4.1 数据预处理
#导入必要的包
import zipfile
import os
import io
import random
import json
import matplotlib.pyplot as plt
import numpy as np
import paddle
import paddle.fluid as fluid
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, Linear, Embedding
from paddle.fluid.dygraph.base import to_variable
from paddle.fluid.dygraph import GRUUnit
import paddle.dataset.imdb as imdb
#加载字典
def load_vocab():
vocab = imdb.word_dict()
return vocab
#定义数据生成器
class SentaProcessor(object):
def __init__(self):
self.vocab = load_vocab()
def data_generator(self, batch_size, phase='train'):
if phase == "train":
return paddle.batch(paddle.reader.shuffle(imdb.train(self.vocab),25000), batch_size, drop_last=True)
elif phase == "eval":
return paddle.batch(imdb.test(self.vocab), batch_size,drop_last=True)
else:
raise ValueError(
"Unknown phase, which should be in ['train', 'eval']")
步骤
-
首先导入必要的第三方库
-
接下来就是数据预处理，需要注意的是：数据是以数据标签的方式表示一个句子，因此，每个句子都是以一串整数来表示的，每个数字都是对应一个单词。当然，数据集就会有一个数据集字典，这个字典是训练数据中出现单词对应的数字标签。
4.2 构建网络
这次的GRU模型分为以下的几个步骤
- 定义网络
- 定义损失函数
- 定义优化算法
具体实现如下
#定义动态GRU
class DynamicGRU(fluid.dygraph.Layer):
def init(self,
size,
param_attr=None,
bias_attr=None,
is_reverse=False,
gate_activation=‘sigmoid’,
candidate_activation=‘relu’,
h_0=None,
origin_mode=False,
):
super(DynamicGRU, self).init()
self.gru_unit = GRUUnit(
size * 3,
param_attr=param_attr,
bias_attr=bias_attr,
activation=candidate_activation,
gate_activation=gate_activation,
origin_mode=origin_mode)
self.size = size
self.h_0 = h_0
self.is_reverse = is_reverse
def forward(self, inputs):
hidden = self.h_0
res = []
for i in range(inputs.shape[1]):
if self.is_reverse:
i = inputs.shape[1] - 1 - i
input_ = inputs[ :, i:i+1, :]
input_ = fluid.layers.reshape(input_, [-1, input_.shape[2]], inplace=False)
hidden, reset, gate = self.gru_unit(input_, hidden)
hidden_ = fluid.layers.reshape(hidden, [-1, 1, hidden.shape[1]], inplace=False)
res.append(hidden_)
if self.is_reverse:
res = res[::-1]
res = fluid.layers.concat(res, axis=1)
return res
class GRU(fluid.dygraph.Layer):
def __init__(self):
super(GRU, self).__init__()
self.dict_dim = train_parameters["vocab_size"]
self.emb_dim = 128
self.hid_dim = 128
self.fc_hid_dim = 96
self.class_dim = 2
self.batch_size = train_parameters["batch_size"]
self.seq_len = train_parameters["padding_size"]
self.embedding = Embedding(
size=[self.dict_dim + 1, self.emb_dim],
dtype='float32',
param_attr=fluid.ParamAttr(learning_rate=30),
is_sparse=False)
h_0 = np.zeros((self.batch_size, self.hid_dim), dtype="float32")
h_0 = to_variable(h_0)
self._fc1 = Linear(input_dim=self.hid_dim, output_dim=self.hid_dim*3)
self._fc2 = Linear(input_dim=self.hid_dim, output_dim=self.fc_hid_dim, act="relu")
self._fc_prediction = Linear(input_dim=self.fc_hid_dim,
output_dim=self.class_dim,
act="softmax")
self._gru = DynamicGRU(size=self.hid_dim, h_0=h_0)
def forward(self, inputs, label=None):
emb = self.embedding(inputs)
o_np_mask =to_variable(inputs.numpy().reshape(-1,1) != self.dict_dim).astype('float32')
mask_emb = fluid.layers.expand(
