在循环神经网络中,神经元不但可以接受其他神经元的信息,也可以接受自身的信息,形成具有环路的网络结构.
和前馈神经网络相比,循环神经网络更加符合生物神经网络的结构.
目前,循环神经网络已经被广泛应用在语音识别、语言模型以及自然语言生成等任务上.
简单循环网络在参数学习时存在长程依赖问题,很难建模长时间间隔(Long Range)的状态之间的依赖关系。
为了测试简单循环网络的记忆能力,本节构建一个【数字求和任务】进行实验。
数字求和任务的输入是一串数字,前两个位置的数字为0-9,其余数字随机生成(主要为0),预测目标是输入序列中前两个数字的加和。图6.3展示了长度为10的数字序列.
循环神经网络的一种简单实现是简单循环网络(Simple Recurrent Network,SRN)
简单循环网络在参数学习时存在长程依赖问题,很难建模长时间间隔(Long Range)的状态之间的依赖关系。为了测试简单循环网络的记忆能力,本节构建一个数字求和任务进行实验。
数字求和任务的输入是一串数字,前两个位置的数字为0-9,其余数字随机生成(主要为0),预测目标是输入序列中前两个数字的加和。图6.3展示了长度为10的数字序列.
如果序列长度越长,准确率越高,则说明网络的记忆能力越好.因此,我们可以构建不同长度的数据集,通过验证简单循环网络在不同长度的数据集上的表现,从而测试简单循环网络的长程依赖能力.
构建不同长度的数字预测数据集DigitSum
由于在本任务中,输入序列的前两位数字为 0 − 9,其组合数是固定的,所以可以穷举所有的前两位数字组合,并在后面默认用0填充到固定长度. 但考虑到数据的多样性,这里对生成的数字序列中的零位置进行随机采样,并将其随机替换成0-9的数字以增加样本的数量.
我们可以通过设置k
的数值来指定一条样本随机生成的数字序列数量.当生成某个指定长度的数据集时,会同时生成训练集、验证集和测试集。当k=3时,生成训练集。当k=1时,生成验证集和测试集. 代码实现如下:
import random
import numpy as np
# 固定随机种子
random.seed(0)
np.random.seed(0)
def generate_data(length, k, save_path):
if length < 3:
raise ValueError("The length of data should be greater than 2.")
if k == 0:
raise ValueError("k should be greater than 0.")
# 生成100条长度为length的数字序列,除前两个字符外,序列其余数字暂用0填充
base_examples = []
for n1 in range(0, 10):
for n2 in range(0, 10):
seq = [n1, n2] + [0] * (length - 2)
label = n1 + n2
base_examples.append((seq, label))
examples = []
# 数据增强:对base_examples中的每条数据,默认生成k条数据,放入examples
for base_example in base_examples:
for _ in range(k):
# 随机生成替换的元素位置和元素
idx = np.random.randint(2, length)
val = np.random.randint(0, 10)
# 对序列中的对应零元素进行替换
seq = base_example[0].copy()
label = base_example[1]
seq[idx] = val
examples.append((seq, label))
# 保存增强后的数据
with open(save_path, "w", encoding="utf-8") as f:
for example in examples:
# 将数据转为字符串类型,方便保存
seq = [str(e) for e in example[0]]
label = str(example[1])
line = " ".join(seq) + "\t" + label + "\n"
f.write(line)
print(f"generate data to: {save_path}.")
