def benchmark_using_loadgen(scenario_str, mode_str, samples_in_mem,
config_filepath):
"Perform the benchmark using python API for the LoadGen librar"
scenario = {
'SingleStream': lg.TestScenario.SingleStream,
'MultiStream': lg.TestScenario.MultiStream,
'Server': lg.TestScenario.Server,
'Offline': lg.TestScenario.Offline,
}[scenario_str]
mode = {
'AccuracyOnly': lg.TestMode.AccuracyOnly,
'PerformanceOnly': lg.TestMode.PerformanceOnly,
'SubmissionRun': lg.TestMode.SubmissionRun,
}[mode_str]
ts = lg.TestSettings()
if (config_filepath):
ts.FromConfig(config_filepath, 'random_model_name', scenario_str)
ts.scenario = scenario
ts.mode = mode
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(dataset_size, samples_in_mem, load_query_samples,
unload_query_samples)
log_settings = lg.LogSettings()
log_settings.enable_trace = False
lg.StartTestWithLogSettings(sut, qsl, ts, log_settings)
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
def benchmark_using_loadgen():
"Perform the benchmark using python API for the LoadGen library"
scenario = {
'SingleStream': lg.TestScenario.SingleStream,
'MultiStream': lg.TestScenario.MultiStream,
'Server': lg.TestScenario.Server,
'Offline': lg.TestScenario.Offline,
}[LOADGEN_SCENARIO]
mode = {
'AccuracyOnly': lg.TestMode.AccuracyOnly,
'PerformanceOnly': lg.TestMode.PerformanceOnly,
'SubmissionRun': lg.TestMode.SubmissionRun,
}[LOADGEN_MODE]
ts = lg.TestSettings()
ts.FromConfig(MLPERF_CONF_PATH, MODEL_NAME, LOADGEN_SCENARIO)
ts.FromConfig(USER_CONF_PATH, MODEL_NAME, LOADGEN_SCENARIO)
ts.scenario = scenario
ts.mode = mode
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(LOADGEN_DATASET_SIZE, LOADGEN_BUFFER_SIZE,
load_query_samples, unload_query_samples)
log_settings = lg.LogSettings()
log_settings.enable_trace = False
lg.StartTestWithLogSettings(sut, qsl, ts, log_settings)
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
def benchmark_using_loadgen():
"Perform the benchmark using python API for the LoadGen library"
global model
# Load the [cached] Torch model
torchvision_version = '' # master by default
try:
import torchvision
torchvision_version = ':v' + torchvision.__version__
except Exception:
pass
model = torch.hub.load('pytorch/vision' + torchvision_version,
MODEL_NAME,
pretrained=True)
model.eval()
# move the model to GPU for speed if available
if USE_CUDA:
model.to('cuda')
scenario = {
'SingleStream': lg.TestScenario.SingleStream,
'MultiStream': lg.TestScenario.MultiStream,
'Server': lg.TestScenario.Server,
'Offline': lg.TestScenario.Offline,
}[LOADGEN_SCENARIO]
mode = {
'AccuracyOnly': lg.TestMode.AccuracyOnly,
'PerformanceOnly': lg.TestMode.PerformanceOnly,
'SubmissionRun': lg.TestMode.SubmissionRun,
}[LOADGEN_MODE]
ts = lg.TestSettings()
ts.FromConfig(MLPERF_CONF_PATH, MODEL_NAME, LOADGEN_SCENARIO)
ts.FromConfig(USER_CONF_PATH, MODEL_NAME, LOADGEN_SCENARIO)
ts.scenario = scenario
ts.mode = mode
if LOADGEN_MULTISTREAMNESS:
ts.multi_stream_samples_per_query = int(LOADGEN_MULTISTREAMNESS)
if LOADGEN_COUNT_OVERRIDE:
ts.min_query_count = int(LOADGEN_COUNT_OVERRIDE)
ts.max_query_count = int(LOADGEN_COUNT_OVERRIDE)
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(LOADGEN_DATASET_SIZE, LOADGEN_BUFFER_SIZE,
load_query_samples, unload_query_samples)
log_settings = lg.LogSettings()
log_settings.enable_trace = False
lg.StartTestWithLogSettings(sut, qsl, ts, log_settings)
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
def main():
args = get_args()
if args.backend == "pytorch":
assert not args.quantized, "Quantized model is only supported by onnxruntime backend!"
assert not args.profile, "Profiling is only supported by onnxruntime backend!"
from pytorch_SUT import get_pytorch_sut
sut = get_pytorch_sut()
elif args.backend == "tf":
assert not args.quantized, "Quantized model is only supported by onnxruntime backend!"
assert not args.profile, "Profiling is only supported by onnxruntime backend!"
from tf_SUT import get_tf_sut
sut = get_tf_sut()
elif args.backend == "tf_estimator":
assert not args.quantized, "Quantized model is only supported by onnxruntime backend!"
assert not args.profile, "Profiling is only supported by onnxruntime backend!"
from tf_estimator_SUT import get_tf_estimator_sut
sut = get_tf_estimator_sut()
elif args.backend == "onnxruntime":
from onnxruntime_SUT import get_onnxruntime_sut
sut = get_onnxruntime_sut(args)
else:
raise ValueError("Unknown backend: {:}".format(args.backend))
settings = lg.TestSettings()
settings.scenario = scenario_map[args.scenario]
settings.FromConfig(args.mlperf_conf, "bert", args.scenario)
settings.FromConfig(args.user_conf, "bert", args.scenario)
if args.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
else:
settings.mode = lg.TestMode.PerformanceOnly
log_path = "build/logs"
if not os.path.exists(log_path):
os.makedirs(log_path)
log_output_settings = lg.LogOutputSettings()
log_output_settings.outdir = log_path
log_output_settings.copy_summary_to_stdout = True
log_settings = lg.LogSettings()
log_settings.log_output = log_output_settings
print("Running LoadGen test...")
lg.StartTestWithLogSettings(sut.sut, sut.qsl.qsl, settings, log_settings)
if args.accuracy:
cmd = "python3 accuracy-squad.py"
subprocess.check_call(cmd, shell=True)
print("Done!")
print("Destroying SUT...")
lg.DestroySUT(sut.sut)
print("Destroying QSL...")
lg.DestroyQSL(sut.qsl.qsl)
def eval_func(model):
print("Running Loadgen test...")
sut.greedy_decoder._model = model
lg.StartTestWithLogSettings(sut.sut, sut.qsl.qsl, settings,
log_settings)
cmd = f"python3 accuracy_eval.py --log_dir {log_path} \
--dataset_dir {args.dataset_dir} --manifest {args.manifest}"
out = subprocess.check_output(cmd, shell=True)
out = out.decode()
regex_accu = re.compile(pattern[0])
accu = float(regex_accu.findall(out)[0].split('=')[1])
print('Accuracy: %.3f ' % (accu))
return accu
def main():
args = get_args()
batch_size = args.offline_batch_size if args.scenario == "Offline" else 1
settings = lg.TestSettings()
settings.scenario = scenario_map[args.scenario]
settings.FromConfig(args.mlperf_conf, "rnnt", args.scenario)
settings.FromConfig(args.user_conf, "rnnt", args.scenario)
issued_query_count = None
if args.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
issued_query_count = 2513
else:
settings.mode = lg.TestMode.PerformanceOnly
issued_query_count = settings.min_query_count
log_path = args.log_dir
os.makedirs(log_path, exist_ok=True)
log_output_settings = lg.LogOutputSettings()
log_output_settings.outdir = log_path
log_output_settings.copy_summary_to_stdout = True
log_settings = lg.LogSettings()
log_settings.log_output = log_output_settings
if args.backend == "pytorch":
from pytorch_SUT import PytorchSUT
sut = PytorchSUT(args.pytorch_config_toml, args.pytorch_checkpoint,
args.dataset_dir, args.manifest, args.perf_count,
issued_query_count, args.scenario, args.machine_conf,
batch_size, args.cores_for_loadgen,
args.cores_per_instance, args.debug, args.cosim,
args.profile, args.ipex, args.bf16, args.warmup)
else:
raise ValueError("Unknown backend: {:}".format(args.backend))
print("Running Loadgen test...")
