六、新闻主题分类任务

以一段新闻报道中的文本描述内容为输入，使用模型帮助我们判断它最有可能属于哪一种类型的新闻，这是典型的文本分类问题。我们这里假定每种类型是互斥的，即文本描述有且只有一种类型，例如一篇新闻不能即是娱乐类又是财经类，只能是一种类别。

一、数据下载与介绍

我们使用的是AG_NEWS数据集，已经被集成在了torchtext中，下面是下载数据集的代码：

注意：

如果没有torchtext时，使用pip安装时会有一个大坑。

torchtext安装时会检查pytorch的版本，如果版本不兼容，它会卸载你的torch，然后安装一个GPU版本的兼容的torch，这个过程是自动的，没有什么提示，或者大部分人不会具体去看提示，这里会非常坑。

我在刚开始安装torchtext后，怎么也无法使用GPU，我还是以为是显卡有问题了，搞了好久最后才发现是torch被变成了CPU版本，刚开始不知道，就卸载torch，然后重装CUDA版本的torch，但是没用，最后装上的还是CPU版本的torch（torchtext真是霸道！），往复了几次都不行，怎么装都是CUP版本的torch，巨坑！！！

怎么寻找正确的torchtext版本？

一个简单的规律是，torchtext的版本号比torch高一个子版本，然后主版本为0, 阶段版本号最好也是对应的。例如：
torch1.13.1 对应的 torchtext 应该torchtext 0.14.1
那么应该使用下面命令安装
pip install torchtext==0.14.1
上面的规律是对应torch主版本为1的，torch主版本为2的可以参考类似的规律。

感谢博客《更新 torchtext 造成的torch版本不匹配的问题》带来的解答。

# 导入有关torch的工具包
import torch as tc
import torchtext
# 导入torchtext.datasets中的文本分类任务
from torchtext.datasets import AG_NEWS
import os# 定义数据下载路径，当前路径的data文件夹
load_data_path = './Datasets/'
# 如果不存在该路径，则创建这个路径
if not os.path.exists(load_data_path):os.makedirs(load_data_path)# 选取torchtext中的文本分类数据集'AG_NEWS'即新闻主题分类数据，保存在指定目录下
# 将数值映射后的训练和验证数据加载到内存中
train_data, test_data = AG_NEWS(root=load_data_path, split=('train', 'test'))# AG_NEWS返回的数据是一个迭代器，每个元素都是一个元组，包含文本和标签
for (label, text) in train_data:print(f"Label: {label}, Text: {text}")
for (label, text) in test_data:print(f"Label: {label}, Text: {text}")

下载完成后，会有两个以.csv结尾的文件，

数据集中的内容如下：

"3","Fears for T N pension after talks","Unions representing workers at Turner Newall say they are 'disappointed' after talks with stricken parent firm Federal Mogul."
"4","The Race is On: Second Private Team Sets Launch Date for Human Spaceflight (SPACE.com)","SPACE.com - TORONTO, Canada -- A second\team of rocketeers competing for the #36;10 million Ansari X Prize, a contest for\privately funded suborbital space flight, has officially announced the first\launch date for its manned rocket."

训练集有12000个样本，测试集有7600个样本。
一共有四种标签{1,2,3,4}对应{World，Sports，Business，SCI/Tech}分别指世界性新闻、体育新闻、商业新闻和技术类新闻。
每条样本有三列，第一列是标签，说明该新闻属于哪一类；第二列是新闻标题；第三列是新闻简述。
test.csv和train.csv中的格式相同