# 定义生成的数字序列长度
lengths = [5, 10, 15, 20, 25, 30, 35]
for length in lengths:
# 生成长度为length的训练数据
save_path = f"./datasets/{length}/train.txt"
k = 3
generate_data(length, k, save_path)
# 生成长度为length的验证数据
save_path = f"./datasets/{length}/dev.txt"
k = 1
generate_data(length, k, save_path)
# 生成长度为length的测试数据
save_path = f"./datasets/{length}/test.txt"
k = 1
generate_data(length, k, save_path)
为方便使用,本实验提前生成了长度分别为5、10、 15、20、25、30和35的7份数据,存放于“./datasets”目录下,读者可以直接加载使用。代码实现如下:
import os
# 加载数据
def load_data(data_path):
# 加载训练集
train_examples = []
train_path = os.path.join(data_path, "train.txt")
with open(train_path, "r", encoding="utf-8") as f:
for line in f.readlines():
# 解析一行数据,将其处理为数字序列seq和标签label
items = line.strip().split("\t")
seq = [int(i) for i in items[0].split(" ")]
label = int(items[1])
train_examples.append((seq, label))
# 加载验证集
dev_examples = []
dev_path = os.path.join(data_path, "dev.txt")
with open(dev_path, "r", encoding="utf-8") as f:
for line in f.readlines():
# 解析一行数据,将其处理为数字序列seq和标签label
items = line.strip().split("\t")
seq = [int(i) for i in items[0].split(" ")]
label = int(items[1])
dev_examples.append((seq, label))
# 加载测试集
test_examples = []
test_path = os.path.join(data_path, "test.txt")
with open(test_path, "r", encoding="utf-8") as f:
for line in f.readlines():
# 解析一行数据,将其处理为数字序列seq和标签label
items = line.strip().split("\t")
seq = [int(i) for i in items[0].split(" ")]
label = int(items[1])
test_examples.append((seq, label))
return train_examples, dev_examples, test_examples
# 设定加载的数据集的长度
length = 5
# 该长度的数据集的存放目录
data_path = f"./datasets/{length}"
# 加载该数据集
train_examples, dev_examples, test_examples = load_data(data_path)
print("dev example:", dev_examples[:2])
print("训练集数量:", len(train_examples))
print("验证集数量:", len(dev_examples))
print("测试集数量:", len(test_examples))
为了方便使用梯度下降法进行优化,我们构造了DigitSum数据集的Dataset类,函数__getitem__负责根据索引读取数据,并将数据转换为张量。代码实现如下:
from torch.utils.data import Dataset,DataLoader
import torch
class DigitSumDataset(Dataset):
def __init__(self, data):
self.data = data
def __getitem__(self, idx):
example = self.data[idx]
seq = torch.tensor(example[0], dtype="int64")
label = torch.tensor(example[1], dtype="int64")
return seq, label
def __len__(self):
return len(self.data)
使用SRN模型进行数字加和任务的模型结构为如图6.4所示.
整个模型由以下几个部分组成:
(1) 嵌入层:将输入的数字序列进行向量化,即将每个数字映射为向量;
(2) SRN 层:接收向量序列,更新循环单元,将最后时刻的隐状态作为整个序列的表示;
(3) 输出层:一个线性层,输出分类的结果.
本任务输入的样本是数字序列,为了更好地表示数字,需要将数字映射为一个嵌入(Embedding)向量。嵌入向量中的每个维度均能用来刻画该数字本身的某种特性。由于向量能够表达该数字更多的信息,利用向量进行数字求和任务,可以使得模型具有更强的拟合能力。
提醒:为了和代码的实现保持一致性,这里使用形状为(样本数量×序列长度×特征维度)的张量来表示一组样本。
基于索引方式的嵌入层的实现如下:
class Embedding(nn.Module):
def __init__(self, num_embeddings, embedding_dim):
super(Embedding, self).__init__()
self.W = nn.init.xavier_uniform_(torch.empty(num_embeddings, embedding_dim),gain=1.0)
def forward(self, inputs):
# 根据索引获取对应词向量
embs = self.W[inputs]
return embs
emb_layer = Embedding(10, 5)
inputs = torch.tensor([0, 1, 2, 3])
emb_layer(inputs)
import torch
import torch.nn as nn
import torch.nn.functional as F
torch.manual_seed(0)
# SRN模型
class SRN(nn.Module):
def __init__(self, input_size, hidden_size, W_attr=None, U_attr=None, b_attr=None):
super(SRN, self).__init__()
# 嵌入向量的维度
self.input_size = input_size
# 隐状态的维度
self.hidden_size = hidden_size
# 定义模型参数W,其shape为 input_size x hidden_size
if W_attr==None:
W=torch.zeros(size=[input_size, hidden_size], dtype=torch.float32)
else:
W=torch.tensor(W_attr,dtype=torch.float32)
self.W = torch.nn.Parameter(W)
# 定义模型参数U,其shape为hidden_size x hidden_size
if U_attr==None:
U=torch.zeros(size=[hidden_size, hidden_size], dtype=torch.float32)
else:
U=torch.tensor(U_attr,dtype=torch.float32)
self.U = torch.nn.Parameter(U)
# 定义模型参数b,其shape为 1 x hidden_size
if b_attr==None:
b=torch.zeros(size=[1, hidden_size], dtype=torch.float32)
