def _create_cache(self):
# Save all data into cache file(s).
self._cache_positions = []
self._position = 0
percent = 0
if single_or_rankzero():
progress(None)
while self._position < self._data_source._size:
if single_or_rankzero():
progress('Create cache',
self._position * 1.0 / self._data_source._size)
self._store_data_to_cache_buffer(self._position)
self._position += 1
if len(self._cache_positions) > 0:
self._save_cache_to_file()
if single_or_rankzero():
progress(None)
# Adjust data size into reseted position. In most case it means
# multiple of bunch(mini-batch) size.
num_of_cache_files = int(
numpy.ceil(float(self._data_source._size) / self._cache_size))
self._cache_file_order = self._cache_file_order[0:num_of_cache_files]
self._cache_file_data_orders = self._cache_file_data_orders[
0:num_of_cache_files]
if self._data_source._size % self._cache_size != 0:
self._cache_file_data_orders[num_of_cache_files -
1] = self._cache_file_data_orders[
num_of_cache_files -
1][0:self._data_source._size %
self._cache_size]
index_filename = os.path.join(self._cache_dir, "cache_index.csv")
with open(index_filename, 'w') as f:
writer = csv.writer(f, lineterminator='\n')
for fn, orders in zip(self._cache_file_names,
self._cache_file_data_orders):
writer.writerow((os.path.basename(fn), len(orders)))
if self._cache_file_format == ".npy":
info_filename = os.path.join(self._cache_dir, "cache_info.csv")
with open(info_filename, 'w') as f:
writer = csv.writer(f, lineterminator='\n')
for variable in self._variables:
writer.writerow((variable, ))
def _create_dataset(uri, batch_size, shuffle, no_image_normalization, cache_dir, overwrite_cache, create_cache_explicitly, prepare_data_iterator):
class Dataset:
pass
dataset = Dataset()
dataset.uri = uri
dataset.normalize = not no_image_normalization
comm = current_communicator()
# use same random state for each process until slice is called
rng = numpy.random.RandomState(0)
use_memory_cache = comm.size == 1 if comm else True
if prepare_data_iterator:
if cache_dir == '':
cache_dir = None
# Disable implicit cache creation when MPI is available.
if cache_dir and (create_cache_explicitly or comm):
cache_index = os.path.join(cache_dir, "cache_index.csv")
if not os.path.exists(cache_index) or overwrite_cache:
if single_or_rankzero():
logger.log(99, 'Creating cache data for "' + uri + '"')
try:
os.makedirs(cache_dir)
except OSError:
pass # python2 does not support exists_ok arg
with data_iterator_csv_dataset(uri, batch_size, shuffle, rng=rng, normalize=False, cache_dir=cache_dir, with_memory_cache=False) as di:
pass
rng = numpy.random.RandomState(0)
dataset.data_iterator = (lambda: data_iterator_cache(
cache_dir, batch_size, shuffle, rng=rng, normalize=dataset.normalize, with_memory_cache=use_memory_cache))
elif not cache_dir or overwrite_cache or not os.path.exists(cache_dir) or len(os.listdir(cache_dir)) == 0:
if comm:
logger.critical(
'Implicit cache creation does not support with MPI')
import sys
sys.exit(-1)
else:
if cache_dir:
try:
os.makedirs(cache_dir)
except OSError:
pass # python2 does not support exists_ok arg
dataset.data_iterator = (lambda: data_iterator_csv_dataset(
uri, batch_size, shuffle, rng=rng, normalize=dataset.normalize, cache_dir=cache_dir))
else:
dataset.data_iterator = (lambda: data_iterator_cache(
cache_dir, batch_size, shuffle, rng=rng, normalize=dataset.normalize, with_memory_cache=use_memory_cache))
else:
dataset.data_iterator = None
return dataset
def train_command(args):
callback.update_status(args)
if single_or_rankzero():
configure_progress(os.path.join(args.outdir, 'progress.txt'))
info = load.load([args.config],
prepare_data_iterator=None,
exclude_parameter=True)
# Check dataset uri is empty.
dataset_error = False
for dataset in info.datasets.values():
if dataset.uri.strip() == '':
dataset_error = True
if dataset_error:
logger.log(99, 'Fatal error. Dataset URI is empty.')