二、构建Dataset类，读取数据

我们使用上面的代码将数据集进行保存后，新建一个Python文件，开始构建读取数据的Dataset类，代码如下：

#!------------------------第一步：数据读取，构建Dataset类--------------------------------
class AG_NEWS_Data(Dataset):def __init__(self, train=True) -> None:super().__init__()data_path = os.path.join(BASE_PATH, 'train.csv') if train else os.path.join(BASE_PATH, 'test.csv')  # 设置数据路径，本实验中只使用了训练集self.data = pd.read_csv(data_path, sep=',', header=None)  # 读取数据# print(self.data.head())sen_len = []  # 每条样本中文本句子长度self.contents = ''  # 所有样本分词后的内容token_number = 0  # 所有文本中有多少个不同的分词label_count = []  # 所有样本的label标签# * 计算每条样本的长度，取出每条样本的标签label，拼接所有样本内容到contents中for i in range(self.__len__()):content, label = self.__getitem__(i)# for content, label in data:sen_len.append(len(content.split(' ')))  # 每条样本的长度label_count.append(label)  # 取出每条样本的标签labelself.contents += ' '+content  # 拼接样本内容到contents中vocab_dict = {v: idx for idx, v in enumerate(set(self.contents.split(' ')))}  # 获取所有分词集合token_number = len(vocab_dict)sen_len_distribution = {str(i): sen_len.count(i) for i in sorted(set(sen_len))}  # 句子长度分布的字典，如{'80':192,'81':689,...},即长度为80的句子有192个...label_n_distribution = {str(i): label_count.count(i) for i in set(label_count)}  # 标签数量分布的字典，如{'1':20000,'2':20000,...},每个标签对应的样本个数self.vocab_dict, self.token_number, self.sen_len_distribution, self.label_n_distribution = vocab_dict, token_number, sen_len_distribution, label_n_distributiondef __len__(self):return self.data.shape[0]def __getitem__(self, index):label = int(self.data.iloc[index, 0])  # 提取标签，并转换为int类型content = self.data.iloc[index, 1]+' ' + \self.data.iloc[index, 2]  # 拼接样本中的题目和内容文本content = content.lower()  # 将所有单词转换为小写类型# 使用正则表达式，只保留文本中的数字和单词，将其余信息替换为空格content = re.sub(r'[^\w\s]', ' ', content)content = re.sub(r'\s+', ' ', content)  # 将多个空格的位置替换为1个空格return content, label

三、构建网络模型

对网络中的每一层都要设置初始化权重值，权重值的初始换范围一般是一个小于1的数，可以接近零，但不能是0，是0的话，模型会变得特别难训练（大量的经验总结到的）。

只设置三层简单的线性层。

#! -------------------第二步：构建网络模型，构建带有Embedding层的文本分类模型-----
class TextSentiment(nn.Module):"""文本分类模型"""def __init__(self, vocab_size, embed_dim, num_class):"""description:类的初始化函数Args:vocab_size (int): 整个语料包含的不同词汇总数embed_dim (int): 指定词嵌入的维度num_class (int): 文本分类的类别总数"""super().__init__()# 实例化Embedding层，sparse=True代表每次对该层求解梯度，只更新部分权重self.embedding = nn.Embedding(vocab_size, embedding_dim=embed_dim, sparse=True)# 实例化线性层，参数分别是embed_dim和num_classself.fc1 = nn.Linear(in_features=LEN_STA*EMBED_DIM, out_features=512)self.fc2 = nn.Linear(in_features=512, out_features=64)self.fc3 = nn.Linear(in_features=64, out_features=num_class)# 为各层初始化权重self.init_weight()def init_weight(self):"""初始化权重函数"""# 指定初始权重的取值范围数init_range = 0.5# 各层的权重参数都是初始化为均匀分布self.embedding.weight.data.uniform_(-init_range, init_range)for fc in [self.fc1, self.fc2, self.fc3]:fc.weight.data.uniform_(-init_range, init_range)# 偏置初始化为0fc.bias.data.zero_()def forward(self, text):"""正向计算过程Args:text (list): 文本数值映射后的结果Returns:tensor: 与类别数尺寸相同的张量，用以判断文本类别"""# 获得embedding的结果embedded# 此时embedded的尺寸为（m,32)其中m是BACTH_SIZE大小的数据中的词汇总数，32为指定词嵌入的维度EMBED_DIM# print(text.shape)embedded = self.embedding(text)# embedded = F.avg_pool1d(embedded, kernel_size=3)x = embedded.view(embedded.size(0), -1)# print(embedded.shape)# print(len_sta*EMBED_DIM)x = self.fc1(x)x = self.fc2(x)x = self.fc3(x)return x

四、序列化和长度标准化

规范输入句子的长度，并进行序列化，即将文本转换为tensor类型的整数，才可以进行Embedding操作。可以使用one-hot编码进行序列化，这里为了方便直接使用了[0,1,2,3,4,...,]这种单纯的数字。

def get_length_standard(rate=0.9):"""计算文本内容标准化长度的函数，根据样本文本长度的分布情况(从小到大),取前rate的分割点处的长度作为标准长度.Args:rate (float, optional): Defaults to 0.9.Returns:int: 样本文本长度分布中前rate的分割点处的长度"""value_sum = 0  # 统计当前符合条件的样本总数sample_len = len(AG_NEWS)  # 数据集总长度# 取出每个长度对应的样本数量key=句子长度，value=该长度下的样本数量for key, value in AG_NEWS.sen_len_distribution.items():value_sum += int(value)if (value_sum/sample_len >= rate):return int(key)def get_sen_ser(sentence, len_sta):"""对样本内容进行标准化和序列化的函数，多删少补(补0)Args:sentence (str): [description]len_sta ([type]): [description]Returns:[type]: [description]"""# 对句子进行序列化vocab_list = [AG_NEWS.vocab_dict[v] for v in sentence.split(' ')]if (len(vocab_list) >= len_sta):return vocab_list[:len_sta]else:vocab_list.extend([0]*(len_sta-len(vocab_list)))return vocab_list

