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),
}
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)
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
ds.load_query_samples([0])
for _ in range(5):
img, _ = ds.get_samples([0])
_ = 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]
log_output_settings = lg.LogOutputSettings()
log_output_settings.outdir = output_dir
log_output_settings.copy_summary_to_stdout = False
log_settings = lg.LogSettings()
log_settings.enable_trace = args.debug
log_settings.log_output = log_output_settings
settings = lg.TestSettings()
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), 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)
lg.StartTestWithLogSettings(sut, qsl, settings, log_settings)
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 main():
global last_timeing
args = get_args()
log.info(args)
# find backend
backend = get_backend(args.backend, args.dataset, args.max_ind_range,
args.data_sub_sample_rate, args.use_gpu)
# dataset to use
wanted_dataset, pre_proc, post_proc, kwargs = SUPPORTED_DATASETS[
args.dataset]
# --count-samples can be used to limit the number of samples used for testing
ds = wanted_dataset(
data_path=args.dataset_path,
name=args.dataset,
pre_process=pre_proc, # currently an identity function
use_cache=args.cache, # currently not used
count=args.count_samples,
samples_to_aggregate_fix=args.samples_to_aggregate_fix,
samples_to_aggregate_min=args.samples_to_aggregate_min,
samples_to_aggregate_max=args.samples_to_aggregate_max,
samples_to_aggregate_quantile_file=args.
samples_to_aggregate_quantile_file,
samples_to_aggregate_trace_file=args.samples_to_aggregate_trace_file,
test_num_workers=args.test_num_workers,
max_ind_range=args.max_ind_range,
sub_sample_rate=args.data_sub_sample_rate,
mlperf_bin_loader=args.mlperf_bin_loader,
**kwargs)
# load model to backend
model = backend.load(args.model_path,
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
ds.load_query_samples([0])
for _ in range(5):
batch_dense_X, batch_lS_o, batch_lS_i, _, _ = ds.get_samples([0])
_ = backend.predict(batch_dense_X, batch_lS_o, batch_lS_i)
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, 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.duration:
settings.min_duration_ms = args.duration
settings.max_duration_ms = args.duration
if args.target_qps:
settings.server_target_qps = float(args.target_qps)
settings.offline_expected_qps = float(args.target_qps)
if args.count_queries:
settings.min_query_count = args.count_queries
settings.max_query_count = args.count_queries
if args.samples_per_query_multistream:
settings.multi_stream_samples_per_query = args.samples_per_query_multistream
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, args.samples_per_query_offline),
ds.load_query_samples, ds.unload_query_samples)
log.info("starting {}".format(scenario))
result_dict = {
"good": 0,
"total": 0,
"roc_auc": 0,
"scenario": str(scenario)
}
runner.start_run(result_dict, args.accuracy)
lg.StartTest(sut, qsl, settings)
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 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()
# 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=args.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),
}
#
# make one pass over the dataset to validate accuracy
#
count = args.count if args.count else ds.get_item_count()
if args.accuracy:
#
# accuracy pass
#
log.info("starting accuracy pass on {} items".format(count))
last_timeing = []
runner = RunnerBase(model, ds, args.threads, post_proc=post_proc)
result_dict = {"good": 0, "total": 0, "scenario": "Accuracy"}
runner.start_run(result_dict, True)
start = time.time()
for idx in range(0, count):
ds.load_query_samples([idx])
data, label = ds.get_samples([idx])
start_one = time.time()
runner.enqueue([idx], [idx], data, label)
last_timeing.append(time.time() - start_one)
