当前位置: 首页 > news >正文

深度学习实验十四 循环神经网络(1)——测试简单循环网络的记忆能力和梯度爆炸实验

目录

一、数据集构建

1.1数据集的构建函数

1.2加载数据集并划分

1.3 构建Dataset类

二、模型构建

2.1嵌入层

2.2SRN层

2.3模型汇总

三、模型训练

3.1 训练指定长度的数字预测模型

3.2 损失曲线展示

四、模型评价

五、修改

附完整可运行代码

实验大体步骤:

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。下图展示了长度为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.

一、数据集构建

我们首先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kk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kk=1时,生成验证集和测试集. 

1.1数据集的构建函数

代码如下:

import os
import random
import numpy as np
from torch import nn# 固定随机种子
random.seed(0)
np.random.seed(0)# ========数据集构建=============================================
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.")# 创建目录(如果不存在)directory = os.path.dirname(save_path)if not os.path.exists(directory):os.makedirs(directory)# 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 + n2base_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examplesfor 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] = valexamples.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 = 3generate_data(length, k, save_path)# 生成长度为length的验证数据save_path = f"./datasets/{length}/dev.txt"k = 1generate_data(length, k, save_path)# 生成长度为length的测试数据save_path = f"./datasets/{length}/test.txt"k = 1generate_data(length, k, save_path)

1.2加载数据集并划分

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

代码如下:

# ===加载数据并进行数据划分=================================================
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labelitems = 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labelitems = 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labelitems = 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))

运行结果:

1.3 构建Dataset类

代码如下:

# =====构造Dataset类=====================================================
from torch.utils.data import Dataset, DataLoader
import torchclass DigitSumDataset(Dataset):def __init__(self, data):self.data = datadef __getitem__(self, idx):example = self.data[idx]seq = torch.tensor(example[0], dtype=torch.int64)label = torch.tensor(example[1], dtype=torch.int64)return seq, labeldef __len__(self):return len(self.data)

二、模型构建

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.

2.1嵌入层

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。

代码如下:

# ==========嵌入层=====================================================
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 embsemb_layer = Embedding(10, 5)
inputs = torch.tensor([0, 1, 2, 3])
emb_layer(inputs)

思考:如果不使用嵌入层,直接将数字作为SRN层输入有什么问题?

①神经网络尤其是递归神经网络(RNN)通常期望接收连续的输入数据。数字(例如:0, 1, 2, 3, 等)是离散的,它们与模型期望的输入方式不太匹配。如果直接将数字作为输入,模型可能难以理解数字之间的相对关系。模型可能会将数字 1 和 2 看作两个独立的、无关的输入,无法理解它们之间的顺序和数值差异。
②数字 100 可能和数字 1 在输入层上的表示差异非常大,而这种差异不一定反映了数字之间的实际关系。在神经网络中,输入特征的尺度差异可能导致模型训练不稳定,且难以收敛。

2.2SRN层

代码如下:

# ==========SRN层============================================================================
import torch
import torch.nn as nn
import torch.nn.functional as Ftorch.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_sizeif 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_sizeif 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_sizeif 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_sizebatch_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_sizeif 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_sizestep_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)hidden_state = 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)

运行结果如下:

PyTorch框架内置了SRN的API torch.nn.RNN

代码如下:

# ====PyTorch框架内置了SRN的API torch.nn.RNN============================================
# 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)# =====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 = 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:

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_sizetorch_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())

运行结果如下:

对比一下运行时间:

# ======两者时间差异===========================================================
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:continueend_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:continueend_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:

我看着感觉是两者并没有差太多。

2.3模型汇总

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。

代码如下:

# ======模型汇总======================================================
# 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SRNself.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:

三、模型训练

3.1 训练指定长度的数字预测模型

代码如下:

# =========模型训练================================================
import os
import random
import torch
import numpy as np
from nndl_3 import Accuracy, RunnerV3# 训练轮次
num_epochs = 500
# 学习率
lr = 0.0001
# 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Runnerrunner = 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 runnersrn_runners = {}# 多组训练
lengths = [10, 15, 20, 25, 30, 35]
for length in lengths:runner = train(length)srn_runners[length] = runner

