def evaluate_test(args, reader, exe, test_prog, test_pyreader, graph_vars,
epoch, steps):
"""evaluate dev set"""
for ds in args.test_set.split(','):
test_pyreader.decorate_tensor_provider(
reader.data_generator(ds,
batch_size=args.predict_batch_size,
epoch=1,
dev_count=1,
shuffle=False))
print("test result of dataset {}:".format(ds))
evaluate_info = evaluate(exe,
test_prog,
test_pyreader,
graph_vars,
"test",
metric=args.metric,
is_classify=args.is_classify,
is_regression=args.is_regression)
print(evaluate_info +
', file: {}, epoch: {}, steps: {}'.format(ds, epoch, steps))
def evaluate_wrapper(args, reader, exe, test_prog, test_pyreader, graph_vars,
epoch, steps):
# evaluate dev set
batch_size = args.batch_size if args.predict_batch_size is None else args.predict_batch_size
for ds in args.dev_set.split(','):
test_pyreader.set_batch_generator(
reader.data_generator(ds,
batch_size=batch_size,
epoch=1,
dev_count=1,
shuffle=False))
log.info("validation result of dataset {}:".format(ds))
evaluate_info = evaluate(exe,
test_prog,
test_pyreader,
graph_vars,
"dev",
metric=args.metric,
is_classify=args.is_classify,
is_regression=args.is_regression)
log.info(evaluate_info +
', file: {}, epoch: {}, steps: {}'.format(ds, epoch, steps))
def evaluate_wrapper(args, reader, exe, test_prog, test_pyreader, graph_vars,
epoch, steps):
# evaluate dev set
# evaluate dev set
batch_size = args.batch_size if args.predict_batch_size is None else args.predict_batch_size
for ds in args.dev_set.split(','):
test_pyreader.set_batch_generator(
reader.data_generator(ds,
batch_size=batch_size,
epoch=1,
dev_count=1,
shuffle=False))
log.info("validation result of dataset {}:".format(ds))
# 保存评估结果
try:
print("++++++++++++++++++++success1+++++++++++++++++++++++++++++")
content = "validation result of dataset {}:" + ds + "\n"
save_evaluation_results(args.output, content)
except:
print("++++++++++++++++++++Error1+++++++++++++++++++++++++++++")
evaluate_info = evaluate(exe,
test_prog,
test_pyreader,
graph_vars,
"dev",
metric=args.metric,
is_classify=args.is_classify,
is_regression=args.is_regression)
log.info(evaluate_info +
', file: {}, epoch: {}, steps: {}'.format(ds, epoch, steps))
# 保存评估结果
try:
print("++++++++++++++++++++success1+++++++++++++++++++++++++++++")
content = "file:" + ds + ",epoch:" + epoch + ",steps:" + steps + "\n"
save_evaluation_results(args.output, content)
except:
print("++++++++++++++++++++Error1+++++++++++++++++++++++++++++")
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.ClassifyReader(
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["probs"].name,
graph_vars["accuracy"].name,
graph_vars["num_seqs"].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_train:
train_pyreader.start()
steps = 0
if warmup_steps > 0:
graph_vars["learning_rate"] = scheduled_lr
if args.save_log and args.log_path:
if os.path.exists(args.log_path):
raise FileExistsError("Logging file already exists!")
with open(args.log_path, 'w') as logfile:
logfile.write('%s\n' % time.asctime())
print('Writing logs into %s' % args.log_path)
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, "train")
if args.verbose:
verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
)
verbose += "learning rate: %f" % (
outputs["learning_rate"]
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, "
"ave loss: %.4f, micro_f1: %.4f, micro_p: %.4f, micro_r: %.4f, "
"speed: %f steps/s" %
(current_epoch, current_example, num_train_examples, steps,
outputs["loss"], outputs["micro_f"], outputs["micro_p"], outputs["micro_r"],
args.skip_steps / used_time))
# Todo: complete logging function
# Todo: print more useful metrics: f1/p/r instead of acc
if args.save_log and args.log_path:
with open(args.log_path, 'a') as logfile:
logfile.write("epoch: %d, progress: %d/%d, step: %d, "
"ave loss: %.4f, ave_acc: %.4f, micro_f1: %.4f, micro_p: %.4f, micro_r: %.4f, "
"speed: %f steps/s\n" %
(current_epoch, current_example, num_train_examples, steps,
outputs["loss"], outputs["accuracy"], outputs["micro_f"], outputs["micro_p"], outputs["micro_r"],
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,
"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,
"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, "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, "test")
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()))
exe = fluid.Executor(place)
reader = task_reader.ClassifyReader(
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,
tokenizer=args.tokenizer,
is_classify=args.is_classify,
is_regression=args.is_regression,
for_cn=args.for_cn,
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.")