五、自定义生成Batch的函数

#! --------------------------第四步：自定义生成batch的函数----------------------def generate_batch(batch):"""生成batch数据的函数Args:batch (list): 由样本张量和对应标签的元组组成的batch_size大小的列表，形如：[(sample1,label1),(sample2,label2),...]Returns:tensor: 样本张量和标签各自的列表形式(张量)，形如：text=tensor([sample1,sample2,....]),label=tensor([label1,label2,....])"""label = []  # 存储样本标签text = []  # 存储样本的文本for t, l in batch:# 从batch中获得标签张量text.append(get_sen_ser(t, len_sta=LEN_STA))  # 对文本进行标准化和序列化处理# 从batch中获得样本张量label.append(int(l)-1)  # 序列化标签# text = tc.cat(text)# text = torch.tensor(np.array(text), device=device)text = torch.tensor(text, device=device)return text, torch.tensor(label, device=device)

六、构建训练函数

#!---------------------------第五步：构建训练函数----------------------------
def train(train_data):"""模型训练函数"""# 初始化训练损失和准确率为0train_loss = 0train_acc = 0# 使用数据加载器生成BATCH_SIZE大小的数据进行批次训练# data就是N多个generate_batch函数处理后的BATCH_SIZE大小的数据生成器data = DataLoader(train_data, batch_size=BATCH_SIZE,shuffle=True, collate_fn=generate_batch)  # 使用自定义的generate_batch函数# 对data进行循环遍历，使用每个batch的数据进行参数更新for text, label in data:# 1、设置优化器初始梯度为0optimizer.zero_grad()# 2、模型输入一个批次数据，获得输出label_pre = model(text)# 3、根据真实标签与模型输出计算损失loss = loss_F(label_pre, label)# 4、误差反向传播loss.backward()# 5、更新参数optimizer.step()# 将该批次的损失加到总损失中train_loss += loss.item()# 将该批次的准确率加到总准确率中train_acc += (label_pre.argmax(1) == label).sum().item()# 使用学习率调节器自动调整学习率scheduler.step()# 返回本轮训练的平均损失和平均准确率return train_loss/len(train_data), train_acc/len(train_data)

七、构建验证函数

#!-----------------------------第六步：构建验证函数------------------------
def val(val_data):model.eval()# 初始化训练损失和准确率为0val_loss = 0val_acc = 0# 使用数据加载器生成BATCH_SIZE大小的数据进行批次训练# data就是N多个generate_batch函数处理后的BATCH_SIZE大小的数据生成器data = DataLoader(val_data, batch_size=BATCH_SIZE,shuffle=True, collate_fn=generate_batch)  # 使用自定义的generate_batch函数with torch.no_grad():for text, label in data:label_pre = model(text)# 根据真实标签与模型输出计算损失loss = loss_F(label_pre, label)# 将该损失加入到总损失中val_loss += loss# 将该次的准确个数加入到总个数中val_acc += (label_pre.argmax(1) == label).sum().item()# 返回本轮训练的平均损失和平均准确率return val_loss/len(val_data), val_acc/len(val_data)

八、模型训练和验证

if __name__ == '__main__':# 设置数据的存储路径BASE_PATH = r'H:\Pytorch学习\Datasets\datasets\AG_NEWS'# 检查显卡是否可用device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')# 加载训练数据AG_NEWS = AG_NEWS_Data(train=True)  # 加载数据集generator = torch.Generator().manual_seed(2024)  # 设置随机数生成器和随机种子AG_NEWS_train, AG_NEWS_val = random_split(  # 划分训练集和验证集AG_NEWS, [0.7, 0.3], generator=generator)VOCAB_SIZE = len(AG_NEWS.vocab_dict)  # 获取train_data语料中包含的不同词汇总数BATCH_SIZE = 1000  # 指定BATCH_SIZE的大小EMBED_DIM = 32  # 指定词嵌入的维度NUN_CLASS = 4  # 类别总数LEARN_RATE = 0.005  # 学习率LEN_STA = get_length_standard(0.9)  # 每句话的规范长度，统一长度，多删少补EPOCH = 100  # 设置数据集迭代次数model = TextSentiment(VOCAB_SIZE, EMBED_DIM, NUN_CLASS).to(device)  # 实例化模型loss_F = nn.CrossEntropyLoss().to(device)  # 设置损失函数optimizer = optim.SGD(model.parameters(), lr=LEARN_RATE)  # 设置优化函数scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.9)  # 设置学习率调整器# 进行模型训练和验证for epoch in range(EPOCH):train_loss, train_acc = train(AG_NEWS_train)print(f'epoch {epoch}:\ttrain_loss:{train_loss:.6f}\ttrain_acc:{train_acc:.6f}', end='\t')val_loss, val_acc = val(AG_NEWS_val)print(f'val_loss:{val_loss:.6f}\tval_acc:{val_acc:.6f}')