return False
class TrainConfig:
pass
config = TrainConfig()
config.timelimit = -1
if args.param:
load.load([args.param], parameter_only=True)
config.timelimit = callback.get_timelimit(args)
config.global_config = info.global_config
config.training_config = info.training_config
if single_or_rankzero():
logger.log(99, 'Train with contexts {}'.format(available_contexts))
class OptConfig:
pass
config.optimizers = OrderedDict()
for name, opt in info.optimizers.items():
o = OptConfig()
o.optimizer = opt
o.data_iterators = []
config.optimizers[name] = o
class MonConfig:
pass
config.monitors = OrderedDict()
for name, mon in info.monitors.items():
m = MonConfig()
m.monitor = mon
m.data_iterators = []
config.monitors[name] = m
# Training
comm = current_communicator()
config.training_config.iter_per_epoch //= comm.size if comm else 1
max_iteration = config.training_config.max_epoch * \
config.training_config.iter_per_epoch
global _save_parameter_info
_save_parameter_info = {}
_, config_ext = os.path.splitext(args.config)
if config_ext == '.prototxt' or config_ext == '.nntxt':
_save_parameter_info['config'] = args.config
elif config_ext == '.nnp':
with zipfile.ZipFile(args.config, 'r') as nnp:
for name in nnp.namelist():
_, ext = os.path.splitext(name)
if ext == '.nntxt' or ext == '.prototxt':
nnp.extract(name, args.outdir)
_save_parameter_info['config'] = os.path.join(
args.outdir, name)
result = False
restart = False
if max_iteration > 0:
rng = np.random.RandomState(comm.rank if comm else 0)
with ExitStack() as stack:
# Create data_iterator instance only once for each dataset in optimizers
optimizer_data_iterators = {}
for name, o in config.optimizers.items():
for di in o.optimizer.data_iterators.values():
if di not in optimizer_data_iterators:
di_instance = stack.enter_context(di())
if comm and comm.size > 1:
di_instance = di_instance.slice(
rng, comm.size, comm.rank)
optimizer_data_iterators[di] = di_instance
else:
di_instance = optimizer_data_iterators[di]
o.data_iterators.append(di_instance)
# Create data_iterator instance only once for each dataset in monitors
monitor_data_iterators = {}
for name, m in config.monitors.items():
for di in m.monitor.data_iterators.values():
if di not in monitor_data_iterators:
di_instance = stack.enter_context(di())
if comm and comm.size > 1:
di_instance = di_instance.slice(
rng, comm.size, comm.rank)
monitor_data_iterators[di] = di_instance
else:
di_instance = monitor_data_iterators[di]
m.data_iterators.append(di_instance)
monitor_data_iterators.update(optimizer_data_iterators)
result, restart = _train(args, config)
else:
# save parameters without training (0 epoch learning)
logger.log(99, '0 epoch learning. (Just save parameter.)')
if single_or_rankzero():
_save_parameters(args, None, 0, config, True)
result = True
if single_or_rankzero() and not restart:
if result:
logger.log(99, 'Training Completed.')
callback.update_status('finished')
else:
logger.log(99, 'Training Incompleted.')
callback.update_status('failed')
if single_or_rankzero():
progress(None)
return True
def _train(args, config):
global _save_parameter_info
comm = current_communicator()
_CGLOAD_LOG_INTERVAL = 20
best_epoch = None
best_error = None
last_epoch = 0
if args.resume:
last_epoch, best_epoch, best_error = _get_current_parameter(args)
if best_epoch is not None:
logger.log(
99, "Best error {} recorded at epoch {} in previous training.".
format(best_error, best_epoch))
if best_epoch > last_epoch:
logger.log(
99,
"Resumed epoch is {} but this training keep this result.".
format(last_epoch))
logger.log(99, "Resume from epoch {}".format(last_epoch + 1))
callback.update_status(('epoch.max', config.training_config.max_epoch))
callback.update_status(
('epoch.current',
last_epoch + 1 if last_epoch < config.training_config.max_epoch else
config.training_config.max_epoch))
max_iteration = config.training_config.max_epoch * \
config.training_config.iter_per_epoch
if single_or_rankzero():
logger.log(
99, 'Training epoch {} of {} begin'.format(
last_epoch + 1, config.training_config.max_epoch))
class Cost:
pass
cost = Cost()
cost.sum_epoch = 0.0
cost.num_iteration = 0
cost.sum_iteration = 0.0
cost.variables = None
class TimeInfo:
pass
timeinfo = TimeInfo()
timeinfo.past_time = 0
timeinfo.estimate_time = 0
timeinfo.last_past_time = None
if max_iteration > 0:
last_iteration = last_epoch * config.training_config.iter_per_epoch
if last_iteration < max_iteration:
timeinfo.start_time = time.time()
timeinfo.last_epoch_start_time = timeinfo.start_time
callback.update_status('processing', True, timeinfo.start_time)
for iteration in range(last_iteration, max_iteration):
# instant load measurement
measure_cpu_gpu_instant_load()
cost = _update(iteration, config, cost)
if np.isnan(cost.sum_epoch) or np.isinf(cost.sum_epoch):
logger.log(99, 'Cost is Nan')
return False, False
timeinfo = _calc_estimate_time(timeinfo, max_iteration,
last_iteration, iteration + 1)
callback.update_time_train(prediction=timeinfo.estimate_time)
if 0 < config.timelimit < timeinfo.estimate_time:
logger.log(
99,
'Expected training time ({:.3f}s) will exceed time limit ({}s).'
