def _load_process(self, executor, main_prog):
"""
读取模型的过程，
如果想从零开始，请将load_model_path设为空字符串，且read_checkpoint，continue_train为false
如果想从预训练模型（或某个基线）开始训练，请设置continue_train为False，
如果想继续训练，请设置请设置continue_train为True，
如果想从断点训练，请设置read_checkpoint为true。
:param executor:
:param main_prog:
:return: 字典，保存当前训练状态, 将保存在self.train_status中
"""
CONTINUE = self.args["continue_train"]
MODEL_PATH = self.args["load_model_path"]
CHECK_POINT = self.args["read_checkpoint"]
total_step = 0
step_in_epoch = 0
total_epoch = 1
# 读取模型现有的参数并为继续训练进行相应处理
if CONTINUE and CHECK_POINT:
info = model_utils.load_train_snapshot(executor, main_prog, MODEL_PATH)
self.logger.info("Model file in {} has been loaded".format(MODEL_PATH))
if info:
total_step = info.get("total_step", 0)
step_in_epoch = info.get("step_in_epoch", 0)
total_epoch = info.get("epoch", 1)
self.logger.info("Load train info: {}".format(info))
elif MODEL_PATH != "":
# 若是第一次训练且预训练模型参数不为空，则加载预训练模型参数
model_utils.load_model_params(exe=executor, program=main_prog, params_path=MODEL_PATH)
self.logger.info("Pre-trained model file in {} has been loaded".format(MODEL_PATH))
return {'total_step': total_step, 'total_epoch': total_epoch, 'step_in_epoch': step_in_epoch}
def init_model(self):
"""
根据模型参数路径读入模型来初始化，包括预测程序编译，模型参数赋值，并行策略
:param vocab_size: 词典大小
:return:
"""
model_path = self.args["load_model_path"]
self.logger.info("Initializing predict model...")
self.exe = fluid.Executor(
TrainEngine.get_executor_run_places(self.args))
with fluid.program_guard(self.predict_program, self.predict_startup):
# 根据gzl的模型来定义网络，输出占位符
loader, probs, qas_id = classifier.create_model_for_cls_merge(
args=self.args_model_build, is_prediction=True)
self.logger.info("Prediction neural network created.")
self.logger.info("Prediction neural network parameter initialized.")
# start_up程序运行初始参数
self.exe.run(self.predict_startup)
# 加载模型参数到网络中
load_model_params(self.exe, model_path, self.predict_program)
# 若并行，用并行编译program
if self.args["use_parallel"]:
build_strategy = fluid.BuildStrategy()
# 并行策略暂时写死
build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
self.predict_program = fluid.CompiledProgram(self.predict_program). \
with_data_parallel(places=TrainEngine.get_data_run_places(self.args),
build_strategy=build_strategy)
self.logger.info("Finish initializing predict model!")
return loader, probs, qas_id
def _load_process(self, executor, main_prog):
"""
读取模型的过程
:param executor:
:param main_prog:
:return: 字典，保存当前训练状态, 将保存在self.train_status中
"""
MODEL_PATH = self.args["load_model_path"]
model_utils.load_model_params(exe=executor, program=main_prog, params_path=MODEL_PATH)
self.logger.info("Pre-trained model file in {} has been loaded".format(MODEL_PATH))
def _load_process(self, executor, main_prog):
"""
读取模型的过程，
如果想从零开始，请将load_model_path设为空字符串，且read_checkpoint，continue_train为false
如果想从预训练模型（或某个基线）开始训练，请设置continue_train为False，
如果想继续训练，请设置请设置continue_train为True，
如果想从断点训练，请设置read_checkpoint为true。
:param executor:
:param main_prog:
:return: 字典，保存当前训练状态, 将保存在self.train_status中
"""
MODEL_PATH = self.args["load_model_path"]
model_utils.load_model_params(exe=executor, program=main_prog, params_path=MODEL_PATH)
self.logger.info("Pre-trained model file in {} has been loaded".format(MODEL_PATH))
return
def train(self):
"""
用于训练流程，根据参数完成训练，并使用验证数据对模型效果进行验证
:return: 无
"""
APP_NAME = self.args["app_name"]
MAX_EPOCH = self.args["max_epoch"]
SNAPSHOT_FREQUENCY = self.args["snapshot_frequency"]
EARLY_STOPPING = self.args["early_stopping"]
if EARLY_STOPPING:
THRESHOLD = self.args["early_stopping_threshold"]
STANDSTILL_STEP = self.args["early_stopping_stand_times"]
CONTINUE = self.args["continue_train"]
MODEL_PATH = self.args["load_model_path"]
CHECK_POINT = self.args["read_checkpoint"]
# 定义执行器
executor = fluid.Executor(self.get_executor_run_places(self.args))
# 执行初始化
executor.run(self.train_startup_prog)
total_step = 0
step_in_epoch = 0
total_epoch = 0
# 读取模型现有的参数并为继续训练进行相应处理
if MODEL_PATH != "":
# 若是第一次训练且预训练模型参数不为空，则加载预训练模型参数
model_utils.load_model_params(exe=executor,
program=self.origin_train_prog,
params_path=MODEL_PATH)
self.logger.info(
"Pre-trained model file in {} has been loaded".format(
MODEL_PATH))
self.logger.info("Ready to train the model.Executing...")