lg.StartTestWithLogSettings(sut.sut, sut.qsl.qsl, settings, log_settings)
if args.accuracy:
cmd = f"python3 accuracy_eval.py --log_dir {log_path} --dataset_dir {args.dataset_dir} --manifest {args.manifest}"
print(f"Running accuracy script: {cmd}")
subprocess.check_call(cmd, shell=True)
lg.DestroySUT(sut.sut)
print("Done!")
def main():
args = get_args()
if args.backend == "pytorch":
from pytorch_SUT import get_pytorch_sut
sut = get_pytorch_sut(args.model_dir, args.preprocessed_data_dir,
args.performance_count)
elif args.backend == "onnxruntime":
from onnxruntime_SUT import get_onnxruntime_sut
sut = get_onnxruntime_sut(args.onnx_model, args.preprocessed_data_dir,
args.performance_count)
else:
raise ValueError("Unknown backend: {:}".format(args.backend))
settings = lg.TestSettings()
settings.scenario = scenario_map[args.scenario]
settings.FromConfig(args.mlperf_conf, "3d-unet", args.scenario)
settings.FromConfig(args.user_conf, "3d-unet", args.scenario)
if args.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
else:
settings.mode = lg.TestMode.PerformanceOnly
log_path = "build/logs"
if not os.path.exists(log_path):
os.makedirs(log_path)
log_output_settings = lg.LogOutputSettings()
log_output_settings.outdir = log_path
log_output_settings.copy_summary_to_stdout = True
log_settings = lg.LogSettings()
log_settings.log_output = log_output_settings
print("Running Loadgen test...")
lg.StartTestWithLogSettings(sut.sut, sut.qsl.qsl, settings, log_settings)
if args.accuracy:
print("Running accuracy script...")
cmd = "python3 brats_eval.py"
subprocess.check_call(cmd, shell=True)
print("Done!")
print("Destroying SUT...")
lg.DestroySUT(sut.sut)
print("Destroying QSL...")
lg.DestroyQSL(sut.qsl.qsl)
def benchmark_using_loadgen():
"Perform the benchmark using python API for the LoadGen library"
global num_classes
global model_output_volume
pycuda_context, max_batch_size, input_volume, model_output_volume, num_layers = initialize_predictor(
)
num_classes = len(class_labels)
scenario = {
'SingleStream': lg.TestScenario.SingleStream,
'MultiStream': lg.TestScenario.MultiStream,
'Server': lg.TestScenario.Server,
'Offline': lg.TestScenario.Offline,
}[LOADGEN_SCENARIO]
mode = {
'AccuracyOnly': lg.TestMode.AccuracyOnly,
'PerformanceOnly': lg.TestMode.PerformanceOnly,
'SubmissionRun': lg.TestMode.SubmissionRun,
}[LOADGEN_MODE]
ts = lg.TestSettings()
ts.FromConfig(MLPERF_CONF_PATH, MODEL_NAME, LOADGEN_SCENARIO)
ts.FromConfig(USER_CONF_PATH, MODEL_NAME, LOADGEN_SCENARIO)
ts.scenario = scenario
ts.mode = mode
if LOADGEN_MULTISTREAMNESS:
ts.multi_stream_samples_per_query = int(LOADGEN_MULTISTREAMNESS)
if LOADGEN_COUNT_OVERRIDE:
ts.min_query_count = int(LOADGEN_COUNT_OVERRIDE)
ts.max_query_count = int(LOADGEN_COUNT_OVERRIDE)
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(LOADGEN_DATASET_SIZE, LOADGEN_BUFFER_SIZE,
load_query_samples, unload_query_samples)
log_settings = lg.LogSettings()
log_settings.enable_trace = False
lg.StartTestWithLogSettings(sut, qsl, ts, log_settings)
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
pycuda_context.pop()
def benchmark(model):
print("Running Loadgen test...")
sut.greedy_decoder._model = model
lg.StartTestWithLogSettings(sut.sut, sut.qsl.qsl, settings,
log_settings)
file_path = os.path.join(log_path, 'mlperf_log_summary.txt')
f = open(file_path, 'r', encoding='UTF-8')
file_content = f.read()
f.close()
regex_batch = re.compile(pattern[1])
regex_late = re.compile(pattern[2])
samples_per_query = int(
regex_batch.findall(file_content)[0].split(': ')[1])
latency_per_sample = int(
regex_late.findall(file_content)[0].split(': ')[1])
print('Batch size = %d' % samples_per_query)
print('Latency: %.3f ms' % (latency_per_sample / 10**6))
print('Throughput: %.3f samples/sec' % (10**9 / latency_per_sample))
def perf_func(model):
print("Running Loadgen test...")
sut.greedy_decoder._model = model
lg.StartTestWithLogSettings(sut.sut, sut.qsl.qsl, settings,
log_settings)
if not args.accuracy:
file_path = os.path.join(log_path, 'mlperf_log_summary.txt')
f = open(file_path, 'r', encoding='UTF-8')
file_content = f.read()
f.close()
regex_batch = re.compile(pattern[1])
regex_thro = re.compile(pattern[2])
samples_per_query = int(
regex_batch.findall(file_content)[0].split(': ')[1])
samples_per_second = float(
regex_thro.findall(file_content)[0].split(': ')[1])
print('Batch size = %d' % samples_per_query)
print('Latency: %.3f ms' % ((1 / samples_per_second) * 1000))
print('Throughput: %.3f samples/sec' % samples_per_second)
def start(self):
"""Starts the load test."""
settings = self.get_test_settings()
log_settings = lg.LogSettings()
log_settings.log_output.outdir = tempfile.mkdtemp()
log_settings.log_output.copy_detail_to_stdout = True
log_settings.log_output.copy_summary_to_stdout = True
log_settings.enable_trace = False
logging.info("Constructing SUT.")
sut = lg.ConstructSUT(self.issue_queries, self.flush_queries,
self.process_metrics)
logging.info("Constructing QSL.")
qsl = lg.ConstructQSL(self.total_sample_count,
self.performance_sample_count, self.load_samples,
self.unload_samples)
logging.info("Starting test.")