to_variable(o_np_mask), [1, self.hid_dim])
emb = emb * mask_emb
emb = fluid.layers.reshape(emb, shape=[self.batch_size, -1, self.hid_dim])
fc_1 = self._fc1(emb)
gru_hidden = self._gru(fc_1)
gru_hidden = fluid.layers.reduce_max(gru_hidden, dim=1)
tanh_1 = fluid.layers.tanh(gru_hidden)
fc_2 = self._fc2(tanh_1)
prediction = self._fc_prediction(fc_2)
if label is not None:
acc = fluid.layers.accuracy(prediction, label=label)
return prediction, acc
else:
return prediction
4.3 训练模型
def train():
with fluid.dygraph.guard(place = fluid.CUDAPlace(0)): # # 因为要进行很大规模的训练，因此我们用的是GPU，如果没有安装GPU的可以使用下面一句，把这句代码注释掉即可
# with fluid.dygraph.guard(place = fluid.CPUPlace()):
processor = SentaProcessor()
train_data_generator = processor.data_generator(batch_size=train_parameters["batch_size"], phase='train')
model = GRU()
sgd_optimizer = fluid.optimizer.Adagrad(learning_rate=train_parameters["lr"],parameter_list=model.parameters())
steps = 0
Iters, total_loss, total_acc = [], [], []
for eop in range(train_parameters["epoch"]):
for batch_id, data in enumerate(train_data_generator()):
steps += 1
doc = to_variable(
np.array([
np.pad(x[0][0:train_parameters["padding_size"]],
(0, train_parameters["padding_size"] - len(x[0][0:train_parameters["padding_size"]])),
'constant',
constant_values=(train_parameters["vocab_size"]))
for x in data
]).astype('int64').reshape(-1))
label = to_variable(
np.array([x[1] for x in data]).astype('int64').reshape(
train_parameters["batch_size"], 1))
model.train()
prediction, acc = model(doc, label)
loss = fluid.layers.cross_entropy(prediction, label)
avg_loss = fluid.layers.mean(loss)
avg_loss.backward()
sgd_optimizer.minimize(avg_loss)
model.clear_gradients()
if steps % train_parameters["skip_steps"] == 0:
Iters.append(steps)
total_loss.append(avg_loss.numpy()[0])
total_acc.append(acc.numpy()[0])
print("step: %d, ave loss: %f, ave acc: %f" %
(steps,avg_loss.numpy(),acc.numpy()))
if steps % train_parameters["save_steps"] == 0:
save_path = train_parameters["checkpoints"]+"/"+"save_dir_" + str(steps)
print('save model to: ' + save_path)
fluid.dygraph.save_dygraph(model.state_dict(),
save_path)
draw_train_process(Iters, total_loss, total_acc)
4.4 模型评估
结果还可以，这里说明的是，刚开始的模型训练评估不可能这么好，很明显是过拟合的问题，这就需要我们调整我们的epoch、batchsize、激活函数的选择以及优化器、学习率等各种参数，通过不断的调试、训练最好可以得到不错的结果，但是，如果还要更好的模型效果，其实可以将GRU模型换为更为合适的RNN中的LSTM以及bi-
LSTM模型会好很多。
4.5 模型预测
train_parameters[“batch_size”] = 1
with fluid.dygraph.guard(place = fluid.CUDAPlace(0)):
sentences = 'this is a great movie'
data = load_data(sentences)
print(sentences)
print(data)
data_np = np.array(data)
data_np = np.array(np.pad(data_np,(0,150-len(data_np)),"constant",constant_values =train_parameters["vocab_size"])).astype('int64').reshape(-1)
infer_np_doc = to_variable(data_np)
model_infer = GRU()
model, _ = fluid.load_dygraph("data/save_dir_750.pdparams")
model_infer.load_dict(model)
model_infer.eval()
result = model_infer(infer_np_doc)
print('预测结果为：正面概率为：%0.5f，负面概率为：%0.5f' % (result.numpy()[0][0],result.numpy()[0][1]))
训练的结果还是挺满意的，到此为止，我们的本次项目实验到此结束。