# -*- coding: utf-8 -*-
# File: prof.py
import multiprocessing as mp
import numpy as np
import os
import time
import tensorflow as tf
from six.moves import map, queue
import psutil
from ..tfutils.common import gpu_available_in_session
from ..utils import logger
from ..utils.timer import Timer
from ..utils.concurrency import ensure_proc_terminate, start_proc_mask_signal
from ..utils.gpu import get_num_gpu
from ..utils.nvml import NVMLContext
from .base import Callback
__all__ = ['GPUUtilizationTracker', 'GraphProfiler', 'PeakMemoryTracker',
'GPUMemoryTracker', 'HostMemoryTracker', 'ThroughputTracker']
class GPUUtilizationTracker(Callback):
""" Summarize the average GPU utilization within an epoch.
It will start a process to obtain GPU utilization through NVML every second
within the epoch (the trigger_epoch time was not included),
and write average utilization to monitors.
This callback creates a process, therefore it's not safe to be used with MPI.
"""
_chief_only = False
def __init__(self, devices=None):
"""
Args:
devices (list[int]): physical GPU ids to monitor. If None, will guess from the environment.
"""
assert os.name != 'nt', "GPUUtilizationTracker does not support windows!"
self._devices = devices
self._enabled = True
def _guess_devices(self):
env = os.environ.get('CUDA_VISIBLE_DEVICES')
if env is None:
devices = list(range(get_num_gpu()))
if len(devices) > 1:
logger.warn("[GPUUtilizationTracker] Both devices and CUDA_VISIBLE_DEVICES are None! "
"Will monitor all {} visible GPUs!".format(len(devices)))
else:
if len(env):
devices = list(map(int, env.split(',')))
else:
devices = []
return devices
def _setup_graph(self):
# special heuristics for Horovod
from ..train import HorovodTrainer
if isinstance(self.trainer, HorovodTrainer):
if self.trainer.mpi_enabled():
logger.warn("GPUUtilizationTracker is disabled under MPI.")
self._enabled = False
return
else:
self._devices = [self.trainer.hvd.local_rank()]
if self._devices is None:
self._devices = self._guess_devices()
assert len(self._devices), "[GPUUtilizationTracker] No GPU device given!"
self._evt = mp.Event()
self._stop_evt = mp.Event()
self._queue = mp.Queue()
self._proc = mp.Process(target=self.worker, args=(
self._evt, self._queue, self._stop_evt, self._devices))
ensure_proc_terminate(self._proc)
start_proc_mask_signal(self._proc)
def _before_train(self):
assert gpu_available_in_session(), "[GPUUtilizationTracker] needs GPU!"
def _before_epoch(self):
if self._enabled:
self._evt.set()
def _after_epoch(self):
if self._enabled:
while self._evt.is_set(): # unlikely, unless the epoch is extremely fast
pass
self._evt.set()
def _trigger_epoch(self):
# Don't do this in after_epoch because
# before,after_epoch are supposed to be extremely fast by design.
if not self._enabled:
return
try:
stats = self._queue.get(timeout=60)
except queue.Empty:
if self._proc.is_alive():
raise RuntimeError("GPUUtilization.worker() is stuck. This is a bug.")
else:
raise RuntimeError("GPUUtilization.worker() process is killed unexpectedly.")
if isinstance(stats, int) and stats == -1:
from ..train.base import StopTraining
raise StopTraining("GPUUtilizationTracker.worker has failed.")
for idx, dev in enumerate(self._devices):
self.trainer.monitors.put_scalar('GPUUtil/{}'.format(dev), stats[idx])
def _after_train(self):
if self._enabled:
self._stop_evt.set()
self._evt.set()
self._proc.terminate()
@staticmethod
def worker(evt, rst_queue, stop_evt, devices):
"""
Args:
devices (list[int])
"""
with NVMLContext() as ctx:
devices = [ctx.device(i) for i in devices]
while True:
try:
evt.wait() # start epoch
evt.clear()
if stop_evt.is_set(): # or on exit
return
stats = np.zeros((len(devices),), dtype='f4')
cnt = 0
while True:
time.sleep(1)
data = [d.utilization()['gpu'] for d in devices]
data = list(map(float, data))
stats += data
cnt += 1
if evt.is_set(): # stop epoch
if stop_evt.is_set(): # or on exit
return
evt.clear()
if cnt > 1:
# Ignore the last datapoint. Usually is zero, makes us underestimate the util.
stats -= data
cnt -= 1
rst_queue.put(stats / cnt)
break
except Exception:
logger.exception("Exception in GPUUtilizationTracker.worker")
rst_queue.put(-1)
return
# Can add more features from tfprof
# https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/profiler/README.md
class GraphProfiler(Callback):
"""
Enable profiling by installing session hooks,
and write tracing files / events / metadata to ``logger.get_logger_dir()``.
The tracing files can be loaded from ``chrome://tracing``.
The metadata files can be processed by
`tfprof command line utils
<https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/profiler/README.md>`_.
The event is viewable from tensorboard.
Tips:
Note that the profiling is by default enabled for every step and is expensive.