lg.StartTestWithLogSettings(sut, qsl, settings, log_settings)
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
def benchmark_using_loadgen():
"Perform the benchmark using python API for the LoadGen library"
global pycuda_context
initialize_predictor()
scenario = {
'SingleStream': lg.TestScenario.SingleStream,
'MultiStream': lg.TestScenario.MultiStream,
'Server': lg.TestScenario.Server,
'Offline': lg.TestScenario.Offline,
}[LOADGEN_SCENARIO]
mode = {
'AccuracyOnly': lg.TestMode.AccuracyOnly,
'PerformanceOnly': lg.TestMode.PerformanceOnly,
'SubmissionRun': lg.TestMode.SubmissionRun,
}[LOADGEN_MODE]
ts = lg.TestSettings()
if LOADGEN_CONF_FILE:
ts.FromConfig(LOADGEN_CONF_FILE, 'random_model_name', LOADGEN_SCENARIO)
ts.scenario = scenario
ts.mode = mode
if LOADGEN_MULTISTREAMNESS:
ts.multi_stream_samples_per_query = int(LOADGEN_MULTISTREAMNESS)
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(LOADGEN_DATASET_SIZE, LOADGEN_BUFFER_SIZE,
load_query_samples, unload_query_samples)
log_settings = lg.LogSettings()
log_settings.enable_trace = False
lg.StartTestWithLogSettings(sut, qsl, ts, log_settings)
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
pycuda_context.pop()
def main():
args = get_args()
if args.backend == "pytorch":
from pytorch_SUT import PytorchSUT
sut = PytorchSUT(args.pytorch_config_toml, args.pytorch_checkpoint,
args.dataset_dir, args.manifest, args.perf_count)
else:
raise ValueError("Unknown backend: {:}".format(args.backend))
settings = lg.TestSettings()
settings.scenario = scenario_map[args.scenario]
settings.FromConfig(args.mlperf_conf, "rnnt", args.scenario)
settings.FromConfig(args.user_conf, "rnnt", args.scenario)
if args.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
else:
settings.mode = lg.TestMode.PerformanceOnly
log_path = args.log_dir
os.makedirs(log_path, exist_ok=True)
log_output_settings = lg.LogOutputSettings()
log_output_settings.outdir = log_path
log_output_settings.copy_summary_to_stdout = True
log_settings = lg.LogSettings()
log_settings.log_output = log_output_settings
print("Running Loadgen test...")
lg.StartTestWithLogSettings(sut.sut, sut.qsl.qsl, settings, log_settings)
if args.accuracy:
cmd = f"python3 accuracy_eval.py --log_dir {log_path} --dataset_dir {args.dataset_dir} --manifest {args.manifest}"
print(f"Running accuracy script: {cmd}")
subprocess.check_call(cmd, shell=True)
print("Done!")
def main(argv):
del argv
global last_timeing
if FLAGS.scenario == "Server":
# Disable garbage collection for realtime performance.
gc.disable()
# define backend
backend = BackendTensorflow()
# override image format if given
image_format = FLAGS.data_format if FLAGS.data_format else backend.image_format(
)
# dataset to use
wanted_dataset, pre_proc, post_proc, kwargs = SUPPORTED_DATASETS[
FLAGS.dataset]
ds = wanted_dataset(data_path=FLAGS.dataset_path,
image_list=FLAGS.dataset_list,
name=FLAGS.dataset,
image_format=image_format,
use_cache=FLAGS.cache,
count=FLAGS.count,
cache_dir=FLAGS.cache_dir,
annotation_file=FLAGS.annotation_file,
use_space_to_depth=FLAGS.use_space_to_depth)
# load model to backend
# TODO(wangtao): parse flags to params.
params = dict(ssd_model.default_hparams().values())
params["conv0_space_to_depth"] = FLAGS.use_space_to_depth
params["use_bfloat16"] = FLAGS.use_bfloat16
params["use_fused_bn"] = FLAGS.use_fused_bn
masters = []
tpu_names = FLAGS.tpu_name
tpu_names = tpu_names.split(",")
for tpu_name in tpu_names:
tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(
tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
masters.append(tpu_cluster_resolver.get_master())
#
# make one pass over the dataset to validate accuracy
#
count = FLAGS.count if FLAGS.count else ds.get_item_count()
#
# warmup
#
log.info("warmup ...")
batch_size = FLAGS.batch_size[0] if FLAGS.scenario == "Offline" else 1
backend_lists = []
for _ in range(len(tpu_names)):
backend = BackendTensorflow()
backend_lists.append(backend)
runner = QueueRunner(backend_lists,
ds,
FLAGS.threads,
post_proc=post_proc,
max_batchsize=batch_size)
runner.start_run({}, FLAGS.accuracy)
def issue_queries(query_samples):
for i in [1]:
runner.enqueue(query_samples)
def flush_queries():
pass
def process_latencies(latencies_ns):
# called by loadgen to show us the recorded latencies
global last_timeing
last_timeing = [t / NANO_SEC for t in latencies_ns]
tf.logging.info("starting {}, latency={}".format(FLAGS.scenario,
FLAGS.max_latency))
settings = lg.TestSettings()
tf.logging.info(FLAGS.scenario)
settings.scenario = SCENARIO_MAP[FLAGS.scenario]
settings.qsl_rng_seed = FLAGS.qsl_rng_seed
settings.sample_index_rng_seed = FLAGS.sample_index_rng_seed
settings.schedule_rng_seed = FLAGS.schedule_rng_seed
if FLAGS.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
else:
settings.mode = lg.TestMode.PerformanceOnly
if FLAGS.qps:
qps = float(FLAGS.qps)
settings.server_target_qps = qps
settings.offline_expected_qps = qps
if FLAGS.time:
settings.min_duration_ms = FLAGS.time * MILLI_SEC
settings.max_duration_ms = 0
qps = FLAGS.qps or 100
settings.min_query_count = qps * FLAGS.time
settings.max_query_count = 0
else:
settings.min_query_count = 270336
settings.max_query_count = 0
target_latency_ns = int(float(FLAGS.max_latency) * NANO_SEC)
settings.single_stream_expected_latency_ns = target_latency_ns
settings.multi_stream_target_latency_ns = target_latency_ns
settings.server_target_latency_ns = target_latency_ns
log_settings = lg.LogSettings()
log_settings.log_output.outdir = tempfile.mkdtemp()
log_settings.log_output.copy_detail_to_stdout = True
log_settings.log_output.copy_summary_to_stdout = True
log_settings.enable_trace = False
def load_query_samples(sample_list):
"""Load query samples and warmup the model."""
ds.load_query_samples(sample_list)
data = ds.get_image_list_inmemory()
def init_fn(cloud_tpu_id):
tf.logging.info("Load model for %dth cloud tpu", cloud_tpu_id)
runner.models[cloud_tpu_id].load(
FLAGS.model,
FLAGS.output_model_dir,
data,
params,
batch_size=FLAGS.batch_size,
master=masters[cloud_tpu_id],
scenario=FLAGS.scenario,
batch_timeout_micros=FLAGS.batch_timeout_micros)