九、完整代码与输出结果

（一）完整代码

import re
import os
import pandas as pd
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader, random_split
from torch.utils.data import DataLoader
import torch.optim as optim
import torch#!------------------------第一步：数据读取，构建Dataset类-----------------------------
class AG_NEWS_Data(Dataset):def __init__(self, train=True) -> None:super().__init__()data_path = os.path.join(BASE_PATH, 'train.csv') if train else os.path.join(BASE_PATH, 'test.csv')  # 设置数据路径，本实验中只使用了训练集self.data = pd.read_csv(data_path, sep=',', header=None)  # 读取数据# print(self.data.head())sen_len = []  # 每条样本中文本句子长度self.contents = ''  # 所有样本分词后的内容token_number = 0  # 所有文本中有多少个不同的分词label_count = []  # 所有样本的label标签# * 计算每条样本的长度，取出每条样本的标签label，拼接所有样本内容到contents中for i in range(self.__len__()):content, label = self.__getitem__(i)# for content, label in data:sen_len.append(len(content.split(' ')))  # 每条样本的长度label_count.append(label)  # 取出每条样本的标签labelself.contents += ' '+content  # 拼接样本内容到contents中vocab_dict = {v: idx for idx, v in enumerate(set(self.contents.split(' ')))}  # 获取所有分词集合token_number = len(vocab_dict)sen_len_distribution = {str(i): sen_len.count(i) for i in sorted(set(sen_len))}  # 句子长度分布的字典，如{'80':192,'81':689,...},即长度为80的句子有192个...label_n_distribution = {str(i): label_count.count(i) for i in set(label_count)}  # 标签数量分布的字典，如{'1':20000,'2':20000,...},每个标签对应的样本个数self.vocab_dict, self.token_number, self.sen_len_distribution, self.label_n_distribution = vocab_dict, token_number, sen_len_distribution, label_n_distributiondef __len__(self):return self.data.shape[0]def __getitem__(self, index):label = int(self.data.iloc[index, 0])  # 提取标签，并转换为int类型content = self.data.iloc[index, 1]+' ' + \self.data.iloc[index, 2]  # 拼接样本中的题目和内容文本content = content.lower()  # 将所有单词转换为小写类型# 使用正则表达式，只保留文本中的数字和单词，将其余信息替换为空格content = re.sub(r'[^\w\s]', ' ', content)content = re.sub(r'\s+', ' ', content)  # 将多个空格的位置替换为1个空格return content, label#! --------------------第二步：构建网络模型，构建带有Embedding层的文本分类模型------
class TextSentiment(nn.Module):"""文本分类模型"""def __init__(self, vocab_size, embed_dim, num_class):"""description:类的初始化函数Args:vocab_size (int): 整个语料包含的不同词汇总数embed_dim (int): 指定词嵌入的维度num_class (int): 文本分类的类别总数"""super().__init__()# 实例化Embedding层，sparse=True代表每次对该层求解梯度，只更新部分权重self.embedding = nn.Embedding(vocab_size, embedding_dim=embed_dim, sparse=True)# 实例化线性层，参数分别是embed_dim和num_classself.fc1 = nn.Linear(in_features=LEN_STA*EMBED_DIM, out_features=512)self.fc2 = nn.Linear(in_features=512, out_features=64)self.fc3 = nn.Linear(in_features=64, out_features=num_class)# 为各层初始化权重self.init_weight()def init_weight(self):"""初始化权重函数"""# 指定初始权重的取值范围数init_range = 0.5# 各层的权重参数都是初始化为均匀分布self.embedding.weight.data.uniform_(-init_range, init_range)for fc in [self.fc1, self.fc2, self.fc3]:fc.weight.data.uniform_(-init_range, init_range)# 偏置初始化为0fc.bias.data.zero_()def forward(self, text):"""正向计算过程Args:text (list): 文本数值映射后的结果Returns:tensor: 与类别数尺寸相同的张量，用以判断文本类别"""# 获得embedding的结果embedded# 此时embedded的尺寸为（m,32)其中m是BACTH_SIZE大小的数据中的词汇总数，32为指定词嵌入的维度EMBED_DIM# print(text.shape)embedded = self.embedding(text)# embedded = F.avg_pool1d(embedded, kernel_size=3)x = embedded.view(embedded.size(0), -1)# print(embedded.shape)# print(len_sta*EMBED_DIM)x = self.fc1(x)x = self.fc2(x)x = self.fc3(x)return x#! ----------------------第三步：将每个样本中的句子进行长度标准化和序列化-----------------