else:
b=torch.tensor(b_attr,dtype=torch.float32)
self.b = torch.nn.Parameter(b)
# 初始化向量
def init_state(self, batch_size):
hidden_state = torch.zeros(size=[batch_size, self.hidden_size], dtype=torch.float32)
return hidden_state
# 定义前向计算
def forward(self, inputs, hidden_state=None):
# inputs: 输入数据, 其shape为batch_size x seq_len x input_size
batch_size, seq_len, input_size = inputs.shape
# 初始化起始状态的隐向量, 其shape为 batch_size x hidden_size
if hidden_state is None:
hidden_state = self.init_state(batch_size)
# 循环执行RNN计算
for step in range(seq_len):
# 获取当前时刻的输入数据step_input, 其shape为 batch_size x input_size
step_input = inputs[:, step, :]
# 获取当前时刻的隐状态向量hidden_state, 其shape为 batch_size x hidden_size
hidden_state = F.tanh(torch.matmul(step_input, self.W) + torch.matmul(hidden_state, self.U) + self.b)
return hidden_state
## 初始化参数并运行
U_attr = [[0.0, 0.1], [0.1,0.0]]
b_attr = [[0.1, 0.1]]
W_attr=[[0.1, 0.2], [0.1,0.2]]
srn = SRN(2, 2, W_attr=W_attr, U_attr=U_attr, b_attr=b_attr)
inputs = torch.tensor([[[1, 0],[0, 2]]], dtype=torch.float32)
hidden_state = srn(inputs)
print("hidden_state", hidden_state)
运行结果:
hidden_state tensor([[0.3177, 0.4775]], grad_fn=<TanhBackward>)
另外,内置SRN API在执行完前向计算后,会返回两个参数:序列向量和最后时刻的隐状态向量。在飞桨实现时,考虑到了双向和多层SRN的因素,返回的向量附带了这些信息。
其中序列向量outputs是指最后一层SRN的输出向量,其shape为[batch_size, seq_len, num_directions * hidden_size];最后时刻的隐状态向量shape为[num_layers * num_directions, batch_size, hidden_size]。
这里我们可以将自己实现的SRN和Paddle框架内置的SRN返回的结果进行打印展示,实现代码如下。
## 初始化参数并运行
U_attr = [[0.0, 0.1], [0.1,0.0]]
b_attr = [[0.1, 0.1]]
W_attr=[[0.1, 0.2], [0.1,0.2]]
srn = SRN(2, 2, W_attr=W_attr, U_attr=U_attr, b_attr=b_attr)
inputs = torch.tensor([[[1, 0],[0, 2]]], dtype=torch.float32)
hidden_state = srn(inputs)
print("hidden_state", hidden_state)
# 这里创建一个随机数组作为测试数据,数据shape为batch_size x seq_len x input_size
batch_size, seq_len, input_size = 8, 20, 32
inputs = torch.randn([batch_size, seq_len, input_size])
# 设置模型的hidden_size
hidden_size = 32
torch_srn = nn.RNN(input_size, hidden_size)
self_srn = SRN(input_size, hidden_size)
self_hidden_state = self_srn(inputs)
torch_outputs, torch_hidden_state = torch_srn(inputs)
print("self_srn hidden_state: ", self_hidden_state.shape)
print("torch_srn outpus:", torch_outputs.shape)
print("torch_srn hidden_state:", torch_hidden_state.shape)
运行结果:
hidden_state tensor([[0.3177, 0.4775]], grad_fn=<TanhBackward>)
self_srn hidden_state: torch.Size([8, 32])
torch_srn outpus: torch.Size([8, 20, 32])
torch_srn hidden_state: torch.Size([1, 20, 32])
可以看到,自己实现的SRN由于没有考虑多层因素,因此没有层次这个维度,因此其输出shape为[8, 32]。同时由于在以上代码使用Paddle内置API实例化SRN时,默认定义的是1层的单向SRN,因此其shape为[1, 8, 32],同时隐状态向量为[8,20, 32].
接下来,我们可以将自己实现的SRN与Paddle内置的SRN在输出值的精度上进行对比,这里首先根据Paddle内置的SRN实例化模型(为了进行对比,在实例化时只保留一个偏置,将偏置bx设置为0),然后提取该模型对应的参数,使用该参数去初始化自己实现的SRN,从而保证两者在参数初始化时是一致的。
将自己实现的SRN和PyTorch框架内置的RNN返回的结果进行打印展示:
# 这里创建一个随机数组作为测试数据,数据shape为batch_size x seq_len x input_size
batch_size, seq_len, input_size, hidden_size = 2, 5, 10, 10
inputs = torch.randn([batch_size, seq_len, input_size])
# 设置模型的hidden_size
torch_srn = nn.RNN(input_size, hidden_size, bias=False)
# 获取torch_srn中的参数,并设置相应的paramAttr,用于初始化SRN
W_attr = torch_srn.weight_ih_l0.T
U_attr = torch_srn.weight_hh_l0.T
self_srn = SRN(input_size, hidden_size, W_attr=W_attr, U_attr=U_attr)
# 进行前向计算,获取隐状态向量,并打印展示
self_hidden_state = self_srn(inputs)
torch_outputs, torch_hidden_state = torch_srn(inputs)
print("torch SRN:\n", torch_hidden_state.detach().numpy().squeeze(0))
print("self SRN:\n", self_hidden_state.detach().numpy())
运行结果:
torch SRN:
[[ 0.19587958 0.14590137 -0.47067317 -0.79495096 0.5207038 -0.40053073
0.30152693 -0.5072332 0.66418064 -0.8378919 ]
[-0.319796 -0.7052159 0.34430254 -0.28116354 -0.83387375 0.09060298
0.40611687 -0.5577904 -0.5175905 0.6714291 ]
