可以做基因通路分析的网站,小程序登录入口官网网址,清苑区建设网站找那家公司,wordpress 转载文章文章最前#xff1a; 我是Octopus#xff0c;这个名字来源于我的中文名–章鱼#xff1b;我热爱编程、热爱算法、热爱开源。所有源码在我的个人github #xff1b;这博客是记录我学习的点点滴滴#xff0c;如果您对 Python、Java、AI、算法有兴趣#xff0c;可以关注我的… 文章最前 我是Octopus这个名字来源于我的中文名–章鱼我热爱编程、热爱算法、热爱开源。所有源码在我的个人github 这博客是记录我学习的点点滴滴如果您对 Python、Java、AI、算法有兴趣可以关注我的动态一起学习共同进步。 import os#mac系统上pytorch和matplotlib在jupyter中同时跑需要更改环境变量
os.environ[KMP_DUPLICATE_LIB_OK]TRUE
!pip install gensim
!pip install torchkerasimport torch
import gensim
import torchkeras
print(torch.__version__ , torch.__version__)
print(gensim.__version__ , gensim.__version__)
print(torchkeras.__version__ , torchkeras.__version__)
torch.__version__ 2.0.1
gensim.__version__ 4.3.1
torchkeras.__version__ 3.9.3公众号 算法美食屋 回复关键词pytorch 获取本项目源码和所用数据集百度云盘下载链接。 一准备数据
imdb数据集的目标是根据电影评论的文本内容预测评论的情感标签。
训练集有20000条电影评论文本测试集有5000条电影评论文本其中正面评论和负面评论都各占一半。
文本数据预处理较为繁琐包括文本切词构建词典编码转换序列填充构建数据管道等等。
此处使用gensim中的词典工具并自定义Dataset。
下面进行演示。 import numpy as np
import pandas as pd
import torch MAX_LEN 200 #每个样本保留200个词的长度
BATCH_SIZE 20 dftrain pd.read_csv(./eat_pytorch_datasets/imdb/train.tsv,sep\t,header None,names [label,text])
dfval pd.read_csv(./eat_pytorch_datasets/imdb/test.tsv,sep\t,header None,names [label,text])from gensim import corpora
import string#1文本切词
def textsplit(text):translator str.maketrans(, , string.punctuation)words text.translate(translator).split( )return words#2构建词典
vocab corpora.Dictionary((textsplit(text) for text in dftrain[text]))
vocab.filter_extremes(no_below5,no_above5000)
special_tokens {pad: 0, unk: 1}
vocab.patch_with_special_tokens(special_tokens)
vocab_size len(vocab.token2id)
print(vocab_size ,vocab_size)#3序列填充
def pad(seq,max_length,pad_value0):n len(seq)result seq[pad_value]*max_lengthreturn result[:max_length]#4编码转换
def text_pipeline(text):tokens vocab.doc2idx(textsplit(text))tokens [x if x0 else special_tokens[unk] for x in tokens ]result pad(tokens,MAX_LEN,special_tokens[pad])return result print(text_pipeline(this is an example!))
vocab_size 29924
[145, 77, 569, 55, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]#5构建管道
from torch.utils.data import Dataset,DataLoaderclass ImdbDataset(Dataset):def __init__(self,df):self.df dfdef __len__(self):return len(self.df)def __getitem__(self,index):text self.df[text].iloc[index]label torch.tensor([self.df[label].iloc[index]]).float()tokens torch.tensor(text_pipeline(text)).int() return tokens,labelds_train ImdbDataset(dftrain)
ds_val ImdbDataset(dfval)
dl_train DataLoader(ds_train,batch_size 50,shuffle True)
dl_val DataLoader(ds_val,batch_size 50,shuffle False)
for features,labels in dl_train:break 二定义模型
使用Pytorch通常有三种方式构建模型使用nn.Sequential按层顺序构建模型继承nn.Module基类构建自定义模型继承nn.Module基类构建模型并辅助应用模型容器(nn.Sequential,nn.ModuleList,nn.ModuleDict)进行封装。
此处选择使用第三种方式进行构建。
import torch
from torch import nn
torch.manual_seed(42)
torch._C.Generator at 0x142700950class Net(nn.Module):def __init__(self):super(Net, self).__init__()#设置padding_idx参数后将在训练过程中将填充的token始终赋值为0向量self.embedding nn.Embedding(num_embeddings vocab_size,embedding_dim 3,padding_idx 0)self.conv nn.Sequential()self.conv.add_module(conv_1,nn.Conv1d(in_channels 3,out_channels 16,kernel_size 5))self.conv.add_module(pool_1,nn.MaxPool1d(kernel_size 2))self.conv.add_module(relu_1,nn.ReLU())self.conv.add_module(conv_2,nn.Conv1d(in_channels 16,out_channels 128,kernel_size 2))self.conv.add_module(pool_2,nn.MaxPool1d(kernel_size 2))self.conv.add_module(relu_2,nn.ReLU())self.dense nn.Sequential()self.dense.add_module(flatten,nn.Flatten())self.dense.add_module(linear,nn.Linear(6144,1))def forward(self,x):x self.embedding(x).transpose(1,2)x self.conv(x)y self.dense(x)return ynet Net()