def get_length_standard(rate=0.9):"""计算文本内容标准化长度的函数，根据样本文本长度的分布情况(从小到大),取前rate的分割点处的长度作为标准长度.Args:rate (float, optional): Defaults to 0.9.Returns:int: 样本文本长度分布中前rate的分割点处的长度"""value_sum = 0  # 统计当前符合条件的样本总数sample_len = len(AG_NEWS)  # 数据集总长度# 取出每个长度对应的样本数量key=句子长度，value=该长度下的样本数量for key, value in AG_NEWS.sen_len_distribution.items():value_sum += int(value)if (value_sum/sample_len >= rate):return int(key)def get_sen_ser(sentence, len_sta):"""对样本内容进行标准化和序列化的函数，多删少补(补0)Args:sentence (str): [description]len_sta ([type]): [description]Returns:[type]: [description]"""# 对句子进行序列化vocab_list = [AG_NEWS.vocab_dict[v] for v in sentence.split(' ')]if (len(vocab_list) >= len_sta):return vocab_list[:len_sta]else:vocab_list.extend([0]*(len_sta-len(vocab_list)))return vocab_list#! --------------------------第四步：自定义生成batch的函数-------------------------
def generate_batch(batch):"""生成batch数据的函数Args:batch (list): 由样本张量和对应标签的元组组成的batch_size大小的列表，形如：[(sample1,label1),(sample2,label2),...]Returns:tensor: 样本张量和标签各自的列表形式(张量)，形如：text=tensor([sample1,sample2,....]),label=tensor([label1,label2,....])"""label = []  # 存储样本标签text = []  # 存储样本的文本for t, l in batch:# 从batch中获得标签张量text.append(get_sen_ser(t, len_sta=LEN_STA))  # 对文本进行标准化和序列化处理# 从batch中获得样本张量label.append(int(l)-1)  # 序列化标签# text = tc.cat(text)# text = torch.tensor(np.array(text), device=device)text = torch.tensor(text, device=device)return text, torch.tensor(label, device=device)#!-----------------------------------第五步：构建训练函数-------------------------
def train(train_data):"""模型训练函数"""# 初始化训练损失和准确率为0train_loss = 0train_acc = 0# 使用数据加载器生成BATCH_SIZE大小的数据进行批次训练# data就是N多个generate_batch函数处理后的BATCH_SIZE大小的数据生成器data = DataLoader(train_data, batch_size=BATCH_SIZE,shuffle=True, collate_fn=generate_batch)  # 使用自定义的generate_batch函数# 对data进行循环遍历，使用每个batch的数据进行参数更新for text, label in data:# 1、设置优化器初始梯度为0optimizer.zero_grad()# 2、模型输入一个批次数据，获得输出label_pre = model(text)# 3、根据真实标签与模型输出计算损失loss = loss_F(label_pre, label)# 4、误差反向传播loss.backward()# 5、更新参数optimizer.step()# 将该批次的损失加到总损失中train_loss += loss.item()# 将该批次的准确率加到总准确率中train_acc += (label_pre.argmax(1) == label).sum().item()# 使用学习率调节器自动调整学习率scheduler.step()# 返回本轮训练的平均损失和平均准确率return train_loss/len(train_data), train_acc/len(train_data)#!-----------------------------第六步：构建验证函数--------------------------
def val(val_data):model.eval()# 初始化训练损失和准确率为0val_loss = 0val_acc = 0# 使用数据加载器生成BATCH_SIZE大小的数据进行批次训练# data就是N多个generate_batch函数处理后的BATCH_SIZE大小的数据生成器data = DataLoader(val_data, batch_size=BATCH_SIZE,shuffle=True, collate_fn=generate_batch)  # 使用自定义的generate_batch函数with torch.no_grad():for text, label in data:label_pre = model(text)# 根据真实标签与模型输出计算损失loss = loss_F(label_pre, label)# 将该损失加入到总损失中val_loss += loss# 将该次的准确个数加入到总个数中val_acc += (label_pre.argmax(1) == label).sum().item()# 返回本轮训练的平均损失和平均准确率return val_loss/len(val_data), val_acc/len(val_data)#! --------------------------第七步：进行模型训练和验证---------------------------------
if __name__ == '__main__':# 设置数据的存储路径BASE_PATH = r'H:\Pytorch学习\Datasets\datasets\AG_NEWS'# 检查显卡是否可用device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')# 加载训练数据AG_NEWS = AG_NEWS_Data(train=True)  # 加载数据集generator = torch.Generator().manual_seed(2024)  # 设置随机数生成器和随机种子AG_NEWS_train, AG_NEWS_val = random_split(  # 划分训练集和验证集AG_NEWS, [0.7, 0.3], generator=generator)VOCAB_SIZE = len(AG_NEWS.vocab_dict)  # 获取train_data语料中包含的不同词汇总数BATCH_SIZE = 1000  # 指定BATCH_SIZE的大小EMBED_DIM = 32  # 指定词嵌入的维度NUN_CLASS = 4  # 类别总数LEARN_RATE = 0.005  # 学习率LEN_STA = get_length_standard(0.9)  # 每句话的规范长度，统一长度，多删少补EPOCH = 100  # 设置数据集迭代次数model = TextSentiment(VOCAB_SIZE, EMBED_DIM, NUN_CLASS).to(device)  # 实例化模型loss_F = nn.CrossEntropyLoss().to(device)  # 设置损失函数optimizer = optim.SGD(model.parameters(), lr=LEARN_RATE)  # 设置优化函数scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.9)  # 设置学习率调整器# 进行模型训练和验证for epoch in range(EPOCH):train_loss, train_acc = train(AG_NEWS_train)print(f'epoch {epoch}:\ttrain_loss:{train_loss:.6f}\ttrain_acc:{train_acc:.6f}', end='\t')val_loss, val_acc = val(AG_NEWS_val)print(f'val_loss:{val_loss:.6f}\tval_acc:{val_acc:.6f}')