[-0.01631957 -0.47080883 0.35313848 0.0725399 -0.17794743 0.31247756
0.40214932 -0.6849591 0.2507963 0.32316315]
[-0.13542189 -0.24074143 0.10872725 0.28753144 0.8345687 0.27860302
-0.5483449 -0.00179329 0.8561251 -0.79913914]
[-0.6216069 -0.24778096 -0.7806171 -0.85125434 0.8160991 0.36891603
0.2517544 -0.6713402 0.7224134 -0.5762565 ]]
self SRN:
[[ 0.20592752 -0.03319736 0.6064573 0.76857877 -0.2848402 0.17541209
-0.2603813 0.22346874 -0.5298694 -0.621583 ]
[-0.7845441 -0.12988532 -0.49171838 -0.2670862 0.7836917 -0.05539662
0.76456785 -0.71611995 0.80632466 -0.5775269 ]]
可以看到,两者的输出基本是一致的。另外,还可以进行对比两者在运算速度方面的差异。代码实现如下:
import time
# 这里创建一个随机数组作为测试数据,数据shape为batch_size x seq_len x input_size
batch_size, seq_len, input_size, hidden_size = 2, 5, 10, 10
inputs = torch.randn([batch_size, seq_len, input_size])
# 实例化模型
self_srn = SRN(input_size, hidden_size)
torch_srn = nn.RNN(input_size, hidden_size)
# 计算自己实现的SRN运算速度
model_time = 0
for i in range(100):
strat_time = time.time()
out = self_srn(inputs)
if i < 10:
continue
end_time = time.time()
model_time += (end_time - strat_time)
avg_model_time = model_time / 90
print('self_srn speed:', avg_model_time, 's')
# 计算torch内置的SRN运算速度
model_time = 0
for i in range(100):
strat_time = time.time()
out = torch_srn(inputs)
# 预热10次运算,不计入最终速度统计
if i < 10:
continue
end_time = time.time()
model_time += (end_time - strat_time)
avg_model_time = model_time / 90
print('torch_srn speed:', avg_model_time, 's')
运行结果:
self_srn speed: 0.000211284325767998047 s
torch_srn speed: 0.00010002178493375217 s
可以看到,由于Paddle内部相关算子由C++实现,Paddle框架实现的SRN的运行效率显著高于自己实现的SRN效率。
提醒:在分类问题的实践中,我们通常只需要模型输出分类的对数几率(Logits),而不用输出每个类的概率。这需要损失函数可以直接接收对数几率来损失计算。
线性层直接使用paddle.nn.Linear算子。
在定义了每一层的算子之后,我们定义一个数字求和模型Model_RNN4SeqClass,该模型会将嵌入层、SRN层和线性层进行组合,以实现数字求和的功能.
具体来讲,Model_RNN4SeqClass会接收一个SRN层实例,用于处理数字序列数据,同时在__init__函数中定义一个Embedding嵌入层,其会将输入的数字作为索引,输出对应的向量,最后会使用paddle.nn.Linear定义一个线性层。
提醒:为了方便进行对比实验,我们将SRN层的实例化放在\code{Model_RNN4SeqClass}类外面。通常情况下,模型内部算子的实例化是放在模型里面。
在forward函数中,调用上文实现的嵌入层、SRN层和线性层处理数字序列,同时返回最后一个位置的隐状态向量。代码实现如下:
# 基于RNN实现数字预测的模型
class Model_RNN4SeqClass(nn.Module):
def __init__(self, model, num_digits, input_size, hidden_size, num_classes):
super(Model_RNN4SeqClass, self).__init__()
# 传入实例化的RNN层,例如SRN
self.rnn_model = model
# 词典大小
self.num_digits = num_digits
# 嵌入向量的维度
self.input_size = input_size
# 定义Embedding层
self.embedding = Embedding(num_digits, input_size)
# 定义线性层
self.linear = nn.Linear(hidden_size, num_classes)
def forward(self, inputs):
# 将数字序列映射为相应向量
inputs_emb = self.embedding(inputs)
# 调用RNN模型
hidden_state = self.rnn_model(inputs_emb)
# 使用最后一个时刻的状态进行数字预测
logits = self.linear(hidden_state)
return logits
# 实例化一个input_size为4, hidden_size为5的SRN
srn = SRN(4, 5)
# 基于srn实例化一个数字预测模型实例
model = Model_RNN4SeqClass(srn, 10, 4, 5, 19)
# 生成一个shape为 2 x 3 的批次数据
inputs = torch.tensor([[1, 2, 3], [2, 3, 4]])
# 进行模型前向预测
logits = model(inputs)
print(logits)
运行结果:
tensor([[-0.0460, 0.0124, -0.0386, 0.0905, 0.2844, 0.4236, 0.2840, 0.4246,
-0.0323, -0.4017, -0.2120, 0.3045, -0.0029, -0.2223, -0.3427, -0.4185,
-0.3775, -0.0907, 0.2453],
[-0.0460, 0.0124, -0.0386, 0.0905, 0.2844, 0.4236, 0.2840, 0.4246,
-0.0323, -0.4017, -0.2120, 0.3045, -0.0029, -0.2223, -0.3427, -0.4185,
-0.3775, -0.0907, 0.2453]], grad_fn=<AddmmBackward>)
基于RunnerV3类进行训练,只需要指定length便可以加载相应的数据。设置超参数,使用Adam优化器,学习率为 0.001,实例化模型,使用第4.5.4节定义的Accuracy计算准确率。使用Runner进行训练,训练回合数设为500。代码实现如下:
import os
import random
import torch
import numpy as np
# 训练轮次
num_epochs = 500
# 学习率
lr = 0.001
# 输入数字的类别数
num_digits = 10
# 将数字映射为向量的维度
input_size = 32
# 隐状态向量的维度
hidden_size = 32
# 预测数字的类别数
num_classes = 19
# 批大小
batch_size = 8
# 模型保存目录
save_dir = "./checkpoints"
# 通过指定length进行不同长度数据的实验
def train(length):
print(f"\n====> Training SRN with data of length {length}.")