runner.finish()
# aggregate results
post_proc.finalize(result_dict, ds, output_dir=os.path.dirname(args.output))
add_results(final_results, "Accuracy", result_dict, last_timeing, time.time() - start)
# warmup
ds.load_query_samples([0])
for _ in range(5):
img, _ = ds.get_samples([0])
_ = backend.predict({backend.inputs[0]: img})
ds.unload_query_samples(None)
for scenario in 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)
def issue_query(query_samples):
# called by loadgen to issue queries
idx = [q.index for q in query_samples]
query_id = [q.id for q in query_samples]
data, label = ds.get_samples(idx)
runner.enqueue(query_id, idx, data, label)
def process_latencies(latencies_ns):
# called by loadgen to show us the recorded latencies
global last_timeing
last_timeing = [t / 1e9 for t in latencies_ns]
settings = lg.TestSettings()
settings.enable_spec_overrides = True
settings.scenario = scenario
settings.mode = lg.TestMode.PerformanceOnly
settings.multi_stream_samples_per_query = 8
if args.time:
# override the time we want to run
settings.enable_spec_overrides = True
settings.override_min_duration_ms = args.time * MILLI_SEC
settings.override_max_duration_ms = args.time * MILLI_SEC
if args.qps:
qps = float(args.qps)
settings.server_target_qps = qps
settings.offline_expected_qps = qps
# mlperf rules - min queries
if scenario == lg.TestScenario.SingleStream:
settings.override_min_query_count = args.queries_single
settings.override_max_query_count = args.queries_single
else:
settings.override_min_query_count = args.queries_multi
settings.override_max_query_count = args.queries_multi
sut = lg.ConstructSUT(issue_query, process_latencies)
qsl = lg.ConstructQSL(count, min(count, 1000), ds.load_query_samples, ds.unload_query_samples)
for target_latency in args.max_latency:
log.info("starting {}, latency={}".format(scenario, target_latency))
settings.single_stream_expected_latency_ns = int(target_latency * NANO_SEC)
settings.override_target_latency_ns = int(target_latency * NANO_SEC)
result_dict = {"good": 0, "total": 0, "scenario": str(scenario)}
runner.start_run(result_dict, False)
lg.StartTest(sut, qsl, settings)
add_results(final_results, "{}-{}".format(scenario, target_latency),
result_dict, last_timeing, time.time() - ds.last_loaded)
runner.finish()
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
#
# write final results
#
if args.output:
with open(args.output, "w") as f:
json.dump(final_results, f, sort_keys=True, indent=4)
.format(self.count, qitem.sample_id[0]))
self.count += 1
return self.count
if __name__ == "__main__":
runner = DummyRunner()
runner.start_worker()
settings = mlperf_loadgen.TestSettings()
settings.scenario = mlperf_loadgen.TestScenario.SingleStream
settings.mode = mlperf_loadgen.TestMode.PerformanceOnly
# Specify exactly how many queries need to be made
settings.min_query_count = 3003
settings.max_query_count = 3003
total_queries = 256 # Maximum sample ID + 1
perf_queries = 8 # TBD: Doesn't seem to have an effect
sut = mlperf_loadgen.ConstructSUT(runner.enqueue, flush_queries,
process_latencies)
qsl = mlperf_loadgen.ConstructQSL(total_queries, perf_queries,
runner.load_samples_to_ram,
runner.unload_samples_from_ram)
mlperf_loadgen.StartTest(sut, qsl, settings)
mlperf_loadgen.DestroyQSL(qsl)
mlperf_loadgen.DestroySUT(sut)
def main():
args = get_args()
print(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(
)
# dataset to use
wanted_dataset, preprocessor, postprocessor, 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=preprocessor,
use_cache=args.cache,
count=args.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),
}
#
# make one pass over the dataset to validate accuracy
#
count = args.count if args.count else ds.get_item_count()
runner = Runner(model, ds, args.threads, post_process=postprocessor)
runner.start_pool()
# warmup
log.info("warmup ...")