（二）输出结果

EPOCH = 100 ，设置数据集迭代了100次，结果如下，可以看出，模型能力有限，有预测能力，但只有一点点。

epoch 0: train_loss:0.017730 train_acc:0.267131 val_loss:0.003573 val_acc:0.266333
epoch 1: train_loss:0.002014 train_acc:0.274238 val_loss:0.001452 val_acc:0.284139
epoch 2: train_loss:0.001417 train_acc:0.289357 val_loss:0.001417 val_acc:0.289583
epoch 3: train_loss:0.001396 train_acc:0.292762 val_loss:0.001392 val_acc:0.293667
epoch 4: train_loss:0.001389 train_acc:0.294369 val_loss:0.001386 val_acc:0.298972
epoch 5: train_loss:0.001383 train_acc:0.298071 val_loss:0.001384 val_acc:0.297611
epoch 6: train_loss:0.001381 train_acc:0.300738 val_loss:0.001382 val_acc:0.303028
epoch 7: train_loss:0.001379 train_acc:0.303667 val_loss:0.001379 val_acc:0.302861
epoch 8: train_loss:0.001376 train_acc:0.304119 val_loss:0.001375 val_acc:0.303528
epoch 9: train_loss:0.001375 train_acc:0.304893 val_loss:0.001376 val_acc:0.300528
epoch 10: train_loss:0.001374 train_acc:0.307119 val_loss:0.001372 val_acc:0.308639
epoch 11: train_loss:0.001372 train_acc:0.308905 val_loss:0.001374 val_acc:0.303667
epoch 12: train_loss:0.001371 train_acc:0.310357 val_loss:0.001372 val_acc:0.309667
epoch 13: train_loss:0.001370 train_acc:0.311393 val_loss:0.001372 val_acc:0.309917
epoch 14: train_loss:0.001369 train_acc:0.311607 val_loss:0.001370 val_acc:0.308667
epoch 15: train_loss:0.001369 train_acc:0.311929 val_loss:0.001370 val_acc:0.312222
epoch 16: train_loss:0.001368 train_acc:0.312952 val_loss:0.001369 val_acc:0.309778
epoch 17: train_loss:0.001368 train_acc:0.313524 val_loss:0.001367 val_acc:0.314528
epoch 18: train_loss:0.001367 train_acc:0.313905 val_loss:0.001368 val_acc:0.315444
epoch 19: train_loss:0.001367 train_acc:0.314810 val_loss:0.001367 val_acc:0.315694
epoch 20: train_loss:0.001366 train_acc:0.315952 val_loss:0.001368 val_acc:0.313333
epoch 21: train_loss:0.001366 train_acc:0.317262 val_loss:0.001367 val_acc:0.314750
epoch 22: train_loss:0.001366 train_acc:0.315976 val_loss:0.001366 val_acc:0.316222
epoch 23: train_loss:0.001365 train_acc:0.317345 val_loss:0.001366 val_acc:0.316139
epoch 24: train_loss:0.001365 train_acc:0.315976 val_loss:0.001366 val_acc:0.316444
epoch 25: train_loss:0.001365 train_acc:0.316786 val_loss:0.001366 val_acc:0.314111
epoch 26: train_loss:0.001365 train_acc:0.316905 val_loss:0.001365 val_acc:0.318611
epoch 27: train_loss:0.001364 train_acc:0.318774 val_loss:0.001365 val_acc:0.316944
epoch 28: train_loss:0.001364 train_acc:0.319036 val_loss:0.001366 val_acc:0.314944
epoch 29: train_loss:0.001364 train_acc:0.318393 val_loss:0.001365 val_acc:0.316111
epoch 30: train_loss:0.001364 train_acc:0.319250 val_loss:0.001365 val_acc:0.316833
epoch 31: train_loss:0.001364 train_acc:0.318440 val_loss:0.001365 val_acc:0.317444
epoch 32: train_loss:0.001364 train_acc:0.319500 val_loss:0.001365 val_acc:0.316444