# Init TPU.
for it in range(FLAGS.init_iterations):
tf.logging.info("Initialize cloud tpu at iteration %d", it)
for batch_size in FLAGS.batch_size:
example, _ = ds.get_indices([sample_list[0]] * batch_size)
_ = runner.models[cloud_tpu_id].predict(example)
threads = []
for i in range(len(tpu_names)):
thread = threading.Thread(target=init_fn, args=(i, ))
threads.append(thread)
thread.start()
for thread in threads:
thread.join()
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(count, min(count, 350), load_query_samples,
ds.unload_query_samples)
lg.StartTestWithLogSettings(sut, qsl, settings, log_settings)
runner.finish()
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
tf.io.gfile.mkdir(FLAGS.outdir)
for oldfile in tf.gfile.Glob(
os.path.join(log_settings.log_output.outdir, "*")):
basename = os.path.basename(oldfile)
newfile = os.path.join(FLAGS.outdir, basename)
tf.gfile.Copy(oldfile, newfile, overwrite=True)
if FLAGS.accuracy:
with tf.gfile.Open(os.path.join(FLAGS.outdir, "results.txt"),
"w") as f:
results = {"mAP": accuracy_coco.main()}
json.dump(results, f, sort_keys=True, indent=4)
def main():
global last_timeing
args = get_args()
log.info(args)
# find backend
backend = get_backend(args.backend)
# override image format if given
image_format = args.data_format if args.data_format else backend.image_format()
# --count applies to accuracy mode only and can be used to limit the number of images
# for testing. For perf model we always limit count to 200.
count_override = False
count = args.count
if count:
count_override = True
# dataset to use
wanted_dataset, pre_proc, post_proc, kwargs = SUPPORTED_DATASETS[args.dataset]
ds = wanted_dataset(data_path=args.dataset_path,
image_list=args.dataset_list,
name=args.dataset,
image_format=image_format,
pre_process=pre_proc,
use_cache=args.cache,
count=count, **kwargs)
# load model to backend
model = backend.load(args.model, inputs=args.inputs, outputs=args.outputs)
final_results = {
"runtime": model.name(),
"version": model.version(),
"time": int(time.time()),
"cmdline": str(args),
}
config = os.path.abspath(args.config)
if not os.path.exists(config):
log.error("{} not found".format(config))
sys.exit(1)
if args.output:
output_dir = os.path.abspath(args.output)
os.makedirs(output_dir, exist_ok=True)
os.chdir(output_dir)
#
# make one pass over the dataset to validate accuracy
#
count = ds.get_item_count()
# warmup
warmup_queries = range(args.max_batchsize)
ds.load_query_samples(warmup_queries)
for _ in range(2):
img, _ = ds.get_samples(warmup_queries)
_ = backend.predict({backend.inputs[0]: img})
ds.unload_query_samples(None)
scenario = SCENARIO_MAP[args.scenario]
runner_map = {
lg.TestScenario.SingleStream: RunnerBase,
lg.TestScenario.MultiStream: QueueRunner,
lg.TestScenario.Server: QueueRunner,
lg.TestScenario.Offline: QueueRunner
}
runner = runner_map[scenario](model, ds, args.threads, post_proc=post_proc, max_batchsize=args.max_batchsize)
def issue_queries(query_samples):
runner.enqueue(query_samples)
def flush_queries():
pass
def process_latencies(latencies_ns):
# called by loadgen to show us the recorded latencies
global last_timeing
last_timeing = [t / NANO_SEC for t in latencies_ns]
settings = lg.TestSettings()
settings.FromConfig(config, args.model_name, args.scenario)
settings.scenario = scenario
settings.mode = lg.TestMode.PerformanceOnly
if args.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
if args.find_peak_performance:
settings.mode = lg.TestMode.FindPeakPerformance
if args.time:
# override the time we want to run
settings.min_duration_ms = args.time * MILLI_SEC
settings.max_duration_ms = args.time * MILLI_SEC
if args.qps:
qps = float(args.qps)
settings.server_target_qps = qps
settings.offline_expected_qps = qps
if count_override:
settings.min_query_count = count
settings.max_query_count = count
if args.samples_per_query:
settings.multi_stream_samples_per_query = args.samples_per_query
if args.max_latency:
settings.server_target_latency_ns = int(args.max_latency * NANO_SEC)
settings.multi_stream_target_latency_ns = int(args.max_latency * NANO_SEC)
# override target latency when it needs to be less than 1ms
if args.model_name == "mobilenet":
settings.single_stream_expected_latency_ns = 200000
elif args.model_name == "resnet50":
settings.single_stream_expected_latency_ns = 900000
elif args.model_name == "ssd-mobilenet":
settings.single_stream_expected_latency_ns = 1000000
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
#qsl = lg.ConstructQSL(count, min(count, 500), ds.load_query_samples, ds.unload_query_samples)
qsl = lg.ConstructQSL(count, min(count, 1024), ds.load_query_samples, ds.unload_query_samples)
log.info("starting {}".format(scenario))
result_dict = {"good": 0, "total": 0, "scenario": str(scenario)}
runner.start_run(result_dict, args.accuracy)
if args.enable_trace:
lg.StartTest(sut, qsl, settings)
else:
logsettings = lg.LogSettings()
logsettings.enable_trace = False
lg.StartTestWithLogSettings(sut, qsl, settings, logsettings)
if not last_timeing:
last_timeing = runner.result_timing
if args.accuracy:
post_proc.finalize(result_dict, ds, output_dir=args.output)
add_results(final_results, "{}".format(scenario),
result_dict, last_timeing, time.time() - ds.last_loaded, args.accuracy)
runner.finish()
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
#
# write final results
#
if args.output:
with open("results.json", "w") as f:
json.dump(final_results, f, sort_keys=True, indent=4)
def benchmark_using_loadgen():
"Perform the benchmark using python API for the LoadGen library"
global funnel_should_be_running, warmup_mode, openme_data
scenario = {
'SingleStream': lg.TestScenario.SingleStream,
'MultiStream': lg.TestScenario.MultiStream,
'Server': lg.TestScenario.Server,
'Offline': lg.TestScenario.Offline,
}[LOADGEN_SCENARIO]
mode = {
'AccuracyOnly': lg.TestMode.AccuracyOnly,
'PerformanceOnly': lg.TestMode.PerformanceOnly,
'SubmissionRun': lg.TestMode.SubmissionRun,
}[LOADGEN_MODE]
ts = lg.TestSettings()
if LOADGEN_CONFIG_FILE:
ts.FromConfig(LOADGEN_CONFIG_FILE, 'random_model_name', LOADGEN_SCENARIO)
ts.scenario = scenario
ts.mode = mode
if LOADGEN_MULTISTREAMNESS:
ts.multi_stream_samples_per_query = int(LOADGEN_MULTISTREAMNESS)
if LOADGEN_MAX_DURATION_S:
ts.max_duration_ms = int(LOADGEN_MAX_DURATION_S)*1000
if LOADGEN_COUNT_OVERRIDE:
ts.min_query_count = int(LOADGEN_COUNT_OVERRIDE)
ts.max_query_count = int(LOADGEN_COUNT_OVERRIDE)
if LOADGEN_TARGET_QPS:
target_qps = float(LOADGEN_TARGET_QPS)
ts.multi_stream_target_qps = target_qps
ts.server_target_qps = target_qps
ts.offline_expected_qps = target_qps
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(LOADGEN_DATASET_SIZE, LOADGEN_BUFFER_SIZE, load_query_samples, unload_query_samples)
log_settings = lg.LogSettings()
log_settings.enable_trace = False
funnel_thread = threading.Thread(target=send_responses, args=())
funnel_should_be_running = True
funnel_thread.start()
if LOADGEN_WARMUP_SAMPLES:
warmup_id_range = list(range(LOADGEN_WARMUP_SAMPLES))
load_query_samples(warmup_id_range)
warmup_mode = True
print("Sending out the warm-up samples, waiting for responses...")