.format(timeinfo.estimate_time, config.timelimit))
return False, False
if (iteration +
1) % config.training_config.iter_per_epoch == 0:
last_past_time = -1
# End of epoch
epoch = iteration // config.training_config.iter_per_epoch + 1
cost_avg_epoch = cost.sum_epoch / cost.num_iteration if cost.num_iteration else 0
cost.sum_epoch = 0.0
cost.num_iteration = 0
monitoring_report = []
# Evaluation
error_str = ''
if epoch % config.training_config.monitor_interval == 0 or epoch <= 5:
best_error, error_str = _evaluate(
args, config, monitoring_report, best_error, epoch)
# Cpu/Gpu average load
cg_load_str = ''
cgload_log = ''
cg_load = get_cpu_gpu_average_load()
if cg_load:
cg_load_str = 'epoch {} average_load_matrix: {}'.format(
epoch, cg_load)
span = _calc_epoch_span(timeinfo)
if span > _CGLOAD_LOG_INTERVAL:
cgload_log = _format_cgload_log(cg_load)
if single_or_rankzero():
# Write to monitoring_report.yml
f = open(
os.path.join(args.outdir, 'monitoring_report.yml'),
'a')
f.write('{}:\n'.format(epoch - 1))
f.write(' cost: {}\n'.format(cost_avg_epoch))
for s in monitoring_report:
f.write(s)
f.close()
callback.update_status(
(['monitoring_report', epoch,
'cost'], cost_avg_epoch))
_save_parameters(args, 'current', epoch, config)
callback.update_status(('epoch.current', epoch))
callback.update_status()
logger.log(
99,
'epoch {} of {} cost={:.6f} {} time=({:.1f}s /{:.1f}s) {}'
.format(epoch, config.training_config.max_epoch,
cost_avg_epoch, error_str,
timeinfo.past_time, timeinfo.estimate_time,
cgload_log))
if cg_load_str:
# cpu_gpu_average_load record at epoch level
callback.update_status(
(['cpu_gpu_epoch_load', epoch], cg_load))
progress(cg_load_str, 1)
if not callback.check_training_time(
args, config, timeinfo, epoch, last_epoch):
_save_parameters(args, 'current', epoch, config,
True)
return False, True
if single_or_rankzero():
_save_parameters(args, 'current', epoch, config, True)
return True, False
def _evaluate(args, config, monitoring_report, best_error, epoch):
comm = current_communicator()
error_str = ''
valid_error = 0.0
def _sum_error(sum, error):
ret = None
if comm:
# logger.log(99, "Calc error with communicator")
var = [nn.NdArray()]
var[0].data = error
_all_reduce(comm, var, division=False, inplace=True)
ret = sum + var[0].data
else:
ret = sum + error
return ret
for name, mon in config.monitors.items():
m = mon.monitor
error_sum_monitor = 0.0
error_count = 0
data_size = max([di.size for di in mon.data_iterators])
batch_size = max([di.batch_size for di in mon.data_iterators])
for i in range(data_size // batch_size):
# Load dataset
data = OrderedDict()
for di in mon.data_iterators:
data.update(zip(di.variables, di.next()))
# Set data to variable
for v, d in m.dataset_assign.items():
dest_context = config.global_config.default_context if not m.forward_sequence or v not in m.forward_sequence[
0].inputs else None
let_data_to_variable(v.variable_instance,
data[d],
ctx=dest_context,
data_name=d,
variable_name=v.name)
# Generate data
for v, generator in m.generator_assign.items():
dest_context = config.global_config.default_context if not m.forward_sequence or v not in m.forward_sequence[
0].inputs else None
let_data_to_variable(v.variable_instance,
data=generator(v.shape),
ctx=dest_context,
variable_name=v.name)
# Sum error before forward to prepare input data while processing
# on GPU
if error_count > 0:
error_sum = 0.0
for v in m.monitor_variables:
error_sum += np.mean(v.variable_instance.d)
# v.variable_instance.data.zero()
error_sum_monitor = _sum_error(error_sum_monitor, error_sum)
if single_or_rankzero():
progress(
'Evaluating "{0}"'.format(name) +
' : error={0:0.6f}'.format(
error_sum_monitor / error_count),
di.position * 1.0 / di.size)
error_count += comm.size if comm else 1
# Forward recursive
m.network.forward(m.forward_sequence)
# Sum error at the end of dataset
error_sum = 0.0
for v in m.monitor_variables:
error_sum += np.mean(v.variable_instance.d)
# v.variable_instance.data.zero()
error_sum_monitor = _sum_error(error_sum_monitor, error_sum)
if error_count == 0:
error = 0
else:
error = error_sum_monitor / error_count
if np.isnan(error) or np.isinf(error):
logger.log(99, 'Validation error is Nan')
error = 0.0
monitoring_report.append(' {}: {}\n'.format(name, error))
callback.update_status((['monitoring_report', epoch, name], error))
callback.update_status((['last', name], error)) # save last value
if error_str != '':
error_str += ', '
else:
error_str = ' {'
error_str += '{}={:.6f}'.format(name, error)
if name == 'valid_error':
valid_error = error
if error_str != '':
error_str += '}'
# Save Parameters
if single_or_rankzero():
