def setup_logger(args):
aggregator_dict = OrderedDict([
('loss', 'average'),
('weighted_loss', 'average'),
('tokens', ('average', 'performance')),
('updates', 'performance'),
('gnorm', 'average')
])
os.makedirs(args.save_dir, exist_ok=True)
log_path = os.path.join(args.save_dir, args.stat_file)
if os.path.exists(log_path):
for i in itertools.count():
s_fname = args.stat_file.split('.')
fname = '.'.join(s_fname[:-1]) + f'_{i}.' + s_fname[-1] if len(s_fname) > 1 else args.stat_file + f'.{i}'
log_path = os.path.join(args.save_dir, fname)
if not os.path.exists(log_path):
break
if not args.distributed_world_size > 1 or args.distributed_rank == 0:
dllogger.init(backends=[JSONStreamBackend(verbosity=1, filename=log_path),
AggregatorBackend(verbosity=0, agg_dict=aggregator_dict),
TensorBoardBackend(verbosity=1, log_dir=args.save_dir)])
else:
dllogger.init(backends=[])
for k, v in vars(args).items():
dllogger.log(step='PARAMETER', data={k: v}, verbosity=0)
container_setup_info = get_framework_env_vars()
dllogger.log(step='PARAMETER', data=container_setup_info, verbosity=0)
dllogger.metadata('loss', {'unit': 'nat', 'GOAL': 'MINIMIZE', 'STAGE': 'TRAIN'})
dllogger.metadata('val_loss', {'unit': 'nat', 'GOAL': 'MINIMIZE', 'STAGE': 'VAL'})
dllogger.metadata('speed', {'unit': 'tokens/s', 'format': ':.3f', 'GOAL': 'MAXIMIZE', 'STAGE': 'TRAIN'})
dllogger.metadata('accuracy', {'unit': 'bleu', 'format': ':.2f', 'GOAL': 'MAXIMIZE', 'STAGE': 'VAL'})
def main():
# setup params
arguments = PARSER.parse_args()
params = Namespace(**{**vars(CONFIG), **vars(arguments)})
# setup logging
# noinspection PyArgumentList
logging.basicConfig(
level=logging.DEBUG if params.verbose else logging.INFO,
format='{asctime} {levelname:.1} {name:15} {message}',
style='{'
)
# remove custom tf handler that logs to stderr
logging.getLogger('tensorflow').setLevel(logging.WARNING)
logging.getLogger('tensorflow').handlers.clear()
# setup dllogger
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE, filename=params.log_file),
LoggingBackend(verbosity=dllogger.Verbosity.VERBOSE)
])
dllogger.log(step='PARAMETER', data=vars(params))
# setup dataset
dataset = Dataset(params)
if params.mode in ['train', 'train_and_eval']:
run_training(dataset, params)
if params.mode == 'eval':
run_evaluation(dataset, params)
if params.mode == 'infer':
run_inference(dataset, params)
def setup_training(args):
assert (torch.cuda.is_available())
if args.local_rank == -1:
device = torch.device("cuda")
args.n_gpu = torch.cuda.device_count()
args.allreduce_post_accumulation = False
args.allreduce_post_accumulation_fp16 = False
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='smddp',
init_method='env://')
args.n_gpu = 1
if args.gradient_accumulation_steps == 1:
args.allreduce_post_accumulation = False
args.allreduce_post_accumulation_fp16 = False
if is_main_process():
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.json_summary),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE,
step_format=format_step)
])
else:
dllogger.init(backends=[])
print(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, args.n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.train_batch_size % args.gradient_accumulation_steps != 0:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, batch size {} should be divisible"
.format(args.gradient_accumulation_steps, args.train_batch_size))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
if not args.do_train:
raise ValueError(" `do_train` must be True.")