# 加载长度为length的数据
data_path = f"./datasets/{length}"
train_examples, dev_examples, test_examples = load_data(data_path)
train_set, dev_set, test_set = DigitSumDataset(train_examples), DigitSumDataset(dev_examples), DigitSumDataset(test_examples)
train_loader = DataLoader(train_set, batch_size=batch_size)
dev_loader = DataLoader(dev_set, batch_size=batch_size)
test_loader = DataLoader(test_set, batch_size=batch_size)
# 实例化模型
base_model = SRN(input_size, hidden_size)
model = Model_RNN4SeqClass(base_model, num_digits, input_size, hidden_size, num_classes)
# 指定优化器
optimizer = torch.optim.Adam(lr=lr, params=model.parameters())
# 定义评价指标
metric = Accuracy()
# 定义损失函数
loss_fn = nn.CrossEntropyLoss()
# 基于以上组件,实例化Runner
runner = RunnerV3(model, optimizer, loss_fn, metric)
# 进行模型训练
model_save_path = os.path.join(save_dir, f"best_srn_model_{length}.pdparams")
runner.train(train_loader, dev_loader, num_epochs=num_epochs, eval_steps=100, log_steps=100, save_path=model_save_path)
return runner
srn_runners = {}
注:RunnerV3类与Accuracy类如下:
RunnerV3:
class RunnerV3(object):
def __init__(self, model, optimizer, loss_fn, metric, **kwargs):
self.model = model
self.optimizer = optimizer
self.loss_fn = loss_fn
self.metric = metric # 只用于计算评价指标
# 记录训练过程中的评价指标变化情况
self.dev_scores = []
# 记录训练过程中的损失函数变化情况
self.train_epoch_losses = [] # 一个epoch记录一次loss
self.train_step_losses = [] # 一个step记录一次loss
self.dev_losses = []
# 记录全局最优指标
self.best_score = 0
def train(self, train_loader, dev_loader=None, **kwargs):
# 将模型切换为训练模式
self.model.train()
# 传入训练轮数,如果没有传入值则默认为0
num_epochs = kwargs.get("num_epochs", 0)
# 传入log打印频率,如果没有传入值则默认为100
log_steps = kwargs.get("log_steps", 100)
# 评价频率
eval_steps = kwargs.get("eval_steps", 0)
# 传入模型保存路径,如果没有传入值则默认为"best_model.pdparams"
save_path = kwargs.get("save_path", "best_model.pdparams")
custom_print_log = kwargs.get("custom_print_log", None)
# 训练总的步数
num_training_steps = num_epochs * len(train_loader)
if eval_steps:
if self.metric is None:
raise RuntimeError('Error: Metric can not be None!')
if dev_loader is None:
raise RuntimeError('Error: dev_loader can not be None!')
