def re_use_some(tri_dict, suf):
ernie_config = tri_dict["ernie_config"]
startup_prog = tri_dict["startup_prog"]
args = tri_dict["args"]
labels_map = tri_dict["labels_map"]
place = tri_dict["place"]
reader = task_reader.RoleSequenceLabelReader(
vocab_path=args.vocab_path,
labels_map=labels_map,
max_seq_len=args.max_seq_len,
do_lower_case=args.do_lower_case,
in_tokens=args.in_tokens,
random_seed=args.random_seed,
task_id=args.task_id)
startup_prog = fluid.Program()
if args.random_seed is not None:
startup_prog.random_seed = args.random_seed
if args.do_val or args.do_test:
test_prog = fluid.Program()
with fluid.program_guard(test_prog, startup_prog):
with fluid.unique_name.guard():
# TODO pyreader_name 再次调整为不同
test_pyreader, graph_vars = create_model(
args,
pyreader_name='test_reader_role' + suf,
ernie_config=ernie_config)
test_prog = test_prog.clone(for_test=True)
nccl2_num_trainers = 1
nccl2_trainer_id = 0
exe = fluid.Executor(place)
exe.run(startup_prog)
if args.do_val or args.do_test:
if not args.init_checkpoint:
raise ValueError("args 'init_checkpoint' should be set if"
"only doing validation or testing!")
init_checkpoint(exe,
args.init_checkpoint,
main_program=startup_prog,
use_fp16=args.use_fp16)
return reader, startup_prog, test_prog, test_pyreader, graph_vars, exe
def get_role_init_dict(args, suf):
"""main"""
# log = logging.getLogger()
# prepare_logger(log)
log = logging.getLogger(__name__)
check_cuda(args.use_cuda)
labels_map = {} # label
for line in utils.read_by_lines(args.label_map_config):
arr = line.split("\t")
labels_map[arr[0]] = int(arr[1])
args.num_labels = len(labels_map)
print("=========ERNIE CONFIG============")
ernie_config = ErnieConfig(args.ernie_config_path)
# ernie_config.print_config()
print("=========ERNIE CONFIG============")
if args.use_cuda:
dev_list = fluid.cuda_places()
place = dev_list[0]
print("==============place==================", place)
# place = dev_list[1]
dev_count = len(dev_list)
else:
place = fluid.CPUPlace()
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
print("==============place, dev_count==================", place, dev_count)
reader = task_reader.RoleSequenceLabelReader(
vocab_path=args.vocab_path,
labels_map=labels_map,
max_seq_len=args.max_seq_len,
do_lower_case=args.do_lower_case,
in_tokens=args.in_tokens,
random_seed=args.random_seed,
task_id=args.task_id)
if not (args.do_train or args.do_val or args.do_test):
raise ValueError("For args `do_train`, `do_val` and `do_test`, at "
"least one of them must be True.")
startup_prog = fluid.Program()
if args.random_seed is not None:
startup_prog.random_seed = args.random_seed
if args.do_val or args.do_test:
test_prog = fluid.Program()
with fluid.program_guard(test_prog, startup_prog):
with fluid.unique_name.guard():
# TODO pyreader_name 再次调整为不同
test_pyreader, graph_vars = create_model(
args,
pyreader_name='test_reader_role' + suf,
ernie_config=ernie_config)
test_prog = test_prog.clone(for_test=True)
nccl2_num_trainers = 1
nccl2_trainer_id = 0
exe = fluid.Executor(place)
exe.run(startup_prog)
if args.do_val or args.do_test:
if not args.init_checkpoint:
raise ValueError("args 'init_checkpoint' should be set if"
"only doing validation or testing!")