if not args.resume_from_checkpoint and os.path.exists(
args.output_dir) and (os.listdir(args.output_dir) and any(
[i.startswith('ckpt') for i in os.listdir(args.output_dir)])):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if (not args.resume_from_checkpoint
or not os.path.exists(args.output_dir)) and is_main_process():
os.makedirs(args.output_dir, exist_ok=True)
return device, args
def init(log_fpath, log_dir, enabled=True, tb_subsets=[], **tb_kw):
if enabled:
backends = [JSONStreamBackend(Verbosity.DEFAULT,
unique_log_fpath(log_fpath)),
StdOutBackend(Verbosity.VERBOSE,
step_format=stdout_step_format,
metric_format=stdout_metric_format)]
else:
backends = []
dllogger.init(backends=backends)
dllogger.metadata("train_lrate", {"name": "lrate", "format": ":>3.2e"})
for id_, pref in [('train', ''), ('train_avg', 'avg train '),
('val', ' avg val '), ('val_ema', ' EMA val ')]:
dllogger.metadata(f"{id_}_loss",
{"name": f"{pref}loss", "format": ":>5.2f"})
dllogger.metadata(f"{id_}_mel_loss",
{"name": f"{pref}mel loss", "format": ":>5.2f"})
dllogger.metadata(f"{id_}_frames/s",
{"name": None, "unit": "frames/s", "format": ":>10.2f"})
dllogger.metadata(f"{id_}_took",
{"name": "took", "unit": "s", "format": ":>3.2f"})
global tb_loggers
tb_loggers = {s: TBLogger(enabled, log_dir, name=s, **tb_kw)
for s in tb_subsets}
def init_dllogger(log_fpath=None, dummy=False):
if dummy:
DLLogger.init(backends=[])
return
DLLogger.init(backends=[
JSONStreamBackend(Verbosity.DEFAULT, log_fpath),
StdOutBackend(Verbosity.VERBOSE, step_format=stdout_step_format,
metric_format=stdout_metric_format)
]
)
DLLogger.metadata("train_loss", {"name": "loss", "format": ":>5.2f"})
DLLogger.metadata("train_mel_loss", {"name": "mel loss", "format": ":>5.2f"})
DLLogger.metadata("avg_train_loss", {"name": "avg train loss", "format": ":>5.2f"})
DLLogger.metadata("avg_train_mel_loss", {"name": "avg train mel loss", "format": ":>5.2f"})
DLLogger.metadata("val_loss", {"name": " avg val loss", "format": ":>5.2f"})
DLLogger.metadata("val_mel_loss", {"name": " avg val mel loss", "format": ":>5.2f"})
DLLogger.metadata(
"val_ema_loss",
{"name": " EMA val loss", "format": ":>5.2f"})
DLLogger.metadata(
"val_ema_mel_loss",
{"name": " EMA val mel loss", "format": ":>5.2f"})
DLLogger.metadata(
"train_frames/s", {"name": None, "unit": "frames/s", "format": ":>10.2f"})
DLLogger.metadata(
"avg_train_frames/s", {"name": None, "unit": "frames/s", "format": ":>10.2f"})
DLLogger.metadata(
"val_frames/s", {"name": None, "unit": "frames/s", "format": ":>10.2f"})
DLLogger.metadata(
"val_ema_frames/s", {"name": None, "unit": "frames/s", "format": ":>10.2f"})
DLLogger.metadata(
"took", {"name": "took", "unit": "s", "format": ":>3.2f"})
DLLogger.metadata("lrate_change", {"name": "lrate"})
def __init__(self, log_file, global_batch_size, warmup_steps: int = 0, profile: bool = False): logger.init(backends=[JSONStreamBackend(Verbosity.VERBOSE, log_file), StdOutBackend(Verbosity.VERBOSE)]) self.warmup_steps = warmup_steps self.global_batch_size = global_batch_size self.step = 0 self.profile = profile self.timestamps = []
def main():
"""
Launches inference benchmark.
Inference is executed on a single GPU.