ds.load_query_samples([0])
for _ in range(100):
img, _ = ds.get_samples([0])
_ = backend.predict({backend.inputs[0]: img})
def issue_query(query_samples):
idx = [q.index for q in query_samples]
query_id = [q.id for q in query_samples]
data, label = ds.get_samples(idx)
runner.enqueue(query_id, data, label)
sut = lg.ConstructSUT(issue_query)
qsl = lg.ConstructQSL(count, args.time, ds.load_query_samples,
ds.unload_query_samples)
scenarios = [
# lg.TestScenario.SingleStream,
lg.TestScenario.MultiStream,
# lg.TestScenario.Cloud,
# lg.TestScenario.Offline,
]
for scenario in scenarios:
for target_latency in args.max_latency:
log.info("starting {}, latency={}".format(scenario,
target_latency))
settings = lg.TestSettings()
settings.scenario = scenario
settings.mode = lg.TestMode.SubmissionRun
settings.samples_per_query = 4 # FIXME: we don't want to know about this
settings.target_qps = 1000 # FIXME: we don't want to know about this
settings.target_latency_ns = int(target_latency * 1000000000)
result_list = []
result_dict = {"good": 0, "total": 0}
runner.start_run(result_list, result_dict)
start = time.time()
lg.StartTest(sut, qsl, settings)
add_results(final_results, "{}-{}".format(scenario,
target_latency),
result_dict, result_list,
time.time() - start)
runner.finish()
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
#
# write final results
#
if args.output:
with open(args.output, "w") as f:
json.dump(final_results, f, sort_keys=True, indent=4)
def main():
global num_ins
global num_phy_cpus
global in_queue_cnt
global out_queue_cnt
args = get_args()
log.info(args)
num_ins = args.num_instance
num_phy_cpus = args.num_phy_cpus
log.info('Run with {} instance on {} cpus'.format(num_ins, num_phy_cpus))
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)
image_format = 'NCHW'
dataset = "imagenet"
wanted_dataset, pre_proc, post_proc, kwargs = SUPPORTED_DATASETS[dataset]
ds = wanted_dataset(data_path=args.dataset_path,
image_list=args.dataset_list,
name=dataset,
image_format=image_format,
pre_process=pre_proc,
use_cache=args.cache,
cache_dir=args.cache_dir,
count=args.count,
use_int8=args.use_int8_dataset,
num_workers=num_phy_cpus,
**kwargs)
# Establish communication queues
log.info('Start comsumer queue and response thread')
lock = multiprocessing.Lock()
init_counter = multiprocessing.Value("i", 0)
in_queue = multiprocessing.JoinableQueue()
out_queue = multiprocessing.Queue()
ds_queue = multiprocessing.Queue()
# Start consumers
consumers = [Consumer(in_queue, out_queue, ds_queue, lock, init_counter, i, args)
for i in range(num_ins)]
for c in consumers:
c.start()
# Wait until all sub-processors are ready
block_until(init_counter, num_ins, 2)
# Start response thread
response_worker = threading.Thread(
target=response_loadgen, args=(out_queue,))
response_worker.daemon = True
response_worker.start()
scenario = SCENARIO_MAP[args.scenario]
runner = QueueRunner(in_queue, args.batch_size)
def issue_queries(response_ids, query_sample_indexes):
runner.put(response_ids, query_sample_indexes)
def flush_queries():
pass
def process_latencies(latencies_ns):
log.info("Average latency: {}".format(np.mean(latencies_ns)))
log.info("Median latency: {}".format(np.percentile(latencies_ns, 50)))
log.info("90 percentile latency: {}".format(np.percentile(latencies_ns, 90)))
def load_query_samples(sample_list):
for _ in range(num_ins):
ds_queue.put(sample_list)
block_until(init_counter, 2 * num_ins, 2)
def unload_query_samples(sample_list):
pass
settings = lg.TestSettings()
settings.FromConfig(mlperf_conf, "resnet50", args.scenario)
settings.FromConfig(user_conf, "resnet50", 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.qps:
qps = float(args.qps)
settings.server_target_qps = qps
settings.offline_expected_qps = qps
count = ds.get_item_count()
perf_count = 1024
if args.accuracy:
perf_count = count
sut = lg.ConstructFastSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(count, perf_count, load_query_samples, unload_query_samples)
log.info("starting {}".format(scenario))
lg.StartTest(sut, qsl, 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('DONE')
for c in consumers:
c.join()
out_queue.put('DONE')
if args.accuracy:
output_file = 'accuracy.txt'
if args.output_file:
output_file = args.output_file
cmd = "python tools/accuracy-imagenet.py " \
"--mlperf-accuracy-file=mlperf_log_accuracy.json " \
"--imagenet-val-file=val_map.txt --output-file={}".format(output_file)
cmd = cmd.split(' ')
subprocess.check_call(cmd)
lg.DestroyQSL(qsl)
lg.DestroyFastSUT(sut)
log.info('Test 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(
)
# 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=args.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),
}
#
# make one pass over the dataset to validate accuracy
#
count = args.count if args.count else ds.get_item_count()
runner = Runner(model, ds, args.threads, post_proc=post_proc)
#
# warmup
#
log.info("warmup ...")
