def start_stage(self):
if self.n_procs > 1:
config = cfg.freeze()
self.inter_comm.bcast(config, root=MPI.ROOT)
seeds = [gen_seed() for _ in range(self.n_procs - 1)]
self.inter_comm.scatter(seeds, root=MPI.ROOT)
def framework_initialize_stage(self, stack):
# Configure and create session and graph for stage.
session_config = tf.ConfigProto()
session_config.intra_op_parallelism_threads = cfg.get(
'intra_op_parallelism_threads', 0)
session_config.inter_op_parallelism_threads = cfg.get(
'inter_op_parallelism_threads', 0)
session_config.log_device_placement = cfg.get('log_device_placement',
0)
if cfg.use_gpu:
per_process_gpu_memory_fraction = getattr(
cfg, 'per_process_gpu_memory_fraction', None)
if per_process_gpu_memory_fraction:
session_config.gpu_options.per_process_gpu_memory_fraction = per_process_gpu_memory_fraction
gpu_allow_growth = getattr(cfg, 'gpu_allow_growth', None)
if gpu_allow_growth:
session_config.gpu_options.allow_growth = gpu_allow_growth
_print("Using GPU if available.")
_print("Using {}% of GPU memory.".format(
100 *
session_config.gpu_options.per_process_gpu_memory_fraction))
_print("Allowing growth of GPU memory: {}".format(
session_config.gpu_options.allow_growth))
graph = tf.Graph()
sess = tf.Session(graph=graph, config=session_config)
# This HAS to come after the creation of the session, otherwise
# it allocates all GPU memory if using the GPU.
_print("\nAvailable devices: ")
from tensorflow.python.client import device_lib
_print(device_lib.list_local_devices())
if not cfg.use_gpu:
_print("Not using GPU.")
stack.enter_context(graph.device("/cpu:0"))
stack.enter_context(graph.as_default())
stack.enter_context(sess)
stack.enter_context(sess.as_default())
# Set the seed for the stage.
tf_seed = gen_seed()
_print(
"Setting tensorflow seed to generated seed: {}\n".format(tf_seed))
tf.set_random_seed(tf_seed)
tf.logging.set_verbosity(tf.logging.ERROR)
def sample_configs(distributions, n_repeats, n_samples=None):
""" Samples configs from a distribution for hyper-parameter search.
Parameters
----------
distributions: dict or None
Mapping from parameter names to distributions (objects with
member function ``rvs`` which accepts a shape and produces
an array of samples with that shape).
n_repeats: int > 0
Number of different seeds to use for each sampled configuration.
n_samples: int > 0
Number of configs to sample.
"""
samples = []
if distributions is None:
samples = [Config()]
elif isinstance(distributions, list):
samples = distributions + []
if n_samples:
samples = list(np.random.permutation(samples)[:n_samples])
else:
if not n_samples:
samples = generate_all(distributions)
else:
samples = nested_sample(distributions, n_samples)
print("Sampled configs:")
pprint(samples)
configs = []
for i, s in enumerate(samples):
s['idx'] = i
for r in range(n_repeats):
_new = copy.deepcopy(s)
_new['repeat'] = r
_new['seed'] = gen_seed()
configs.append(_new)
return configs
def framework_initialize_stage(self, stack):
# Set the seed for the stage.
torch_seed = gen_seed()
_print(
"Setting pytorch seed to generated seed: {}\n".format(torch_seed))
torch.manual_seed(torch_seed)
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = cfg.pytorch_cudnn_benchmark
torch.backends.cudnn.deterministic = cfg.pytorch_cudnn_deterministic
if cfg.use_gpu:
_print("Trying to use GPU...")
try:
device = torch.cuda.current_device()
use_gpu = True
except AssertionError:
tb.print_exc()
use_gpu = False
else:
use_gpu = False
if use_gpu:
_print("Using GPU.")