issue_queries([lg.QuerySample(id,id) for id in warmup_id_range])
while len(in_progress)>0: # waiting for the in_progress queue to clear up
time.sleep(1)
print(" Done!")
warmup_mode = False
lg.StartTestWithLogSettings(sut, qsl, ts, log_settings)
funnel_should_be_running = False # politely ask the funnel_thread to end
funnel_thread.join() # wait for it to actually end
from_workers.close()
to_workers.close()
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
if SIDELOAD_JSON:
with open(SIDELOAD_JSON, 'w') as sideload_fd:
json.dump(openme_data, sideload_fd, indent=4, sort_keys=True)
def eval_func(model):
args = get_args()
if args.backend == "pytorch":
from pytorch_SUT import get_pytorch_sut
sut = get_pytorch_sut(model, args.preprocessed_data_dir,
args.performance_count)
elif args.backend == "onnxruntime":
from onnxruntime_SUT import get_onnxruntime_sut
sut = get_onnxruntime_sut(args.model, args.preprocessed_data_dir,
args.performance_count)
elif args.backend == "tf":
from tf_SUT import get_tf_sut
sut = get_tf_sut(args.model, args.preprocessed_data_dir,
args.performance_count)
elif args.backend == "ov":
from ov_SUT import get_ov_sut
sut = get_ov_sut(args.model, args.preprocessed_data_dir,
args.performance_count)
else:
raise ValueError("Unknown backend: {:}".format(args.backend))
settings = lg.TestSettings()
settings.scenario = scenario_map[args.scenario]
settings.FromConfig(args.mlperf_conf, "3d-unet", args.scenario)
settings.FromConfig(args.user_conf, "3d-unet", args.scenario)
if args.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
else:
settings.mode = lg.TestMode.PerformanceOnly
log_path = "build/logs"
if not os.path.exists(log_path):
os.makedirs(log_path)
log_output_settings = lg.LogOutputSettings()
log_output_settings.outdir = log_path
log_output_settings.copy_summary_to_stdout = True
log_settings = lg.LogSettings()
log_settings.log_output = log_output_settings
print("Running Loadgen test...")
if args.benchmark:
start = time.time()
lg.StartTestWithLogSettings(sut.sut, sut.qsl.qsl, settings, log_settings)
if args.benchmark:
end = time.time()
if args.accuracy:
print("Running accuracy script...")
process = subprocess.Popen(['python3', 'accuracy-brats.py'],
stdout=subprocess.PIPE,
stderr=subprocess.PIPE)
out, err = process.communicate()
print(out)
print("Done!", float(err))
if args.benchmark:
print('Batch size = 1')
print('Latency: %.3f ms' % ((end - start) * 1000 / sut.qsl.count))
print('Throughput: %.3f images/sec' % (sut.qsl.count /
(end - start)))
print('Accuracy: {mean:.5f}'.format(mean=float(err)))
print("Destroying SUT...")
lg.DestroySUT(sut.sut)
print("Destroying QSL...")
lg.DestroyQSL(sut.qsl.qsl)
return float(err)
def main():
global num_ins
global num_cpus
global in_queue_cnt
global out_queue_cnt
global batching
global queries_so_far
global Latencies
queries_so_far = 0
args = get_args()
log.info(args)
scenario = args.scenario
accuracy_mode = args.accuracy
perf_count = args.perf_count
batch_size = args.batch_size
num_ins = args.num_instance
num_cpus = args.num_phy_cpus
batching = args.batching
# Read Loadgen and workload config parameters
settings = lg.TestSettings()
settings.scenario = scenario_map[scenario]
settings.FromConfig(args.mlperf_conf, "bert", scenario)
settings.FromConfig(args.user_conf, "bert", scenario)
settings.mode = lg.TestMode.AccuracyOnly if accuracy_mode else lg.TestMode.PerformanceOnly
# Establish communication queues
lock = multiprocessing.Lock()
init_counter = multiprocessing.Value("i", 0)
calibrate_counter = multiprocessing.Value("i", 0)
out_queue = multiprocessing.Queue()
# Create consumers
consumers = []
if scenario == "Server":
from parse_server_config import configParser
buckets = configParser("machine_conf.json")
cutoffs = list(buckets.keys())
batch_sizes = {}
in_queue = {j: multiprocessing.JoinableQueue() for j in buckets}
proc_idx = 0
num_cpus = 0
total_ins = 0
for cutoff in list(buckets.keys()):
batch_sizes[cutoff] = buckets[cutoff]["batch_size"]
num_ins = buckets[cutoff]["instances"]
cpus_per_instance = buckets[cutoff]["cpus_per_instance"]
num_cpus = num_ins * cpus_per_instance
total_ins += num_ins
for j in range(num_ins):
consumer = Consumer(in_queue[cutoff], out_queue, lock,
init_counter, calibrate_counter, proc_idx,
num_ins, args, cutoff)
consumer.start_core_idx = proc_idx
consumer.end_core_idx = proc_idx + cpus_per_instance - 1
consumers.append(consumer)
proc_idx = consumer.end_core_idx + 1
num_ins = total_ins
else:
total_ins = num_ins
in_queue = MultiprocessShapeBasedQueue()
consumers = [
Consumer(in_queue, out_queue, lock, init_counter,
calibrate_counter, i, num_ins, args)
for i in range(num_ins)
]
for c in consumers:
c.start()
# Dataset object used by constructQSL
data_set = BERTDataSet(args.vocab, args.perf_count)
if scenario == "Server":
issue_queue = InQueueServer(in_queue, batch_sizes, data_set,
settings.min_query_count)
else:
issue_queue = InQueue(in_queue, batch_size, data_set)
# Wait until all sub-processors are ready
block_until(init_counter, total_ins, 2)
# Start response thread
response_worker = threading.Thread(target=response_loadgen,
args=(out_queue, ))
response_worker.daemon = True
response_worker.start()
def issue_queries(query_samples):
# It's called by loadgen to send query to SUT
issue_queue.put(query_samples)
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(data_set.count, data_set.perf_count,
load_query_samples, unload_query_samples)
log_path = "build/logs"
if not os.path.exists(log_path):
os.makedirs(log_path)
log_output_settings = lg.LogOutputSettings()
log_output_settings.outdir = log_path
log_output_settings.copy_summary_to_stdout = True
log_settings = lg.LogSettings()
log_settings.log_output = log_output_settings
lg.StartTestWithLogSettings(sut, qsl, settings, log_settings)
# Wait until outQueue done
while out_queue_cnt < in_queue_cnt:
time.sleep(0.2)
if scenario == "Server":
for i in in_queue:
in_queue[i].join()
for j in range(buckets[i]["cpus_per_instance"]):
in_queue[i].put(None)
else:
for i in range(num_ins):
in_queue.put(None)
for c in consumers:
c.join()
out_queue.put(None)
if accuracy_mode:
cmd = "python accuracy-squad.py --log_file={}/mlperf_log_accuracy.json".format(
log_path)
subprocess.check_call(cmd, shell=True)
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
def main():
global num_ins
global num_cpus
global in_queue_cnt
global out_queue_cnt
global batching
global bs_step
args = get_args()
log.info(args)
scenario = args.scenario
accuracy_mode = args.accuracy
perf_count = args.perf_count
batch_size = args.batch_size
num_ins = args.num_instance
num_cpus = args.num_phy_cpus
batching = args.batching
## TODO, remove
log.info('Run with {} instance on {} cpus: '.format(num_ins, num_cpus))
# Establish communication queues
lock = multiprocessing.Lock()
init_counter = multiprocessing.Value("i", 0)
calibrate_counter = multiprocessing.Value("i", 0)
out_queue = multiprocessing.Queue()
in_queue = MultiprocessShapeBasedQueue()
if args.perf_calibrate:
with open('prof_new.py', 'w') as f:
print('prof_bs_step = {}'.format(bs_step), file=f)
print('prof_map = {', file=f)
# Start consumers
consumers = [
Consumer(in_queue, out_queue, lock, init_counter, calibrate_counter, i,
num_ins, args) for i in range(num_ins)
]
for c in consumers:
c.start()
# used by constructQSL
data_set = BERTDataSet(args.vocab, args.perf_count)
issue_queue = InQueue(in_queue, batch_size, data_set)
# Wait until all sub-processors ready to do calibration
block_until(calibrate_counter, num_ins)
# Wait until all sub-processors done calibration
block_until(calibrate_counter, 2 * num_ins)
if args.perf_calibrate:
with open('prof_new.py', 'a') as f:
print('}', file=f)
sys.exit(0)
# Wait until all sub-processors are ready
block_until(init_counter, num_ins)
# Start response thread
response_worker = threading.Thread(target=response_loadgen,
args=(out_queue, ))
response_worker.daemon = True
response_worker.start()
# Start loadgen
settings = lg.TestSettings()
settings.scenario = scenario_map[scenario]
settings.FromConfig(args.mlperf_conf, "bert", scenario)
settings.FromConfig(args.user_conf, "bert", scenario)
settings.mode = lg.TestMode.AccuracyOnly if accuracy_mode else lg.TestMode.PerformanceOnly
# TODO, for debug, remove
#settings.server_target_qps = 40
#settings.server_target_latency_ns = 100000000
#settings.min_query_count = 100
#settings.min_duration_ms = 10000
def issue_queries(query_samples):
# It's called by loadgen to send query to SUT
issue_queue.put(query_samples)
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(data_set.count, data_set.perf_count,
load_query_samples, unload_query_samples)
log_path = "build/logs"
if not os.path.exists(log_path):
os.makedirs(log_path)
log_output_settings = lg.LogOutputSettings()
log_output_settings.outdir = log_path
log_output_settings.copy_summary_to_stdout = True
log_settings = lg.LogSettings()
log_settings.log_output = log_output_settings
#lg.StartTest(sut, qsl, settings)
lg.StartTestWithLogSettings(sut, qsl, settings, log_settings)
# Wait until outQueue done
while out_queue_cnt < in_queue_cnt:
time.sleep(0.2)
in_queue.join()
for i in range(num_ins):
in_queue.put(None)
for c in consumers:
c.join()
out_queue.put(None)
if accuracy_mode:
cmd = "python accuracy-squad.py --log_file={}/mlperf_log_accuracy.json".format(
log_path)
subprocess.check_call(cmd, shell=True)
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
def start_run(self):
lg.StartTestWithLogSettings(self.sut, self.qsl, self.settings, self.log_settings)
def main():
"""
Runs 3D UNet performing KiTS19 Kidney Tumore Segmentation task as below:
1. instantiate SUT and QSL for the chosen backend
2. configure LoadGen for the chosen scenario
3. configure MLPerf logger
4. start LoadGen
5. collect logs and if needed evaluate inference results
6. clean up
"""
# scenarios in LoadGen
scenario_map = {
"SingleStream": lg.TestScenario.SingleStream,
"Offline": lg.TestScenario.Offline,
"Server": lg.TestScenario.Server,
"MultiStream": lg.TestScenario.MultiStream
}
args = get_args()
# instantiate SUT as per requested backend; QSL is also instantiated
if args.backend == "pytorch":
from pytorch_SUT import get_sut
elif args.backend == "pytorch_checkpoint":
from pytorch_checkpoint_SUT import get_sut
elif args.backend == "onnxruntime":
from onnxruntime_SUT import get_sut
elif args.backend == "tensorflow":
from tensorflow_SUT import get_sut
else:
raise ValueError("Unknown backend: {:}".format(args.backend))
sut = get_sut(args.model, args.preprocessed_data_dir,
args.performance_count)
# setup LoadGen
settings = lg.TestSettings()
settings.scenario = scenario_map[args.scenario]
settings.FromConfig(args.mlperf_conf, "3d-unet", args.scenario)
settings.FromConfig(args.user_conf, "3d-unet", args.scenario)
if args.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
else:
settings.mode = lg.TestMode.PerformanceOnly
# set up mlperf logger
log_path = Path("build", "logs").absolute()
log_path.mkdir(parents=True, exist_ok=True)
log_output_settings = lg.LogOutputSettings()
log_output_settings.outdir = str(log_path)
log_output_settings.copy_summary_to_stdout = True
log_settings = lg.LogSettings()
log_settings.log_output = log_output_settings
# start running test, from LoadGen
print("Running Loadgen test...")
lg.StartTestWithLogSettings(sut.sut, sut.qsl.qsl, settings, log_settings)
# if needed check accuracy
if args.accuracy:
print("Checking accuracy...")
cmd = "python3 accuracy_kits.py"
subprocess.check_call(cmd, shell=True)
# all done
print("Done!")
# cleanup
print("Destroying SUT...")
lg.DestroySUT(sut.sut)
print("Destroying QSL...")
lg.DestroyQSL(sut.qsl.qsl)
def main():
global so
global last_timeing
global last_loaded
global result_timeing
args = get_args()
log.info(args)
# find backend
backend = get_backend(args.backend)
