def train_model(opt):
"""
Run through a TrainLoop.
If model_file is not in opt, then this helper will create a temporary
directory to store the model, dict, etc.
:return: (stdout, valid_results, test_results)
:rtype: (str, dict, dict)
"""
import parlai.scripts.train_model as tms
with capture_output() as output:
with tempdir() as tmpdir:
if 'model_file' not in opt:
opt['model_file'] = os.path.join(tmpdir, 'model')
if 'dict_file' not in opt:
opt['dict_file'] = os.path.join(tmpdir, 'model.dict')
parser = tms.setup_args()
# needed at the very least to set the overrides.
parser.set_params(**opt)
parser.set_params(log_every_n_secs=10)
popt = parser.parse_args(print_args=False)
# in some rare cases, like for instance if the model class also
# overrides its default params, the params override will not
# be taken into account.
for k, v in opt.items():
popt[k] = v
tl = tms.TrainLoop(popt)
valid, test = tl.train()
return (output.getvalue(), valid, test)
def get_tl(tmpdir):
final_opt = Opt({
'task': 'integration_tests',
'datatype': 'valid',
'validation_max_exs': 30,
'short_final_eval': True,
})
final_opt.save(os.path.join(tmpdir, "final_opt.opt"))
opt = Opt({
'task':
'integration_tests',
'validation_max_exs':
10,
'model':
'repeat_label',
'model_file':
os.path.join(tmpdir, 'model'),
'short_final_eval':
True,
'num_epochs':
1.0,
'final_extra_opt':
str(os.path.join(tmpdir, "final_opt.opt")),
})
parser = tms.setup_args()
parser.set_params(**opt)
popt = parser.parse_args([])
for k, v in opt.items():
popt[k] = v
return tms.TrainLoop(popt)
def get_popt_and_tl(opt):
parser = tms.setup_args()
parser.set_params(**opt)
popt = parser.parse_args([], print_args=False)
for k, v in opt.items():
popt[k] = v
return popt, tms.TrainLoop(popt)
def train_model(opt):
"""
Runs through a TrainLoop.
If model_file is not in opt, then this helper will create a temporary
directory to store the model, dict, etc.
:return: (stdout, stderr, valid_results, test_results)
:rtype: (str, str, dict, dict)
"""
import parlai.scripts.train_model as tms
with capture_output() as output:
with tempdir() as tmpdir:
if 'model_file' not in opt:
opt['model_file'] = os.path.join(tmpdir, 'model')
if 'dict_file' not in opt:
opt['dict_file'] = os.path.join(tmpdir, 'model.dict')
parser = tms.setup_args()
parser.set_params(**opt)
popt = parser.parse_args(print_args=False)
tl = tms.TrainLoop(popt)
valid, test = tl.train()
return (
output.getvalue(),
valid,
test,
)
def run(self):
with distributed_utils.slurm_distributed_context(self.opt) as opt:
self.train_loop = single_train.TrainLoop(opt)
self.parser = self.parser
self.parser.opt = self.train_loop.agent.opt
self.parser.print_args()
return self.train_loop.train()
def multiprocess_train(
rank, opt, port=61337, rank_offset=0, gpu=None, hostname='localhost'
):
"""
Subprocess which initializes distributed training, and begins training.
This should be launched n times for n GPUs; this is handled either in main
or via srun.
:param int rank: This process's rank - 1. (Starts at -1 ... n - 2). See comments.
:param opt: command line options
:param int port: A TCP port to use. This will need to be changed to run
multiple distributed training setups on the same machine.
:param int gpu: Which GPU to use. Defaults to using rank and local devices,
but must be manually specified when using many-hosts.
:param str hostname: Hostname of the main server.
"""
# Set per-host options
opt = copy.deepcopy(opt)
# we need to manually adjust the rank differently in multiprocessing
# and distributed train
rank = rank + rank_offset
opt['rank'] = rank
if gpu is None:
# default assumption is local GPUs
gpu = rank % torch.cuda.device_count()
opt['gpu'] = gpu
# make sure we don't just use whatever GPU was saved in the model file
if 'override' not in opt:
opt['override'] = {}
opt['override']['gpu'] = gpu
# Suppress output of workers except the main host.
if opt.get('verbose') or rank != 0:
print_prefix = '[rank:{:3d}]'.format(rank)
else:
print_prefix = None
suppress_output = not opt.get('verbose') and rank != 0
with distributed_utils.override_print(suppress_output, print_prefix):
# perform distributed setup, ensuring all hosts are ready
if opt['gpu'] != -1:
torch.cuda.set_device(opt['gpu'])
dist.init_process_group(
backend="nccl",
init_method="tcp://{}:{}".format(hostname, port),
world_size=opt['distributed_world_size'],
rank=rank,
)
logging.info("Distributed group initialized")
# manual_seed can be a noop without this
torch.cuda.init()
# make sure all parameters will be in sync
torch.manual_seed(42)
# force a sync so that no one gets ahead, and all are seeded together
distributed_utils.sync_object(None)
# Run the actual training
return single_train.TrainLoop(opt).train()
def multiprocess_train(
rank, opt, port=61337, rank_offset=0, gpu=None, hostname='localhost'
):
with distributed_utils.distributed_context(
rank, opt, port, rank_offset, gpu, hostname
) as opt:
# Run the actual training
opt['multiprocessing'] = True
return single_train.TrainLoop(opt).train()
def test_train_model_with_no_dict_file(self):
"""Ensure training a model requires a dict_file or model_file."""