_print("Device count: {}".format(torch.cuda.device_count()))
_print("Device idx: {}".format(device))
_print("Device name: {}".format(
torch.cuda.get_device_name(device)))
_print("Device capability: {}".format(
torch.cuda.get_device_capability(device)))
set_pytorch_device('cuda')
else:
_print("Not using GPU.")
set_pytorch_device('cpu')
torch.set_printoptions(profile='full')
def build_and_submit(category,
exp_name,
config,
distributions,
n_param_settings=0,
n_repeats=1,
do_local_test=False,
kind="local",
readme="",
tasks_per_gpu=1,
**run_kwargs):
""" Build a job and submit it. Meant to be called from within a script.
Parameters
----------
category: str
High-level category of the experiment. Determines the ExperimentStore
where the experiment data will be stored.
exp_name: str
Low-level name of the experiment.
config: Config instance or dict
Configuration to use as the base config for all jobs.
distributions: dict
Object used to generate variations of the base config (so that different
jobs test different parameters).
n_param_settings: int
Number of different configurations to sample from `distributions`. If not
supplied, it is assumed that `distributions` actually specifies a grid
search, and an attempt is made to generate all possible configurations int
that grid search.
n_repeats: int
Number of experiments to run (with different random seeds) for each
generated configuration.
do_local_test: bool
If True, sample one of the generated configurations and use it to run a
short test locally, to ensure that the jobs will run properly.
kind: str
One of pbs, slurm, slurm-local, parallel, local. Specifies which method
should be used to run the jobs in parallel.
readme: str
A string outlining the purpose/context for the created experiment.
**run_kwargs:
Additional arguments that are ultimately passed to `ParallelSession` in
order to run the job.
"""
# Get run_kwargs from command line
sig = inspect.signature(ParallelSession.__init__)
default_run_kwargs = sig.bind_partial()
default_run_kwargs.apply_defaults()
cl_run_kwargs = clify.command_line(default_run_kwargs.arguments).parse()
run_kwargs.update(cl_run_kwargs)
if config.seed is None or config.seed < 0:
config.seed = gen_seed()
assert kind in "pbs slurm slurm-local parallel local".split()
assert 'build_command' not in config
config['build_command'] = ' '.join(sys.argv)
print(config['build_command'])
if kind == "local":
with config:
from dps.train import training_loop
return training_loop()
else:
config.name = category
config = config.copy()
if readme == "_vim_":
readme = edit_text(prefix="dps_readme_",
editor="vim",
initial_text="README.md: \n")
scratch = os.path.join(cfg.parallel_experiments_build_dir, category)
archive_path, n_tasks = build_search(scratch,
exp_name,
distributions,
config,
add_date=1,
_zip=True,
do_local_test=do_local_test,
n_param_settings=n_param_settings,
n_repeats=n_repeats,
readme=readme)
run_kwargs.update(archive_path=archive_path,
category=category,
exp_name=exp_name,
kind=kind)
gpu_kind = run_kwargs.get('gpu_kind', None)
resources = compute_required_resources(n_tasks, tasks_per_gpu,
gpu_kind)
run_kwargs.update(resources)
parallel_session = submit_job(**run_kwargs)
return parallel_session
def build_search(path,
name,
distributions,
config,
n_repeats,
n_param_settings=None,
_zip=True,
add_date=0,
do_local_test=True,
readme=""):
""" Create a job implementing a hyper-parameter search.
Parameters
----------
path: str
Path to the directory where the search archive will be saved.
name: str
Name for the search.
distributions: dict (str -> (list or distribution))
Distributions to sample from. Can also be a list of samples.
config: Config instance
The base configuration.
n_repeats: int
Number of different random seeds to run each sample with.
n_param_settings: int
Number of parameter settings to sample. If not supplied, all
possibilities are generated.
_zip: bool
Whether to zip the created search directory.
add_date: bool
Whether to add time to name of experiment directory.
do_local_test: bool
If True, run a short test using one of the sampled
configs on the local machine to catch any dumb errors
before starting the real experiment.
readme: str
String specifiying context/purpose of search.