if (not config.training_config.save_best) or \
(not best_error) or \
(best_error is not None and valid_error <= best_error):
best_error = valid_error
callback.update_status(('best.valid_error', best_error))
callback.update_status(('best.epoch', epoch))
_save_parameters(args, 'best', epoch, config, True)
return best_error, error_str
def _update(iter, config, cost):
comm = current_communicator()
loaded_data = {}
is_first_optimizer = True
def _sum_cost():
if comm:
# logger.log(99, "Calc cost with communicator")
var = [nn.NdArray()]
var[0].data = cost.sum_iteration
_all_reduce(comm, var, division=False, inplace=True)
cost.sum_epoch += var[0].data
cost.num_iteration += comm.size
else:
cost.sum_epoch += cost.sum_iteration
cost.num_iteration += 1
def _get_reserved_variable(shape, reserved_variable_name, iter,
iter_per_epoch, max_epoch):
if reserved_variable_name == "%iter":
value = iter
elif reserved_variable_name == "%max_iter":
value = max_epoch * iter_per_epoch
elif reserved_variable_name == "%epoch":
value = iter // iter_per_epoch
elif reserved_variable_name == "%epochf":
value = iter * 1.0 / iter_per_epoch
elif reserved_variable_name == "%max_epoch":
value = max_epoch
elif reserved_variable_name == "%progress":
value = (iter * 1.0 / iter_per_epoch) / max_epoch
else:
raise ValueError(
"Unknown reserved variable {}".format(reserved_variable_name))
return value
for opt in config.optimizers.values():
o = opt.optimizer
if (o.start_iter == 0
or iter + 1 >= o.start_iter) and (o.end_iter == 0
or iter + 1 <= o.end_iter):
# Load dataset
data = OrderedDict()
for di in opt.data_iterators:
if di not in loaded_data:
loaded_data[di] = di.next()
data.update(zip(di.variables, loaded_data[di]))
for v, d in o.dataset_assign.items():
dest_context = config.global_config.default_context if not o.forward_sequence or v not in o.forward_sequence[
0].inputs else None
if d not in data and d[0] == "%":
value = _get_reserved_variable(
v.variable_instance.shape, d, iter,
config.training_config.iter_per_epoch,
config.training_config.max_epoch)
v.variable_instance.data.fill(value)
elif d in data:
let_data_to_variable(v.variable_instance,
data[d],
ctx=dest_context,
data_name=d,
variable_name=v.name)
else:
raise ValueError(
'Variable "{}" is not found in dataset "{}", optimizer "{}"'
.format(d, ', '.join(o.data_iterators.keys()), o.name))
# Generate data
for v, generator in o.generator_assign.items():
dest_context = config.global_config.default_context if not o.forward_sequence or v not in o.forward_sequence[
0].inputs else None
let_data_to_variable(v.variable_instance,
data=generator(v.shape),
ctx=dest_context,
variable_name=v.name)
# Monitor loss before forward to prepare input data while processing on
# GPU
if cost.variables:
for l in cost.variables:
cost.sum_iteration += np.mean(l.variable_instance.d)
# l.variable_instance.data.zero()
if is_first_optimizer:
is_first_optimizer = False
_sum_cost()
if single_or_rankzero():
progress(
"Training : cost={0:0.6f}".format(
cost.sum_iteration),
(iter % config.training_config.iter_per_epoch) *
1.0 / config.training_config.iter_per_epoch)
cost.sum_iteration = 0.0
with nodeTimeCollector.collect_cost_time(comm, iter):
# Forward
o.network.forward(o.forward_sequence)
# Backward
o.network.backward(o.backward_sequence,
iter % o.update_interval == 0)
# Update
if iter % o.update_interval == o.update_interval - 1:
if o.weight_decay > 0:
o.solver.weight_decay(o.weight_decay)
if o.comm: # Updated param with communicator
params = [x.grad for x in o.parameters.values()]
_all_reduce(o.comm, params, division=True, inplace=True)
if o.scheduler is not None:
o.solver.set_learning_rate(
o.scheduler.get_learning_rate(iter))
o.solver.update()
# Sync w sometimes
if iter % 10 == 9: # TODO: change the interval
if o.comm:
params = [x.data for x in o.parameters.values()]
_all_reduce(o.comm, params, division=True, inplace=True)
# Reserve monitor loss
cost.variables = o.loss_variables
# Monitor loss at the end of epoch
if iter % config.training_config.iter_per_epoch == config.training_config.iter_per_epoch - 1 and cost.variables:
for l in cost.variables:
cost.sum_iteration += np.mean(l.variable_instance.d)
# l.variable_instance.data.zero()
_sum_cost()
cost.variables = None
cost.sum_iteration = 0.0
return cost
def train_command(args):
if single_or_rankzero():
configure_progress(os.path.join(args.outdir, 'progress.txt'))
info = load.load([args.config], exclude_parameter=True)
# Check dataset uri is empty.
dataset_error = False
for dataset in info.datasets.values():
if dataset.uri.strip() == '':
dataset_error = True
if dataset_error:
logger.log(99, 'Fatal error. Dataset URI is empty.')