if args.do_test:
assert args.test_save is not None
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,
dev_count=dev_count,
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)
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()
if args.random_seed is not None and args.enable_ce:
train_program.random_seed = args.random_seed
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,
is_classify=args.is_classify,
is_regression=args.is_regression)
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,
is_classify=args.is_classify,
is_regression=args.is_regression)
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.warning(
"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
test_exe = exe
if args.do_val or args.do_test:
if args.use_multi_gpu_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
ce_info = []
time_begin = time.time()
last_epoch = 0
current_epoch = 0
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,
"train",
metric=args.metric,
is_classify=args.is_classify,
is_regression=args.is_regression)
if args.verbose:
verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
)
verbose += "learning rate: %f" % (
outputs["learning_rate"]
if warmup_steps > 0 else args.learning_rate)
log.info(verbose)
current_example, current_epoch = reader.get_train_progress()
time_end = time.time()
used_time = time_end - time_begin
if args.is_classify:
log.info(
"epoch: %d, progress: %d/%d, step: %d, ave loss: %f, "
"ave acc: %f, speed: %f steps/s" %
(current_epoch, current_example, num_train_examples,
steps, outputs["loss"], outputs["accuracy"],
args.skip_steps / used_time))
ce_info.append(
[outputs["loss"], outputs["accuracy"], used_time])
if args.is_regression:
log.info(
"epoch: %d, progress: %d/%d, step: %d, ave loss: %f, "
" speed: %f steps/s" %
(current_epoch, current_example, num_train_examples,
steps, outputs["loss"],
args.skip_steps / used_time))
time_begin = time.time()
if nccl2_trainer_id == 0:
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 or last_epoch != current_epoch:
# evaluate dev set
if args.do_val:
evaluate_wrapper(args, reader, exe, test_prog,
test_pyreader, graph_vars,
current_epoch, steps)
if args.do_test:
predict_wrapper(args, reader, exe, test_prog,
test_pyreader, graph_vars,
current_epoch, steps)
if last_epoch != current_epoch:
last_epoch = current_epoch
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
if args.enable_ce:
card_num = get_cards()
ce_loss = 0
ce_acc = 0
ce_time = 0
try:
ce_loss = ce_info[-2][0]
ce_acc = ce_info[-2][1]
ce_time = ce_info[-2][2]
except:
log.info("ce info error")
log.info("kpis\ttrain_duration_card%s\t%s" % (card_num, ce_time))
log.info("kpis\ttrain_loss_card%s\t%f" % (card_num, ce_loss))
log.info("kpis\ttrain_acc_card%s\t%f" % (card_num, ce_acc))
# final eval on dev set
if args.do_val:
evaluate_wrapper(args, reader, exe, test_prog, test_pyreader,
graph_vars, current_epoch, steps)
# final eval on test set
if args.do_test:
predict_wrapper(args, reader, exe, test_prog, test_pyreader, graph_vars,
current_epoch, steps)
# final eval on dianostic, hack for glue-ax
if args.diagnostic:
test_pyreader.decorate_tensor_provider(
reader.data_generator(
args.diagnostic,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
log.info("Final diagnostic")
qids, preds, probs = predict(
test_exe,
test_prog,
test_pyreader,
graph_vars,
is_classify=args.is_classify,
is_regression=args.is_regression)
assert len(qids) == len(preds), '{} v.s. {}'.format(
len(qids), len(preds))
with open(args.diagnostic_save, 'w') as f:
for id, s, p in zip(qids, preds, probs):
f.write('{}\t{}\t{}\n'.format(id, s, p))
log.info("Done final diagnostic, saving to {}".format(
args.diagnostic_save))
def main(args):
"""main function"""
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.ClassifyReader(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,
tokenizer=args.tokenizer,
is_classify=args.is_classify,
is_regression=args.is_regression,
for_cn=args.for_cn,
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.")