"""
if config.get('seed', None) is None:
config.seed = gen_seed()
with NumpySeed(config.seed):
es = ExperimentStore(path, prefix="build_search")
count = 0
base_name = name
has_built = False
while not has_built:
try:
exp_dir = es.new_experiment(name,
config.seed,
add_date=add_date,
force_fresh=1)
has_built = True
except FileExistsError:
name = "{}_{}".format(base_name, count)
count += 1
if readme:
with open(exp_dir.path_for('README.md'), 'w') as f:
f.write(readme)
print(config)
exp_dir.record_environment(config=config)
print("Building parameter search at {}.".format(exp_dir.path))
job = Job(exp_dir.path)
new_configs = sample_configs(distributions, n_repeats,
n_param_settings)
with open(exp_dir.path_for("sampled_configs.txt"), "w") as f:
f.write("\n".join("idx={}: {}".format(c["idx"], pformat(c))
for c in new_configs))
print("{} configs were sampled for parameter search.".format(
len(new_configs)))
if do_local_test:
print("\nStarting local test " + ("=" * 80))
test_config = new_configs[0].copy()
test_config.update(max_steps=1000, render_hook=None)
_RunTrainingLoop(config)(test_config)
print("Done local test " + ("=" * 80) + "\n")
job.map(_RunTrainingLoop(config.copy()), new_configs)
job.save_object('metadata', 'distributions', distributions)
job.save_object('metadata', 'config', config)
print(job.summary())
if _zip:
path = job.zip(delete=True)
else:
path = exp_dir.path
print("Zipped {} as {}.".format(exp_dir.path, path))
return path, len(new_configs)
def seed(self, seed=None):
np.random.seed(seed)
for env in self._env_copies:
s = gen_seed()
env.seed(s)
def _run(self):
print(cfg.to_string())
threshold_reached = True
self.global_step = 0
self.n_global_experiences = 0
self.curriculum_remaining = self.curriculum + []
self.curriculum_complete = []
stage_idx = 0
while self.curriculum_remaining:
print("\n" + "=" * 50)
self.timestamp("Starting stage {}".format(stage_idx))
print("\n")
if cfg.start_tensorboard:
restart_tensorboard(self.experiment_store.path, cfg.tbport, cfg.reload_interval)
stage_config = self.curriculum_remaining.pop(0)
stage_config = Config(stage_config)
self.data.start_stage(stage_idx, stage_config)
with ExitStack() as stack:
# --------------- Stage set-up -------------------
print("\n" + "-" * 10 + " Stage set-up " + "-" * 10)
print("\nNew config values for this stage are: \n{}\n".format(pformat(stage_config)))
stack.enter_context(stage_config)
stage_prepare_func = cfg.get("stage_prepare_func", None)
if callable(stage_prepare_func):
stage_prepare_func() # Modify the stage config in arbitrary ways before starting stage
self.mpi_context.start_stage()
# Configure and create session and graph for stage.
session_config = tf.ConfigProto()
session_config.intra_op_parallelism_threads = cfg.get('intra_op_parallelism_threads', 0)
session_config.inter_op_parallelism_threads = cfg.get('inter_op_parallelism_threads', 0)
session_config.log_device_placement = cfg.get('log_device_placement', 0)
if cfg.use_gpu:
per_process_gpu_memory_fraction = getattr(cfg, 'per_process_gpu_memory_fraction', None)
if per_process_gpu_memory_fraction:
session_config.gpu_options.per_process_gpu_memory_fraction = per_process_gpu_memory_fraction
gpu_allow_growth = getattr(cfg, 'gpu_allow_growth', None)
if gpu_allow_growth:
session_config.gpu_options.allow_growth = gpu_allow_growth
if cfg.use_gpu:
print("Using GPU if available.")
print("Using {}% of GPU memory.".format(
100 * session_config.gpu_options.per_process_gpu_memory_fraction))
print("Allowing growth of GPU memory: {}".format(session_config.gpu_options.allow_growth))
graph = tf.Graph()
sess = tf.Session(graph=graph, config=session_config)
# This HAS to come after the creation of the session, otherwise
# it allocates all GPU memory if using the GPU.
print("\nAvailable devices: ")
from tensorflow.python.client import device_lib
print(device_lib.list_local_devices())
if not cfg.use_gpu:
print("Not using GPU.")