# --count applies to accuracy mode only and can be used to limit the number of images
# for testing. For perf model we always limit count to 200.
count_override = False
count = args.count
if count:
count_override = True
"""
Python signature
go_initialize(backend, model_path, dataset_path, count, use_gpu, gpu_id, trace_level, max_batchsize)
"""
count, err = go_initialize(backend, args.model_path, args.dataset_path,
count, args.use_gpu, args.gpu_id,
args.trace_level, args.max_batchsize)
if (err != 'nil'):
print(err)
raise RuntimeError('initialization in go failed')
mlperf_conf = os.path.abspath(args.mlperf_conf)
if not os.path.exists(mlperf_conf):
log.error("{} not found".format(mlperf_conf))
sys.exit(1)
user_conf = os.path.abspath(args.user_conf)
if not os.path.exists(user_conf):
log.error("{} not found".format(user_conf))
sys.exit(1)
log_dir = None
if args.log_dir:
log_dir = os.path.abspath(args.log_dir)
os.makedirs(log_dir, exist_ok=True)
scenario = SCENARIO_MAP[args.scenario]
def issue_queries(query_samples):
global so
global last_timeing
global result_timeing
idx = np.array([q.index for q in query_samples]).astype(np.int32)
query_id = [q.id for q in query_samples]
if args.dataset == 'brats2019':
start = time.time()
response_array_refs = []
response = []
for i, qid in enumerate(query_id):
processed_results = so.IssueQuery(1, idx[i][np.newaxis])
processed_results = json.loads(
processed_results.decode('utf-8'))
response_array = array.array(
"B",
np.array(processed_results[0], np.float16).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(qid, bi[0], bi[1]))
result_timeing.append(time.time() - start)
lg.QuerySamplesComplete(response)
else:
start = time.time()
processed_results = so.IssueQuery(len(idx), idx)
result_timeing.append(time.time() - start)
processed_results = json.loads(processed_results.decode('utf-8'))
response_array_refs = []
response = []
for idx, qid in enumerate(query_id):
response_array = array.array(
"B",
np.array(processed_results[idx], np.float32).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(qid, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
def flush_queries():
pass
def process_latencies(latencies_ns):
# called by loadgen to show us the recorded latencies
global last_timeing
last_timeing = [t / NANO_SEC for t in latencies_ns]
def load_query_samples(sample_list):
global so
global last_loaded
err = go_load_query_samples(sample_list, so)
last_loaded = time.time()
if (err != ''):
print(err)
raise RuntimeError('load query samples failed')
def unload_query_samples(sample_list):
global so
err = go_unload_query_samples(sample_list, so)
if (err != ''):
print(err)
raise RuntimeError('unload query samples failed')
settings = lg.TestSettings()
if args.model_name != "":
settings.FromConfig(mlperf_conf, args.model_name, args.scenario)
settings.FromConfig(user_conf, args.model_name, args.scenario)
settings.scenario = scenario
settings.mode = lg.TestMode.PerformanceOnly
if args.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
if args.find_peak_performance:
settings.mode = lg.TestMode.FindPeakPerformance
if args.time:
# override the time we want to run
settings.min_duration_ms = args.time * MILLI_SEC
settings.max_duration_ms = args.time * MILLI_SEC
if args.qps:
qps = float(args.qps)
settings.server_target_qps = qps
settings.offline_expected_qps = qps
if count_override:
settings.min_query_count = count
settings.max_query_count = count
if args.samples_per_query:
settings.multi_stream_samples_per_query = args.samples_per_query
if args.max_latency:
settings.server_target_latency_ns = int(args.max_latency * NANO_SEC)
settings.multi_stream_target_latency_ns = int(args.max_latency *
NANO_SEC)
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(count, min(count, 500), load_query_samples,
unload_query_samples)
log.info("starting {}".format(scenario))
log_path = os.path.realpath(args.log_dir)
log_output_settings = lg.LogOutputSettings()
log_output_settings.outdir = log_path
log_output_settings.copy_summary_to_stdout = True
log_settings = lg.LogSettings()
log_settings.log_output = log_output_settings
# log_settings.enable_trace = True
# lg.StartTest(sut, qsl, settings)
lg.StartTestWithLogSettings(sut, qsl, settings, log_settings)
if not last_timeing:
last_timeing = result_timeing
if args.accuracy:
accuracy_script_paths = {
'coco':
os.path.realpath(
'../inference/vision/classification_and_detection/tools/accuracy-coco.py'
),
'imagenet':
os.path.realpath(
'../inference/vision/classification_and_detection/tools/accuracy-imagenet.py'
),
'squad':
os.path.realpath('../inference/language/bert/accuracy-squad.py'),
'brats2019':
os.path.realpath(
'../inference/vision/medical_imaging/3d-unet/accuracy-brats.py'
),
}
accuracy_script_path = accuracy_script_paths[args.dataset]
accuracy_file_path = os.path.join(log_dir, 'mlperf_log_accuracy.json')
data_dir = os.environ['DATA_DIR']
if args.dataset == 'coco':
if args.use_inv_map:
subprocess.check_call(
'python3 {} --mlperf-accuracy-file {} --coco-dir {} --use-inv-map'
.format(accuracy_script_path, accuracy_file_path,
data_dir),
shell=True)
else:
subprocess.check_call(
'python3 {} --mlperf-accuracy-file {} --coco-dir {}'.
format(accuracy_script_path, accuracy_file_path, data_dir),
shell=True)
elif args.dataset == 'imagenet': # imagenet
subprocess.check_call(
'python3 {} --mlperf-accuracy-file {} --imagenet-val-file {}'.
format(accuracy_script_path, accuracy_file_path,
os.path.join(data_dir, 'val_map.txt')),
shell=True)
elif args.dataset == 'squad': # squad
vocab_path = os.path.join(data_dir, 'vocab.txt')
val_path = os.path.join(data_dir, 'dev-v1.1.json')
out_path = os.path.join(log_dir, 'predictions.json')
cache_path = os.path.join(data_dir, 'eval_features.pickle')
subprocess.check_call(
'python3 {} --vocab_file {} --val_data {} --log_file {} --out_file {} --features_cache_file {} --max_examples {}'
.format(accuracy_script_path, vocab_path, val_path,
accuracy_file_path, out_path, cache_path, count),
shell=True)
elif args.dataset == 'brats2019': # brats2019
base_dir = os.path.realpath(
'../inference/vision/medical_imaging/3d-unet/build')
post_dir = os.path.join(base_dir, 'postprocessed_data')
label_dir = os.path.join(
base_dir,
'raw_data/nnUNet_raw_data/Task043_BraTS2019/labelsTr')
os.makedirs(post_dir, exist_ok=True)
subprocess.check_call(
'python3 {} --log_file {} --preprocessed_data_dir {} --postprocessed_data_dir {} --label_data_dir {}'
.format(accuracy_script_path, accuracy_file_path, data_dir,
post_dir, label_dir),
shell=True)
else:
raise RuntimeError('Dataset not Implemented.')
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
"""
Python signature
go_finalize(so)
"""
err = go_finalize(so)
if (err != ''):
print(err)
raise RuntimeError('finialize in go failed')
def run():
"""Runs the offline mode."""
global last_timing
# Initiazation
final_results, count, runner = setup()
#
# run the benchmark with timing
#
runner.start_pool()
def issue_query_offline(query_samples):
"""Adds query to the queue."""
for i in [1]:
idx = np.array([q.index for q in query_samples])
query_id = np.array([q.id for q in query_samples])
batch_size = FLAGS.batch_size[0]
for i in range(0, len(query_samples), batch_size):
runner.enqueue(query_id[i:i + batch_size], idx[i:i + batch_size])
def flush_queries():
pass
def process_latencies(latencies_ns):
global last_timing
last_timing = [t / 1e9 for t in latencies_ns]
sut = lg.ConstructSUT(issue_query_offline, flush_queries, process_latencies)
masters = []
outdir = FLAGS.outdir if FLAGS.outdir else tempfile.mkdtemp()
export_outdir = FLAGS.export_outdir if FLAGS.export_outdir else outdir
export_outdir = os.path.join(export_outdir, "export_model")
def load_query_samples(sample_list):
"""Load query samples."""
runner.ds.load_query_samples(sample_list)
# Find tpu master.
if FLAGS.num_tpus == 1:
runner.model.update_qsl(runner.ds.get_image_list_inmemory())
else:
for i in range(FLAGS.num_tpus):
runner.models[i].update_qsl(runner.ds.get_image_list_inmemory())
def warmup():
"""Warmup the TPUs."""
load_query_samples([0])
if FLAGS.num_tpus == 1:
log.info("warmup ...")