# 运行的step数目
global_step = 0
# 进行num_epochs轮训练
for epoch in range(num_epochs):
# 用于统计训练集的损失
total_loss = 0
for step, data in enumerate(train_loader):
X, y = data
# 获取模型预测
logits = self.model(X)
loss = self.loss_fn(logits, y.long()) # 默认求mean
total_loss += loss
# 训练过程中,每个step的loss进行保存
self.train_step_losses.append((global_step, loss.item()))
if log_steps and global_step % log_steps == 0:
print(
f"[Train] epoch: {epoch}/{num_epochs}, step: {global_step}/{num_training_steps}, loss: {loss.item():.5f}")
# 梯度反向传播,计算每个参数的梯度值
loss.backward()
if custom_print_log:
custom_print_log(self)
# 小批量梯度下降进行参数更新
self.optimizer.step()
# 梯度归零
self.optimizer.zero_grad()
# 判断是否需要评价
if eval_steps > 0 and global_step > 0 and \
(global_step % eval_steps == 0 or global_step == (num_training_steps - 1)):
dev_score, dev_loss = self.evaluate(dev_loader, global_step=global_step)
print(f"[Evaluate] dev score: {dev_score:.5f}, dev loss: {dev_loss:.5f}")
# 将模型切换为训练模式
self.model.train()
# 如果当前指标为最优指标,保存该模型
if dev_score > self.best_score:
self.save_model(save_path)
print(
f"[Evaluate] best accuracy performence has been updated: {self.best_score:.5f} --> {dev_score:.5f}")
self.best_score = dev_score
global_step += 1
# 当前epoch 训练loss累计值
trn_loss = (total_loss / len(train_loader)).item()
# epoch粒度的训练loss保存
self.train_epoch_losses.append(trn_loss)
print("[Train] Training done!")
# 模型评估阶段,使用'torch.no_grad()'控制不计算和存储梯度
@torch.no_grad()
def evaluate(self, dev_loader, **kwargs):
assert self.metric is not None
# 将模型设置为评估模式
self.model.eval()
global_step = kwargs.get("global_step", -1)
# 用于统计训练集的损失
total_loss = 0
# 重置评价
self.metric.reset()
# 遍历验证集每个批次
for batch_id, data in enumerate(dev_loader):
X, y = data
# 计算模型输出
logits = self.model(X)
# 计算损失函数
loss = self.loss_fn(logits, y.long()).item()
# 累积损失
total_loss += loss
# 累积评价
self.metric.update(logits, y)
dev_loss = (total_loss / len(dev_loader))
dev_score = self.metric.accumulate()
# 记录验证集loss
if global_step != -1:
self.dev_losses.append((global_step, dev_loss))
self.dev_scores.append(dev_score)
return dev_score, dev_loss
# 模型评估阶段,使用'torch.no_grad()'控制不计算和存储梯度
@torch.no_grad()
def predict(self, x, **kwargs):
# 将模型设置为评估模式
self.model.eval()
# 运行模型前向计算,得到预测值
logits = self.model(x)
return logits
def save_model(self, save_path):
torch.save(self.model.state_dict(), save_path)
def load_model(self, model_path):
state_dict = torch.load(model_path)
self.model.load_state_dict(state_dict)
Accuracy:
class Accuracy():
def __init__(self, is_logist=True):
# 用于统计正确的样本个数
self.num_correct = 0
# 用于统计样本的总数
self.num_count = 0
self.is_logist = is_logist
def update(self, outputs, labels):
# 判断是二分类任务还是多分类任务,shape[1]=1时为二分类任务,shape[1]>1时为多分类任务
if outputs.shape[1] == 1: # 二分类
outputs = torch.squeeze(outputs, dim=-1)
if self.is_logist:
# logist判断是否大于0
preds = torch.tensor((outputs >= 0), dtype=torch.float32)
else:
# 如果不是logist,判断每个概率值是否大于0.5,当大于0.5时,类别为1,否则类别为0
preds = torch.tensor((outputs >= 0.5), dtype=torch.float32)
else:
# 多分类时,使用'torch.argmax'计算最大元素索引作为类别
preds = torch.argmax(outputs, dim=1)
# 获取本批数据中预测正确的样本个数
labels = torch.squeeze(labels, dim=-1)
batch_correct = torch.sum(torch.tensor(preds == labels, dtype=torch.float32)).cpu().numpy()
batch_count = len(labels)
# 更新num_correct 和 num_count
self.num_correct += batch_correct
self.num_count += batch_count
def accumulate(self):
# 使用累计的数据,计算总的指标
if self.num_count == 0:
return 0
return self.num_correct / self.num_count
def reset(self):
# 重置正确的数目和总数
self.num_correct = 0
self.num_count = 0
def name(self):
return "Accuracy"
接下来,分别进行数据长度为10, 15, 20, 25, 30, 35的数字预测模型训练实验,训练后的runner保存至runners字典中。
srn_runners = {}
lengths = [10, 15, 20, 25, 30, 35]
for length in lengths:
runner = train(length)
srn_runners[length] = runner
运行结果:
====> Training SRN with data of length 10.
[Train] epoch: 489/500, step: 18600/19000, loss: 0.05225
[Evaluate] dev score: 0.50000, dev loss: 2.47189
[Train] epoch: 492/500, step: 18700/19000, loss: 0.05124
[Evaluate] dev score: 0.51000, dev loss: 2.48123
[Train] epoch: 494/500, step: 18800/19000, loss: 0.05769
[Evaluate] dev score: 0.50000, dev loss: 2.47833
[Train] epoch: 497/500, step: 18900/19000, loss: 0.04340
[Evaluate] dev score: 0.50000, dev loss: 2.47491
[Evaluate] dev score: 0.51000, dev loss: 2.47137
[Train] Training done!