stack.enter_context(graph.device("/cpu:0"))
stack.enter_context(graph.as_default())
stack.enter_context(sess)
stack.enter_context(sess.as_default())
# Set the seed for the stage. Notice we generate a new tf seed for each stage.
tf_seed = gen_seed()
print("Setting tensorflow seed to generated seed: {}\n".format(tf_seed))
tf.set_random_seed(tf_seed)
# Set limit on CPU RAM for the stage
cpu_ram_limit_mb = cfg.get("cpu_ram_limit_mb", None)
if cpu_ram_limit_mb is not None:
stack.enter_context(memory_limit(cfg.cpu_ram_limit_mb))
print("Building env...\n")
# Maybe build env
if stage_idx == 0 or not cfg.preserve_env:
if getattr(self, 'env', None):
self.env.close()
self.env = cfg.build_env()
if hasattr(self.env, "print_memory_footprint"):
self.env.print_memory_footprint()
print("\nDone building env.\n")
print("Building updater...\n")
import warnings
with warnings.catch_warnings():
warnings.simplefilter('once')
if cfg.n_procs > 1:
updater = cfg.get_updater(self.env, mpi_context=self.mpi_context)
else:
updater = cfg.get_updater(self.env)
updater.stage_idx = stage_idx
updater.exp_dir = self.exp_dir
updater.build_graph()
print("\nDone building updater.\n")
walk_variable_scopes(max_depth=3)
# Maybe initialize network weights.
# Let a *path_specification* be one of three things:
# 1. An integer specifying a stage to load the best hypothesis from.
# 2. A string of format: "stage_idx,kind" where `stage_idx` specifies a stage to load from
# and `kind` is either "final" or "best", specifying whether to load final or best
# hypothesis from that stage.
# 3. A path on the filesystem that gives a prefix for a tensorflow checkpoint file to load from.
#
# Then cfg.load_path can either be a path_specification itself, in which case all variables
# in the network will be loaded from that path_specification, or a dictionary mapping from
# variable scope names to path specifications, in which case all variables in each supplied
# variable scope name will be loaded from the path_specification paired with that scope name.
load_path = cfg.load_path
if load_path is not None:
if isinstance(load_path, str) or isinstance(load_path, int):
load_path = {"": load_path}
load_path = dict(load_path)
# Sort in increasing order, so that it if one variable scope lies within another scope,
# the outer scope gets loaded before the inner scope, rather than having the outer scope
# wipe out the inner scope.
items = sorted(load_path.items())
for var_scope, path in items:
variables = {v.name: v for v in trainable_variables(var_scope, for_opt=False)}
if not variables:
print("No variables to load in scope {}.".format(str(var_scope)))
continue
saver = tf.train.Saver(variables)
load_stage, kind = None, None
if isinstance(path, int):
load_stage = path
kind = "best"
elif isinstance(path, str):
try:
split = path.split(',')
load_stage = int(split[0])
kind = 'best' if len(split) > 1 else split[1]
assert kind in 'best final'.split(), "path={}".format(path)
except Exception:
load_stage, kind = None, None
if load_stage is not None:
if stage_idx == 0:
print(
"Not loading var scope \"{}\" from stage {}, "
"currently in stage 0.".format(var_scope, load_stage))
continue
else:
key = kind + '_path'
completed_history = self.data.history[:-1]
path = completed_history[load_stage][key]
path = os.path.realpath(path)
saver.restore(tf.get_default_session(), path)
print("Loading var scope \"{}\" from {}.".format(var_scope, path))
else:
print("Using a fresh set of weights, not loading anything.")