init_checkpoint(exe,
args.init_checkpoint,
main_program=startup_prog,
use_fp16=args.use_fp16)
trigger_dict = dict()
trigger_dict['log'] = log
trigger_dict['args'] = args
trigger_dict['labels_map'] = labels_map
trigger_dict['ernie_config'] = ernie_config
trigger_dict['place'] = place
trigger_dict['dev_count'] = dev_count
trigger_dict['reader'] = reader
trigger_dict['startup_prog'] = startup_prog
trigger_dict['test_prog'] = test_prog
trigger_dict['test_pyreader'] = test_pyreader
trigger_dict['graph_vars'] = graph_vars
trigger_dict['nccl2_num_trainers'] = nccl2_num_trainers
trigger_dict['nccl2_trainer_id'] = nccl2_trainer_id
trigger_dict['exe'] = exe
return trigger_dict
def main(args):
ernie_config = ErnieConfig(args.ernie_config_path)
ernie_config.print_config()
if args.use_cuda:
place = fluid.CUDAPlace(int(os.getenv('FLAGS_selected_gpus', '0')))
dev_count = fluid.core.get_cuda_device_count()
else:
place = fluid.CPUPlace()
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
exe = fluid.Executor(place)
reader = task_reader.SequenceLabelReader(
vocab_path=args.vocab_path,
label_map_config=args.label_map_config,
max_seq_len=args.max_seq_len,
do_lower_case=args.do_lower_case,
in_tokens=args.in_tokens,
random_seed=args.random_seed)
if not (args.do_train or args.do_val or args.do_test):
raise ValueError("For args `do_train`, `do_val` and `do_test`, at "
"least one of them must be True.")
startup_prog = fluid.Program()
if args.random_seed is not None:
startup_prog.random_seed = args.random_seed
if args.do_train:
train_data_generator = reader.data_generator(
input_file=args.train_set,
batch_size=args.batch_size,
epoch=args.epoch,
shuffle=True,
phase="train")
num_train_examples = reader.get_num_examples(args.train_set)
if args.in_tokens:
max_train_steps = args.epoch * num_train_examples // (
args.batch_size // args.max_seq_len) // dev_count
else:
max_train_steps = args.epoch * num_train_examples // args.batch_size // dev_count
warmup_steps = int(max_train_steps * args.warmup_proportion)
print("Device count: %d" % dev_count)
print("Num train examples: %d" % num_train_examples)
print("Max train steps: %d" % max_train_steps)
print("Num warmup steps: %d" % warmup_steps)
train_program = fluid.Program()
with fluid.program_guard(train_program, startup_prog):
with fluid.unique_name.guard():
train_pyreader, graph_vars = create_model(
args,
pyreader_name='train_reader',
ernie_config=ernie_config)
scheduled_lr = optimization(loss=graph_vars["loss"],
warmup_steps=warmup_steps,
num_train_steps=max_train_steps,
learning_rate=args.learning_rate,
train_program=train_program,
startup_prog=startup_prog,
weight_decay=args.weight_decay,
scheduler=args.lr_scheduler,
use_fp16=args.use_fp16,
loss_scaling=args.loss_scaling)
fluid.memory_optimize(input_program=train_program,
skip_opt_set=[
graph_vars["loss"].name,
graph_vars["labels"].name,
graph_vars["infers"].name,
graph_vars["seq_lens"].name
])
if args.verbose:
if args.in_tokens:
lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
program=train_program,
batch_size=args.batch_size // args.max_seq_len)
else:
lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
program=train_program, batch_size=args.batch_size)
print("Theoretical memory usage in training: %.3f - %.3f %s" %
(lower_mem, upper_mem, unit))
if args.do_val or args.do_test:
test_prog = fluid.Program()
with fluid.program_guard(test_prog, startup_prog):
with fluid.unique_name.guard():
test_pyreader, graph_vars = create_model(
args,
pyreader_name='test_reader',
ernie_config=ernie_config)
test_prog = test_prog.clone(for_test=True)
exe.run(startup_prog)
if args.do_train:
if args.init_checkpoint and args.init_pretraining_params:
print(
"WARNING: args 'init_checkpoint' and 'init_pretraining_params' "
"both are set! Only arg 'init_checkpoint' is made valid.")
