maxmustermann/DeepLearningExamples
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

[Transformer/PyT] Removing obsolete profiling code

Browse source
This commit is contained in:
Jan Baczek 2021-08-13 16:36:58 +00:00 committed by Krzysztof Kudrynski
parent 793027c9da
commit 233287038c
4 changed files with 3 additions and 35 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
PyTorch/Translation/Transformer/README.md
View file

@ -283,8 +283,7 @@ The following (partial) output is printed when running the sample:
```
usage: train.py [-h] [--no-progress-bar] [--log-interval N]
[--log-format {json,none,simple,tqdm}] [--seed N] [--fp16]
[--profile PROFILE] [--task TASK]
[--skip-invalid-size-inputs-valid-test] [--max-tokens N]
[--task TASK] [--skip-invalid-size-inputs-valid-test] [--max-tokens N]
[--max-sentences N] [--sentencepiece] [--train-subset SPLIT]
[--valid-subset SPLIT] [--max-sentences-valid N]
[--gen-subset SPLIT] [--num-shards N] [--shard-id ID]

13
PyTorch/Translation/Transformer/fairseq/options.py
View file

@ -40,7 +40,6 @@ def get_training_parser():
add_checkpoint_args(parser)
add_inference_args(parser)
add_perf_args(parser)
add_profiling_args(parser)
return parser
@ -304,18 +303,6 @@ def add_inference_args(parser):
group.add_argument('--fp16', action='store_true', help='use fp16 precision')
return group
def add_profiling_args(parser):
group = parser.add_argument_group('Profiling')
group.add_argument('--profile', action='store_true',
help='Run profiler')
group.add_argument('--profiler-steps', type=int, default=10,
help='Override to the max steps argument')
group.add_argument('--profiler-start-iter', type=int, default=20,
help='Start profiling on this iteration')
return group
def add_model_args(parser):
group = parser.add_argument_group('Model configuration')

2
PyTorch/Translation/Transformer/inference.py
View file

@ -127,7 +127,7 @@ def setup_logger(args):
def main(args):
setup_logger(args)
args.interactive = sys.stdin.isatty() # Just make the code more understendable
args.interactive = sys.stdin.isatty() and not args.file # Just make the code more understendable
if args.file:
data_descriptor = open(args.file, 'r')

20
PyTorch/Translation/Transformer/train.py
View file

@ -29,9 +29,6 @@ import ctypes
from copy import deepcopy
import pyprof
import torch.cuda.profiler as profiler
import torch
import sacrebleu
import dllogger as DLLogger
@ -44,12 +41,8 @@ from fairseq.data import data_utils, load_dataset_splits
from fairseq.models import build_model
from fairseq.log_helper import setup_logger, reset_perf_meters
def main(args):
if args.profile:
pyprof.init(enable_function_stack=True)
print(args)
setup_logger(args)
@ -125,8 +118,7 @@ def main(args):
while lr >= args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates() < max_update and current_bleu < tgt_bleu:
DLLogger.log(step=trainer.get_num_updates()+1, data={'epoch': epoch_itr.epoch}, verbosity=0)
# train for one epoch
with torch.autograd.profiler.emit_nvtx(enabled=args.profile):
train(args, trainer, epoch_itr)
train(args, trainer, epoch_itr)
DLLogger.log(step=trainer.get_num_updates(), data={'walltime': train_meter.sum}, verbosity=1)
DLLogger.log(step=trainer.get_num_updates(),
data={'avg_epoch_loss': trainer.avg_loss_meter.avg}, verbosity=1)
@ -179,11 +171,6 @@ def train(args, trainer, epoch_itr):
trainer.get_throughput_meter().reset()
for i, sample in enumerate(itr):
# Profiling ---------
if trainer.get_num_updates() == args.profiler_start_iter:
profiler.start()
# -------------------
if i < num_batches - 1 and (i + 1) % update_freq > 0:
# buffer updates according to --update-freq
trainer.train_step(sample, update_params=False, last_step=(i == len(itr)-1))
@ -191,11 +178,6 @@ def train(args, trainer, epoch_itr):
else:
trainer.train_step(sample, update_params=True, last_step=(i == len(itr)-1))
# Profiling ---------
if trainer.get_num_updates() == args.profiler_start_iter + args.profiler_steps:
profiler.stop()
# -------------------
# ignore the first mini-batch in words-per-second calculation
if i == 0:
trainer.get_throughput_meter().reset()