ds.load_query_samples([0])
for _ in range(5):
img, _ = ds.get_samples([0])
_ = backend.predict({backend.inputs[0]: img})
ds.unload_query_samples(None)
if args.accuracy:
#
# accuracy pass
#
log.info("starting accuracy pass on {} items".format(count))
runner.start_pool(nolg=True)
result_dict = {
"good": 0,
"total": 0,
"scenario": "Accuracy",
"timing": []
}
runner.start_run(result_dict, True)
start = time.time()
for idx in range(0, count):
ds.load_query_samples([idx])
data, label = ds.get_samples([idx])
runner.enqueue([idx], [idx], data, label)
runner.finish()
# aggregate results
post_proc.finalize(result_dict,
ds,
output_dir=os.path.dirname(args.output))
last_timeing = result_dict["timing"]
del result_dict["timing"]
add_results(final_results, "Accuracy", result_dict, last_timeing,
time.time() - start)
#
# run the benchmark with timing
#
runner.start_pool()
def issue_query(query_samples):
idx = [q.index for q in query_samples]
query_id = [q.id for q in query_samples]
data, label = ds.get_samples(idx)
runner.enqueue(query_id, idx, data, label)
def process_latencies(latencies_ns):
global last_timeing
last_timeing = [t / 1e9 for t in latencies_ns]
sut = lg.ConstructSUT(issue_query, process_latencies)
qsl = lg.ConstructQSL(count, min(count, 1000), ds.load_query_samples,
ds.unload_query_samples)
for scenario in args.scenario:
for target_latency in args.max_latency:
log.info("starting {}, latency={}".format(scenario,
target_latency))
settings = lg.TestSettings()
log.info(scenario)
if str(scenario) == 'TestMode.AccuracyOnly':
settings.mode = scenario
else:
settings.scenario = scenario
if args.qps:
settings.enable_spec_overrides = True
qps = float(args.qps)
settings.server_target_qps = qps
settings.offline_expected_qps = qps
if args.time:
settings.enable_spec_overrides = True
settings.override_min_duration_ms = args.time * MILLI_SEC
settings.override_max_duration_ms = args.time * MILLI_SEC
qps = args.qps or 100
settings.override_min_query_count = qps * args.time
settings.override_max_query_count = qps * args.time
if args.time or args.qps and str(
scenario) != 'TestMode.AccuracyOnly':
settings.mode = lg.TestMode.PerformanceOnly
# FIXME: add SubmissionRun once available
settings.enable_spec_overrides = True
settings.single_stream_expected_latency_ns = int(target_latency *
NANO_SEC)
settings.override_target_latency_ns = int(target_latency *
NANO_SEC)
result_dict = {"good": 0, "total": 0, "scenario": str(scenario)}
runner.start_run(result_dict, False)
lg.StartTest(sut, qsl, settings)
add_results(final_results, "{}-{}".format(scenario,
target_latency),
result_dict, last_timeing,
time.time() - ds.last_loaded)
#
# write final results
#
if args.output:
with open(args.output, "w") as f:
json.dump(final_results, f, sort_keys=True, indent=4)
runner.finish()
lg.DestroyQSL(qsl)
lg.DestroySUT(sut)
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 = args.count
if not count:
if not args.accuracy:
count = 200
# 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),
}
#
# make one pass over the dataset to validate accuracy
#
count = ds.get_item_count()
# warmup
ds.load_query_samples([0])
for _ in range(5):
img, _ = ds.get_samples([0])
_ = backend.predict({backend.inputs[0]: img})
ds.unload_query_samples(None)
for scenario in 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.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 scenario == lg.TestScenario.SingleStream:
settings.min_query_count = args.queries_single
settings.max_query_count = args.queries_single
elif scenario == lg.TestScenario.MultiStream:
settings.min_query_count = args.queries_multi
settings.max_query_count = args.queries_multi
settings.multi_stream_samples_per_query = 4
elif scenario == lg.TestScenario.Server:
max_latency = args.max_latency