print(net)Net((embedding): Embedding(29924, 3, padding_idx0)(conv): Sequential((conv_1): Conv1d(3, 16, kernel_size(5,), stride(1,))(pool_1): MaxPool1d(kernel_size2, stride2, padding0, dilation1, ceil_modeFalse)(relu_1): ReLU()(conv_2): Conv1d(16, 128, kernel_size(2,), stride(1,))(pool_2): MaxPool1d(kernel_size2, stride2, padding0, dilation1, ceil_modeFalse)(relu_2): ReLU())(dense): Sequential((flatten): Flatten(start_dim1, end_dim-1)(linear): Linear(in_features6144, out_features1, biasTrue))
)Net((embedding): Embedding(8813, 3, padding_idx0)(conv): Sequential((conv_1): Conv1d(3, 16, kernel_size(5,), stride(1,))(pool_1): MaxPool1d(kernel_size2, stride2, padding0, dilation1, ceil_modeFalse)(relu_1): ReLU()(conv_2): Conv1d(16, 128, kernel_size(2,), stride(1,))(pool_2): MaxPool1d(kernel_size2, stride2, padding0, dilation1, ceil_modeFalse)(relu_2): ReLU())(dense): Sequential((flatten): Flatten(start_dim1, end_dim-1)(linear): Linear(in_features6144, out_features1, biasTrue))
)from torchkeras import summary
summary(net,input_datafeatures);
--------------------------------------------------------------------------
Layer (type) Output Shape Param #Embedding-1 [-1, 200, 3] 89,772
Conv1d-2 [-1, 16, 196] 256
MaxPool1d-3 [-1, 16, 98] 0
ReLU-4 [-1, 16, 98] 0
Conv1d-5 [-1, 128, 97] 4,224
MaxPool1d-6 [-1, 128, 48] 0
ReLU-7 [-1, 128, 48] 0
Flatten-8 [-1, 6144] 0
Linear-9 [-1, 1] 6,145Total params: 100,397
Trainable params: 100,397
Non-trainable params: 0
--------------------------------------------------------------------------
Input size (MB): 0.000069
Forward/backward pass size (MB): 0.287788
Params size (MB): 0.382984
Estimated Total Size (MB): 0.670841
--------------------------------------------------------------------------三训练模型
训练Pytorch通常需要用户编写自定义训练循环训练循环的代码风格因人而异。
有3类典型的训练循环代码风格脚本形式训练循环函数形式训练循环类形式训练循环。
此处介绍一种较通用的仿照Keras风格的类形式的训练循环。
该训练循环的代码也是torchkeras库的核心代码。
torchkeras详情: https://github.com/lyhue1991/torchkeras
import os,sys,time
import numpy as np
import pandas as pd
import datetime
from tqdm import tqdm import torch
from torch import nn
from copy import deepcopydef printlog(info):nowtime datetime.datetime.now().strftime(%Y-%m-%d %H:%M:%S)print(\n*8 %s%nowtime)print(str(info)\n)class StepRunner:def __init__(self, net, loss_fn,stage train, metrics_dict None, optimizer None, lr_scheduler None):self.net,self.loss_fn,self.metrics_dict,self.stage net,loss_fn,metrics_dict,stageself.optimizer,self.lr_scheduler optimizer,lr_schedulerdef __call__(self, features, labels):#losspreds self.net(features)loss self.loss_fn(preds,labels)#backward()if self.optimizer is not None and self.stagetrain:loss.backward()self.optimizer.step()if self.lr_scheduler is not None:self.lr_scheduler.step()self.optimizer.zero_grad()#metricsstep_metrics {self.stage_name:metric_fn(preds, labels).item() for name,metric_fn in self.metrics_dict.items()}return loss.item(),step_metricsclass EpochRunner:def __init__(self,steprunner):self.steprunner steprunnerself.stage steprunner.stageself.steprunner.net.train() if self.stagetrain else self.steprunner.net.eval()def __call__(self,dataloader):total_loss,step 0,0loop tqdm(enumerate(dataloader), total len(dataloader))for i, batch in loop: if self.stagetrain:loss, step_metrics self.steprunner(*batch)else:with torch.no_grad():loss, step_metrics self.steprunner(*batch)step_log dict({self.stage_loss:loss},**step_metrics)total_loss lossstep1if i!len(dataloader)-1:loop.set_postfix(**step_log)else:epoch_loss total_loss/stepepoch_metrics {self.stage_name:metric_fn.compute().item() for name,metric_fn in self.steprunner.metrics_dict.items()}epoch_log dict({self.stage_loss:epoch_loss},**epoch_metrics)loop.set_postfix(**epoch_log)for name,metric_fn in self.steprunner.metrics_dict.items():metric_fn.reset()return epoch_logclass KerasModel(torch.nn.Module):def __init__(self,net,loss_fn,metrics_dictNone,optimizerNone,lr_scheduler None):super().