if args.init_checkpoint:
init_checkpoint(exe,
args.init_checkpoint,
main_program=startup_prog,
use_fp16=args.use_fp16)
elif args.init_pretraining_params:
init_pretraining_params(exe,
args.init_pretraining_params,
main_program=startup_prog,
use_fp16=args.use_fp16)
elif args.do_val or args.do_test:
if not args.init_checkpoint:
raise ValueError("args 'init_checkpoint' should be set if"
"only doing validation or testing!")
init_checkpoint(exe,
args.init_checkpoint,
main_program=startup_prog,
use_fp16=args.use_fp16)
if args.do_train:
exec_strategy = fluid.ExecutionStrategy()
if args.use_fast_executor:
exec_strategy.use_experimental_executor = True
exec_strategy.num_threads = dev_count
exec_strategy.num_iteration_per_drop_scope = args.num_iteration_per_drop_scope
train_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
loss_name=graph_vars["loss"].name,
exec_strategy=exec_strategy,
main_program=train_program)
train_pyreader.decorate_tensor_provider(train_data_generator)
else:
train_exe = None
if args.do_val or args.do_test:
test_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
main_program=test_prog,
share_vars_from=train_exe)
if args.do_train:
train_pyreader.start()
steps = 0
if warmup_steps > 0:
graph_vars["learning_rate"] = scheduled_lr
time_begin = time.time()
while True:
try:
steps += 1
if steps % args.skip_steps != 0:
train_exe.run(fetch_list=[])
else:
outputs = evaluate(train_exe, train_program,
train_pyreader, graph_vars,
args.num_labels, "train", dev_count)
if args.verbose:
verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
)
verbose += "learning rate: %f" % (outputs["lr"] if
warmup_steps > 0 else
args.learning_rate)
print(verbose)
current_example, current_epoch = reader.get_train_progress(
)
time_end = time.time()
used_time = time_end - time_begin
print(
"epoch: %d, progress: %d/%d, step: %d, loss: %f, "
"f1: %f, precision: %f, recall: %f, speed: %f steps/s"
% (current_epoch, current_example, num_train_examples,
steps, outputs["loss"], outputs["f1"],
outputs["precision"], outputs["recall"],
args.skip_steps / used_time))
time_begin = time.time()
if steps % args.save_steps == 0:
save_path = os.path.join(args.checkpoints,
"step_" + str(steps))
fluid.io.save_persistables(exe, save_path, train_program)
if steps % args.validation_steps == 0:
# evaluate dev set
if args.do_val:
test_pyreader.decorate_tensor_provider(
reader.data_generator(args.dev_set,
batch_size=args.batch_size,
epoch=1,
shuffle=False))
evaluate(exe, test_prog, test_pyreader, graph_vars,
args.num_labels, "dev")
# evaluate test set
if args.do_test:
test_pyreader.decorate_tensor_provider(
reader.data_generator(args.test_set,
batch_size=args.batch_size,
epoch=1,
shuffle=False))
evaluate(exe, test_prog, test_pyreader, graph_vars,
args.num_labels, "test")
except fluid.core.EOFException:
save_path = os.path.join(args.checkpoints,
"step_" + str(steps))
fluid.io.save_persistables(exe, save_path, train_program)
train_pyreader.reset()
break
# final eval on dev set
if args.do_val:
test_pyreader.decorate_tensor_provider(
reader.data_generator(args.dev_set,
batch_size=args.batch_size,
epoch=1,
shuffle=False))
print("Final validation result:")
evaluate(exe, test_prog, test_pyreader, graph_vars, args.num_labels,
"dev")
# final eval on test set
if args.do_test:
test_pyreader.decorate_tensor_provider(
reader.data_generator(args.test_set,
batch_size=args.batch_size,
epoch=1,
shuffle=False))
print("Final test result:")
evaluate(exe, test_prog, test_pyreader, graph_vars, args.num_labels,
"test")
def main(args):
"""main"""
reader = task_reader.RoleSequenceLabelReader(
vocab_path=args.vocab_path,
labels_map=labels_map,
max_seq_len=args.max_seq_len,
do_lower_case=args.do_lower_case,
in_tokens=args.in_tokens,
random_seed=args.random_seed,
task_id=args.task_id)
if not (args.do_train or args.do_val or args.do_test):
raise ValueError("For args `do_train`, `do_val` and `do_test`, at "
"least one of them must be True.")
