def test(args):
ernie_config = ErnieConfig(args.ernie_config_path)
ernie_config.print_config()
test_prog = fluid.Program()
test_startup = fluid.Program()
with fluid.program_guard(test_prog, test_startup):
with fluid.unique_name.guard():
test_pyreader, next_sent_acc, mask_lm_loss, total_loss = create_model(
pyreader_name='test_reader', ernie_config=ernie_config)
test_prog = test_prog.clone(for_test=True)
place = fluid.CUDAPlace(0) if args.use_cuda == True else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(test_startup)
predict = predict_wrapper(
args,
exe,
ernie_config,
test_prog=test_prog,
pyreader=test_pyreader,
fetch_list=[next_sent_acc.name, mask_lm_loss.name, total_loss.name])
log.info("test begin")
loss, lm_loss, acc, steps, speed = predict()
log.info(
"[test_set] loss: %f, global ppl: %f, next_sent_acc: %f, speed: %f steps/s"
% (np.mean(np.array(loss) / steps),
np.exp(np.mean(np.array(lm_loss) / steps)),
np.mean(np.array(acc) / steps), speed))
def extract_weights(args):
# add ERNIR to environment
print('extract weights start'.center(60, '='))
startup_prog = fluid.Program()
test_prog = fluid.Program()
place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(startup_prog)
args.max_seq_len = 512
args.use_fp16 = False
args.num_labels = 2
args.loss_scaling = 1.0
ernie_config = ErnieConfig(args.ernie_config_path)
ernie_config.print_config()
with fluid.program_guard(test_prog, startup_prog):
with fluid.unique_name.guard():
create_model(args,
pyreader_name='train',
ernie_config=ernie_config)
fluid.io.load_vars(exe,
args.init_pretraining_params,
main_program=test_prog,
predicate=if_exist)
state_dict = collections.OrderedDict()
weight_map = build_weight_map()
for ernie_name, pytorch_name in weight_map.items():
fluid_tensor = fluid.global_scope().find_var(ernie_name).get_tensor()
fluid_array = np.array(fluid_tensor, dtype=np.float32)
if 'w_0' in ernie_name:
fluid_array = fluid_array.transpose()
state_dict[pytorch_name] = fluid_array
print(f'{ernie_name} -> {pytorch_name} {fluid_array.shape}')
print('extract weights done!'.center(60, '='))
return state_dict
def save_model(state_dict, dump_path):
print('save model start'.center(60, '='))
if not os.path.exists(dump_path):
os.makedirs(dump_path)
# save model
for key in state_dict:
state_dict[key] = torch.FloatTensor(state_dict[key])
torch.save(state_dict, os.path.join(dump_path, "pytorch_model.bin"))
print('finish save model')
# save config
ernie_config = ErnieConfig(args.ernie_config_path)._config_dict
# set layer_norm_eps, more detail see: https://github.com/PaddlePaddle/LARK/issues/75
ernie_config['layer_norm_eps'] = 1e-5
with open(os.path.join(dump_path, "config.json"), 'wt',
encoding='utf-8') as f:
json.dump(ernie_config, f, indent=4)
print('finish save config')
# save vocab.txt
vocab_f = open(os.path.join(dump_path, "vocab.txt"),
"wt",
encoding='utf-8')
with open("./ERNIE/config/vocab.txt", "rt", encoding='utf-8') as f:
for line in f:
data = line.strip().split("\t")
vocab_f.writelines(data[0] + "\n")
vocab_f.close()
print('finish save vocab')
print('save model done!'.center(60, '='))
def main(args):
args = parser.parse_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.ExtractEmbeddingReader(
vocab_path=args.vocab_path,
max_seq_len=args.max_seq_len,
do_lower_case=args.do_lower_case)
startup_prog = fluid.Program()
data_generator = reader.data_generator(input_file=args.data_set,
batch_size=args.batch_size,
epoch=1,
shuffle=False)
total_examples = reader.get_num_examples(args.data_set)
print("Device count: %d" % dev_count)
print("Total num examples: %d" % total_examples)
infer_program = fluid.Program()
with fluid.program_guard(infer_program, startup_prog):
with fluid.unique_name.guard():
pyreader, graph_vars = create_model(args,
pyreader_name='reader',
ernie_config=ernie_config)
infer_program = infer_program.clone(for_test=True)
exe.run(startup_prog)
if args.init_pretraining_params:
init_pretraining_params(exe,
args.init_pretraining_params,
main_program=startup_prog)
else:
raise ValueError(
"WARNING: args 'init_pretraining_params' must be specified")
exec_strategy = fluid.ExecutionStrategy()
exec_strategy.num_threads = dev_count
pyreader.set_batch_generator(data_generator)
pyreader.start()
total_cls_emb = []
total_top_layer_emb = []
total_labels = []
while True:
try:
cls_emb, unpad_top_layer_emb = exe.run(
program=infer_program,
fetch_list=[
graph_vars["cls_embeddings"].name,
graph_vars["top_layer_embeddings"].name
],
return_numpy=False)
# batch_size * embedding_size
total_cls_emb.append(np.array(cls_emb))
total_top_layer_emb.append(np.array(unpad_top_layer_emb))
except fluid.core.EOFException:
break
total_cls_emb = np.concatenate(total_cls_emb)
total_top_layer_emb = np.concatenate(total_top_layer_emb)
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
else:
raise RuntimeError('output dir exists: %s' % args.output_dir)
with open(os.path.join(args.output_dir, "cls_emb.npy"),
"wb") as cls_emb_file:
np.save(cls_emb_file, total_cls_emb)