====> Training SRN with data of length 15.
[Train] epoch: 489/500, step: 18600/19000, loss: 0.11159
[Evaluate] dev score: 0.42000, dev loss: 3.65576
[Train] epoch: 492/500, step: 18700/19000, loss: 1.34253
[Evaluate] dev score: 0.35000, dev loss: 4.08752
[Train] epoch: 494/500, step: 18800/19000, loss: 0.29696
[Evaluate] dev score: 0.25000, dev loss: 4.61034
[Train] epoch: 497/500, step: 18900/19000, loss: 0.12404
[Evaluate] dev score: 0.41000, dev loss: 3.44996
[Evaluate] dev score: 0.42000, dev loss: 3.38767
[Train] Training done!
====> Training SRN with data of length 20.
[Train] epoch: 489/500, step: 18600/19000, loss: 0.94918
[Evaluate] dev score: 0.28000, dev loss: 3.03929
[Train] epoch: 492/500, step: 18700/19000, loss: 1.78990
[Evaluate] dev score: 0.30000, dev loss: 2.91098
[Train] epoch: 494/500, step: 18800/19000, loss: 0.58229
[Evaluate] dev score: 0.34000, dev loss: 2.88835
[Train] epoch: 497/500, step: 18900/19000, loss: 0.83346
[Evaluate] dev score: 0.28000, dev loss: 2.91086
[Evaluate] dev score: 0.26000, dev loss: 3.06699
[Train] Training done!
====> Training SRN with data of length 25.
[Train] epoch: 489/500, step: 18600/19000, loss: 1.25829
[Evaluate] dev score: 0.10000, dev loss: 5.29650
[Train] epoch: 492/500, step: 18700/19000, loss: 0.49421
[Evaluate] dev score: 0.10000, dev loss: 5.00213
[Train] epoch: 494/500, step: 18800/19000, loss: 0.48700
[Evaluate] dev score: 0.13000, dev loss: 5.15761
[Train] epoch: 497/500, step: 18900/19000, loss: 0.51467
[Evaluate] dev score: 0.13000, dev loss: 5.27542
[Evaluate] dev score: 0.12000, dev loss: 5.33683
[Train] Training done!
====> Training SRN with data of length 30.
[Train] epoch: 489/500, step: 18600/19000, loss: 0.12149
[Evaluate] dev score: 0.18000, dev loss: 4.28667
[Train] epoch: 492/500, step: 18700/19000, loss: 0.64496
[Evaluate] dev score: 0.19000, dev loss: 4.32516
[Train] epoch: 494/500, step: 18800/19000, loss: 0.36210
[Evaluate] dev score: 0.22000, dev loss: 4.37720
[Train] epoch: 497/500, step: 18900/19000, loss: 0.97052
[Evaluate] dev score: 0.20000, dev loss: 4.29592
[Evaluate] dev score: 0.12000, dev loss: 4.48994
[Train] Training done!
====> Training SRN with data of length 35.
[Train] epoch: 489/500, step: 18600/19000, loss: 1.99054
[Evaluate] dev score: 0.05000, dev loss: 4.47362
[Train] epoch: 492/500, step: 18700/19000, loss: 1.63049
[Evaluate] dev score: 0.07000, dev loss: 4.44207
[Train] epoch: 494/500, step: 18800/19000, loss: 1.58705
[Evaluate] dev score: 0.07000, dev loss: 4.48929
[Train] epoch: 497/500, step: 18900/19000, loss: 1.63524
[Evaluate] dev score: 0.03000, dev loss: 4.50457
[Evaluate] dev score: 0.06000, dev loss: 4.08931
[Train] Training done!
定义plot_training_loss函数,分别画出各个长度的数字预测模型训练过程中,在训练集和验证集上的损失曲线,实现代码实现如下:
import matplotlib.pyplot as plt
def plot_training_loss(runner, fig_name, sample_step):
plt.figure()
train_items = runner.train_step_losses[::sample_step]
train_steps = [x[0] for x in train_items]
train_losses = [x[1] for x in train_items]
plt.plot(train_steps, train_losses, color='#e4007f', label="Train loss")
dev_steps = [x[0] for x in runner.dev_losses]
dev_losses = [x[1] for x in runner.dev_losses]
plt.plot(dev_steps, dev_losses, color='#f19ec2', linestyle='--', label="Dev loss")
# 绘制坐标轴和图例
plt.ylabel("loss", fontsize='large')
plt.xlabel("step", fontsize='large')
plt.legend(loc='upper right', fontsize='x-large')
plt.savefig(fig_name)
plt.show()
# 画出训练过程中的损失图
for length in lengths:
runner = srn_runners[length]
fig_name = f"./images/6.6_{length}.pdf"
plot_training_loss(runner, fig_name, sample_step=100)
运行结果:
K=10
K=15
K=20
K=25
K=30
K=35
图6.6展示了在6个数据集上的损失变化情况,数据集的长度分别为10、15、20、25、30和35. 从输出结果看,随着数据序列长度的增加,虽然训练集损失逐渐逼近于0,但是验证集损失整体趋向越来越大,这表明当序列变长时,SRN模型保持序列长期依赖能力在逐渐变弱,越来越无法学习到有用的知识.