startup_prog = fluid.Program()
if args.random_seed is not None:
startup_prog.random_seed = args.random_seed
if args.do_train:
train_data_generator = reader.data_generator(
input_file=args.train_set,
batch_size=args.batch_size,
epoch=args.epoch,
shuffle=True,
phase="train")
num_train_examples = reader.get_num_examples(args.train_set)
if args.in_tokens:
if args.batch_size < args.max_seq_len:
raise ValueError(
'if in_tokens=True, batch_size should greater than max_sqelen, got batch_size:%d seqlen:%d'
% (args.batch_size, args.max_seq_len))
max_train_steps = args.epoch * num_train_examples // (
args.batch_size // args.max_seq_len) // dev_count
else:
max_train_steps = args.epoch * num_train_examples // args.batch_size // dev_count
warmup_steps = int(max_train_steps * args.warmup_proportion)
print("Device count: %d" % dev_count)
print("Num train examples: %d" % num_train_examples)
print("Max train steps: %d" % max_train_steps)
print("Num warmup steps: %d" % warmup_steps)
train_program = fluid.Program()
with fluid.program_guard(train_program, startup_prog):
with fluid.unique_name.guard():
train_pyreader, graph_vars = create_model(
args,
pyreader_name='train_reader',
ernie_config=ernie_config)
scheduled_lr, loss_scaling = optimization(
loss=graph_vars["loss"],
warmup_steps=warmup_steps,
num_train_steps=max_train_steps,
learning_rate=args.learning_rate,
train_program=train_program,
startup_prog=startup_prog,
weight_decay=args.weight_decay,
scheduler=args.lr_scheduler,
use_fp16=args.use_fp16,
use_dynamic_loss_scaling=args.use_dynamic_loss_scaling,
init_loss_scaling=args.init_loss_scaling,
incr_every_n_steps=args.incr_every_n_steps,
decr_every_n_nan_or_inf=args.decr_every_n_nan_or_inf,
incr_ratio=args.incr_ratio,
decr_ratio=args.decr_ratio)
if args.verbose:
if args.in_tokens:
lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
program=train_program,
batch_size=args.batch_size // args.max_seq_len)
else:
lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
program=train_program, batch_size=args.batch_size)
print("Theoretical memory usage in training: %.3f - %.3f %s" %
(lower_mem, upper_mem, unit))
if args.do_val or args.do_test:
test_prog = fluid.Program()
with fluid.program_guard(test_prog, startup_prog):
with fluid.unique_name.guard():
test_pyreader, graph_vars = create_model(
args,
pyreader_name='test_reader',
ernie_config=ernie_config)
test_prog = test_prog.clone(for_test=True)
nccl2_num_trainers = 1
nccl2_trainer_id = 0
exe = fluid.Executor(place)
exe.run(startup_prog)
if args.do_train:
if args.init_checkpoint and args.init_pretraining_params:
print(
"WARNING: args 'init_checkpoint' and 'init_pretraining_params' "
"both are set! Only arg 'init_checkpoint' is made valid.")
if args.init_checkpoint:
init_checkpoint(exe,
args.init_checkpoint,
main_program=startup_prog,
use_fp16=args.use_fp16)
elif args.init_pretraining_params:
init_pretraining_params(exe,
args.init_pretraining_params,
main_program=startup_prog,
use_fp16=args.use_fp16)
elif args.do_val or args.do_test:
if not args.init_checkpoint:
raise ValueError("args 'init_checkpoint' should be set if"
"only doing validation or testing!")