3.2 损失曲线展示

代码如下:

# =======损失函数绘制===================================================
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:(分别是长度为10、15、20、25、30、35,lr=0.001的情况)

可以发现是有点过拟合的,Train loss不断减小,但是dev loss却在上升。

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。SRN模型过拟合到序列结尾的信息,而遗忘了序列开始位置的信息。说明,SRN模型在建模长程依赖方面的能力比较弱。

四、模型评价

代码如下:

# ======模型评价===========================================
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}")# ========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 pltplt.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:

可以看出实验结果并不是很好。

五、修改

于是参看书上的动手练习6.1:

参考《神经网络与深度学习》中的公式(6.50),改进SRN的循环单元,加入隐状态之间的残差连接,并重复数字求和实验。观察是否可以缓解长程依赖问题。

这样h_th_{t-1}之间为既有线性关系,也有非线性关系,并且可以缓解梯度消失问题.于是尝试加入残差连接,并重新进行实验。根据这个公式可知,只需要将原来的hidden_state加上上一个时刻的hidden_state即可。

原始代码:

  hidden_state = F.tanh(torch.matmul(step_input, self.W) + torch.matmul(hidden_state, self.U) + self.b)

修改为:

 hidden_state =hidden_state + F.tanh(torch.matmul(step_input, self.W) + torch.matmul(hidden_state, self.U) + self.b)

运行结果:

从以上的loss图可以看出,train的loss基本和dev的loss重合,减轻了过拟合。

但是这种改进方法还存在着两个问题:

于是可以通过引入门控机制来进一步改进模型,也就是LSTM。

造成简单循环网络较难建模长程依赖问题的原因有两个:梯度爆炸和梯度消失。

一般来讲,循环网络的梯度爆炸问题比较容易解决,一般通过权重衰减或梯度截断可以较好地来避免;对于梯度消失问题,更加有效的方式是改变模型,比如通过长短期记忆网络LSTM来进行缓解。

 各个长度数据集的准确率可以看出,并不是长度越长,准确率越低,往往在25或者30处存在一个拐点(中点低两边高或者是中间高两边低),那这种情况是为什么呢?

查阅了一下资料做出的一些解释(可能有点不太准确):
①网络的训练时间对准确率也有影响。较长的序列需要更多的训练时间来优化网络参数,且较长序列可能更容易陷入局部最优解。30长度的序列可能正好处于训练过程中较难收敛的区间,而25和35的序列可能更容易收敛到较好的解。
②在较短的序列的情况下,网络可能会过拟合或过于依赖序列中的少量信息,网络可能会过度“记住”某些样本的特征,而不是学习到更通用的规律。随着数据集长度增加,网络获得的信息变得更加冗余,可以帮助它做出更泛化的预测。过短的序列会使模型过度依赖训练集中的特定样本,而较长的序列可以让网络更好地学习到前两个数字的关系,避免过拟合。从图xlen=25的loss也可以看出,他的train loss和dev loss相差最多,也说明了为什么len=25是准确率最低的情况。

调了调学习率

lr=0.01

lr=0.1

(这个也太不好了吧)

lr=0.0001

这个就感觉有点欠拟合的意思了。

我发现SRN的准确率不宜调的过大,当lr=0.001时所有长度数据集准确率基本都在0.8以上,但是lr=0.01和0.1时准确率就不是很好看了。所以较长的序列可能需要更低的学习率或更高的隐藏层单元数。如果超参数设置不合理,可能会导致模型在某些长度下的性能波动。
但是从lr=0.0001的图中可以看出过小的准确率也不是很好,train和dev的loss都在逐渐减小,但是可能还没有训练好,产生了欠拟合,所以出现了len=35比任何长度的准确率都高的情况。

思考题:如果不使用嵌入层,直接将数字作为SRN层输入有什么问题? 