tf.train.get_or_create_global_step()
sess.run(uninitialized_variables_initializer())
sess.run(tf.assert_variables_initialized())
for hook in cfg.hooks:
assert isinstance(hook, Hook)
hook.start_stage(self, updater, stage_idx)
threshold_reached = False
reason = None
try:
# --------------- Run stage -------------------
start = time.time()
phys_memory_before = memory_usage(physical=True)
gpu_memory_before = gpu_memory_usage()
threshold_reached, reason = self._run_stage(stage_idx, updater)
except KeyboardInterrupt:
reason = "User interrupt"
except NotImplementedError as e:
# There is a bug in pdb_postmortem that prevents instances of `NotImplementedError`
# from being handled properly, so replace it with an instance of `Exception`.
if cfg.robust:
traceback.print_exc()
reason = "Exception occurred ({})".format(repr(e))
else:
raise Exception("NotImplemented") from e
except Exception as e:
reason = "Exception occurred ({})".format(repr(e))
if cfg.robust:
traceback.print_exc()
else:
raise
except Alarm:
reason = "Time limit exceeded"
raise
finally:
phys_memory_after = memory_usage(physical=True)
gpu_memory_after = gpu_memory_usage()
self.data.record_values_for_stage(
stage_duration=time.time()-start,
phys_memory_before_mb=phys_memory_before,
phys_memory_delta_mb=phys_memory_after - phys_memory_before,
gpu_memory_before_mb=gpu_memory_before,
gpu_memory_delta_mb=gpu_memory_after - gpu_memory_before
)
self.data.record_values_for_stage(reason=reason)
print("\n" + "-" * 10 + " Optimization complete " + "-" * 10)
print("\nReason: {}.\n".format(reason))
final_path = self.data.path_for('weights/final_for_stage_{}'.format(stage_idx))
final_path = cfg.get('save_path', final_path)
final_path = updater.save(tf.get_default_session(), final_path)
self.data.record_values_for_stage(final_path=final_path)
# --------------- Maybe render performance of best hypothesis -------------------
do_final_testing = (
"Exception occurred" not in reason
and reason != "Time limit exceeded"
and 'best_path' in self.data.current_stage_record)
if do_final_testing:
try:
print("\n" + "-" * 10 + " Final testing/rendering " + "-" * 10)
print("Best hypothesis for this stage was found on "
"step (l: {best_local_step}, g: {best_global_step}) "
"with stopping criteria ({sc_name}) of {best_stopping_criteria}.".format(
sc_name=self.stopping_criteria_name, **self.data.current_stage_record))
best_path = self.data.current_stage_record['best_path']
print("Loading best hypothesis for this stage "
"from file {}...".format(best_path))
updater.restore(sess, best_path)
test_record = updater.evaluate(cfg.batch_size, mode="test")
for hook in cfg.hooks:
if hook.call_per_timestep and hook.final:
hook_record = hook.step(self, updater)
if hook_record:
assert len(hook_record) == 1
for k, d in dict(hook_record).items():
test_record.update(d)
self.data.record_values_for_stage(
**{'_test_' + k: v for k, v in test_record.items()})
if cfg.render_step > 0 and cfg.render_hook is not None:
print("Rendering...")
cfg.render_hook(updater)
print("Done rendering.")
except BaseException:
print("Exception occurred while performing final testing/rendering: ")
traceback.print_exc()
else:
print("\n" + "-" * 10 + " Skipping final testing/rendering " + "-" * 10)
# --------------- Finish up the stage -------------------
self.data.end_stage(updater.n_updates)
print("\n" + "-" * 10 + " Running end-of-stage hooks " + "-" * 10 + "\n")
for hook in cfg.hooks:
hook.end_stage(self, stage_idx)
print()
self.timestamp("Done stage {}".format(stage_idx))
print("=" * 50)
stage_idx += 1
self.curriculum_complete.append(stage_config)
if not (threshold_reached or cfg.power_through):
print("Failed to reach stopping criteria threshold on stage {} "
"of the curriculum, terminating.".format(stage_idx))
break
def run(self, start_time):
""" Run the training loop.
Parameters
----------
start_time: int
Start time (in seconds since epoch) for measuring elapsed time for
purposes of interrupting the training loop.