if args.do_test:
assert args.test_save is not None
startup_prog = fluid.Program()
if args.random_seed is not None:
startup_prog.random_seed = args.random_seed
if args.predict_batch_size is None:
args.predict_batch_size = args.batch_size
if args.do_train:
train_data_generator = reader.data_generator(
input_file=args.train_set,
batch_size=args.batch_size,
epoch=args.epoch,
dev_count=dev_count,
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()
"""
if args.random_seed is not None and args.enable_ce:
train_program.random_seed = args.random_seed
"""
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,
is_classify=args.is_classify,
is_regression=args.is_regression)
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)
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,
is_classify=args.is_classify,
is_regression=args.is_regression)
test_prog = test_prog.clone(for_test=True)
nccl2_num_trainers = 1
nccl2_trainer_id = 0
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
test_exe = exe
if args.do_val or args.do_test:
if args.use_multi_gpu_test:
test_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
main_program=test_prog,
share_vars_from=train_exe)
steps = 10000
current_epoch = 1
if args.do_train:
train_pyreader.start()
steps = 0
if warmup_steps > 0:
graph_vars["learning_rate"] = scheduled_lr
ce_info = []
time_begin = time.time()
last_epoch = 0
current_epoch = 0
previous_eval_acc = 0.80
previous_train_acc = 0.90
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,
"train",
metric=args.metric,
is_classify=args.is_classify,
is_regression=args.is_regression)
acc = outputs["accuracy"]
if acc > previous_train_acc or acc > 0.95:
print(
"previous train accuracy is %f and current train accuracy is %f "
% (previous_train_acc, acc))
previous_train_acc = acc
eval_acc = evaluate_wrapper(args, reader, exe,
test_prog, test_pyreader,
graph_vars, current_epoch,
steps)
print(
"previous evaluate accuracy is %f and current evaluate accuracy is %f "
% (previous_eval_acc, eval_acc))
if eval_acc > previous_eval_acc:
previous_eval_acc = eval_acc
save_path = os.path.join(
args.checkpoints,
"evalacc_" + str(eval_acc).split('.')[1])
fluid.io.save_persistables(exe, save_path,
train_program)
predict_wrapper(args,
reader,
exe,
test_prog,
test_pyreader,
graph_vars,
current_epoch,
steps="evalacc_" +
str(eval_acc).split('.')[1])
print(
"predict and save model!!!!!!!!!!!!!!!!!!!!!!!!!! in %s"
% (save_path))
if args.verbose:
verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
)
verbose += "learning rate: %f" % (
outputs["learning_rate"]
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, ave loss: %f, "
"ave acc: %f, speed: %f steps/s" %
(current_epoch, current_example, num_train_examples,
steps, outputs["loss"], outputs["accuracy"],
args.skip_steps / used_time))
ce_info.append(
[outputs["loss"], outputs["accuracy"], 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 or last_epoch != current_epoch:
# # evaluate dev set
# if args.do_val:
# ret=evaluate_wrapper(args, reader, exe, test_prog,
# test_pyreader, graph_vars,
# current_epoch, steps)
# if args.do_test:
# predict_wrapper(args, reader, exe,
# test_prog, test_pyreader, graph_vars,
# current_epoch, steps)
if last_epoch != current_epoch:
last_epoch = current_epoch
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:
# evaluate_wrapper(args, reader, exe, test_prog, test_pyreader,
# graph_vars, current_epoch, steps)
# final eval on test set
steps = 0
# if args.do_test:
# current_epoch = 0
# predict_wrapper(args, reader, exe, test_prog, test_pyreader, graph_vars,
# current_epoch, steps)
# final eval on dianostic, hack for glue-ax
if args.diagnostic:
test_pyreader.decorate_tensor_provider(
reader.data_generator(args.diagnostic,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
print("Final diagnostic")
qids, preds, probs = predict(test_exe,
test_prog,
test_pyreader,
graph_vars,
is_classify=args.is_classify,
is_regression=args.is_regression)
assert len(qids) == len(preds), '{} v.s. {}'.format(
len(qids), len(preds))
with open(args.diagnostic_save, 'w') as f:
for id, s, p in zip(qids, preds, probs):
f.write('{}\t{}\t{}\n'.format(id, s, p))
print("Done final diagnostic, saving to {}".format(
args.diagnostic_save))
def main(args):
"""main"""
model_config = UNIMOConfig(args.unimo_config_path)
model_config.print_config()
gpu_id = 0
gpus = fluid.core.get_cuda_device_count()
if args.is_distributed and os.getenv("FLAGS_selected_gpus") is not None:
gpu_list = os.getenv("FLAGS_selected_gpus").split(",")
gpus = len(gpu_list)
gpu_id = int(gpu_list[0])
if args.use_cuda:
place = fluid.CUDAPlace(gpu_id)
dev_count = gpus
else:
place = fluid.CPUPlace()
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
tokenizer = GptBpeTokenizer(vocab_file=args.unimo_vocab_file,
encoder_json_file=args.encoder_json_file,
vocab_bpe_file=args.vocab_bpe_file,
do_lower_case=args.do_lower_case)
data_reader = ClassifyReader(tokenizer, args)
if not (args.do_train or args.do_val or args.do_val_hard \
or args.do_test or args.do_test_hard or args.do_diagnostic):
raise ValueError("For args `do_train`, `do_val`, `do_val_hard`, `do_test`," \
" `do_test_hard` and `do_diagnostic`, 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:
trainers_num = int(os.getenv("PADDLE_TRAINERS_NUM", "1"))
train_data_generator = data_reader.data_generator(
input_file=args.train_set,
batch_size=args.batch_size,
epoch=args.epoch,
dev_count=trainers_num,
shuffle=True,
phase="train")
num_train_examples = data_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) // trainers_num
else:
max_train_steps = args.epoch * num_train_examples // args.batch_size // trainers_num
warmup_steps = int(max_train_steps * args.warmup_proportion)
print("Device count: %d, gpu_id: %d" % (dev_count, gpu_id))
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', config=model_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,
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,
beta1=args.beta1,
beta2=args.beta2,
epsilon=args.epsilon)
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_val_hard or args.do_test or args.do_test_hard \
or args.do_pred or args.do_pred_hard or args.do_diagnostic:
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', config=model_config)
test_prog = test_prog.clone(for_test=True)
nccl2_num_trainers = 1
nccl2_trainer_id = 0
print("args.is_distributed:", args.is_distributed)
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)
print("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"
if args.nccl_comm_num > 1:
config.nccl_comm_num = args.nccl_comm_num
if args.use_hierarchical_allreduce and trainers_num > args.hierarchical_allreduce_inter_nranks:
config.use_hierarchical_allreduce = args.use_hierarchical_allreduce
config.hierarchical_allreduce_inter_nranks = args.hierarchical_allreduce_inter_nranks
assert config.hierarchical_allreduce_inter_nranks > 1
assert trainers_num % config.hierarchical_allreduce_inter_nranks == 0
config.hierarchical_allreduce_exter_nranks = \
trainers_num / config.hierarchical_allreduce_inter_nranks
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:
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=train_program)
elif args.init_pretraining_params:
init_pretraining_params(exe,
args.init_pretraining_params,
main_program=train_program)
elif args.do_val or args.do_val_hard or args.do_test or args.do_test_hard \
or args.do_pred or args.do_pred_hard or args.do_diagnostic:
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)
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
test_exe = exe
if args.do_val or args.do_val_hard or args.do_test or args.do_test_hard \
or args.do_pred or args.do_pred_hard or args.do_diagnostic:
if args.use_multi_gpu_test:
test_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
main_program=test_prog,
share_vars_from=train_exe)
dev_ret_history = [] # (steps, key_eval, eval)
dev_hard_ret_history = [] # (steps, key_eval, eval)
test_ret_history = [] # (steps, key_eval, eval)
test_hard_ret_history = [] # (steps, key_eval, eval)
if args.do_train:
train_pyreader.start()
steps = 0
if warmup_steps > 0:
graph_vars["learning_rate"] = scheduled_lr
time_begin = time.time()
skip_steps = args.skip_steps
while True:
try:
steps += 1
if steps % skip_steps == 0:
train_fetch_list = [
graph_vars["loss"].name, graph_vars["accuracy"].name,
graph_vars["num_seqs"].name
]
if "learning_rate" in graph_vars:
train_fetch_list.append(
graph_vars["learning_rate"].name)
res = train_exe.run(fetch_list=train_fetch_list)
outputs = {"loss": np.mean(res[0])}
if "learning_rate" in graph_vars:
outputs["learning_rate"] = float(res[3][0])
if args.verbose:
verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
)
verbose += "learning rate: %f" % (
outputs["learning_rate"]
if warmup_steps > 0 else args.learning_rate)
print(verbose)
current_example, current_epoch = data_reader.get_train_progress(
)
time_end = time.time()
used_time = time_end - time_begin
print("%s - epoch: %d, progress: %d/%d, step: %d, ave loss: %f, speed: %f steps/s" % \
(get_time(), current_epoch, current_example, num_train_examples, steps, \
outputs["loss"], args.skip_steps / used_time))
time_begin = time.time()
else:
train_exe.run(fetch_list=[])
if nccl2_trainer_id == 0:
if steps % args.save_steps == 0 and args.save_checkpoints:
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(
data_reader.data_generator(
args.dev_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
outputs = evaluate(args, test_exe, test_prog,
test_pyreader, graph_vars,
"dev")
dev_ret_history.append(
(steps, outputs['key_eval'],
outputs[outputs['key_eval']]))
# evaluate dev_hard set
if args.do_val_hard:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(
args.dev_hard_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
outputs = evaluate(args, test_exe, test_prog,
test_pyreader, graph_vars,
"dev_hard")
dev_hard_ret_history.append(
(steps, outputs['key_eval'],
outputs[outputs['key_eval']]))
# evaluate test set
if args.do_test:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(
args.test_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
outputs = evaluate(args, test_exe, test_prog,
test_pyreader, graph_vars,
"test")
test_ret_history.append(
(steps, outputs['key_eval'],
outputs[outputs['key_eval']]))
# evaluate test_hard set
if args.do_test_hard:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(
args.test_hard_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
outputs = evaluate(args, test_exe, test_prog,
test_pyreader, graph_vars,
"test_hard")
test_hard_ret_history.append(
(steps, outputs['key_eval'],
outputs[outputs['key_eval']]))
# pred diagnostic set
if args.do_diagnostic:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(
args.diagnostic_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
qids, preds, probs = predict(test_exe,
test_prog,
test_pyreader,
graph_vars,
dev_count=1)
save_path = args.pred_save + '.diagnostic.' + str(
steps) + '.txt'
print("testing {}, save to {}".format(
args.diagnostic_set, save_path))
with open(save_path, 'w') as f:
for id, s, p in zip(qids, preds, probs):
f.write('{}\t{}\t{}\n'.format(id, s, p))
# pred test set
if args.do_pred:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(
args.test_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
qids, preds, probs = predict(test_exe,
test_prog,
test_pyreader,
graph_vars,
dev_count=1)
save_path = args.pred_save + '.test.' + str(
steps) + '.txt'
print("testing {}, save to {}".format(
args.test_set, save_path))
with open(save_path, 'w') as f:
for id, s, p in zip(qids, preds, probs):
f.write('{}\t{}\t{}\n'.format(id, s, p))
# pred test hard set
if args.do_pred_hard:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(
args.test_hard_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
qids, preds, probs = predict(test_exe,
test_prog,
test_pyreader,
graph_vars,
dev_count=1)
save_path = args.pred_save + '.test_hard.' + str(
steps) + '.txt'
print("testing {}, save to {}".format(
args.test_hard_set, save_path))
with open(save_path, 'w') as f:
for id, s, p in zip(qids, preds, probs):
f.write('{}\t{}\t{}\n'.format(id, s, p))
except fluid.core.EOFException:
if args.save_checkpoints:
save_path = os.path.join(args.checkpoints,
"step_" + str(steps))
fluid.io.save_persistables(exe, save_path, train_program)
train_pyreader.reset()
break
if nccl2_trainer_id == 0:
# final pred on diagnostic set
if args.do_diagnostic:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(args.diagnostic_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
qids, preds, probs = predict(test_exe,
test_prog,
test_pyreader,
graph_vars,
dev_count=1)
save_path = args.pred_save + '.diagnostic.' + str(steps) + '.txt'
print("testing {}, save to {}".format(args.diagnostic_set,
save_path))
with open(save_path, 'w') as f:
for id, s, p in zip(qids, preds, probs):
f.write('{}\t{}\t{}\n'.format(id, s, p))
# final pred on test set
if args.do_pred:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(args.test_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
qids, preds, probs = predict(test_exe,
test_prog,
test_pyreader,
graph_vars,
dev_count=1)
save_path = args.pred_save + '.test.' + str(steps) + '.txt'
print("testing {}, save to {}".format(args.test_set, save_path))
with open(save_path, 'w') as f:
for id, s, p in zip(qids, preds, probs):
f.write('{}\t{}\t{}\n'.format(id, s, p))
# final pred on test_hard set
if args.do_pred_hard:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(args.test_hard_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
qids, preds, probs = predict(test_exe,
test_prog,
test_pyreader,
graph_vars,
dev_count=1)
save_path = args.pred_save + '.test_hard.' + str(steps) + '.txt'
print("testing {}, save to {}".format(args.test_hard_set,
save_path))
with open(save_path, 'w') as f:
for id, s, p in zip(qids, preds, probs):
f.write('{}\t{}\t{}\n'.format(id, s, p))
# final eval on test set
if args.do_test:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(args.test_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
print("Final test result:")
outputs = evaluate(args, test_exe, test_prog, test_pyreader,
graph_vars, "test")
test_ret_history.append(
(steps, outputs['key_eval'], outputs[outputs['key_eval']]))
test_ret_history = sorted(test_ret_history,
key=lambda a: a[2],
reverse=True)
print("Best testing result: step %d %s %f" %
(test_ret_history[0][0], test_ret_history[0][1],
test_ret_history[0][2]))
# final eval on test hard set
if args.do_test_hard:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(args.test_hard_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
print("Final test_hard result:")
outputs = evaluate(args, test_exe, test_prog, test_pyreader,
graph_vars, "test_hard")
test_hard_ret_history.append(
(steps, outputs['key_eval'], outputs[outputs['key_eval']]))
test_hard_ret_history = sorted(test_hard_ret_history,
key=lambda a: a[2],
reverse=True)
print("Best testing hard result: step %d %s %f" %
(test_hard_ret_history[0][0], test_hard_ret_history[0][1],
test_hard_ret_history[0][2]))
# final eval on dev set
if args.do_val:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(args.dev_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
print("Final validation result:")
outputs = evaluate(args, test_exe, test_prog, test_pyreader,
graph_vars, "dev")
dev_ret_history.append(
(steps, outputs['key_eval'], outputs[outputs['key_eval']]))
dev_ret_history = sorted(dev_ret_history,
key=lambda a: a[2],
reverse=True)
print("Best validation result: step %d %s %f" %
(dev_ret_history[0][0], dev_ret_history[0][1],
dev_ret_history[0][2]))
# final eval on dev hard set
if args.do_val_hard:
test_pyreader.decorate_tensor_provider(
data_reader.data_generator(args.dev_hard_set,
batch_size=args.batch_size,
epoch=1,
dev_count=1,
shuffle=False))
print("Final validation_hard result:")
outputs = evaluate(args, test_exe, test_prog, test_pyreader,
graph_vars, "dev_hard")
dev_hard_ret_history.append(
(steps, outputs['key_eval'], outputs[outputs['key_eval']]))
dev_hard_ret_history = sorted(dev_hard_ret_history,
key=lambda a: a[2],
reverse=True)
print("Best validation_hard result: step %d %s %f" %
(dev_hard_ret_history[0][0], dev_hard_ret_history[0][1],
dev_hard_ret_history[0][2]))