"""
parser = argparse.ArgumentParser(
description='PyTorch FastPitch Inference Benchmark')
parser = parse_args(parser)
args, _ = parser.parse_known_args()
log_file = args.log_file
DLLogger.init(backends=[
JSONStreamBackend(Verbosity.DEFAULT, args.log_file),
StdOutBackend(Verbosity.VERBOSE)
])
for k, v in vars(args).items():
DLLogger.log(step="PARAMETER", data={k: v})
DLLogger.log(step="PARAMETER", data={'model_name': 'FastPitch_PyT'})
model = load_and_setup_model('FastPitch',
parser,
None,
args.amp_run,
'cuda',
unk_args=[],
forward_is_infer=True,
ema=False,
jitable=True)
# FIXME Temporarily disabled due to nn.LayerNorm fp16 casting bug in pytorch:20.02-py3 and 20.03
# model = torch.jit.script(model)
warmup_iters = 3
iters = 1
gen_measures = MeasureTime()
all_frames = 0
for i in range(-warmup_iters, iters):
text_padded = torch.randint(low=0,
high=148,
size=(args.batch_size, 128),
dtype=torch.long).to('cuda')
input_lengths = torch.IntTensor([text_padded.size(1)] *
args.batch_size).to('cuda')
durs = torch.ones_like(text_padded).mul_(4).to('cuda')
with torch.no_grad(), gen_measures:
mels, *_ = model(text_padded, input_lengths, dur_tgt=durs)
num_frames = mels.size(0) * mels.size(2)
if i >= 0:
all_frames += num_frames
DLLogger.log(step=(i, ), data={"latency": gen_measures[-1]})
DLLogger.log(step=(i, ),
data={"frames/s": num_frames / gen_measures[-1]})
measures = gen_measures[warmup_iters:]
DLLogger.log(step=(), data={'avg latency': np.mean(measures)})
DLLogger.log(step=(), data={'avg frames/s': all_frames / np.sum(measures)})
DLLogger.flush()
def get_logger(params):
backends = []
if hvd.rank() == 0:
backends += [StdOutBackend(Verbosity.VERBOSE)]
if params.log_dir:
backends += [JSONStreamBackend(Verbosity.VERBOSE, params.log_dir)]
logger.init(backends=backends)
return logger
def _initialize_dllogger(self, log_dir, filename, append):
backends = [
JSONStreamBackend(Verbosity.VERBOSE,
os.path.join(log_dir, filename),
append=append),
StdOutBackend(Verbosity.VERBOSE),
]
logger.init(backends=backends)
def __init__(self, print_interval, backends, verbose=False):
self.epoch = -1
self.iteration = -1
self.val_iteration = -1
self.metrics = OrderedDict()
self.backends = backends
self.print_interval = print_interval
self.verbose = verbose
dllogger.init(backends)
def __init__(self, log_dir, global_batch_size, mode, warmup, dim, profile):
logger.init(backends=[JSONStreamBackend(Verbosity.VERBOSE, log_dir), StdOutBackend(Verbosity.VERBOSE)])
self.warmup_steps = warmup
self.global_batch_size = global_batch_size
self.step = 0
self.dim = dim
self.mode = mode
self.profile = profile
self.timestamps = []
def main():
args = parse_args()
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.log_path),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)
])
dllogger.log(data=vars(args), step='PARAMETER')
model = NeuMF(nb_users=args.n_users,
nb_items=args.n_items,
mf_dim=args.factors,
mlp_layer_sizes=args.layers,
dropout=args.dropout)
model = model.cuda()
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
model.load_state_dict(state_dict)
if args.opt_level == "O2":
model = amp.initialize(model,
opt_level=args.opt_level,
keep_batchnorm_fp32=False,
loss_scale='dynamic')
model.eval()
users = torch.cuda.LongTensor(args.batch_size).random_(0, args.n_users)