import parlai.scripts.train_model as tms
with testing_utils.capture_output():
parser = tms.setup_args()
parser.set_params(task='babi:task1k:1', model='seq2seq')
popt = parser.parse_args(print_args=False)
with self.assertRaises(RuntimeError):
tms.TrainLoop(popt)
def test_train_model_with_no_dict_file(self):
"""Check that attempting to train a model without specifying a dict_file
or model_file fails
"""
import parlai.scripts.train_model as tms
with testing_utils.capture_output():
parser = tms.setup_args()
parser.set_params(task='babi:task1k:1', model='seq2seq')
popt = parser.parse_args(print_args=False)
with self.assertRaises(RuntimeError):
tms.TrainLoop(popt)
def multiprocess_train(rank, opt, port=61337, gpu=None, hostname='localhost'):
"""
Subprocess which initializes distributed training, and begins training.
This should be launched n times for n GPUs; this is handled either in main
or via srun.
:param int rank: This process's rank
:param opt: command line options
:param int port: A TCP port to use. This will need to be changed to run
multiple distributed training setups on the same machine.
:param int gpu: Which GPU to use. Defaults to using rank and local devices,
but must be manually specified when using many-hosts.
:param str hostname: Hostname of the main server.
"""
# Set per-host options
opt = copy.deepcopy(opt)
opt['rank'] = rank
if gpu is None:
# default assumption is local GPUs
gpu = rank % torch.cuda.device_count()
opt['gpu'] = gpu
# make sure we don't just use whatever GPU was saved in the model file
if 'override' not in opt:
opt['override'] = {}
opt['override']['gpu'] = gpu
# Suppress output of workers except the main host.
if opt.get('verbose') or rank != 0:
print_prefix = '[rank:{:2d}]'.format(rank)
else:
print_prefix = None
distributed_utils.override_print(
suppress=(not opt.get('verbose') and rank != 0),
prefix=print_prefix
)
# perform distributed setup, ensuring all hosts are ready
torch.cuda.set_device(opt['gpu'])
dist.init_process_group(
backend="nccl",
init_method="tcp://{}:{}".format(hostname, port),
world_size=opt['distributed_world_size'],
rank=rank,
)
print("Distributed group initialized")
# Run the actual training
return single_train.TrainLoop(opt).train()
def get_tl(tmpdir):
opt = {
'task': 'integration_tests',
'model': 'parlai.agents.test_agents.test_agents:MockTrainUpdatesAgent',
'model_file': os.path.join(tmpdir, 'model'),
'dict_file': os.path.join(tmpdir, 'model.dict'),
# step opts
'max_train_steps': num_train_steps,
'validation_every_n_steps': int(num_train_steps / num_validations),
'log_every_n_steps': int(num_train_steps / num_logs),
'update_freq': update_freq,
}
parser = tms.setup_args()
parser.set_params(**opt)
popt = parser.parse_args([])
for k, v in opt.items():
popt[k] = v
return tms.TrainLoop(popt)
def multiprocess_train(rank,
opt,
port=61337,
rank_offset=0,
gpu=None,
hostname='localhost'):
init_method = f"tcp://{hostname}:{port}"
with distributed_utils.distributed_context(rank,
opt,
rank_offset,
gpu,
init_method=init_method) as opt:
# Run the actual training
opt['multiprocessing'] = True
try:
return single_train.TrainLoop(opt).train()
except Exception:
import parlai.utils.logging as logging
logging.critical(traceback.format_exc())
logging.critical(
f"Got the above exception on worker {rank + rank_offset}. "
"This may cause hangs requiring manual killing of processes.")
raise
def run(self):
with distributed_utils.slurm_distributed_context(self.opt) as opt:
self.train_loop = single_train.TrainLoop(opt)
return self.train_loop.train()
<<<<<<< HEAD
popt = parser.parse_args([], print_args=False)
=======
popt = parser.parse_args(print_args=False)
>>>>>>> 4f6b99642d60aff1a41b9eae8bd2ccd9e40ebba4
>>>>>>> origin/master
=======
popt = parser.parse_args(print_args=False)
>>>>>>> 4f6b99642d60aff1a41b9eae8bd2ccd9e40ebba4
>>>>>>> ef574cebef2a8d5aa38b73176b1e71a919d6670f
# in some rare cases, like for instance if the model class also
# overrides its default params, the params override will not
# be taken into account.
for k, v in opt.items():
popt[k] = v
tl = tms.TrainLoop(popt)
valid, test = tl.train()
return (output.getvalue(), valid, test)
=======
with tempdir() as tmpdir:
if 'model_file' not in opt:
opt['model_file'] = os.path.join(tmpdir, 'model')
if 'dict_file' not in opt:
opt['dict_file'] = os.path.join(tmpdir, 'model.dict')
parser = tms.setup_args()
# needed at the very least to set the overrides.
parser.set_params(**opt)
parser.set_params(log_every_n_secs=10)
popt = parser.parse_args([], print_args=False)
# in some rare cases, like for instance if the model class also