init_checkpoint(exe,
args.init_checkpoint,
main_program=startup_prog,
use_fp16=args.use_fp16)
if args.do_train:
exec_strategy = fluid.ExecutionStrategy()
if args.use_fast_executor:
exec_strategy.use_experimental_executor = True
exec_strategy.num_threads = dev_count
exec_strategy.num_iteration_per_drop_scope = args.num_iteration_per_drop_scope
train_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
loss_name=graph_vars["loss"].name,
exec_strategy=exec_strategy,
main_program=train_program,
num_trainers=nccl2_num_trainers,
trainer_id=nccl2_trainer_id)
train_pyreader.decorate_tensor_provider(train_data_generator)
else:
train_exe = None
if args.do_val or args.do_test:
test_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
main_program=test_prog,
share_vars_from=train_exe)
if args.do_train:
train_pyreader.start()
steps = 0
graph_vars["learning_rate"] = scheduled_lr
time_begin = time.time()
while True:
try:
steps += 1
if steps % args.skip_steps != 0:
train_exe.run(fetch_list=[])
else:
fetch_list = [
graph_vars["num_infer"].name,
graph_vars["num_label"].name,
graph_vars["num_correct"].name,
graph_vars["loss"].name,
graph_vars['learning_rate'].name,
]
out = train_exe.run(fetch_list=fetch_list)
num_infer, num_label, num_correct, np_loss, np_lr = out
lr = float(np_lr[0])
loss = np_loss.mean()
precision, recall, f1 = calculate_f1(
num_label, num_infer, num_correct)
if args.verbose:
print(
"train pyreader queue size: %d, learning rate: %f"
% (train_pyreader.queue.size(),
lr if warmup_steps > 0 else args.learning_rate))
current_example, current_epoch = reader.get_train_progress(
)
time_end = time.time()
used_time = time_end - time_begin
print(
u"【train】epoch: {}, step: {}, loss: {:.6f}, "
"f1: {:.4f}, precision: {:.4f}, recall: {:.4f}, speed: {:.3f} steps/s"
.format(current_epoch, steps, float(loss), float(f1),
float(precision), float(recall),
args.skip_steps / used_time))
time_begin = time.time()
if steps % args.save_steps == 0:
save_path = os.path.join(args.checkpoints,
"step_" + str(steps))
fluid.io.save_persistables(exe, save_path, train_program)
if steps % args.validation_steps == 0:
# evaluate dev set
if args.do_val:
precision, recall, f1 = evaluate_wrapper(
reader, exe, test_prog, test_pyreader, graph_vars,
current_epoch, steps)
print(
u"【dev】precision {:.4f} , recall {:.4f}, f1-score {:.4f}"
.format(float(precision), float(recall),
float(f1)))
# evaluate test set
if args.do_test:
precision, recall, f1 = evaluate_wrapper(
reader, exe, test_prog, test_pyreader, graph_vars,
current_epoch, steps)
print(
u"【test】precision {:.4f} , recall {:.4f}, f1-score {:.4f}"
.format(float(precision), float(recall),
float(f1)))
except fluid.core.EOFException:
save_path = os.path.join(args.checkpoints, "final_model")
fluid.io.save_persistables(exe, save_path, train_program)
train_pyreader.reset()
break
# final eval on dev set
if args.do_val:
precision, recall, f1 = evaluate_wrapper(reader, exe, test_prog,
test_pyreader, graph_vars, 1,
'final')
print(u"【dev】precision {:.4f} , recall {:.4f}, f1-score {:.4f}".format(
float(precision), float(recall), float(f1)))
if args.do_test:
test_ret = predict_wrapper(reader, exe, test_prog, test_pyreader,
graph_vars, 1, 'final')
utils.write_by_lines(args.trigger_pred_save_path, test_ret)
def main(args):
ernie_config = ErnieConfig(args.ernie_config_path)
ernie_config.print_config()
if args.use_cuda:
dev_list = fluid.cuda_places()
place = dev_list[0]
dev_count = len(dev_list)
else:
place = fluid.CPUPlace()
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
reader = task_reader.SequenceLabelReader(
vocab_path=args.vocab_path,
label_map_config=args.label_map_config,
max_seq_len=args.max_seq_len,
do_lower_case=args.do_lower_case,
in_tokens=args.in_tokens,
random_seed=args.random_seed,
task_id=args.task_id)
if not (args.do_train or args.do_val or args.do_test):
raise ValueError("For args `do_train`, `do_val` and `do_test`, at "
"least one of them must be True.")