self.__run_train_iterable(executor, total_step, 0, 0)
def train(self):
"""
用于训练流程，根据参数完成训练，并使用验证数据对模型效果进行验证
:return: 无
"""
APP_NAME = self.args["app_name"]
MAX_EPOCH = self.args["max_epoch"]
SNAPSHOT_FREQUENCY = self.args["snapshot_frequency"]
EARLY_STOPPING = self.args["early_stopping"]
if EARLY_STOPPING:
THRESHOLD = self.args["early_stopping_threshold"]
STANDSTILL_STEP = self.args["early_stopping_stand_times"]
CONTINUE = self.args["continue_train"]
if CONTINUE:
MODEL_PATH = self.args["load_model_path"]
PRETRAIN_MODEL = self.args["pretrained_model_path"]
# 定义执行器
executor = fluid.Executor(self.get_executor_run_places(self.args))
# 执行初始化
executor.run(self.train_startup_prog)
total_step = 0
step_in_epoch = 0
total_epoch = 0
# 读取模型现有的参数并为继续训练进行相应处理
if CONTINUE:
info = model_utils.load_train_snapshot(executor,
self.origin_train_prog,
MODEL_PATH)
self.logger.info(
"Model file in {} has been loaded".format(MODEL_PATH))
if info:
total_step == info.get("total_step", 0)
step_in_epoch = info.get("step_in_epoch", 0)
total_epoch = info.get("epoch", 0)
elif PRETRAIN_MODEL != "":
# 若是第一次训练且预训练模型参数不为空，则加载预训练模型参数
model_utils.load_model_params(exe=executor,
program=self.origin_train_prog,
params_path=PRETRAIN_MODEL)
self.logger.info(
"Pre-trained model file in {} has been loaded".format(
PRETRAIN_MODEL))
self.logger.info("Ready to train the model.Executing...")
# 执行MAX_EPOCH次迭代save_train_snapshot
for epoch_id in range(MAX_EPOCH):
if epoch_id == 0:
epoch_id += total_epoch
# 一个epoch的训练过程，一个迭代
total_step, loss = self.__run_train_iterable(
executor, total_step, epoch_id, step_in_epoch)
step_in_epoch = 0
self.logger.info(
'Epoch {epoch} done, train mean loss is {loss}'.format(
epoch=epoch_id, loss=loss))
# 进行一次验证集上的验证
info = {"total_step": total_step, "epoch": epoch_id}
file_path = model_utils.save_train_snapshot(
executor,
self.origin_train_prog,
file_name="{}_epoch{}".format(APP_NAME, epoch_id),
train_info=info)
self.logger.info(
"Snapshot of training process has been saved as folder {}".
format(file_path))
# 保存现有模型
file_path = model_utils.save_train_snapshot(executor,
self.origin_train_prog,
APP_NAME)
self.logger.info(
"Training process completed. model saved in {}".format(file_path))
def train(self):
"""
用于训练流程，根据参数完成训练，并使用验证数据对模型效果进行验证
:return: 无
"""
APP_NAME = self.args["app_name"]
MAX_EPOCH = self.args["max_epoch"]
SNAPSHOT_FREQUENCY = self.args["snapshot_frequency"]
EARLY_STOPPING = self.args["early_stopping"]
if EARLY_STOPPING:
THRESHOLD = self.args["early_stopping_threshold"]
STANDSTILL_STEP = self.args["early_stopping_stand_times"]
CONTINUE = self.args["continue_train"]
MODEL_PATH = self.args["load_model_path"]
CHECK_POINT = self.args["read_checkpoint"]
# 定义执行器
executor = fluid.Executor(self.get_executor_run_places(self.args))
# 执行初始化
executor.run(self.train_startup_prog)
total_step = 0
step_in_epoch = 0
total_epoch = 0
# 读取模型现有的参数并为继续训练进行相应处理
if CONTINUE and CHECK_POINT:
info = model_utils.load_train_snapshot(executor,
self.origin_train_prog,
MODEL_PATH)
self.logger.info(
"Model file in {} has been loaded".format(MODEL_PATH))
if info:
total_step = info.get("total_step", 0)
step_in_epoch = info.get("step_in_epoch", 0)
total_epoch = info.get("epoch", 0)
self.logger.info("Load train info: {}".format(info))
elif MODEL_PATH != "":
# 若是第一次训练且预训练模型参数不为空，则加载预训练模型参数
model_utils.load_model_params(exe=executor,
program=self.origin_train_prog,
params_path=MODEL_PATH)
self.logger.info(
"Pre-trained model file in {} has been loaded".format(
MODEL_PATH))
self.logger.info("Ready to train the model.Executing...")
# 执行MAX_EPOCH次迭代save_train_snapshot
for epoch_id in range(total_epoch, MAX_EPOCH):
# 一个epoch的训练过程，一个迭代
total_step, loss = self.__run_train_iterable(
executor, total_step, epoch_id, step_in_epoch)
step_in_epoch = 0
self.logger.info(
'Epoch {epoch} done, train mean loss is {loss}'.format(
epoch=epoch_id, loss=loss))
# 进行一次验证集上的验证
valid_loss, valid_acc = self.__valid(executor)
self.logger.info(' Epoch {epoch} Validated'.format(epoch=epoch_id))
# 进行保存
info = {"total_step": total_step, "epoch": epoch_id}
file_path = model_utils.save_train_snapshot(
executor,
self.origin_train_prog,
file_name="{}_epoch{}".format(APP_NAME, epoch_id),
train_info=info)
self.logger.info(
"Snapshot of training process has been saved as folder {}".
format(file_path))
# 应用早停策略
if EARLY_STOPPING:
need_stop = self.early_stopping_strategy(
-valid_acc,
threshold=THRESHOLD,
standstill_step=STANDSTILL_STEP)
if need_stop:
self.logger.info(
"Performance improvement stalled, ending the training process"
)
break
# 保存现有模型
file_path = model_utils.save_train_snapshot(executor,
self.origin_train_prog,
APP_NAME)
self.logger.info(
"Training process completed. model saved in {}".format(file_path))