epoch 33: train_loss:0.001364 train_acc:0.319333 val_loss:0.001365 val_acc:0.315972
epoch 34: train_loss:0.001363 train_acc:0.319786 val_loss:0.001365 val_acc:0.315389
epoch 35: train_loss:0.001363 train_acc:0.319560 val_loss:0.001365 val_acc:0.316583
epoch 36: train_loss:0.001363 train_acc:0.320024 val_loss:0.001365 val_acc:0.316556
epoch 37: train_loss:0.001363 train_acc:0.320774 val_loss:0.001365 val_acc:0.316639
epoch 38: train_loss:0.001363 train_acc:0.320179 val_loss:0.001365 val_acc:0.315889
epoch 39: train_loss:0.001363 train_acc:0.320393 val_loss:0.001365 val_acc:0.315139
epoch 40: train_loss:0.001363 train_acc:0.320774 val_loss:0.001365 val_acc:0.316278
epoch 41: train_loss:0.001363 train_acc:0.320821 val_loss:0.001365 val_acc:0.315167
epoch 42: train_loss:0.001363 train_acc:0.321167 val_loss:0.001365 val_acc:0.315667
epoch 43: train_loss:0.001363 train_acc:0.320619 val_loss:0.001365 val_acc:0.316167
epoch 44: train_loss:0.001363 train_acc:0.320571 val_loss:0.001365 val_acc:0.316778
epoch 45: train_loss:0.001363 train_acc:0.321714 val_loss:0.001365 val_acc:0.316611
epoch 46: train_loss:0.001363 train_acc:0.321143 val_loss:0.001365 val_acc:0.316000
epoch 47: train_loss:0.001363 train_acc:0.321262 val_loss:0.001365 val_acc:0.316056
epoch 48: train_loss:0.001363 train_acc:0.321429 val_loss:0.001365 val_acc:0.315722
epoch 49: train_loss:0.001363 train_acc:0.321036 val_loss:0.001365 val_acc:0.315917
epoch 50: train_loss:0.001363 train_acc:0.321417 val_loss:0.001365 val_acc:0.315639
epoch 51: train_loss:0.001362 train_acc:0.321560 val_loss:0.001365 val_acc:0.315889
epoch 52: train_loss:0.001362 train_acc:0.321524 val_loss:0.001365 val_acc:0.316056
epoch 53: train_loss:0.001362 train_acc:0.321690 val_loss:0.001365 val_acc:0.315889
epoch 54: train_loss:0.001362 train_acc:0.321429 val_loss:0.001365 val_acc:0.316028
epoch 55: train_loss:0.001362 train_acc:0.321536 val_loss:0.001365 val_acc:0.316083
epoch 56: train_loss:0.001362 train_acc:0.321417 val_loss:0.001365 val_acc:0.315639
epoch 57: train_loss:0.001362 train_acc:0.321476 val_loss:0.001365 val_acc:0.315750
epoch 58: train_loss:0.001362 train_acc:0.321512 val_loss:0.001365 val_acc:0.315806
epoch 59: train_loss:0.001362 train_acc:0.321452 val_loss:0.001365 val_acc:0.315861
epoch 60: train_loss:0.001362 train_acc:0.321750 val_loss:0.001365 val_acc:0.316000
epoch 61: train_loss:0.001362 train_acc:0.321298 val_loss:0.001365 val_acc:0.315889
epoch 62: train_loss:0.001362 train_acc:0.321405 val_loss:0.001365 val_acc:0.316000
epoch 63: train_loss:0.001362 train_acc:0.321607 val_loss:0.001365 val_acc:0.315972
epoch 64: train_loss:0.001362 train_acc:0.321583 val_loss:0.001365 val_acc:0.316111
epoch 65: train_loss:0.001362 train_acc:0.321452 val_loss:0.001365 val_acc:0.316056