with open(os.path.join(args.output_dir, "top_layer_emb.npy"),
"wb") as top_layer_emb_file:
np.save(top_layer_emb_file, total_top_layer_emb)
def train(args):
log.info("pretraining start")
ernie_config = ErnieConfig(args.ernie_config_path)
ernie_config.print_config()
train_program = fluid.Program()
startup_prog = fluid.Program()
with fluid.program_guard(train_program, startup_prog):
with fluid.unique_name.guard():
train_pyreader, next_sent_acc, mask_lm_loss, total_loss = create_model(
pyreader_name='train_reader', ernie_config=ernie_config)
scheduled_lr, _ = optimization(
loss=total_loss,
warmup_steps=args.warmup_steps,
num_train_steps=args.num_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)
test_prog = fluid.Program()
with fluid.program_guard(test_prog, startup_prog):
with fluid.unique_name.guard():
test_pyreader, next_sent_acc, mask_lm_loss, total_loss = create_model(
pyreader_name='test_reader', ernie_config=ernie_config)
test_prog = test_prog.clone(for_test=True)
if len(fluid.cuda_places()) == 0:
raise RuntimeError('not cuda device cound, check ur env setting')
if args.use_cuda:
place = fluid.cuda_places()[0]
dev_count = fluid.core.get_cuda_device_count()
else:
place = fluid.CPUPlace()
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
log.info("Device count %d" % dev_count)
log.info("theoretical memory usage: ")
log.info(
fluid.contrib.memory_usage(program=train_program,
batch_size=args.batch_size //
args.max_seq_len))
nccl2_num_trainers = 1
nccl2_trainer_id = 0
log.info("args.is_distributed: %s" % args.is_distributed)
if args.is_distributed:
worker_endpoints_env = os.getenv("PADDLE_TRAINER_ENDPOINTS")
worker_endpoints = worker_endpoints_env.split(",")
trainers_num = len(worker_endpoints)
current_endpoint = os.getenv("PADDLE_CURRENT_ENDPOINT")
trainer_id = worker_endpoints.index(current_endpoint)
if trainer_id == 0:
log.info("train_id == 0, sleep 60s")
time.sleep(60)
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,
startup_program=startup_prog)
nccl2_num_trainers = trainers_num
nccl2_trainer_id = trainer_id
exe = fluid.Executor(place)
exe.run(startup_prog)
if args.init_checkpoint and args.init_checkpoint != "":
init_checkpoint(exe, args.init_checkpoint, train_program,
args.use_fp16)
data_reader = ErnieDataReader(filelist=args.train_filelist,
batch_size=args.batch_size,
vocab_path=args.vocab_path,
voc_size=ernie_config['vocab_size'],
epoch=args.epoch,
max_seq_len=args.max_seq_len,
generate_neg_sample=args.generate_neg_sample)
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 = min(10, args.skip_steps)
build_strategy = fluid.BuildStrategy()
build_strategy.remove_unnecessary_lock = False
train_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
loss_name=total_loss.name,
build_strategy=build_strategy,
exec_strategy=exec_strategy,
main_program=train_program,
num_trainers=nccl2_num_trainers,
trainer_id=nccl2_trainer_id)
if args.valid_filelist and args.valid_filelist != "":
predict = predict_wrapper(args,
exe,
ernie_config,
test_prog=test_prog,
pyreader=test_pyreader,
fetch_list=[
next_sent_acc.name, mask_lm_loss.name,
total_loss.name
])
train_pyreader.set_batch_generator(data_reader.data_generator())
train_pyreader.start()
steps = 0
cost = []
lm_cost = []
acc = []
time_begin = time.time()
while steps < args.num_train_steps:
try:
steps += nccl2_num_trainers
skip_steps = args.skip_steps * nccl2_num_trainers
if nccl2_trainer_id != 0:
train_exe.run(fetch_list=[])
continue
if steps % skip_steps != 0:
train_exe.run(fetch_list=[])
else:
each_next_acc, each_mask_lm_cost, each_total_cost, np_lr = train_exe.run(
fetch_list=[
next_sent_acc.name, mask_lm_loss.name, total_loss.name,
scheduled_lr.name
])
acc.extend(each_next_acc)
lm_cost.extend(each_mask_lm_cost)
cost.extend(each_total_cost)
log.info("feed_queue size %d" % train_pyreader.queue.size())
time_end = time.time()
used_time = time_end - time_begin
epoch, current_file_index, total_file, current_file, mask_type = data_reader.get_progress(
)
log.info("current learning_rate:%f" % np_lr[0])
log.info(
"epoch: %d, progress: %d/%d, step: %d, loss: %f, "
"ppl: %f, next_sent_acc: %f, speed: %f steps/s, file: %s, mask_type: %s"
%
(epoch, current_file_index, total_file, steps,
np.mean(np.array(cost)), np.mean(np.exp(
np.array(lm_cost))), np.mean(np.array(acc)),
skip_steps / used_time, current_file, mask_type))
cost = []
lm_cost = []
acc = []
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 args.valid_filelist and steps % args.validation_steps == 0:
vali_cost, vali_lm_cost, vali_acc, vali_steps, vali_speed = predict(
)
log.info(
"[validation_set] epoch: %d, step: %d, "
"loss: %f, global ppl: %f, batch-averged ppl: %f, "
"next_sent_acc: %f, speed: %f steps/s" %
(epoch, steps, np.mean(np.array(vali_cost) / vali_steps),
np.exp(np.mean(np.array(vali_lm_cost) / vali_steps)),
np.mean(np.exp(np.array(vali_lm_cost) / vali_steps)),
np.mean(np.array(vali_acc) / vali_steps), vali_speed))
except fluid.core.EOFException:
train_pyreader.reset()
break