return False
class TrainConfig:
pass
config = TrainConfig()
config.timelimit = -1
if args.param:
load.load([args.param], parameter_only=True)
config.global_config = info.global_config
config.training_config = info.training_config
if single_or_rankzero():
logger.log(99, 'Train with contexts {}'.format(available_contexts))
class OptConfig:
pass
config.optimizers = OrderedDict()
for name, opt in info.optimizers.items():
o = OptConfig()
o.optimizer = opt
o.data_iterator = None
config.optimizers[name] = o
class MonConfig:
pass
config.monitors = OrderedDict()
for name, mon in info.monitors.items():
m = MonConfig()
m.monitor = mon
m.data_iterator = None
config.monitors[name] = m
# Training
comm = current_communicator()
config.training_config.iter_per_epoch //= comm.size if comm else 1
max_iteration = config.training_config.max_epoch * \
config.training_config.iter_per_epoch
global _save_parameter_info
_save_parameter_info = {}
_, config_ext = os.path.splitext(args.config)
if config_ext == '.prototxt' or config_ext == '.nntxt':
_save_parameter_info['config'] = args.config
elif config_ext == '.nnp':
with zipfile.ZipFile(args.config, 'r') as nnp:
for name in nnp.namelist():
_, ext = os.path.splitext(name)
if ext == '.nntxt' or ext == '.prototxt':
nnp.extract(name, args.outdir)
_save_parameter_info['config'] = os.path.join(
args.outdir, name)
result = False
if max_iteration > 0:
data_iterators = {'optimizer': {}, 'monitor': {}}
rng = np.random.RandomState(comm.rank if comm else 0)
with ExitStack() as stack:
for name, o in config.optimizers.items():
o.data_iterator = stack.enter_context(
o.optimizer.data_iterator())
if comm and comm.size > 1:
o.data_iterator = o.data_iterator.slice(
rng, comm.size, comm.rank)
for name, m in config.monitors.items():
m.data_iterator = stack.enter_context(
m.monitor.data_iterator())
if comm and comm.size > 1:
m.data_iterator = m.data_iterator.slice(
rng, comm.size, comm.rank)
result = _train(args, config)
else:
# save parameters without training (0 epoch learning)
logger.log(99, '0 epoch learning. (Just save parameter.)')
if single_or_rankzero():
_save_parameters(args, 'current', 0, True)
result = True
if single_or_rankzero():
if result:
logger.log(99, 'Training Completed.')
else:
logger.log(99, 'Training Incompleted.')
if single_or_rankzero():
progress(None)
return True
def _train(args, config):
global _save_parameter_info
comm = current_communicator()
last_epoch = 0
if args.resume:
last_epoch = _get_current_parameter(args)
logger.log(99, "Resume from epoch {}".format(last_epoch + 1))
max_iteration = config.training_config.max_epoch * \
config.training_config.iter_per_epoch
if single_or_rankzero():
logger.log(
99, 'Training epoch {} of {} begin'.format(
last_epoch + 1, config.training_config.max_epoch))
class Cost:
pass
cost = Cost()
cost.sum_epoch = 0.0
cost.num_iteration = 0
cost.sum_iteration = 0.0
cost.variables = None
best_error = None
class TimeInfo:
pass
timeinfo = TimeInfo()
timeinfo.last_past_time = None
if max_iteration > 0:
last_iteration = last_epoch * config.training_config.iter_per_epoch
if last_iteration < max_iteration:
timeinfo.start_time = time.time()
for iteration in range(last_iteration, max_iteration):
cost = _update(iteration, config, cost)
if (iteration - last_iteration) > 0:
timeinfo = _calc_estimate_time(timeinfo, max_iteration,
last_iteration, iteration)
if config.timelimit > 0 and timeinfo.estimate_time > config.timelimit:
logger.log(
99,
'Expected training time ({:.3f}s) will exceed time limit ({}s).'
.format(timeinfo.estimate_time, config.timelimit))
return False
if (iteration +
1) % config.training_config.iter_per_epoch == 0:
last_past_time = -1
# End of epoch
epoch = iteration // config.training_config.iter_per_epoch + 1
cost_avg_epoch = cost.sum_epoch / cost.num_iteration
cost.sum_epoch = 0.0
cost.num_iteration = 0
monitoring_report = []
# Evaluation
error_str = ''
if epoch % config.training_config.monitor_interval == 0 or epoch <= 5:
best_error, error_str = _evaluate(
args, config, monitoring_report, best_error, epoch)
if single_or_rankzero():
# Write to monitoring_report.yml
f = open(
os.path.join(args.outdir, 'monitoring_report.yml'),
'a')
f.write('{}:\n'.format(epoch - 1))
f.write(' cost: {}\n'.format(cost_avg_epoch))
for s in monitoring_report:
f.write(s)
f.close()
_save_parameters(args, 'current', epoch)
logger.log(
99,
'epoch {} of {} cost={:.6f} {} time=({:.1f}s /{:.1f}s)'
.format(epoch, config.training_config.max_epoch,
cost_avg_epoch, error_str,
timeinfo.past_time,
timeinfo.estimate_time))
if single_or_rankzero():
_save_parameters(args, 'current', epoch, True)
return True
def _update(iter, config, cost):
comm = current_communicator()
loaded_data = {}
is_first_optimizer = True
def _sum_cost():
if comm:
# logger.log(99, "Calc cost with communicator")
var = [nn.NdArray()]
var[0].data = cost.sum_iteration
_all_reduce(comm, var, division=False, inplace=True)
cost.sum_epoch += var[0].data
cost.num_iteration += comm.size
else:
cost.sum_epoch += cost.sum_iteration
cost.num_iteration += 1
for opt in config.optimizers.values():
o = opt.optimizer
# Load dataset
di = opt.data_iterator
if o.data_iterator not in loaded_data:
loaded_data[o.data_iterator] = di.next()
data = loaded_data[o.data_iterator]
for v, d in o.dataset_assign.items():
dest_context = config.global_config.default_context if not o.forward_sequence or v not in o.forward_sequence[
0].inputs else None
let_data_to_variable(v.variable_instance,