epoch 66: train_loss:0.001362 train_acc:0.321452 val_loss:0.001365 val_acc:0.316111
epoch 67: train_loss:0.001362 train_acc:0.321583 val_loss:0.001365 val_acc:0.316083
epoch 68: train_loss:0.001362 train_acc:0.321464 val_loss:0.001365 val_acc:0.316111
epoch 69: train_loss:0.001362 train_acc:0.321679 val_loss:0.001365 val_acc:0.316139
epoch 70: train_loss:0.001362 train_acc:0.321476 val_loss:0.001365 val_acc:0.316139
epoch 71: train_loss:0.001362 train_acc:0.321714 val_loss:0.001365 val_acc:0.316111
epoch 72: train_loss:0.001362 train_acc:0.321679 val_loss:0.001365 val_acc:0.316056
epoch 73: train_loss:0.001362 train_acc:0.321560 val_loss:0.001365 val_acc:0.316056
epoch 74: train_loss:0.001362 train_acc:0.321583 val_loss:0.001365 val_acc:0.316028
epoch 75: train_loss:0.001362 train_acc:0.321548 val_loss:0.001365 val_acc:0.316028
epoch 76: train_loss:0.001362 train_acc:0.321500 val_loss:0.001365 val_acc:0.316083
epoch 77: train_loss:0.001362 train_acc:0.321548 val_loss:0.001365 val_acc:0.316056
epoch 78: train_loss:0.001362 train_acc:0.321571 val_loss:0.001365 val_acc:0.316056
epoch 79: train_loss:0.001362 train_acc:0.321548 val_loss:0.001365 val_acc:0.316028
epoch 80: train_loss:0.001362 train_acc:0.321631 val_loss:0.001365 val_acc:0.316028
epoch 81: train_loss:0.001362 train_acc:0.321512 val_loss:0.001365 val_acc:0.316028
epoch 82: train_loss:0.001362 train_acc:0.321536 val_loss:0.001365 val_acc:0.316056
epoch 83: train_loss:0.001362 train_acc:0.321583 val_loss:0.001365 val_acc:0.316056
epoch 84: train_loss:0.001362 train_acc:0.321512 val_loss:0.001365 val_acc:0.316056
epoch 85: train_loss:0.001362 train_acc:0.321560 val_loss:0.001365 val_acc:0.316056
epoch 86: train_loss:0.001362 train_acc:0.321583 val_loss:0.001365 val_acc:0.316056
epoch 87: train_loss:0.001362 train_acc:0.321536 val_loss:0.001365 val_acc:0.316056
epoch 88: train_loss:0.001362 train_acc:0.321548 val_loss:0.001365 val_acc:0.316056
epoch 89: train_loss:0.001362 train_acc:0.321571 val_loss:0.001365 val_acc:0.316056
epoch 90: train_loss:0.001362 train_acc:0.321536 val_loss:0.001365 val_acc:0.316056
epoch 91: train_loss:0.001362 train_acc:0.321560 val_loss:0.001365 val_acc:0.316056
epoch 92: train_loss:0.001362 train_acc:0.321560 val_loss:0.001365 val_acc:0.316083
epoch 93: train_loss:0.001362 train_acc:0.321583 val_loss:0.001365 val_acc:0.316083
epoch 94: train_loss:0.001362 train_acc:0.321571 val_loss:0.001365 val_acc:0.316056
epoch 95: train_loss:0.001362 train_acc:0.321583 val_loss:0.001365 val_acc:0.316056
epoch 96: train_loss:0.001362 train_acc:0.321548 val_loss:0.001365 val_acc:0.316056
epoch 97: train_loss:0.001362 train_acc:0.321536 val_loss:0.001365 val_acc:0.316056
epoch 98: train_loss:0.001362 train_acc:0.321548 val_loss:0.001365 val_acc:0.316056
epoch 99: train_loss:0.001362 train_acc:0.321560 val_loss:0.001365 val_acc:0.316056