items = torch.cuda.LongTensor(args.batch_size).random_(0, args.n_items)
latencies = []
for _ in range(args.num_batches):
torch.cuda.synchronize()
start = time.time()
predictions = model(users, items, sigmoid=True)
torch.cuda.synchronize()
latencies.append(time.time() - start)
dllogger.log(data={
'batch_size':
args.batch_size,
'best_inference_throughput':
args.batch_size / min(latencies),
'best_inference_latency':
min(latencies),
'mean_inference_throughput':
args.batch_size / np.mean(latencies),
'mean_inference_latency':
np.mean(latencies),
'inference_latencies':
latencies
},
step=tuple())
dllogger.flush()
return
def setup_logger(args):
aggregator_dict = OrderedDict([('loss', 'average'),
('weighted_loss', 'average'),
('tokens', ('average', 'performance')),
('updates', 'performance'),
('gnorm', 'average')])
os.makedirs(args.save_dir, exist_ok=True)
log_path = os.path.join(args.save_dir, args.stat_file)
os.makedirs(args.save_dir, exist_ok=True)
if not args.distributed_world_size > 1 or args.distributed_rank == 0:
dllogger.init(backends=[
JSONStreamBackend(verbosity=1, filename=log_path),
AggregatorBackend(verbosity=0, agg_dict=aggregator_dict),
TensorBoardBackend(verbosity=1, log_dir=args.save_dir)
])
else:
dllogger.init(backends=[])
for k, v in vars(args).items():
dllogger.log(step='PARAMETER', data={k: v}, verbosity=0)
container_setup_info = {
'NVIDIA_PYTORCH_VERSION': os.environ.get('NVIDIA_PYTORCH_VERSION'),
'PYTORCH_VERSION': os.environ.get('PYTORCH_VERSION'),
'CUBLAS_VERSION': os.environ.get('CUBLAS_VERSION'),
'NCCL_VERSION': os.environ.get('NCCL_VERSION'),
'CUDA_DRIVER_VERSION': os.environ.get('CUDA_DRIVER_VERSION'),
'CUDNN_VERSION': os.environ.get('CUDNN_VERSION'),
'CUDA_VERSION': os.environ.get('CUDA_VERSION'),
'NVIDIA_PIPELINE_ID': os.environ.get('NVIDIA_PIPELINE_ID'),
'NVIDIA_BUILD_ID': os.environ.get('NVIDIA_BUILD_ID'),
'NVIDIA_TF32_OVERRIDE': os.environ.get('NVIDIA_TF32_OVERRIDE'),
}
dllogger.log(step='PARAMETER', data=container_setup_info, verbosity=0)
dllogger.metadata('loss', {
'unit': 'nat',
'GOAL': 'MINIMIZE',
'STAGE': 'TRAIN'
})
dllogger.metadata('val_loss', {
'unit': 'nat',
'GOAL': 'MINIMIZE',
'STAGE': 'VAL'
})
dllogger.metadata('speed', {
'unit': 'tokens/s',
'format': ':.3f',
'GOAL': 'MAXIMIZE',
'STAGE': 'TRAIN'
})
dllogger.metadata('accuracy', {
'unit': 'bleu',
'format': ':.2f',
'GOAL': 'MAXIMIZE',
'STAGE': 'VAL'
})
def get_logger(params):
backends = []
if params.worker_id == 0 or params.log_all_workers:
backends += [StdOutBackend(Verbosity.VERBOSE)]
if params.log_dir:
os.makedirs(params.log_dir, exist_ok=True)
log_file = f"{params.log_dir}/log.json"
backends += [JSONStreamBackend(Verbosity.VERBOSE, log_file)]
logger.init(backends=backends)
return logger
def get_logger(params):
backends = []
worker_id = hvd_rank() if horovod_enabled() else 0
if worker_id == 0:
backends += [StdOutBackend(Verbosity.VERBOSE)]
if params.log_dir:
os.makedirs(params.log_dir, exist_ok=True)
log_file = f"{params.log_dir}/log.json"
backends += [JSONStreamBackend(Verbosity.VERBOSE, log_file)]
logger.init(backends=backends)
return logger
def setup_logger(args):
os.makedirs(args.results, exist_ok=True)
log_path = os.path.join(args.results, args.log_file)
if os.path.exists(log_path):
for i in itertools.count():
s_fname = args.log_file.split('.')