在模型评价时,加载不同长度的效果最好的模型,然后使用测试集对该模型进行评价,观察模型在测试集上预测的准确度. 同时记录一下不同长度模型在训练过程中,在验证集上最好的效果。代码实现如下:
srn_dev_scores = []
srn_test_scores = []
for length in lengths:
print(f"Evaluate SRN with data length {length}.")
runner = srn_runners[length]
# 加载训练过程中效果最好的模型
model_path = os.path.join(save_dir, f"best_srn_model_{length}.pdparams")
runner.load_model(model_path)
# 加载长度为length的数据
data_path = f"./datasets/{length}"
train_examples, dev_examples, test_examples = load_data(data_path)
test_set = DigitSumDataset(test_examples)
test_loader = DataLoader(test_set, batch_size=batch_size)
# 使用测试集评价模型,获取测试集上的预测准确率
score, _ = runner.evaluate(test_loader)
srn_test_scores.append(score)
srn_dev_scores.append(max(runner.dev_scores))
for length, dev_score, test_score in zip(lengths, srn_dev_scores, srn_test_scores):
print(f"[SRN] length:{length}, dev_score: {dev_score}, test_score: {test_score: .5f}")
运行结果:
Evaluate SRN with data length 15.
Evaluate SRN with data length 20.
Evaluate SRN with data length 25.
Evaluate SRN with data length 30.
Evaluate SRN with data length 35.
[SRN] length:10, dev_score: 0.41, test_score: 0.22000
[SRN] length:15, dev_score: 0.2, test_score: 0.18000
[SRN] length:20, dev_score: 0.18, test_score: 0.13000
[SRN] length:25, dev_score: 0.12, test_score: 0.17000
[SRN] length:30, dev_score: 0.16, test_score: 0.05000
[SRN] length:35, dev_score: 0.18, test_score: 0.07000
接下来,将SRN在不同长度的验证集和测试集数据上的表现,绘制成图片进行观察。
import matplotlib.pyplot as plt
plt.plot(lengths, srn_dev_scores, '-o', color='#e4007f', label="Dev Accuracy")
plt.plot(lengths, srn_test_scores,'-o', color='#f19ec2', label="Test Accuracy")
#绘制坐标轴和图例
plt.ylabel("accuracy", fontsize='large')
plt.xlabel("sequence length", fontsize='large')
plt.legend(loc='upper right', fontsize='x-large')
fig_name = "./images/6.7.pdf"
plt.savefig(fig_name)
plt.show()
运行结果:
图6.7 展示了SRN模型在不同长度数据训练出来的最好模型在验证集和测试集上的表现。可以看到,随着序列长度的增加,验证集和测试集的准确度整体趋势是降低的,这同样说明SRN模型保持长期依赖的能力在不断降低.
将代码一行改为:
hidden_state =hidden_state + F.tanh(torch.matmul(step_input, self.W) + torch.matmul(hidden_state, self.U) + self.b)
运行结果:
Evaluate SRN with data length 10.
Evaluate SRN with data length 15.
Evaluate SRN with data length 20.
Evaluate SRN with data length 25.
Evaluate SRN with data length 30.
Evaluate SRN with data length 35.
[SRN] length:10, dev_score: 0.97, test_score: 0.96000
[SRN] length:15, dev_score: 0.98, test_score: 0.96000
[SRN] length:20, dev_score: 0.94, test_score: 0.96000
[SRN] length:25, dev_score: 0.98, test_score: 0.97000
[SRN] length:30, dev_score: 0.97, test_score: 0.96000
[SRN] length:35, dev_score: 0.87, test_score: 0.92000
显然,加了残差之后,准确率大幅升高,有效缓解了长程依赖问题。
通过此次实验,了解了长程依赖问题,认识了循环神经网络的模型结构,对其中的嵌入层和SRN层,有了比较深的了解,对于RNN的计算图理解更加深刻了。实验发现改进SRN的循环单元,加入隐状态之间的残差连接,准确率大幅升高,有效缓解了长程依赖问题。
NNDL 实验6(上) - HBU_DAVID - 博客园 (cnblogs.com)
NNDL 实验七 循环神经网络(1)RNN记忆能力实验