"""
if start_time is None:
start_time = time.time()
self.start_time = start_time
self.timestamp("Entering TrainingLoop.run")
prepare_func = cfg.get("prepare_func", None)
if callable(prepare_func):
prepare_func() # Modify the config in arbitrary ways before training
else:
try:
prepare_funcs = list(prepare_func)
except (TypeError, ValueError):
pass
else:
for f in prepare_funcs:
if callable(f):
f()
self.curriculum = cfg.curriculum + []
if cfg.seed is None or cfg.seed < 0:
cfg.seed = gen_seed()
# Create a directory to store the results of the training session.
self.experiment_store = ExperimentStore(os.path.join(cfg.local_experiments_dir, cfg.env_name))
exp_dir = self.experiment_store.new_experiment(
self.exp_name, cfg.seed, add_date=1, force_fresh=1, update_latest=False)
self.exp_dir = exp_dir
cfg.path = exp_dir.path
breaker = "-" * 40
header = "{}\nREADME.md - {}\n{}\n\n\n".format(breaker, os.path.basename(exp_dir.path), breaker)
readme = header + (cfg.readme if cfg.readme else "") + "\n\n"
with open(exp_dir.path_for('README.md'), 'w') as f:
f.write(readme)
self.data = _TrainingLoopData(exp_dir)
self.data.setup()
frozen_data = None
with ExitStack() as stack:
if cfg.pdb:
stack.enter_context(pdb_postmortem())
print("`pdb` is turned on, so forcing setting robust=False")
cfg.robust = False
stack.enter_context(redirect_stream('stdout', self.data.path_for('stdout'), tee=cfg.tee))
stack.enter_context(redirect_stream('stderr', self.data.path_for('stderr'), tee=cfg.tee))
print("\n\n" + "=" * 80)
self.timestamp("Starting training run (name={})".format(self.exp_name))
print("\nDirectory for this training run is {}.".format(exp_dir.path))
stack.enter_context(NumpySeed(cfg.seed))
print("\nSet numpy random seed to {}.\n".format(cfg.seed))
limiter = time_limit(
self.time_remaining, verbose=True,
timeout_callback=lambda limiter: print("Training run exceeded its time limit."))
self.mpi_context = MPI_MasterContext(cfg.get('n_procs', 1), exp_dir)
try:
with limiter:
self._run()
finally:
self.data.summarize()
self.timestamp("Done training run (name={})".format(self.exp_name))
print("=" * 80)
print("\n\n")
frozen_data = self.data.freeze()
self.timestamp("Leaving TrainingLoop.run")
return frozen_data
def make_dataset_in_parallel(run_kwargs, dataset_cls, param_values=None):
""" Uses dps.hyper.parallel_session.ParallelSession to create a dataset in parallel. """
# Get run_kwargs from command line
sig = inspect.signature(ParallelSession.__init__)
default_run_kwargs = sig.bind_partial()
default_run_kwargs.apply_defaults()
cl_run_kwargs = clify.command_line(default_run_kwargs.arguments).parse()
run_kwargs.update(cl_run_kwargs)
param_values = param_values or dataset_cls._capture_param_values()
param_values = Config(param_values)
seed = param_values["seed"]
if seed is None or seed < 0:
seed = gen_seed()
n_examples = param_values["n_examples"]
n_examples_per_shard = run_kwargs["n_examples_per_shard"]
experiment_store = ExperimentStore(
cfg.parallel_experiments_build_dir, prefix="build_{}".format(dataset_cls.__name__))
count = 0
name = "attempt=0"
has_built = False
while not has_built:
try:
exp_dir = experiment_store.new_experiment(name, seed, add_date=True, force_fresh=True)
has_built = True
except FileExistsError:
count += 1
name = "attempt_{}".format(count)
print("Building dataset.")