data[di.variables.index(d)],
ctx=dest_context,
data_name=d,
variable_name=v.name)
# Generate data
for v, generator in o.generator_assign.items():
dest_context = config.global_config.default_context if not o.forward_sequence or v not in o.forward_sequence[
0].inputs else None
let_data_to_variable(v.variable_instance,
data=generator(v.shape),
ctx=dest_context,
variable_name=v.name)
# Monitor loss before forward to prepare input data while processing on
# GPU
if cost.variables:
for l in cost.variables:
cost.sum_iteration += np.mean(l.variable_instance.d)
l.variable_instance.data.zero()
if is_first_optimizer:
is_first_optimizer = False
_sum_cost()
if single_or_rankzero():
progress(
"Training : cost={0:0.6f}".format(cost.sum_iteration),
(iter % config.training_config.iter_per_epoch) * 1.0 /
config.training_config.iter_per_epoch)
cost.sum_iteration = 0.0
# Forward
o.network.forward(o.forward_sequence)
# Backward
o.network.backward(o.backward_sequence, iter % o.update_interval == 0)
# Update
if iter % o.update_interval == o.update_interval - 1:
if o.weight_decay > 0:
o.solver.weight_decay(o.weight_decay)
if o.comm: # Updated param with communicator
params = [x.grad for x in o.parameters.values()]
_all_reduce(o.comm, params, division=True, inplace=True)
if o.scheduler is not None:
o.solver.set_learning_rate(o.scheduler.get_learning_rate(iter))
o.solver.update()
# Sync w sometimes
if iter % 10 == 9: # TODO: change the interval
if o.comm:
params = [x.data for x in o.parameters.values()]
_all_reduce(o.comm, params, division=True, inplace=True)
# Reserve monitor loss
cost.variables = o.loss_variables
# Monitor loss at the end of iteration
if iter % config.training_config.iter_per_epoch == config.training_config.iter_per_epoch - 1 and cost.variables:
for l in cost.variables:
cost.sum_iteration += np.mean(l.variable_instance.d)
l.variable_instance.data.zero()
_sum_cost()
cost.variables = None
cost.sum_iteration = 0.0
return cost
def _create_cache(self):
# Save all data into cache file(s).
self._cache_positions = []
self._position = 0
percent = 0
if single_or_rankzero():
progress(None)
while self._position < self._data_source._size:
if single_or_rankzero():
progress('Create cache',
self._position * 1.0 / self._data_source._size)
self._store_data_to_cache_buffer(self._position)
self._position += 1
if len(self._cache_positions) > 0:
self._save_cache_to_file()
if single_or_rankzero():
progress(None)
# Adjust data size into reseted position. In most case it means
# multiple of bunch(mini-batch) size.
num_of_cache_files = int(
numpy.ceil(float(self._data_source._size) / self._cache_size))
self._cache_file_order = self._cache_file_order[0:num_of_cache_files]
self._cache_file_data_orders = self._cache_file_data_orders[
0:num_of_cache_files]
if self._data_source._size % self._cache_size != 0:
self._cache_file_data_orders[num_of_cache_files -
1] = self._cache_file_data_orders[
num_of_cache_files -
1][0:self._data_source._size %
self._cache_size]
# Create Index
index_filename = os.path.join(self._cache_dir, "cache_index.csv")
with open(index_filename, 'w') as f:
writer = csv.writer(f, lineterminator='\n')
for fn, orders in zip(self._cache_file_names,
self._cache_file_data_orders):
writer.writerow((os.path.basename(fn), len(orders)))
# Create Info
if self._cache_file_format == ".npy":
info_filename = os.path.join(self._cache_dir, "cache_info.csv")
with open(info_filename, 'w') as f:
writer = csv.writer(f, lineterminator='\n')
for variable in self._variables:
writer.writerow((variable, ))
# Create original.csv
if self._data_source._original_source_uri is not None:
fr = FileReader(self._data_source._original_source_uri)
with fr.open() as f:
csv_lines = [x.decode('utf-8') for x in f.readlines()]
with open(os.path.join(self._cache_dir, "original.csv"),
'w') as o:
for l in csv_lines:
o.write(l)
# Create order.csv
if self._data_source._order is not None and \
self._data_source._original_order is not None:
with open(os.path.join(self._cache_dir, "order.csv"), 'w') as o:
writer = csv.writer(o, lineterminator='\n')
for orders in zip(self._data_source._original_order,
self._data_source._order):
writer.writerow(list(orders))
def _save_cache(self, args):
position = args[0]
cache_csv = args[1]
# conv dataset
cache_data = [tuple(self._process_row(row)) for row in cache_csv]
start_position = position + 1 - len(cache_data)
end_position = position
cache_filename = os.path.join(
self._cache_dir, '{}_{:08d}_{:08d}{}'.format(self._cache_file_name_prefix,
start_position,
end_position,
self._cache_file_format))
logger.info('Creating cache file {}'.format(cache_filename))
data = collections.OrderedDict(
[(n, []) for n in self._variables])
for _, cd in enumerate(cache_data):
for i, n in enumerate(self._variables):
if isinstance(cd[i], numpy.ndarray):
d = cd[i]
else:
d = numpy.array(cd[i]).astype(numpy.float32)
data[n].append(d)
try:
if self._cache_file_format == ".h5":
h5 = h5py.File(cache_filename, 'w')
for k, v in data.items():
h5.create_dataset(k, data=v)
h5.close()
else:
retry_count = 1
is_create_cache_incomplete = True
while is_create_cache_incomplete:
try:
with open(cache_filename, 'wb') as f:
for v in data.values():
numpy.save(f, v)
is_create_cache_incomplete = False
except OSError:
retry_count += 1
if retry_count > 10:
raise
logger.info(
'Creating cache retry {}/10'.format(retry_count))
except:
logger.critical(
'An error occurred while creating cache file from dataset.')