__init__()self.history {}self.net netself.loss_fn loss_fnself.metrics_dict nn.ModuleDict(metrics_dict) self.optimizer optimizer if optimizer is not None else torch.optim.Adam(self.parameters(), lr1e-2)self.lr_scheduler lr_schedulerdef forward(self, x):if self.net:return self.net.forward(x)else:raise NotImplementedErrordef fit(self, train_data, val_dataNone, epochs10, ckpt_pathcheckpoint.pt, patience5, monitorval_loss, modemin):for epoch in range(1, epochs1):printlog(Epoch {0} / {1}.format(epoch, epochs))# 1train ------------------------------------------------- train_step_runner StepRunner(net self.net,stagetrain,loss_fn self.loss_fn,metrics_dictdeepcopy(self.metrics_dict),optimizer self.optimizer, lr_scheduler self.lr_scheduler)train_epoch_runner EpochRunner(train_step_runner)train_metrics train_epoch_runner(train_data)for name, metric in train_metrics.items():self.history[name] self.history.get(name, []) [metric]# 2validate -------------------------------------------------if val_data:val_step_runner StepRunner(net self.net,stageval,loss_fn self.loss_fn,metrics_dictdeepcopy(self.metrics_dict))val_epoch_runner EpochRunner(val_step_runner)with torch.no_grad():val_metrics val_epoch_runner(val_data)val_metrics[epoch] epochfor name, metric in val_metrics.items():self.history[name] self.history.get(name, []) [metric]# 3early-stopping -------------------------------------------------if not val_data:continuearr_scores self.history[monitor]best_score_idx np.argmax(arr_scores) if modemax else np.argmin(arr_scores)if best_score_idxlen(arr_scores)-1:torch.save(self.net.state_dict(),ckpt_path)print( reach best {0} : {1} .format(monitor,arr_scores[best_score_idx]),filesys.stderr)if len(arr_scores)-best_score_idxpatience:print( {} without improvement in {} epoch, early stopping .format(monitor,patience),filesys.stderr)break self.net.load_state_dict(torch.load(ckpt_path)) return pd.DataFrame(self.history)torch.no_grad()def evaluate(self, val_data):val_step_runner StepRunner(net self.net,stageval,loss_fn self.loss_fn,metrics_dictdeepcopy(self.metrics_dict))val_epoch_runner EpochRunner(val_step_runner)val_metrics val_epoch_runner(val_data)return val_metricstorch.no_grad()def predict(self, dataloader):self.net.eval()result torch.cat([self.forward(t[0]) for t in dataloader])return result.data
from torchmetrics import Accuracynet Net()
model KerasModel(net,loss_fn nn.BCEWithLogitsLoss(),optimizer torch.optim.Adam(net.parameters(),lr 0.01), metrics_dict {acc:Accuracy(taskbinary)})
model.fit(dl_train,val_datadl_val,epochs10,ckpt_pathcheckpoint,patience3,monitorval_acc,modemax)
2023-08-02 14:20:21
Epoch 1 / 10100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 400/400 [00:1000:00, 39.28it/s, train_acc0.496, train_loss0.701]
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:0100:00, 51.21it/s, val_acc0.518, val_loss0.693]reach best val_acc : 0.5180000066757202 2023-08-02 14:20:33
Epoch 2 / 10100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 400/400 [00:0900:00, 40.14it/s, train_acc0.503, train_loss0.693]
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:0100:00, 54.22it/s, val_acc0.58, val_loss0.689]reach best val_acc : 0.5803999900817871 2023-08-02 14:20:45