①神经网络尤其是递归神经网络(RNN)通常期望接收连续的输入数据。数字(例如:0, 1, 2, 3, 等)是离散的,它们与模型期望的输入方式不太匹配。如果直接将数字作为输入,模型可能难以理解数字之间的相对关系。模型可能会将数字 1 和 2 看作两个独立的、无关的输入,无法理解它们之间的顺序和数值差异。
②数字 100 可能和数字 1 在输入层上的表示差异非常大,而这种差异不一定反映了数字之间的实际关系。在神经网络中,输入特征的尺度差异可能导致模型训练不稳定,且难以收敛。

 什么是词向量呢?

之前讲过的word2vec是词向量的一种方式。词向量是一种将词汇映射为固定长度连续向量的技术,每个词被映射成一个固定维度的向量,通常是几十到几百维。在一个好的词向量空间中,语义相似的词会有相似的向量表示。例如,“猫”和“狗”的词向量应该比“猫”和“汽车”的词向量更相似。同时,通过词向量还可以捕捉词汇之间的语义关系(如“国王” - “男人” + “女人” ≈ “女王”)

附完整可运行代码

主程序:

import os
import random
import numpy as np
from torch import nn# 固定随机种子
random.seed(0)
np.random.seed(0)# ========数据集构建=============================================
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.")# 创建目录(如果不存在)directory = os.path.dirname(save_path)if not os.path.exists(directory):os.makedirs(directory)# 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 + n2base_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examplesfor 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] = valexamples.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 = 3generate_data(length, k, save_path)# 生成长度为length的验证数据save_path = f"./datasets/{length}/dev.txt"k = 1generate_data(length, k, save_path)# 生成长度为length的测试数据save_path = f"./datasets/{length}/test.txt"k = 1generate_data(length, k, save_path)# ===加载数据并进行数据划分=================================================
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labelitems = 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labelitems = 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labelitems = 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))# =====构造Dataset类=====================================================
from torch.utils.data import Dataset, DataLoader
import torchclass DigitSumDataset(Dataset):def __init__(self, data):self.data = datadef __getitem__(self, idx):example = self.data[idx]seq = torch.tensor(example[0], dtype=torch.int64)label = torch.tensor(example[1], dtype=torch.int64)return seq, labeldef __len__(self):return len(self.data)# ==========嵌入层=====================================================
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 embsemb_layer = Embedding(10, 5)
inputs = torch.tensor([0, 1, 2, 3])
emb_layer(inputs)# ==========SRN层============================================================================
import torch
import torch.nn as nn
import torch.nn.functional as Ftorch.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_sizeif 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_sizeif 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_sizeif 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_sizebatch_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_sizeif 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_sizestep_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)hidden_state = 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)# ====PyTorch框架内置了SRN的API torch.nn.RNN============================================
# 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)# =====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 = 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)# 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_sizetorch_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())# ======两者时间差异===========================================================
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:continueend_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:continueend_time = time.time()model_time += (end_time - strat_time)
avg_model_time = model_time / 90
print('torch_srn speed:', avg_model_time, 's')# ======模型汇总======================================================
# 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SRNself.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)# =========模型训练================================================
import os
import random
import torch
import numpy as np
from nndl_3 import Accuracy, RunnerV3# 训练轮次
num_epochs = 500
# 学习率
lr = 0.0001
# 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Runnerrunner = 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 runnersrn_runners = {}# 多组训练
lengths = [10, 15, 20, 25, 30, 35]
for length in lengths:runner = train(length)srn_runners[length] = runner# =======损失函数绘制===================================================
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)# ======模型评价===========================================
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}")# ========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 pltplt.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()

 nndl_3:

import torch
from matplotlib import pyplot as plt
from torch import nnclass Op(object):def __init__(self):passdef __call__(self, inputs):return self.forward(inputs)def forward(self, inputs):raise NotImplementedErrordef backward(self, inputs):raise NotImplementedError# 实现一个两层前馈神经网络
class Model_MLP_L2_V3(torch.nn.Module):def __init__(self, input_size, hidden_size, output_size):super(Model_MLP_L2_V3, self).__init__()self.fc1 = torch.nn.Linear(input_size, hidden_size)w_ = torch.normal(0, 0.01, size=(hidden_size, input_size), requires_grad=True)self.fc1.weight = torch.nn.Parameter(w_)self.fc1.bias = torch.nn.init.constant_(self.fc1.bias, val=1.0)self.fc2 = torch.nn.Linear(hidden_size, output_size)w2 = torch.normal(0, 0.01, size=(output_size, hidden_size), requires_grad=True)self.fc2.weight = nn.Parameter(w2)self.fc2.bias = torch.nn.init.constant_(self.fc2.bias, val=1.0)self.act = torch.sigmoiddef forward(self, inputs):outputs = self.fc1(inputs)outputs = self.act(outputs)outputs = self.fc2(outputs)return outputsclass RunnerV3(object):def __init__(self, model, optimizer, loss_fn, metric, **kwargs):self.model = modelself.optimizer = optimizerself.loss_fn = loss_fnself.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lossself.train_step_losses = []  # 一个step记录一次lossself.dev_losses = []# 记录全局最优指标self.best_score = 0def 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0num_epochs = kwargs.get("num_epochs", 0)# 传入log打印频率,如果没有传入值则默认为100log_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 = 0for 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meantotal_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_scoreglobal_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).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lossif 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 logitsdef save_model(self, save_path):torch.save(self.model.state_dict(), save_path)def load_model(self, model_path):model_state_dict = torch.load(model_path)self.model.load_state_dict(model_state_dict)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 = 0self.is_logist = is_logistdef update(self, outputs, labels):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0preds = 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0preds = 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_countself.num_correct += batch_correctself.num_count += batch_countdef accumulate(self):# 使用累计的数据,计算总的指标if self.num_count == 0:return 0return self.num_correct / self.num_countdef reset(self):# 重置正确的数目和总数self.num_correct = 0self.num_count = 0def name(self):return "Accuracy"# 可视化
def plot(runner, fig_name):plt.figure(figsize=(10, 5))plt.subplot(1, 2, 1)train_items = runner.train_step_losses[::30]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='#8E004D', label="Train loss")if runner.dev_losses[0][0] != -1: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='#E20079', linestyle='--', label="Dev loss")# 绘制坐标轴和图例plt.ylabel("loss", fontsize='x-large')plt.xlabel("step", fontsize='x-large')plt.legend(loc='upper right', fontsize='x-large')plt.subplot(1, 2, 2)# 绘制评价准确率变化曲线if runner.dev_losses[0][0] != -1:plt.plot(dev_steps, runner.dev_scores,color='#E20079', linestyle="--", label="Dev accuracy")else:plt.plot(list(range(len(runner.dev_scores))), runner.dev_scores,color='#E20079', linestyle="--", label="Dev accuracy")# 绘制坐标轴和图例plt.ylabel("score", fontsize='x-large')plt.xlabel("step", fontsize='x-large')plt.legend(loc='lower right', fontsize='x-large')plt.savefig(fig_name)plt.show()

补充梯度截断

(写到下一个实验才发现这里忘记写了)

1.梯度打印函数

代码如下:

import torch
from torch import nn# ====梯度打印函数======================================
W_list = []
U_list = []
b_list = []
# 计算梯度范数
def custom_print_log(runner):model = runner.modelW_grad_l2, U_grad_l2, b_grad_l2 = 0, 0, 0for name, param in model.named_parameters():if name == "rnn_model.W":W_grad_l2 = torch.norm(param.grad, p=2).numpy()if name == "rnn_model.U":U_grad_l2 = torch.norm(param.grad, p=2).numpy()if name == "rnn_model.b":b_grad_l2 = torch.norm(param.grad, p=2).numpy()print(f"[Training] W_grad_l2: {W_grad_l2:.5f}, U_grad_l2: {U_grad_l2:.5f}, b_grad_l2: {b_grad_l2:.5f} ")W_list.append(W_grad_l2)U_list.append(U_grad_l2)b_list.append(b_grad_l2)

2.复现梯度爆炸现象

为了更好地复现梯度爆炸问题,使用SGD优化器将批大小和学习率调大,学习率为0.2,同时在计算交叉熵损失时,将reduction设置为sum,表示将损失进行累加。

代码如下:

# ======复现梯度爆炸现象===================================================
import os
import random
import torch
import numpy as np
from nndl_3 import Accuracy, RunnerV3
from RNNandLSTM import DigitSumDataset, Model_RNN4SeqClass, SRN, load_data
from torch.utils.data import Dataset, DataLoadernp.random.seed(0)
random.seed(0)
torch.seed()# 训练轮次
num_epochs = 50
# 学习率
lr = 0.2
# 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 = 64
# 模型保存目录
save_dir = "./checkpoints"# 可以设置不同的length进行不同长度数据的预测实验
length = 20
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.SGD(lr=lr, params=model.parameters())
# 定义评价指标
metric = Accuracy()
# 定义损失函数
loss_fn = nn.CrossEntropyLoss(reduction="sum")# 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"srn_explosion_model_{length}.pdparams")
runner.train(train_loader, dev_loader, num_epochs=num_epochs, eval_steps=100, log_steps=1,save_path=model_save_path, custom_print_log=custom_print_log)

3.把l2范数变化图绘制出来

代码如下:

# =====绘制图片====================================================
import matplotlib.pyplot as plt
def plot_grad(W_list, U_list, b_list, save_path, keep_steps=40):# 开始绘制图片plt.figure()# 默认保留前40步的结果steps = list(range(keep_steps))plt.plot(steps, W_list[:keep_steps], "r-", color="#e4007f", label="W_grad_l2")plt.plot(steps, U_list[:keep_steps], "-.", color="#f19ec2", label="U_grad_l2")plt.plot(steps, b_list[:keep_steps], "--", color="#000000", label="b_grad_l2")plt.xlabel("step")plt.ylabel("L2 Norm")plt.legend(loc="upper right")plt.show()plt.savefig(save_path)print("image has been saved to: ", save_path)save_path = f"./images/6.8.pdf"
plot_grad(W_list, U_list, b_list, save_path)# 加载训练过程中效果最好的模型
model_path = os.path.join(save_dir, "srn_explosion_model_20.pdparams")
runner.load_model(model_path)# 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)
print(f"[SRN] length:{length}, Score: {score: .5f}")

运行结果如下:

4.使用梯度截断解决

只需要在RunnerV3的优化器参数更新之前加一行即可:

 nn.utils.clip_grad_norm_(parameters=model.parameters(), max_norm=20, norm_type=2)

其余代码不变

运行结果:

这次的分享就到这里了,下次再见~


http://www.mrgr.cn/news/79629.html

相关文章:

  • Ubuntu中配置交叉编译工具的三条命令的详细研究
  • 共享GitLab中CICD自动生成的软件包
  • 高并发数据采集场景下Nginx代理Netty服务的优化配置
  • selinux
  • 新能源汽车 “能量侠”:移动充电机器人开启便捷补电新征程
  • vue2 vue3 无限滚动
  • 深入了解架构中常见的4种缓存模式及其实现
  • 在VMWare上安装openEuler 22.03-LTS
  • Mysql索引原理及优化——岁月云实战笔记
  • 嵌入式开发 - 工具记录
  • 【mysql】数据库存量数据双主实现
  • 北京大学《操作系统原理》课堂笔记(一)
  • LLM - 多模态大模型的开源评估工具 VLMEvalKit 部署与测试 教程
  • leetcode-54.螺旋矩阵-day1
  • Adobe Premiere Pro 2024 [24.6.1]
  • 2022 年“泰迪杯”数据分析技能赛A 题竞赛作品的自动评判
  • MySQL-DML之数据表操作
  • 递归算法题(1)
  • C++小小复习一下
  • SpringBoot3整合MyBatis
  • 2020 年“泰迪杯”数据分析职业技能大赛A 题教育平台的线上课程智能推荐策略
  • NanoLog起步笔记-4-Server端的两个线程
  • BottomNavigation
  • NanoLog起步笔记-1
  • ubuntu16.04部署dify教程
  • ESP32开发 云调试