def load():
""" load model for predict """
config = model_config()
config.vocab_size = len(open(config.vocab_path).readlines())
final_score, final_ids, final_index = knowledge_seq2seq(config)
final_score.persistable = True
final_ids.persistable = True
final_index.persistable = True
main_program = fluid.default_main_program()
if config.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
exe = Executor(place)
exe.run(framework.default_startup_program())
fluid.io.load_params(executor=exe, dirname=config.model_path, main_program=main_program)
processors = KnowledgeCorpus(
data_dir=config.data_dir,
data_prefix=config.data_prefix,
vocab_path=config.vocab_path,
min_len=config.min_len,
max_len=config.max_len)
# load dict
id_dict_array = load_id2str_dict(config.vocab_path)
model_handle = [exe, place, final_score, final_ids, final_index, processors, id_dict_array]
return model_handle
def train(config):
""" model training """
config.vocab_size = len(open(config.vocab_path).readlines())
bow_loss, kl_loss, nll_loss, final_loss= knowledge_seq2seq(config)
bow_loss.persistable = True
kl_loss.persistable = True
nll_loss.persistable = True
final_loss.persistable = True
main_program = fluid.default_main_program()
inference_program = fluid.default_main_program().clone(for_test=True)
fluid.clip.set_gradient_clip(
clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=config.grad_clip))
optimizer = fluid.optimizer.Adam(learning_rate=config.lr)
if config.stage == 0:
print("stage 0")
optimizer.minimize(bow_loss)
else:
print("stage 1")
optimizer.minimize(final_loss)
opt_var_name_list = optimizer.get_opti_var_name_list()
if config.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
exe = Executor(place)
exe.run(framework.default_startup_program())
param_list = main_program.block(0).all_parameters()
param_name_list = [p.name for p in param_list]
init_model(config, param_name_list, place)
processors = KnowledgeCorpus(
data_dir=config.data_dir,
data_prefix=config.data_prefix,
vocab_path=config.vocab_path,
min_len=config.min_len,
max_len=config.max_len)
train_generator = processors.data_generator(
batch_size=config.batch_size,
phase="train",
shuffle=True)
valid_generator = processors.data_generator(
batch_size=config.batch_size,
phase="dev",
shuffle=False)
model_handle = [exe, place, bow_loss, kl_loss, nll_loss, final_loss]
train_loop(config,
train_generator, valid_generator,
main_program, inference_program,
model_handle, param_name_list, opt_var_name_list)
def load():
""" load model for predict """
config = model_config()
config.vocab_size = len(open(config.vocab_path).readlines())
final_score, final_ids, final_index = knowledge_seq2seq(config)
#基于知识库生成对话的回复
final_score.persistable = True
final_ids.persistable = True
final_index.persistable = True
#设置是否为永久变量
main_program = fluid.default_main_program()
def train(config):
""" model training """
config.vocab_size = len(open(config.vocab_path).readlines())
#搭建网络：Bi-GRU Utterance encoder +Bi-GRU KG encoder +层级GRU decoder
bow_loss, kl_loss, nll_loss, final_loss= knowledge_seq2seq(config)
#持久性变量（Persistables）是一种在每次迭代结束后均不会被删除的变量
bow_loss.persistable = True
kl_loss.persistable = True
nll_loss.persistable = True
final_loss.persistable = True
#fluid.layers 接口中添加的op和variable会存储在 default main program 中
main_program = fluid.default_main_program()
inference_program = fluid.default_main_program().clone(for_test=True)
#给指定参数做梯度裁剪
fluid.clip.set_gradient_clip(
clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=config.grad_clip))
optimizer = fluid.optimizer.Adam(learning_rate=config.lr)
if config.stage == 0:
print("stage 0")
optimizer.minimize(bow_loss)
else:
print("stage 1")
optimizer.minimize(final_loss)
#优化器的训练参数如lr
opt_var_name_list = optimizer.get_opti_var_name_list()
if config.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
exe = Executor(place)
#初始化 default_startup_program函数可以获取默认/全局 startup Program (初始化启动程序)。
exe.run(framework.default_startup_program())
#block0 表示一段代码的最外层块
param_list = main_program.block(0).all_parameters()
param_name_list = [p.name for p in param_list]
#TODO：init包含
init_model(config, param_name_list, place)
processors = KnowledgeCorpus(
data_dir=config.data_dir,
data_prefix=config.data_prefix,
vocab_path=config.vocab_path,
min_len=config.min_len,
max_len=config.max_len)
#train_generator为yeild 生成函数
#进行了如下操作：
#读取stream record file
#对读入的文本进行tokennize；根据max min len 进行过滤
#并根据词汇表把src\tgt\cue文本串（chatpath+knowledge+":"+history\ response\ KG cue）转为数字串
##进行padding并返回padding后的串和每个串的原长
train_generator = processors.train_generator(
batch_size=config.batch_size,
phase="train",
shuffle=True)