job = Job(exp_dir.path)
n_examples_remaining = n_examples
with NumpySeed(seed):
inputs = []
idx = 0
while n_examples_remaining:
seed = gen_seed()
cur_n_examples = min(n_examples_remaining, n_examples_per_shard)
n_examples_remaining -= cur_n_examples
inputs.append((idx, seed, cur_n_examples))
idx += 1
job.map(_BuildDataset(dataset_cls, param_values), inputs)
job.save_object('metadata', 'param_values', param_values)
print(job.summary())
archive_path = job.zip(delete=True)
print("Zipped {} as {}.".format(exp_dir.path, archive_path))
run_kwargs = run_kwargs.copy()
del run_kwargs['n_examples_per_shard']
run_kwargs.update(
archive_path=archive_path, name=name, kind="parallel",
parallel_exe=cfg.parallel_exe)
parallel_session = submit_job(**run_kwargs)
with cd(os.path.join(parallel_session.job_path, 'experiments')):
dataset_files = []
for dir_path, dirs, files in os.walk('.'):
if not dir_path.startswith("./exp__seed="):
continue
df = [f for f in files if not f.endswith('.cfg')]
assert len(df) == 1
dataset_files.append(os.path.join(dir_path, df[0]))
cached_filename = os.path.join(cfg.data_dir, "cached_datasets", dataset_cls.__name__, str(get_param_hash(param_values)))
command = "cat " + " ".join(dataset_files) + " > " + cached_filename
print("Running command: \n" + command)
subprocess.run(command, shell=True, check=True)
print("Done.")
with open(cached_filename + ".cfg", 'w') as f:
f.write(pprint.pformat(param_values))
return parallel_session
def run_stage(mpi_context, env, stage_idx, exp_dir):
config, seed = mpi_context.start_stage()
with ExitStack() as stack:
stack.enter_context(config)
stack.enter_context(NumpySeed(seed))
# Accept config for new stage
print("\n" + "-" * 10 + " Stage set-up " + "-" * 10)
print(cfg.to_string())
# Configure and create session and graph for stage.
session_config = tf.ConfigProto()
session_config.intra_op_parallelism_threads = cfg.get(
'intra_op_parallelism_threads', 0)
session_config.inter_op_parallelism_threads = cfg.get(
'inter_op_parallelism_threads', 0)
# if cfg.use_gpu:
# per_process_gpu_memory_fraction = getattr(cfg, 'per_process_gpu_memory_fraction', None)
# if per_process_gpu_memory_fraction:
# session_config.gpu_options.per_process_gpu_memory_fraction = \
# per_process_gpu_memory_fraction
# gpu_allow_growth = getattr(cfg, 'gpu_allow_growth', None)
# if gpu_allow_growth:
# session_config.gpu_options.allow_growth = gpu_allow_growth
# if cfg.use_gpu:
# print("Using GPU if available.")
# print("Using {}% of GPU memory.".format(
# 100 * session_config.gpu_options.per_process_gpu_memory_fraction))
# print("Allowing growth of GPU memory: {}".format(session_config.gpu_options.allow_growth))
graph = tf.Graph()
sess = tf.Session(graph=graph, config=session_config)
# This HAS to come after the creation of the session, otherwise
# it allocates all GPU memory if using the GPU.
print("\nAvailable devices:")
from tensorflow.python.client import device_lib
print(device_lib.list_local_devices())
# if not cfg.use_gpu:
# print("Not using GPU.")
# stack.enter_context(graph.device("/cpu:0"))
stack.enter_context(graph.device("/cpu:0"))
stack.enter_context(graph.as_default())
stack.enter_context(sess)
stack.enter_context(sess.as_default())
tf_seed = gen_seed()
print(
"Setting tensorflow seed to generated seed: {}\n".format(tf_seed))
tf.set_random_seed(tf_seed)
# Set limit on CPU RAM for the stage
cpu_ram_limit_mb = cfg.get("cpu_ram_limit_mb", None)
if cpu_ram_limit_mb is not None:
stack.enter_context(memory_limit(cfg.cpu_ram_limit_mb))
print("Building env...\n")
# Maybe build env
if stage_idx == 0 or not cfg.preserve_env:
if env is not None:
env.close()
env = cfg.build_env()
if hasattr(env, "print_memory_footprint"):
env.print_memory_footprint()
print("\nDone building env.\n")
print("Building updater...\n")
updater = cfg.get_updater(env, mpi_context=mpi_context)
updater.stage_idx = stage_idx
updater.exp_dir = exp_dir
updater.build_graph()
print("\nDone building updater.\n")
# walk_variable_scopes(max_depth=3)
tf.train.get_or_create_global_step()
sess.run(uninitialized_variables_initializer())
sess.run(tf.assert_variables_initialized())
updater.worker_code()
stage_idx += 1
return env