You probably want to schedule it less frequently, e.g.:
.. code-block:: none
EnableCallbackIf(
GraphProfiler(dump_tracing=True, dump_event=True),
lambda self: self.trainer.global_step > 20 and self.trainer.global_step < 30)
"""
def __init__(self, dump_metadata=False, dump_tracing=True, dump_event=False):
"""
Args:
dump_metadata(bool): Dump :class:`tf.RunMetadata` to be used with tfprof.
dump_tracing(bool): Dump chrome tracing files.
dump_event(bool): Dump to an event processed by FileWriter and
will be shown in TensorBoard.
"""
self._dir = logger.get_logger_dir()
self._dump_meta = bool(dump_metadata)
self._dump_tracing = bool(dump_tracing)
self._dump_event = bool(dump_event)
assert os.path.isdir(self._dir), self._dir
def _before_run(self, _):
opt = tf.RunOptions()
opt.trace_level = tf.RunOptions.FULL_TRACE
return tf.train.SessionRunArgs(fetches=None, options=opt)
def _after_run(self, _, run_values):
meta = run_values.run_metadata
if self._dump_meta:
self._write_meta(meta)
if self._dump_tracing:
self._write_tracing(meta)
if self._dump_event:
self._write_event(meta)
def _write_meta(self, metadata):
fname = os.path.join(
self._dir, 'runmetadata-{}.pb'.format(self.global_step))
with open(fname, 'wb') as f:
f.write(metadata.SerializeToString())
def _write_tracing(self, metadata):
from tensorflow.python.client import timeline
tl = timeline.Timeline(step_stats=metadata.step_stats)
fname = os.path.join(
self._dir, 'chrome-trace-{}.json'.format(self.global_step))
with open(fname, 'w') as f:
f.write(tl.generate_chrome_trace_format(
show_dataflow=True, show_memory=True))
def _write_event(self, metadata):
evt = tf.Event()
evt.tagged_run_metadata.tag = 'trace-{}'.format(self.global_step)
evt.tagged_run_metadata.run_metadata = metadata.SerializeToString()
self.trainer.monitors.put_event(evt)
class GPUMemoryTracker(Callback):
"""
Track peak memory used on each GPU device every epoch, by :mod:`tf.contrib.memory_stats`.
The peak memory comes from the ``MaxBytesInUse`` op, which is the peak memory used
in recent ``session.run`` calls.
See https://github.com/tensorflow/tensorflow/pull/13107.
"""
_chief_only = False
def __init__(self, devices=(0,)):
"""
Args:
devices([int] or [str]): list of GPU devices to track memory on.
"""
assert isinstance(devices, (list, tuple)), devices
devices = ['/gpu:{}'.format(x) if isinstance(x, int) else x for x in devices]
self._devices = devices
def _setup_graph(self):
from tensorflow.contrib.memory_stats import MaxBytesInUse
ops = []
for dev in self._devices:
with tf.device(dev):
ops.append(MaxBytesInUse())
self._fetches = tf.train.SessionRunArgs(fetches=ops)
def _before_train(self):
assert gpu_available_in_session(), "PeakMemoryTracker only supports GPU!"
def _before_run(self, _):
if self.local_step == self.trainer.steps_per_epoch - 1:
return self._fetches
return None
def _after_run(self, _, rv):
results = rv.results
if results is not None:
for mem, dev in zip(results, self._devices):
self.trainer.monitors.put_scalar('PeakMemory(MB)' + dev, mem / 1e6)
PeakMemoryTracker = GPUMemoryTracker
class HostMemoryTracker(Callback):
"""
Track free RAM on the host.
When triggered, it writes the size of free RAM into monitors.
"""
_chief_only = False
def _setup_graph(self):
logger.info("[HostMemoryTracker] Free RAM in setup_graph() is {:.2f} GB.".format(self._free_ram_gb()))
def _before_train(self):
logger.info("[HostMemoryTracker] Free RAM in before_train() is {:.2f} GB.".format(self._free_ram_gb()))
def _trigger(self):
ram_gb = self._free_ram_gb()
self.trainer.monitors.put_scalar('HostFreeMemory (GB)', ram_gb)
def _free_ram_gb(self):
return psutil.virtual_memory().available / 1024**3
class ThroughputTracker(Callback):
"""
This callback writes the training throughput (in terms of either steps/sec, or samples/sec)
to the monitors everytime it is triggered.
The throughput is computed based on the duration between the consecutive triggers.
The time spent on callbacks after each epoch is excluded.
"""
_chief_only = False
def __init__(self, samples_per_step=None):
"""
Args:
samples_per_step (int or None): total number of samples processed in each step
(i.e., your total batch size in each step).
If not provided, this callback will record "steps/sec" instead of "samples/sec".
"""
if samples_per_step is not None:
samples_per_step = int(samples_per_step)
self._samples_per_step = samples_per_step
self._timer = Timer()
self._timer.pause()
# only include the time between before_epoch/after_epoch
def _before_epoch(self):
self._timer.resume()
def _after_epoch(self):
self._timer.pause()
def _before_train(self):
self._update_last()
def _update_last(self):
old_pause = self._timer.is_paused()
self._timer.reset()
if old_pause:
self._timer.pause()
self._last_step = self.global_step
def _trigger(self):
steps_per_sec = (self.global_step - self._last_step) / self._timer.seconds()
self._update_last()
if self._samples_per_step is None:
self.trainer.monitors.put_scalar("Throughput (steps/sec)", steps_per_sec)
else:
self.trainer.monitors.put_scalar("Throughput (samples/sec)", steps_per_sec * self._samples_per_step)
