-threads=" + \
str(self.args.num_intra_threads) + \
" --batch-size=" + \
str(self.args.batch_size) + \
" --num-of-parallel-batches=" + \
str(num_of_parallel_batches) + \
" --kmp-blocktime=" + \
str(kmp_blocktime)
# if the data location directory is not empty, then include the arg
if self.args.data_location:
self.benchmark_command += " --datafile-path=" + \
self.args.data_location
# if output results is enabled, generate a results file name and pass it to the inference script
if self.args.output_results:
self.results_filename = "{}_{}_{}_results_{}.txt".format(
self.args.model_name, self.args.precision, self.args.mode,
time.strftime("%Y%m%d_%H%M%S", time.gmtime()))
self.results_file_path = os.path.join(self.args.output_dir,
self.results_filename)
self.benchmark_command += " --results-file-path {}".format(
self.results_file_path)
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args, platform_util)
# Set the num_inter_threads and num_intra_threads
self.set_num_inter_intra_threads()
# Set env vars, if they haven't already been set
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads, overwrite_existing=True)
def __init__(self, args, custom_args=[], platform_util=None):
self.args = args
self.custom_args = custom_args
self.platform_util = platform_util
# set num_inter_threads and num_intra_threads
self.set_default_inter_intra_threads(self.platform_util)
# Set KMP env vars, if they haven't already been set
self.set_kmp_vars()
set_env_var("KMP_HW_SUBSET", "1T")
benchmark_script = os.path.join(self.args.intelai_models, args.mode,
args.precision, "inference_bench.py")
self.benchmark_command = self.get_numactl_command(args.socket_id) + \
"python " + benchmark_script
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
self.cifar10_dir = os.path.join(args.model_source_dir, "research",
"gan", "cifar")
self.benchmark_command = self.benchmark_command + \
" -ckpt " + str(self.args.checkpoint) + \
" -dl " + str(self.args.data_location) + \
" --num_inter_threads " + str(self.args.num_inter_threads) + \
" --num_intra_threads " + str(self.args.num_intra_threads) + \
" -nw 100 -nb 500" + \
" --bs " + str(self.args.batch_size) + \
" --kmp_blocktime " + os.environ["KMP_BLOCKTIME"] + \
" --kmp_settings " + os.environ["KMP_SETTINGS"]
def __init__(self, args, custom_args, platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
# set num_inter_threads and num_intra_threads
self.set_num_inter_intra_threads()
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
benchmark_script = os.path.join(self.args.intelai_models,
self.args.mode, self.args.precision,
"one_image_test.py")
self.command_prefix = \
self.get_command_prefix(self.args.socket_id) + \
"{} ".format(self.python_exe) + benchmark_script
self.run_cmd = \
self.command_prefix + \
" --num_inter_threads " + str(self.args.num_inter_threads) + \
" --num_intra_threads " + str(self.args.num_intra_threads) + \
" -ckpt " + self.args.checkpoint + \
" -dl " + self.args.data_location
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
# use default batch size if -1
if self.args.batch_size == -1:
self.args.batch_size = 128
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
# set num_inter_threads and num_intra_threads (override inter threads to 2)
self.set_num_inter_intra_threads(num_inter_threads=2)
script_name = "accuracy.py" if self.args.accuracy_only \
else "eval_image_classifier.py"
script_path = os.path.join(self.args.intelai_models, self.args.mode,
self.args.precision, script_name)
self.command_prefix = "{} {}".format(self.python_exe, script_path)
if self.args.socket_id != -1:
self.command_prefix = "numactl --cpunodebind={} -l {}".format(
str(self.args.socket_id), self.command_prefix)
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
if not self.args.accuracy_only:
self.command_prefix = ("{prefix} "
"--dataset_name imagenet "
"--checkpoint_path {checkpoint} "
"--dataset_split_name=validation "
"--clone_on_cpu=True "
"--model_name {model} "
"--inter_op_parallelism_threads {inter} "
"--intra_op_parallelism_threads {intra} "
"--batch_size {bz}").format(
prefix=self.command_prefix,
checkpoint=self.args.checkpoint,
model=self.args.model_name,
inter=self.args.num_inter_threads,
intra=self.args.num_intra_threads,
bz=self.args.batch_size)
if self.args.data_location:
self.command_prefix += " --dataset_dir {}".format(
self.args.data_location)
else:
# add args for the accuracy script
script_args_list = [
"input_graph", "data_location", "input_height", "input_width",
"batch_size", "input_layer", "output_layer",
"num_inter_threads", "num_intra_threads"
]
self.command_prefix = self.add_args_to_command(
self.command_prefix, script_args_list)
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.cmd = self.get_numactl_command(self.args.socket_id) + \
self.python_exe + " "
# Set KMP env vars, if they haven't already been set
self.set_kmp_vars()
pairs_file = os.path.join(self.args.model_source_dir, "data/pairs.txt")
arg_parser = ArgumentParser(description='Parse custom args')