for k, v in data.items():
size = v[0].shape
for d in v:
if size != d.shape:
logger.critical('The sizes of data "{}" are not the same. ({} != {})'.format(
k, size, d.shape))
raise
self.current_cache_position += 1
if single_or_rankzero():
if self.current_cache_position % int(self.num_of_cache_file/20+1) == 0:
progress('Create cache', self.current_cache_position /
self.num_of_cache_file)
return cache_filename, len(cache_data)
def create(self, output_cache_dirname, normalize=True, cache_file_name_prefix='cache'):
self._normalize = normalize
self._cache_file_name_prefix = cache_file_name_prefix
self._cache_dir = output_cache_dirname
self._cache_file_format = nnabla_config.get(
'DATA_ITERATOR', 'cache_file_format')
logger.info('Cache file format is {}'.format(self._cache_file_format))
progress(None)
csv_position_and_data = []
csv_row = []
for _position in range(self._size):
csv_row.append(self._csv_data[self._order[_position]])
if len(csv_row) == self._cache_size:
csv_position_and_data.append((_position, csv_row))
csv_row = []
if len(csv_row):
csv_position_and_data.append((self._size-1, csv_row))
self.num_of_cache_file = len(csv_position_and_data)
self.current_cache_position = 0
if single_or_rankzero():
progress('Create cache', 0)
with closing(ThreadPool(processes=self._num_of_threads)) as pool:
cache_index_rows = pool.map(
self._save_cache, csv_position_and_data)
if single_or_rankzero():
progress('Create cache', 1.0)
# Create Index
index_filename = os.path.join(output_cache_dirname, "cache_index.csv")
with open(index_filename, 'w') as f:
writer = csv.writer(f, lineterminator='\n')
for row in cache_index_rows:
if row:
# row: (file_path, data_nums)
writer.writerow((os.path.basename(row[0]), row[1]))
# Create Info
if self._cache_file_format == ".npy":
info_filename = os.path.join(
output_cache_dirname, "cache_info.csv")
with open(info_filename, 'w') as f:
writer = csv.writer(f, lineterminator='\n')
for variable in self._variables:
writer.writerow((variable, ))
# Create original.csv
if self._original_source_uri is not None:
shutil.copy(self._original_source_uri, os.path.join(
output_cache_dirname, "original.csv"))
# Create order.csv
if self._order is not None and \
self._original_order is not None:
with open(os.path.join(output_cache_dirname, "order.csv"), 'w') as o:
writer = csv.writer(o, lineterminator='\n')
for orders in zip(self._original_order, self._order):
writer.writerow(list(orders))
def train(args):
"""
Multi-Device Training
NOTE: the communicator exposes low-level interfaces
Steps:
* Instantiate a communicator and set parameter variables.
* Specify contexts for computation.
* Initialize DataIterator.
* Construct a computation graph for training and one for validation.
* Initialize solver and set parameter variables to that.
* Load checkpoint to resume previous training.
* Create monitor instances for saving and displaying training stats.
* Training loop
* Computate error rate for validation data (periodically)
* Get a next minibatch.
* Execute forwardprop
* Set parameter gradients zero
* Execute backprop.
* AllReduce for gradients
* Solver updates parameters by using gradients computed by backprop and all reduce.
* Compute training error
"""
# Create Communicator and Context
comm = create_communicator(ignore_error=True)
if comm:
n_devices = comm.size
mpi_rank = comm.rank
device_id = comm.local_rank
else:
n_devices = 1
mpi_rank = 0
device_id = args.device_id
if args.context == 'cpu':
import nnabla_ext.cpu
context = nnabla_ext.cpu.context()
else:
import nnabla_ext.cudnn
context = nnabla_ext.cudnn.context(device_id=device_id)
nn.set_default_context(context)
n_train_samples = 50000
n_valid_samples = 10000
bs_valid = args.batch_size
iter_per_epoch = int(n_train_samples / args.batch_size / n_devices)
# Model
rng = np.random.RandomState(313)
comm_syncbn = comm if args.sync_bn else None
if args.net == "cifar10_resnet23":
prediction = functools.partial(resnet23_prediction,
rng=rng,
ncls=10,
nmaps=64,
act=F.relu,
comm=comm_syncbn)
data_iterator = data_iterator_cifar10
if args.net == "cifar100_resnet23":
prediction = functools.partial(resnet23_prediction,
rng=rng,
ncls=100,
nmaps=384,
act=F.elu,
comm=comm_syncbn)
data_iterator = data_iterator_cifar100
# Create training graphs
image_train = nn.Variable((args.batch_size, 3, 32, 32))
label_train = nn.Variable((args.batch_size, 1))
pred_train = prediction(image_train, test=False)
pred_train.persistent = True
loss_train = (loss_function(pred_train, label_train) /
n_devices).apply(persistent=True)
error_train = F.mean(F.top_n_error(pred_train, label_train,
axis=1)).apply(persistent=True)
loss_error_train = F.sink(loss_train, error_train)
# Create validation graphs
image_valid = nn.Variable((bs_valid, 3, 32, 32))
label_valid = nn.Variable((bs_valid, 1))
pred_valid = prediction(image_valid, test=True)
error_valid = F.mean(F.top_n_error(pred_valid, label_valid, axis=1))
# Solvers
solver = S.Adam()
solver.set_parameters(nn.get_parameters())
base_lr = args.learning_rate
warmup_iter = iter_per_epoch * args.warmup_epoch
warmup_slope = base_lr * (n_devices - 1) / warmup_iter
solver.set_learning_rate(base_lr)