startup_prog = fluid.Program()
if args.random_seed is not None:
startup_prog.random_seed = args.random_seed
if args.do_train:
train_data_generator = reader.data_generator(
input_file=args.train_set,
batch_size=args.batch_size,
epoch=args.epoch,
shuffle=True,
phase="train")
num_train_examples = reader.get_num_examples(args.train_set)
if args.in_tokens:
if args.batch_size < args.max_seq_len:
raise ValueError(
'if in_tokens=True, batch_size should greater than max_sqelen, got batch_size:%d seqlen:%d'
% (args.batch_size, args.max_seq_len))
max_train_steps = args.epoch * num_train_examples // (
args.batch_size // args.max_seq_len) // dev_count
else:
max_train_steps = args.epoch * num_train_examples // args.batch_size // dev_count
warmup_steps = int(max_train_steps * args.warmup_proportion)
log.info("Device count: %d" % dev_count)
log.info("Num train examples: %d" % num_train_examples)
log.info("Max train steps: %d" % max_train_steps)
log.info("Num warmup steps: %d" % warmup_steps)
train_program = fluid.Program()
with fluid.program_guard(train_program, startup_prog):
with fluid.unique_name.guard():
train_pyreader, graph_vars = create_model(
args,
pyreader_name='train_reader',
ernie_config=ernie_config)
scheduled_lr, loss_scaling = optimization(
loss=graph_vars["loss"],
warmup_steps=warmup_steps,
num_train_steps=max_train_steps,
learning_rate=args.learning_rate,
train_program=train_program,
startup_prog=startup_prog,
weight_decay=args.weight_decay,
scheduler=args.lr_scheduler,
use_fp16=args.use_fp16,
use_dynamic_loss_scaling=args.use_dynamic_loss_scaling,
init_loss_scaling=args.init_loss_scaling,
incr_every_n_steps=args.incr_every_n_steps,
decr_every_n_nan_or_inf=args.decr_every_n_nan_or_inf,
incr_ratio=args.incr_ratio,
decr_ratio=args.decr_ratio)
if args.verbose:
if args.in_tokens:
lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
program=train_program,
batch_size=args.batch_size // args.max_seq_len)
else:
lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
program=train_program, batch_size=args.batch_size)
log.info("Theoretical memory usage in training: %.3f - %.3f %s" %
(lower_mem, upper_mem, unit))
if args.do_val or args.do_test:
test_prog = fluid.Program()
with fluid.program_guard(test_prog, startup_prog):
with fluid.unique_name.guard():
test_pyreader, graph_vars = create_model(
args,
pyreader_name='test_reader',
ernie_config=ernie_config)
test_prog = test_prog.clone(for_test=True)
nccl2_num_trainers = 1
nccl2_trainer_id = 0
if args.is_distributed:
trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
worker_endpoints_env = os.getenv("PADDLE_TRAINER_ENDPOINTS")
current_endpoint = os.getenv("PADDLE_CURRENT_ENDPOINT")
worker_endpoints = worker_endpoints_env.split(",")
trainers_num = len(worker_endpoints)
log.info("worker_endpoints:{} trainers_num:{} current_endpoint:{} \
trainer_id:{}".format(worker_endpoints, trainers_num,
current_endpoint, trainer_id))
# prepare nccl2 env.
config = fluid.DistributeTranspilerConfig()
config.mode = "nccl2"
t = fluid.DistributeTranspiler(config=config)
t.transpile(trainer_id,
trainers=worker_endpoints_env,
current_endpoint=current_endpoint,
program=train_program if args.do_train else test_prog,
startup_program=startup_prog)
nccl2_num_trainers = trainers_num
nccl2_trainer_id = trainer_id
exe = fluid.Executor(place)
exe.run(startup_prog)
if args.do_train:
if args.init_checkpoint and args.init_pretraining_params:
log.info(
"WARNING: args 'init_checkpoint' and 'init_pretraining_params' "
"both are set! Only arg 'init_checkpoint' is made valid.")