fname = '.'.join(s_fname[:-1]) + f'_{i}.' + s_fname[-1] if len(s_fname) > 1 else args.stat_file + f'.{i}'
log_path = os.path.join(args.results, fname)
if not os.path.exists(log_path):
break
def metric_format(metric, metadata, value):
return "{}: {}".format(metric, f'{value:.5f}' if isinstance(value, float) else value)
def step_format(step):
if step == ():
return "Finished |"
elif isinstance(step, int):
return "Step {0: <5} |".format(step)
return "Step {} |".format(step)
if not dist.is_initialized() or not args.distributed_world_size > 1 or args.distributed_rank == 0:
dllogger.init(backends=[JSONStreamBackend(verbosity=1, filename=log_path),
TensorBoardBackend(verbosity=1, log_dir=args.results),
StdOutBackend(verbosity=2,
step_format=step_format,
prefix_format=lambda x: "")#,
#metric_format=metric_format)
])
else:
dllogger.init(backends=[])
dllogger.log(step='PARAMETER', data=vars(args), verbosity=0)
container_setup_info = {**get_framework_env_vars(), **get_system_info()}
dllogger.log(step='ENVIRONMENT', data=container_setup_info, verbosity=0)
dllogger.metadata('loss', {'GOAL': 'MINIMIZE', 'STAGE': 'TRAIN', 'format': ':5f'})
dllogger.metadata('P10', {'GOAL': 'MINIMIZE', 'STAGE': 'TRAIN', 'format': ':5f'})
dllogger.metadata('P50', {'GOAL': 'MINIMIZE', 'STAGE': 'TRAIN', 'format': ':5f'})
dllogger.metadata('P90', {'GOAL': 'MINIMIZE', 'STAGE': 'TRAIN', 'format': ':5f'})
dllogger.metadata('items/s', {'GOAL': 'MAXIMIZE', 'STAGE': 'TRAIN', 'format': ':1f'})
dllogger.metadata('val_loss', {'GOAL': 'MINIMIZE', 'STAGE': 'VAL', 'format':':5f'})
dllogger.metadata('val_P10', {'GOAL': 'MINIMIZE', 'STAGE': 'VAL', 'format': ':5f'})
dllogger.metadata('val_P50', {'GOAL': 'MINIMIZE', 'STAGE': 'VAL', 'format': ':5f'})
dllogger.metadata('val_P90', {'GOAL': 'MINIMIZE', 'STAGE': 'VAL', 'format': ':5f'})
dllogger.metadata('val_items/s', {'GOAL': 'MAXIMIZE', 'STAGE': 'VAL', 'format': ':1f'})
dllogger.metadata('test_P10', {'GOAL': 'MINIMIZE', 'STAGE': 'TEST', 'format': ':5f'})
dllogger.metadata('test_P50', {'GOAL': 'MINIMIZE', 'STAGE': 'TEST', 'format': ':5f'})
dllogger.metadata('test_P90', {'GOAL': 'MINIMIZE', 'STAGE': 'TEST', 'format': ':5f'})
dllogger.metadata('throughput', {'GOAL': 'MAXIMIZE', 'STAGE': 'TEST', 'format': ':1f'})
dllogger.metadata('latency_p90', {'GOAL': 'MIMIMIZE', 'STAGE': 'TEST', 'format': ':5f'})
dllogger.metadata('latency_p95', {'GOAL': 'MIMIMIZE', 'STAGE': 'TEST', 'format': ':5f'})
dllogger.metadata('latency_p99', {'GOAL': 'MIMIMIZE', 'STAGE': 'TEST', 'format': ':5f'})
def setup_dllogger(enabled=True, filename=os.devnull, rank=0):
if enabled and rank == 0:
backends = [
dllogger.JSONStreamBackend(
dllogger.Verbosity.VERBOSE,
filename,
),
]
dllogger.init(backends)
else:
dllogger.init([])
def init_logging(log_path):
json_backend = dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=log_path)
stdout_backend = dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)
stdout_backend._metadata['best_auc'].update({'format': '0:.5f'})
stdout_backend._metadata['best_epoch'].update({'format': '0:.2f'})
stdout_backend._metadata['average_train_throughput'].update({'format': ':.2e'})
stdout_backend._metadata['average_test_throughput'].update({'format': ':.2e'})
dllogger.init(backends=[json_backend, stdout_backend])
def get_dllogger(params):
backends = []
if is_main_process():
backends += [StdOutBackend(Verbosity.VERBOSE)]