arg_parser.add_argument(
"--lfw_pairs",
type=str,
help="The file containing the pairs to use for validation.",
dest="lfw_pairs",
default=pairs_file)
self.args = arg_parser.parse_args(self.custom_args,
namespace=self.args)
# use default batch size if -1
if self.args.batch_size == -1 or self.args.accuracy_only:
self.args.batch_size = 100
# set num_inter_threads and num_intra_threads
if self.args.batch_size > 32:
self.set_num_inter_intra_threads(num_inter_threads=2)
else:
self.set_num_inter_intra_threads(num_inter_threads=1)
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
run_script = os.path.join(self.args.model_source_dir,
"src/validate_on_lfw.py")
warmup_steps = 40
max_steps = 1000
if self.args.batch_size == 1:
warmup_steps = 200
cmd_args = ' ' + self.args.data_location + \
' ' + self.args.checkpoint + ' --distance_metric 1' + \
' --use_flipped_images' + ' --subtract_mean' + \
' --use_fixed_image_standardization' + \
' --num_inter_threads=' + \
str(self.args.num_inter_threads) + \
' --num_intra_threads=' + \
str(self.args.num_intra_threads) + \
' --lfw_batch_size=' + str(self.args.batch_size) + \
' --lfw_pairs=' + self.args.lfw_pairs + \
' --warmup_steps=' + str(warmup_steps) + \
' --max_steps=' + str(max_steps)
self.cmd = self.cmd + run_script + cmd_args
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args, platform_util)
# Set the num_inter_threads and num_intra_threads
self.set_num_inter_intra_threads(num_inter_threads=platform_util.num_cores_per_socket,
num_intra_threads=1)
# Use default KMP AFFINITY values, override KMP_BLOCKTIME & enable KMP SETTINGS
self.set_kmp_vars(kmp_settings="1", kmp_blocktime="0",
kmp_affinity="noverbose,warnings,respect,granularity=core,none")
# Set env vars, if they haven't already been set
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args, platform_util)
# Set the num_inter_threads and num_intra_threads
self.set_num_inter_intra_threads()
# Set env vars, if they haven't already been set
set_env_var("OMP_NUM_THREADS", platform_util.num_cores_per_socket()
if args.num_cores == -1 else args.num_cores)
# Set KMP env vars, but override default KMP_BLOCKTIME value
self.set_kmp_vars(kmp_blocktime="0")
def __init__(self, args, custom_args, platform_util):
super(SSDVGG16ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.parse_custom_args()
self.run_inference_sanity_checks(self.args, self.custom_args)
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
self.set_num_inter_intra_threads(
num_inter_threads=self.args.num_inter_threads,
num_intra_threads=self.args.num_intra_threads)
omp_num_threads = str(int(platform_util.num_cores_per_socket / 2))\
if self.args.precision == "int8" else platform_util.num_cores_per_socket
set_env_var(
"OMP_NUM_THREADS", omp_num_threads
if self.args.num_cores == -1 else self.args.num_cores)
script_path = os.path.join(self.args.intelai_models, self.args.mode,
"eval_ssd.py")
self.run_cmd = self.get_command_prefix(
self.args.socket_id) + "{} {}".format(self.python_exe, script_path)
self.run_cmd += " --input-graph={} " \
" --num-inter-threads={} --num-intra-threads={} ". \
format(self.args.input_graph, self.args.num_inter_threads,
self.args.num_intra_threads)
if self.args.data_num_inter_threads:
self.run_cmd += " --data-num-inter-threads={} ".format(
self.args.data_num_inter_threads)
if self.args.data_num_intra_threads:
self.run_cmd += " --data-num-intra-threads={} ".format(
self.args.data_num_intra_threads)
if self.args.benchmark_only:
self.run_cmd += " --warmup-steps={} --steps={} ". \
format(self.args.warmup_steps, self.args.steps)
# if the data location directory is not empty, then include the arg
if self.args.data_location and os.listdir(self.args.data_location):
self.run_cmd += " --data-location={} ".format(
self.args.data_location)
if self.args.accuracy_only:
self.run_cmd += "--accuracy-only "
def run_accuracy_command(self):
num_cores = str(self.platform_util.num_cores_per_socket())
if self.args.num_cores != -1:
num_cores = str(self.args.num_cores)
set_env_var("OMP_NUM_THREADS", num_cores)
command = "{} {} {} {}".format(self.accuracy_script_path,
self.args.input_graph,
self.args.data_location,
self.args.model_source_dir)
self.run_command(command)
def __init__(self, args, custom_args=[], platform_util=None):
super(InceptionV4ModelInitializer,
self).__init__(args, custom_args, platform_util)
# Environment variables
set_env_var(
"OMP_NUM_THREADS", platform_util.num_cores_per_socket
if self.args.num_cores == -1 else self.args.num_cores)
self.set_kmp_vars(kmp_blocktime="0")
self.set_num_inter_intra_threads(
num_inter_threads=platform_util.num_threads_per_core,
num_intra_threads=platform_util.num_cores_per_socket)
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args, platform_util)
# Set the num_inter_threads and num_intra_threads
self.set_num_inter_intra_threads(num_inter_threads=platform_util.num_cores_per_socket,
num_intra_threads=1)
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