# load checkpoint if file exist.
start_point = 0
if args.use_latest_checkpoint:
files = glob.glob(f'{args.model_save_path}/checkpoint_*.json')
if len(files) != 0:
index = max([
int(n) for n in
[re.sub(r'.*checkpoint_(\d+).json', '\\1', f) for f in files]
])
# load weights and solver state info from specified checkpoint file.
start_point = load_checkpoint(
f'{args.model_save_path}/checkpoint_{index}.json', solver)
print(f'checkpoint is loaded. start iteration from {start_point}')
# Create monitor
monitor = Monitor(args.monitor_path)
monitor_loss = MonitorSeries("Training loss", monitor, interval=10)
monitor_err = MonitorSeries("Training error", monitor, interval=10)
monitor_time = MonitorTimeElapsed("Training time", monitor, interval=10)
monitor_verr = MonitorSeries("Validation error", monitor, interval=1)
monitor_vtime = MonitorTimeElapsed("Validation time", monitor, interval=1)
# Data Iterator
# If the data does not exist, it will try to download it from the server
# and prepare it. When executing multiple processes on the same host, it is
# necessary to execute initial data preparation by the representative
# process (rank is 0) on the host.
# Download dataset by rank-0 process
if single_or_rankzero():
rng = np.random.RandomState(mpi_rank)
_, tdata = data_iterator(args.batch_size, True, rng)
vsource, vdata = data_iterator(bs_valid, False)
# Wait for data to be prepared without watchdog
if comm:
comm.barrier()
# Prepare dataset for remaining process
if not single_or_rankzero():
rng = np.random.RandomState(mpi_rank)
_, tdata = data_iterator(args.batch_size, True, rng)
vsource, vdata = data_iterator(bs_valid, False)
# Training-loop
ve = nn.Variable()
for i in range(start_point // n_devices, args.epochs * iter_per_epoch):
# Validation
if i % iter_per_epoch == 0:
ve_local = 0.
k = 0
idx = np.random.permutation(n_valid_samples)
val_images = vsource.images[idx]
val_labels = vsource.labels[idx]
for j in range(int(n_valid_samples / n_devices * mpi_rank),
int(n_valid_samples / n_devices * (mpi_rank + 1)),
bs_valid):
image = val_images[j:j + bs_valid]
label = val_labels[j:j + bs_valid]
if len(image
) != bs_valid: # note that smaller batch is ignored
continue
image_valid.d = image
label_valid.d = label
error_valid.forward(clear_buffer=True)
ve_local += error_valid.d.copy()
k += 1
ve_local /= k
ve.d = ve_local
if comm:
comm.all_reduce(ve.data, division=True, inplace=True)
# Monitoring error and elapsed time
if single_or_rankzero():
monitor_verr.add(i * n_devices, ve.d.copy())
monitor_vtime.add(i * n_devices)
# Save model
if single_or_rankzero():
if i % (args.model_save_interval // n_devices) == 0:
iter = i * n_devices
nn.save_parameters(
os.path.join(args.model_save_path,
'params_%06d.h5' % iter))
if args.use_latest_checkpoint:
save_checkpoint(args.model_save_path, iter, solver)
# Forward/Zerograd
image, label = tdata.next()
image_train.d = image
label_train.d = label
loss_error_train.forward(clear_no_need_grad=True)
solver.zero_grad()
# Backward/AllReduce
backward_and_all_reduce(
loss_error_train,
comm,
with_all_reduce_callback=args.with_all_reduce_callback)
# Solvers update
solver.update()
# Linear Warmup
if i <= warmup_iter:
lr = base_lr + warmup_slope * i
solver.set_learning_rate(lr)
# Monitoring loss, error and elapsed time
if single_or_rankzero():
monitor_loss.add(i * n_devices, loss_train.d.copy())
monitor_err.add(i * n_devices, error_train.d.copy())
monitor_time.add(i * n_devices)
# Save nnp last epoch
if single_or_rankzero():
runtime_contents = {
'networks': [{
'name': 'Validation',
'batch_size': args.batch_size,
'outputs': {
'y': pred_valid
},
'names': {
'x': image_valid
}
}],
'executors': [{
'name': 'Runtime',
'network': 'Validation',
'data': ['x'],
'output': ['y']
}]
}
iter = args.epochs * iter_per_epoch
nn.save_parameters(
os.path.join(args.model_save_path, 'params_%06d.h5' % iter))
nnabla.utils.save.save(
os.path.join(args.model_save_path, f'{args.net}_result.nnp'),
runtime_contents)
if comm:
comm.barrier()