Epoch 3 / 10100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 400/400 [00:1000:00, 39.46it/s, train_acc0.69, train_loss0.58]
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:0100:00, 53.84it/s, val_acc0.781, val_loss0.47]reach best val_acc : 0.7807999849319458 2023-08-02 14:20:57
Epoch 4 / 10100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 400/400 [00:0900:00, 40.33it/s, train_acc0.83, train_loss0.386]
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:0100:00, 54.18it/s, val_acc0.819, val_loss0.408]reach best val_acc : 0.8194000124931335 2023-08-02 14:21:09
Epoch 5 / 10100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 400/400 [00:0900:00, 40.63it/s, train_acc0.893, train_loss0.262]
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:0100:00, 55.69it/s, val_acc0.836, val_loss0.395]reach best val_acc : 0.8357999920845032 2023-08-02 14:21:21
Epoch 6 / 10100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 400/400 [00:0900:00, 40.58it/s, train_acc0.932, train_loss0.176]
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:0100:00, 50.93it/s, val_acc0.828, val_loss0.456]2023-08-02 14:21:33
Epoch 7 / 10100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 400/400 [00:1000:00, 39.62it/s, train_acc0.956, train_loss0.119]
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:0100:00, 55.26it/s, val_acc0.829, val_loss0.558]2023-08-02 14:21:44
Epoch 8 / 10100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 400/400 [00:0900:00, 40.58it/s, train_acc0.973, train_loss0.0754]
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:0100:00, 52.91it/s, val_acc0.823, val_loss0.67]val_acc without improvement in 3 epoch, early stopping train_losstrain_accval_lossval_accepoch00.7010640.495800.6930450.5180110.6930600.503350.6886560.5804220.5798670.690100.4695740.7808330.3856250.829900.4076330.8194440.2616530.892600.3949010.8358550.1759210.932100.4556040.8284660.1191780.956100.5584300.8286770.0754090.973300.6701720.82328 四评估模型
import pandas as pd history model.history
dfhistory pd.DataFrame(history)
dfhistory train_losstrain_accval_lossval_accepoch00.7010640.495800.6930450.5180110.6930600.503350.6886560.5804220.5798670.690100.4695740.7808330.3856250.829900.4076330.8194440.2616530.892600.3949010.8358550.1759210.932100.4556040.8284660.1191780.956100.5584300.8286770.0754090.973300.6701720.82328
%matplotlib inline
%config InlineBackend.figure_format svgimport matplotlib.pyplot as pltdef plot_metric(dfhistory, metric):train_metrics dfhistory[train_metric]val_metrics dfhistory[val_metric]epochs range(1, len(train_metrics) 1)plt.plot(epochs, train_metrics, bo--)plt.plot(epochs, val_metrics, ro-)plt.title(Training and validation metric)plt.xlabel(Epochs)plt.ylabel(metric)plt.legend([train_metric, val_metric])plt.show()
plot_metric(dfhistory,loss)plot_metric(dfhistory,acc)# 评估
model.evaluate(dl_val)
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:0100:00, 50.26it/s, val_acc0.836, val_loss0.395]{val_loss: 0.39490113019943235, val_acc: 0.8357999920845032}五使用模型
def predict(net,dl):net.eval()with torch.no_grad():result nn.Sigmoid()(torch.cat([net.forward(t[0]) for t in dl]))return(result.data)
y_pred_probs predict(net,dl_val)
y_pred_probstensor([[0.9372],[1.0000],[0.8672],...,[0.5141],[0.4756],[0.9998]])六保存模型
#模型权重已经被保存在了ckpt_pathcheckpoint.
net_clone Net()
net_clone.load_state_dict(torch.load(checkpoint))
All keys matched successfully如果本书对你有所帮助想鼓励一下作者记得给本项目加一颗星星star⭐️并分享给你的朋友们喔!
如果对本书内容理解上有需要进一步和作者交流的地方欢迎在公众号算法美食屋下留言。作者时间和精力有限会酌情予以回复。
也可以在公众号后台回复关键字加群加入读者交流群和大家讨论。