# Set env vars, if they haven't already been set
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
def __init__(self, args, custom_args, platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.research_dir = os.path.join(self.args.model_source_dir,
"research")
self.run_inference_sanity_checks(self.args, self.custom_args)
# set num_inter_threads and num_intra_threads
self.set_num_inter_intra_threads()
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
if self.args.accuracy_only:
accuracy_script = os.path.join(self.args.intelai_models,
self.args.mode, self.args.precision,
"coco_accuracy.sh")
if not os.path.exists(accuracy_script):
raise ValueError("Unable to locate the Faster R-CNN accuracy "
"script: {}".format(accuracy_script))
self.run_cmd = "sh {} {} {}/coco_val.record {}".format(
accuracy_script, self.args.input_graph,
self.args.data_location, self.args.model_source_dir)
else:
self.parse_custom_args()
benchmark_script = os.path.join(self.args.intelai_models,
self.args.mode,
self.args.precision, "eval.py")
self.command_prefix = \
self.get_command_prefix(self.args.socket_id) + self.python_exe + " " + \
benchmark_script
config_file_path = os.path.join(self.args.checkpoint,
self.args.config_file)
self.run_cmd = \
self.command_prefix + \
" --num_inter_threads " + \
str(self.args.num_inter_threads) + \
" --num_intra_threads " + \
str(self.args.num_intra_threads) + \
" --pipeline_config_path " + \
config_file_path + \
" --checkpoint_dir " + \
str(args.checkpoint) + \
" --eval_dir " + self.research_dir + \
"/object_detection/log/eval"
def __init__(self, args, custom_args, platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.benchmark_command = "" # use default batch size if -1
self.results_file_path = ""
if self.args.batch_size == -1:
self.args.batch_size = 128
# Set KMP env vars, if they haven't already been set
self.set_kmp_vars()
# set num_inter_threads and num_intra_threads
self.set_num_inter_intra_threads()
benchmark_script = os.path.join(self.args.intelai_models,
self.args.precision,
"eval_image_classifier_inference.py")
self.benchmark_command = self.get_numactl_command(self.args.socket_id)\
+ "python " + benchmark_script
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
self.benchmark_command += " --input-graph=" + \
self.args.input_graph + \
" --model-name=" + \
str(self.args.model_name) + \
" --inter-op-parallelism-threads=" + \
str(self.args.num_inter_threads) + \
" --intra-op-parallelism-threads=" + \
str(self.args.num_intra_threads) + \
" --batch-size=" + \
str(self.args.batch_size)
# if the data location directory is not empty, then include the arg
if self.args.data_location and os.listdir(self.args.data_location):
self.benchmark_command += " --data-location=" + \
self.args.data_location
if self.args.accuracy_only:
self.benchmark_command += " --accuracy-only"
# if output results is enabled, generate a results file name and pass it to the inference script
if self.args.output_results:
self.results_filename = "{}_{}_{}_results_{}.txt".format(
self.args.model_name, self.args.precision, self.args.mode,
time.strftime("%Y%m%d_%H%M%S", time.gmtime()))
self.results_file_path = os.path.join(self.args.output_dir,
self.results_filename)
self.benchmark_command += " --results-file-path {}".format(
self.results_file_path)
def __init__(self, args, custom_args, platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
self.research_dir = os.path.join(self.args.model_source_dir,
"research")
self.run_inference_sanity_checks(self.args, self.custom_args)
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
# set num_inter_threads and num_intra_threds
self.set_num_inter_intra_threads(
num_inter_threads=self.args.num_inter_threads,
num_intra_threads=self.args.num_intra_threads)
omp_num_threads = self.args.num_intra_threads if self.args.num_cores == -1 else self.args.num_cores
set_env_var("OMP_NUM_THREADS", omp_num_threads)
self.parse_custom_args()
if self.args.accuracy_only:
accuracy_script = os.path.join(self.args.intelai_models,
self.args.mode, self.args.precision,
"coco_accuracy.sh")
if not os.path.exists(accuracy_script):
raise ValueError("Unable to locate the Faster R-CNN accuracy "
"script: {}".format(accuracy_script))
self.run_cmd = "sh {} {} {} {}".format(accuracy_script,
self.args.input_graph,
self.args.data_location,
self.args.model_source_dir)
else:
benchmark_script = os.path.join(self.args.intelai_models,
self.args.mode,
self.args.precision,
"run_frozen_graph_rcnn.py")
self.command_prefix = \
self.get_command_prefix(self.args.socket_id) + self.python_exe + " " + \
benchmark_script
self.run_cmd = \
self.command_prefix + \
" --num-inter-threads " + \
str(self.args.num_inter_threads) + \
" --num-intra-threads " + \
str(self.args.num_intra_threads) + \
" -g " + self.args.input_graph + \
" -d " + self.args.data_location + \
" -n " + str(self.args.steps)
def __init__(self, args, custom_args=[], platform_util=None):
self.args = args
self.custom_args = custom_args
if not platform_util:
raise ValueError("Did not find any platform info.")