runner.warmup(0)
log.info("warmup done")
else:
for cloud_tpu_id in range(FLAGS.num_tpus):
log.info("warmup %d...", cloud_tpu_id)
runner.warmup(0, cloud_tpu_id)
log.info("warmup %d done", cloud_tpu_id)
# After warmup, give the system a moment to quiesce before putting it under
# load.
time.sleep(1)
if FLAGS.num_tpus == 1:
tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(
FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
master = tpu_cluster_resolver.get_master()
runner.model.build_and_export(
FLAGS.model,
export_model_path=export_outdir,
batch_size=FLAGS.batch_size,
master=master,
scenario=FLAGS.scenario)
runner.model.load(export_model_path=export_outdir, master=master)
else:
# Use the first TPU instance to build and export the graph.
tpu_names = FLAGS.tpu_name
tpu_names = tpu_names.split(",")
for tpu_name in tpu_names:
tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(
tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
masters.append(tpu_cluster_resolver.get_master())
runner.models[0].build_and_export(
FLAGS.model,
export_model_path=export_outdir,
batch_size=FLAGS.batch_size,
master=masters[0],
scenario=FLAGS.scenario)
def init_fn(cloud_tpu_id):
"""Init and warmup each cloud tpu."""
runner.models[cloud_tpu_id].load(
export_model_path=export_outdir, master=masters[cloud_tpu_id])
threads = []
for i in range(FLAGS.num_tpus):
thread = threading.Thread(target=init_fn, args=(i,))
threads.append(thread)
thread.start()
for thread in threads:
thread.join()
warmup()
qsl = lg.ConstructQSL(count, min(count, 1024), load_query_samples,
runner.ds.unload_query_samples)
test_scenarios = FLAGS.scenario
if test_scenarios is None:
test_scenarios_list = []
else:
test_scenarios_list = test_scenarios.split(",")
max_latency = FLAGS.max_latency
max_latency_list = max_latency.split(",")
for scenario in test_scenarios_list:
for target_latency in max_latency_list:
log.info("starting %s, latency=%s", scenario, target_latency)
settings = lg.TestSettings()
log.info(scenario)
if FLAGS.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
settings.scenario = utils.SCENARIO_MAP[scenario]
if FLAGS.qps:
qps = float(FLAGS.qps)
settings.server_target_qps = qps
settings.offline_expected_qps = qps
if FLAGS.time:
settings.min_duration_ms = 60 * MILLI_SEC
settings.max_duration_ms = 0
qps = FLAGS.qps or 100
settings.min_query_count = qps * FLAGS.time
settings.max_query_count = int(1.1 * qps * FLAGS.time)
else:
settings.min_query_count = (1 << 21)
if FLAGS.time or FLAGS.qps and FLAGS.accuracy:
settings.mode = lg.TestMode.PerformanceOnly
# FIXME: add SubmissionRun once available
target_latency_ns = int(float(target_latency) * (NANO_SEC / MILLI_SEC))
settings.single_stream_expected_latency_ns = target_latency_ns
settings.multi_stream_target_latency_ns = target_latency_ns
settings.server_target_latency_ns = target_latency_ns
log_settings = lg.LogSettings()
# TODO(brianderson): figure out how to use internal file path.
log_settings.log_output.outdir = tempfile.mkdtemp()
log_settings.log_output.copy_detail_to_stdout = True
log_settings.log_output.copy_summary_to_stdout = True
log_settings.enable_trace = False
result_dict = {"good": 0, "total": 0, "scenario": str(scenario)}
runner.start_run(result_dict, FLAGS.accuracy)
lg.StartTestWithLogSettings(sut, qsl, settings, log_settings)
if FLAGS.accuracy:
runner.get_post_process().finalize(result_dict, runner.ds)
utils.add_results(
final_results, "{}-{}".format(scenario, target_latency),
result_dict, last_timing,
time.time() - runner.ds.last_loaded)
#
# write final results
#
if FLAGS.outdir:
outfile = os.path.join(FLAGS.outdir, "results.txt")
with tf.gfile.Open(outfile, "w") as f:
json.dump(final_results, f, sort_keys=True, indent=4)
else:
json.dump(final_results, sys.stdout, sort_keys=True, indent=4)
runner.finish()
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
def main(argv):
del argv
settings = mlperf_loadgen.TestSettings()
settings.qsl_rng_seed = FLAGS.qsl_rng_seed
settings.sample_index_rng_seed = FLAGS.sample_index_rng_seed
settings.schedule_rng_seed = FLAGS.schedule_rng_seed
if FLAGS.accuracy_mode:
settings.mode = mlperf_loadgen.TestMode.AccuracyOnly
else:
settings.mode = mlperf_loadgen.TestMode.PerformanceOnly
settings.scenario = SCENARIO_MAP[FLAGS.scenario]
if FLAGS.qps:
qps = float(FLAGS.qps)
settings.server_target_qps = qps
settings.offline_expected_qps = qps
if FLAGS.scenario == "Offline" or FLAGS.scenario == "Server":
masters = FLAGS.master
masters = masters.split(",")
if len(masters) < 1:
masters = [FLAGS.master]
runner = loadgen_gnmt.GNMTRunner(input_file=FLAGS.input_file,
ckpt_path=FLAGS.ckpt_path,
hparams_path=FLAGS.hparams_path,
vocab_prefix=FLAGS.vocab_prefix,
outdir=FLAGS.outdir,
batch_size=FLAGS.batch_size,
verbose=FLAGS.verbose,
masters=masters,
scenario=FLAGS.scenario)
runner.load(FLAGS.batch_timeout_micros)
# Specify exactly how many queries need to be made
settings.min_query_count = FLAGS.qps * FLAGS.time
settings.max_query_count = 0
settings.min_duration_ms = 60 * MILLI_SEC
settings.max_duration_ms = 0
settings.server_target_latency_ns = int(0.25 * NANO_SEC)
settings.server_target_latency_percentile = 0.97
else:
print("Invalid scenario selected")
assert False
# Create a thread in the GNMTRunner to start accepting work
runner.start_worker()
# Maximum sample ID + 1
total_queries = FLAGS.query_count
# Select the same subset of $perf_queries samples
perf_queries = FLAGS.query_count
sut = mlperf_loadgen.ConstructSUT(runner.enqueue, flush_queries,
generic_loadgen.process_latencies)
qsl = mlperf_loadgen.ConstructQSL(total_queries, perf_queries,
runner.load_samples_to_ram,
runner.unload_samples_from_ram)
log_settings = mlperf_loadgen.LogSettings()
log_settings.log_output.outdir = tempfile.mkdtemp()
# Disable detail logs to prevent it from stepping on the summary
# log in stdout on some systems.
log_settings.log_output.copy_detail_to_stdout = False
log_settings.log_output.copy_summary_to_stdout = True
log_settings.enable_trace = False
mlperf_loadgen.StartTestWithLogSettings(sut, qsl, settings, log_settings)
runner.finish()
mlperf_loadgen.DestroyQSL(qsl)
mlperf_loadgen.DestroySUT(sut)
for oldfile in tf.gfile.Glob(
os.path.join(log_settings.log_output.outdir, "*")):
basename = os.path.basename(oldfile)
newfile = os.path.join(FLAGS.outdir, basename)
tf.gfile.Copy(oldfile, newfile, overwrite=True)
if FLAGS.accuracy_mode:
log_accuracy = os.path.join(log_settings.log_output.outdir,
"mlperf_log_accuracy.json")
tf.gfile.Copy(FLAGS.reference, "/tmp/reference")
bleu = process_accuracy.get_accuracy("/tmp/reference", log_accuracy)
print("BLEU: %.2f" % (bleu * 100)) # pylint: disable=superfluous-parens