elif scenario == lg.TestScenario.Offline:
settings.min_query_count = args.queries_offline
settings.max_query_count = args.queries_offline
sut = lg.ConstructSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(count, min(count, 1000), ds.load_query_samples, ds.unload_query_samples)
if scenario == lg.TestScenario.Server:
for target_latency in max_latency:
log.info("starting {}, latency={}".format(scenario, target_latency))
settings.server_target_latency_ns = int(target_latency * NANO_SEC)
result_dict = {"good": 0, "total": 0, "scenario": str(scenario)}
runner.start_run(result_dict, args.accuracy)
lg.StartTest(sut, qsl, settings)
if not last_timeing:
last_timeing = runner.result_timing
if args.accuracy:
post_proc.finalize(result_dict, ds, output_dir=os.path.dirname(args.output))
add_results(final_results, "{}-{}".format(scenario, target_latency),
result_dict, last_timeing, time.time() - ds.last_loaded, args.accuracy)
else:
log.info("starting {}".format(scenario))
result_dict = {"good": 0, "total": 0, "scenario": str(scenario)}
runner.start_run(result_dict, args.accuracy)
lg.StartTest(sut, qsl, settings)
if not last_timeing:
last_timeing = runner.result_timing
if args.accuracy:
post_proc.finalize(result_dict, ds, output_dir=os.path.dirname(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(args.output, "w") as f:
json.dump(final_results, f, sort_keys=True, indent=4)
def __del__(self):
lg.DestroyQSL(self.qsl)
print("Finished destroying QSL.")
def main():
global last_timeing
args = get_args()
log.info(args)
backend = BackendTensorRT()
ds = Imagenet(
data_path=args.dataset_path,
use_cache=args.cache,
batch_size=args.batch_size,
image_size=args.image_size,
calib_file='cal_image_list_option_%d.txt' % args.calib_file)
model = backend.load(args, ds=ds)
final_results = {
"runtime": model.name(),
"version": model.version(),
"time": int(time.time()),
"cmdline": str(args),
}
config = os.path.abspath(args.config)
assert(os.path.exists(config)), "%s not existed!" % config
user_config = os.path.abspath(args.user_config)
assert(os.path.exists(user_config)), "%s not existed!" % user_config
base_path = os.path.dirname(os.path.realpath(__file__))
if args.output:
output_dir = os.path.abspath(args.output)
os.makedirs(output_dir, exist_ok=True)
os.chdir(output_dir)
post_proc = PostProcessCommon(offset=0)
runner = QueueRunner(
model, ds, args.threads, post_proc=post_proc, batch_size=args.batch_size)
def issue_queries(ids, indices):
runner.enqueue(ids, indices)
def flush_queries():
pass
def process_latencies(latencies_ns):
global last_timeing
last_timeing = [t / NANO_SEC for t in latencies_ns]
settings = lg.TestSettings()
model_name = 'OFAnet-AutoSinian'
settings.FromConfig(config, model_name, args.scenario)
settings.FromConfig(user_config, model_name, args.scenario)
if args.audit_test:
audit_config_path = base_path + '/audit%s.config' % args.audit_test
settings.FromConfig(audit_config_path, model_name, args.scenario)
scenario = SCENARIO_MAP[args.scenario]
settings.scenario = scenario
settings.mode = lg.TestMode.PerformanceOnly
if args.accuracy:
settings.mode = lg.TestMode.AccuracyOnly
sut = lg.ConstructFastSUT(issue_queries, flush_queries, process_latencies)
qsl = lg.ConstructQSL(ds.get_item_count(), args.batch_size,
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)
start = time.time()
lg.StartTest(sut, qsl, settings)
post_proc.finalize(result_dict)
add_results(final_results, "{}".format(scenario), result_dict, last_timeing,
runner.finishTime - ds.last_loaded, args)
runner.finish()
lg.DestroyQSL(qsl)
lg.DestroyFastSUT(sut)
if args.output:
with open("results.json", "w") as f:
json.dump(final_results, f, sort_keys=True, indent=2)
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