if args.init_checkpoint:
init_checkpoint(exe,
args.init_checkpoint,
main_program=startup_prog,
use_fp16=args.use_fp16)
elif args.init_pretraining_params:
init_pretraining_params(exe,
args.init_pretraining_params,
main_program=startup_prog,
use_fp16=args.use_fp16)
elif args.do_val or args.do_test:
if not args.init_checkpoint:
raise ValueError("args 'init_checkpoint' should be set if"
"only doing validation or testing!")
init_checkpoint(exe,
args.init_checkpoint,
main_program=startup_prog,
use_fp16=args.use_fp16)
if args.do_train:
exec_strategy = fluid.ExecutionStrategy()
if args.use_fast_executor:
exec_strategy.use_experimental_executor = True
exec_strategy.num_threads = dev_count
exec_strategy.num_iteration_per_drop_scope = args.num_iteration_per_drop_scope
train_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
loss_name=graph_vars["loss"].name,
exec_strategy=exec_strategy,
main_program=train_program,
num_trainers=nccl2_num_trainers,
trainer_id=nccl2_trainer_id)
train_pyreader.set_batch_generator(train_data_generator)
else:
train_exe = None
if args.do_val or args.do_test:
test_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
main_program=test_prog,
share_vars_from=train_exe)
if args.do_train:
train_pyreader.start()
steps = 0
graph_vars["learning_rate"] = scheduled_lr
time_begin = time.time()
while True:
try:
steps += 1
if steps % args.skip_steps != 0:
train_exe.run(fetch_list=[])
else:
fetch_list = [
graph_vars["num_infer"].name,
graph_vars["num_label"].name,
graph_vars["num_correct"].name,
graph_vars["loss"].name,
graph_vars['learning_rate'].name,
]
out = train_exe.run(fetch_list=fetch_list)
num_infer, num_label, num_correct, np_loss, np_lr = out
lr = float(np_lr[0])
loss = np_loss.mean()
precision, recall, f1 = calculate_f1(
num_label, num_infer, num_correct)
if args.verbose:
log.info(
"train pyreader queue size: %d, learning rate: %f"
% (train_pyreader.queue.size(),
lr if warmup_steps > 0 else args.learning_rate))
current_example, current_epoch = reader.get_train_progress(
)
time_end = time.time()
used_time = time_end - time_begin
log.info(
"epoch: %d, progress: %d/%d, step: %d, loss: %f, "
"f1: %f, precision: %f, recall: %f, speed: %f steps/s"
% (current_epoch, current_example, num_train_examples,
steps, loss, f1, precision, recall,
args.skip_steps / used_time))
time_begin = time.time()
if nccl2_trainer_id == 0 and steps % args.save_steps == 0:
save_path = os.path.join(args.checkpoints,
"step_" + str(steps))
fluid.io.save_persistables(exe, save_path, train_program)
if nccl2_trainer_id == 0 and steps % args.validation_steps == 0:
# evaluate dev set
if args.do_val:
evaluate_wrapper(reader, exe, test_prog, test_pyreader,
graph_vars, current_epoch, steps)
# evaluate test set
if args.do_test:
predict_wrapper(reader, exe, test_prog, test_pyreader,
graph_vars, current_epoch, steps)
except fluid.core.EOFException:
save_path = os.path.join(args.checkpoints,
"step_" + str(steps))
fluid.io.save_persistables(exe, save_path, train_program)
train_pyreader.reset()
break
# final eval on dev set
if nccl2_trainer_id == 0 and args.do_val:
if not args.do_train:
current_example, current_epoch = reader.get_train_progress()
evaluate_wrapper(reader, exe, test_prog, test_pyreader, graph_vars,
current_epoch, 'final')
if nccl2_trainer_id == 0 and args.do_test:
if not args.do_train:
current_example, current_epoch = reader.get_train_progress()
predict_wrapper(reader, exe, test_prog, test_pyreader, graph_vars,
current_epoch, 'final')