valid_generator = processors.data_generator(
batch_size=config.batch_size,
phase="dev",
shuffle=False)
model_handle = [exe, place, bow_loss, kl_loss, nll_loss, final_loss]
#在训练过程中，每config.valid_steps 个batch（步）进行一次valid，并储存一次最好模型
train_loop(config,
train_generator, valid_generator,
main_program, inference_program,
model_handle, param_name_list, opt_var_name_list)
def test(config):
""" test """
batch_size = config.batch_size
config.vocab_size = len(open(config.vocab_path).readlines())
final_score, final_ids, final_index = knowledge_seq2seq(config)
final_score.persistable = True
final_ids.persistable = True
final_index.persistable = True
main_program = fluid.default_main_program()
if config.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
exe = Executor(place)
exe.run(framework.default_startup_program())
fluid.io.load_params(executor=exe, dirname=config.model_path,
main_program=main_program)
print("laod params finsihed")
# test data generator
processors = KnowledgeCorpus(
data_dir=config.data_dir,
data_prefix=config.data_prefix,
vocab_path=config.vocab_path,
min_len=config.min_len,
max_len=config.max_len)
test_generator = processors.data_generator(
batch_size=config.batch_size,
phase="test",
shuffle=False)
# load dict
id_dict_array = load_id2str_dict(config.vocab_path)
out_file = config.output
fout = open(out_file, 'w')
for batch_id, data in enumerate(test_generator()):
data_feed, sent_num = build_data_feed(data, place, batch_size=batch_size)
if data_feed is None:
break
out = exe.run(feed=data_feed,
fetch_list=[final_score.name, final_ids.name, final_index.name])
batch_score = out[0]
batch_ids = out[1]
batch_pre_index = out[2]
batch_score_arr = np.split(batch_score, batch_size, axis=1)
batch_ids_arr = np.split(batch_ids, batch_size, axis=1)
batch_pre_index_arr = np.split(batch_pre_index, batch_size, axis=1)
index = 0
for (score, ids, pre_index) in zip(batch_score_arr, batch_ids_arr, batch_pre_index_arr):
trace_ids, trace_score = trace_fianl_result(score, ids, pre_index, topk=1, EOS=3)
fout.write(id_to_text(trace_ids[0][:-1], id_dict_array))
fout.write('\n')
index += 1
if index >= sent_num:
break
fout.close()
def train(config):
""" model training """
config.vocab_size = len(open(config.vocab_path).readlines())
#计算行数，即词的个数
bow_loss, kl_loss, nll_loss, final_loss= knowledge_seq2seq(config)
bow_loss.persistable = True
kl_loss.persistable = True
nll_loss.persistable = True
final_loss.persistable = True
main_program = fluid.default_main_program()
inference_program = fluid.default_main_program().clone(for_test=True)
#Program是Paddle Fluid对于计算图的一种静态描述
#1. Program.clone() 方法不会克隆例如 DataLoader 这样的数据读取相关的部分，这可能会造成的数据读取部分在克隆后丢失
#2. 此API当 for_test=True 时将会裁剪部分OP和变量。为防止错误的裁剪，推荐在 append_backward 和执行优化器之前使用 clone(for_test=True) 。
#当 for_test=True 时创建一个新的、仅包含当前Program前向内容的Program。否则创建一个新的，和当前Program完全相同的Program
fluid.clip.set_gradient_clip(
clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=config.grad_clip))
optimizer = fluid.optimizer.Adam(learning_rate=config.lr)
#给指定参数做梯度裁剪。
# 此API对位置使用的要求较高，其必须位于组建网络之后， minimize 之前，因此在未来版本中可能被删除，故不推荐使用。
# 推荐在 optimizer 初始化时设置梯度裁剪。 有三种裁剪策略： GradientClipByGlobalNorm 、 GradientClipByNorm 、
# GradientClipByValue 。 如果在 optimizer 中设置过梯度裁剪，又使用了 set_gradient_clip ，set_gradient_clip 将不会生效。
#深度神经网络采用链式法则进行参数求导的方式并不是绝对安全的，有时会出现梯度消失或者梯度爆炸的情况。
# 其中梯度爆炸问题的常见应对方式为“梯度裁剪”，也就是通过“clip”方式来防止迭代中梯度值过大。
if config.stage == 0:
print("stage 0")
optimizer.minimize(bow_loss)
else:
print("stage 1")
optimizer.minimize(final_loss)
#指定优化器最小化的loss
opt_var_name_list = optimizer.get_opti_var_name_list()
if config.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
#Gpu or Cpu
exe = Executor(place)
exe.run(framework.default_startup_program())
#执行程序
param_list = main_program.block(0).all_parameters()
#Block 是高级语言中变量作用域的概念，在编程语言中，Block是一对大括号，其中包含局部变量定义和一系列指令或
# 操作符。编程语言中的控制流结构 if-else 和 for 在深度学习中可以被等效为：
# Fluid 中的 Block 描述了一组以顺序、选择或是循环执行的 Operator 以及 Operator 操作的对象：Tensor。
param_name_list = [p.name for p in param_list]
#模型列表
init_model(config, param_name_list, place)
processors = KnowledgeCorpus(
data_dir=config.data_dir,
data_prefix=config.data_prefix,
vocab_path=config.vocab_path,
min_len=config.min_len,
max_len=config.max_len)
#知识库对象
train_generator = processors.data_generator(
batch_size=config.batch_size,
phase="train",
shuffle=True)
#训练数据生成器
valid_generator = processors.data_generator(
batch_size=config.batch_size,
phase="dev",
shuffle=False)
#验证数据生成器
model_handle = [exe, place, bow_loss, kl_loss, nll_loss, final_loss]
train_loop(config,
train_generator, valid_generator,
main_program, inference_program,
model_handle, param_name_list, opt_var_name_list)