if params.log_dir:
backends += [
JSONStreamBackend(Verbosity.VERBOSE,
os.path.join(params.log_dir, "log.json"))
]
logger.init(backends=backends)
return logger
def init_log(args):
enabled = not dist.is_initialized() or dist.get_rank() == 0
if enabled:
fpath = args.log_file or os.path.join(args.output_dir, 'nvlog.json')
backends = [
JSONStreamBackend(Verbosity.DEFAULT, unique_log_fpath(fpath)),
StdOutBackend(Verbosity.VERBOSE,
step_format=stdout_step_format,
metric_format=stdout_metric_format)
]
else:
backends = []
dllogger.init(backends=backends)
dllogger.metadata("train_lrate", {"name": "lrate", "format": ":>3.2e"})
for id_, pref in [('train', ''), ('train_avg', 'avg train '),
('dev_ema', ' dev ema ')]:
dllogger.metadata(f"{id_}_loss", {
"name": f"{pref}loss",
"format": ":>7.2f"
})
dllogger.metadata(f"{id_}_wer", {
"name": f"{pref}wer",
"format": ":>6.2f"
})
dllogger.metadata(f"{id_}_pplx", {
"name": f"{pref}pplx",
"format": ":>6.2f"
})
dllogger.metadata(f"{id_}_throughput", {
"name": f"{pref}utts/s",
"format": ":>5.0f"
})
dllogger.metadata(f"{id_}_took", {
"name": "took",
"unit": "s",
"format": ":>5.2f"
})
tb_subsets = ['train', 'dev_ema']
global tb_loggers
tb_loggers = {
s: TBLogger(enabled, args.output_dir, name=s)
for s in tb_subsets
}
log_parameters(vars(args), tb_subset='train')
def setup_dllogger(rank, enabled=True, filename='log.json'):
if enabled and rank == 0:
backends = [
StdOutBackend(Verbosity.DEFAULT),
JSONStreamBackend(
Verbosity.VERBOSE,
filename,
),
]
DLLogger.init(backends)
else:
DLLogger.init([])
def get_logger(params):
""" Get logger object
:param params: Dict with additional parameters
:return: logger
"""
backends = []
if hvd.rank() == 0:
backends += [StdOutBackend(Verbosity.VERBOSE)]
if params.log_dir:
backends += [JSONStreamBackend(Verbosity.VERBOSE, params.log_dir)]
logger.init(backends=backends)
return logger
def setup_dllogger(enabled=True, filename=os.devnull):
rank = utils.distributed.get_rank()
if enabled and rank == 0:
backends = [
dllogger.JSONStreamBackend(
dllogger.Verbosity.VERBOSE,
filename,
),
]
dllogger.init(backends)
else:
dllogger.init([])
def setup_logging(args):
logging.basicConfig(level=logging.DEBUG,
format='{asctime}:{levelname}: {message}',
style='{')
if hvd.rank() == 0:
dllogger.init(backends=[
dllogger.StdOutBackend(dllogger.Verbosity.DEFAULT,
step_format=format_step),
dllogger.JSONStreamBackend(
dllogger.Verbosity.VERBOSE,
os.path.join(args.workspace, args.dllogger_log)),
])
else:
dllogger.init([])
def setup_training(args):
assert (torch.cuda.is_available())
global ort_supplement
import ort_supplement.ort_supplement as ort_supplement
device = ort_supplement.setup_onnxruntime_with_mpi(args)
if is_main_process(args):
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.json_summary),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE,
step_format=format_step)
])
else:
dllogger.init(backends=[])
print(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, args.n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.train_batch_size % args.gradient_accumulation_steps != 0:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, batch size {} should be divisible"
.format(args.gradient_accumulation_steps, args.train_batch_size))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
if not args.do_train:
raise ValueError(" `do_train` must be True.")