# Set env vars, if they haven't already been set
set_env_var(
"OMP_NUM_THREADS",
platform_util.num_cores_per_socket()
if self.args.num_cores == -1 else self.args.num_cores)
# Use default KMP variable values, but override the default KMP_BLOCKTIME value
self.set_kmp_vars(kmp_blocktime="0")
def run_perf_command(self):
# Get the command previx, but numactl is added later in run_perf_command()
self.command.append(
self.get_command_prefix(self.args.socket_id, numactl=False))
num_cores = str(self.platform_util.num_cores_per_socket)
if self.args.num_cores != -1:
num_cores = str(self.args.num_cores)
set_env_var("OMP_NUM_THREADS", num_cores)
if self.args.socket_id != -1:
self.command.append("numactl")
if self.args.socket_id:
socket_id = self.args.socket_id
else:
socket_id = "0"
if self.args.num_cores != -1:
self.command.append("-C")
self.command.append("+0")
i = 1
while i < self.args.num_cores:
self.command.append(",{}".format(i))
i += i
self.command.append("-N")
self.command.append("{}".format(socket_id))
self.command.append("-m")
self.command.append("{}".format(socket_id))
self.command += (self.python_exe, self.perf_script_path)
self.command += ("-m", self.args.model_source_dir)
self.command += ("-g", self.args.input_graph)
self.command += ("--num-intra-threads",
str(self.args.num_intra_threads))
self.command += ("--num-inter-threads",
str(self.args.num_inter_threads))
if self.args.number_of_steps:
self.command += ("-x", "{}".format(self.args.number_of_steps))
if self.args.visualize:
self.command += ("-v")
if self.args.timeline:
self.command += ("-t", self.args.timeline)
if self.args.data_location:
self.command += ("-d", self.args.data_location)
if self.args.evaluate_tensor:
self.command += ("-e", self.args.evaluate_tensor)
if self.args.print_accuracy:
self.command += ("-p")
self.run_command(" ".join(self.command))
def __init__(self, args, custom_args=[], platform_util=None):
super(ModelInitializer, self).__init__(args, custom_args,
platform_util)
# use default batch size if -1
if self.args.batch_size == -1:
self.args.batch_size = 128
# Set KMP env vars, if they haven't already been set
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "config.json")
self.set_kmp_vars(config_file_path)
# set num_inter_threads and num_intra_threads (override inter threads to 2)
self.set_num_inter_intra_threads(num_inter_threads=2)
script_name = "accuracy.py" if self.args.accuracy_only \
else "benchmark.py"
script_path = os.path.join(self.args.intelai_models, self.args.mode,
script_name)
self.command_prefix = "{} {}".format(self.python_exe, script_path)
num_numas = self.platform_util.num_numa_nodes
if self.args.socket_id != -1 and num_numas > 0:
self.command_prefix = "numactl --cpunodebind={} -l {}".format(
str(self.args.socket_id), self.command_prefix)
set_env_var("OMP_NUM_THREADS", self.args.num_intra_threads)
self.parse_args()
if not self.args.accuracy_only:
# add args for the benchmark script
script_args_list = [
"input_graph", "input_height", "input_width", "batch_size",
"input_layer", "output_layer", "num_inter_threads",
"num_intra_threads", "warmup_steps", "steps", "precision"
]
self.command_prefix = self.add_args_to_command(
self.command_prefix, script_args_list)
else:
# add args for the accuracy script
script_args_list = [
"input_graph", "data_location", "input_height", "input_width",
"batch_size", "input_layer", "output_layer",
"num_inter_threads", "num_intra_threads", "precision"
]
self.command_prefix = self.add_args_to_command(
self.command_prefix, script_args_list)