if not args.resume_from_checkpoint and os.path.exists(
args.output_dir) and (os.listdir(args.output_dir) and any(
[i.startswith('ckpt') for i in os.listdir(args.output_dir)])):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if (not args.resume_from_checkpoint
or not os.path.exists(args.output_dir)) and is_main_process(args):
os.makedirs(args.output_dir, exist_ok=True)
return device, args
def main():
args = parse_args()
dllogger.init(backends=[dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.log_path),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)])
dllogger.log(data=vars(args), step='PARAMETER')
model = NeuMF(nb_users=args.n_users, nb_items=args.n_items, mf_dim=args.factors,
mlp_layer_sizes=args.layers, dropout=args.dropout)
model = model.cuda()
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
model.load_state_dict(state_dict)
if args.fp16:
model.half()
model.eval()
batch_sizes = args.batch_sizes.split(',')
batch_sizes = [int(s) for s in batch_sizes]
result_data = {}
for batch_size in batch_sizes:
print('benchmarking batch size: ', batch_size)
users = torch.cuda.LongTensor(batch_size).random_(0, args.n_users)
items = torch.cuda.LongTensor(batch_size).random_(0, args.n_items)
latencies = []
for _ in range(args.num_batches):
torch.cuda.synchronize()
start = time.time()
_ = model(users, items, sigmoid=True)
torch.cuda.synchronize()
latencies.append(time.time() - start)
result_data[f'batch_{batch_size}_mean_throughput'] = batch_size / np.mean(latencies)
result_data[f'batch_{batch_size}_mean_latency'] = np.mean(latencies)
result_data[f'batch_{batch_size}_p90_latency'] = np.percentile(latencies, 0.90)
result_data[f'batch_{batch_size}_p95_latency'] = np.percentile(latencies, 0.95)
result_data[f'batch_{batch_size}_p99_latency'] = np.percentile(latencies, 0.99)
dllogger.log(data=result_data, step=tuple())
dllogger.flush()
return
def __init__(self, name, json_output=None, print_freq=20):
self.name = name
self.train_loss_logger = IterationAverageMeter("Training loss")
self.train_epoch_time_logger = EpochMeter("Training 1 epoch time")
self.val_acc_logger = EpochMeter("Validation accuracy")
self.print_freq = print_freq
backends = [DLLogger.StdOutBackend(DLLogger.Verbosity.DEFAULT)]
if json_output:
backends.append(
DLLogger.JSONStreamBackend(DLLogger.Verbosity.VERBOSE,
json_output))
DLLogger.init(backends)
self.epoch = 0
self.train_iter = 0
self.summary = {}
def setup_logger(args):
if not args.no_dllogger:
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=1, filename=args.stat_file)
])
for k, v in vars(args).items():
dllogger.log(step='PARAMETER', data={k: v}, verbosity=0)
container_setup_info = log_helper.get_framework_env_vars()
dllogger.log(step='PARAMETER', data=container_setup_info, verbosity=0)
dllogger.metadata(
'throughput', {
'unit': 'tokens/s',
'format': ':/3f',
'GOAL': 'MAXIMIZE',
'STAGE': 'INFER'
})
else:
dllogger.init(backends=[])
def init_logger(args, full, logger):
if full:
logger.setLevel(logging.INFO)
log_path = os.path.join(args.results_dir, args.log_filename)
os.makedirs(args.results_dir, exist_ok=True)
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=log_path),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)])
logger.warning('command line arguments: {}'.format(json.dumps(vars(args))))
if not os.path.exists(args.results_dir):
os.mkdir(args.results_dir)
with open('{}/args.json'.format(args.results_dir), 'w') as f:
json.dump(vars(args), f, indent=4)
else:
logger.setLevel(logging.ERROR)
dllogger.init(backends=[])
dllogger.log(data=vars(args), step='PARAMETER')
def setup_logger(args):
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=1, filename=args.stat_file)
])
for k, v in vars(args).items():
dllogger.log(step='PARAMETER', data={k: v}, verbosity=0)
container_setup_info = {
'NVIDIA_PYTORCH_VERSION': os.environ.get('NVIDIA_PYTORCH_VERSION'),
'PYTORCH_VERSION': os.environ.get('PYTORCH_VERSION'),
'CUBLAS_VERSION': os.environ.get('CUBLAS_VERSION'),
'NCCL_VERSION': os.environ.get('NCCL_VERSION'),
'CUDA_DRIVER_VERSION': os.environ.get('CUDA_DRIVER_VERSION'),
'CUDNN_VERSION': os.environ.get('CUDNN_VERSION'),
'CUDA_VERSION': os.environ.get('CUDA_VERSION')
}
dllogger.log(step='PARAMETER', data=container_setup_info, verbosity=0)
dllogger.metadata('throughput', {
'unit': 'tokens/s',
'format': ':/3f',
'GOAL': 'MAXIMIZE',
'STAGE': 'INFER'
})