从训练打印的log中获取loss等参数绘制曲线

weixin_54533760

已于 2023-03-20 11:18:45 修改

阅读量1k

点赞数 1

文章标签： pytorch 深度学习 Powered by 金山文档

于 2023-03-14 19:28:57 首次发布

本文链接：https://blog.csdn.net/weixin_54533760/article/details/129531885

版权

该代码段展示了如何从训练日志文件中读取数据，特别是关注每个epoch的loss、平均loss和学习率。它使用matplotlib库来绘制这些值随训练进程的变化，包括训练loss、学习率以及对象损失、类别损失和边界框损失的曲线图。

摘要由CSDN通过智能技术生成

在训练中加入log，记录每轮epoch的loss值。

# 打印训练log
log_file.write("Epoch %d/%d | training loss = %.3f, avg_loss = %.3f, lr= %f\n" %
                       (epoch, opts.epoch, loss.item(), losssum / num_train,
                        optimizer.state_dict()['param_groups'][0]['lr']))
log_file.write("TrainLoss:obj loss = %.2f, cls loss = %.2f, bbox loss = %.2f\n" %
                       (conf_loss.item(), cls_loss.item(), bbox_loss.item()))
# 打印的示例如下：
2023-03-17 11:42:04
======================training epoch 1======================
Epoch 1/300 | training loss = 4.946, avg_loss = 0.118, lr= 0.000100
TrainLoss:obj loss = 175.42, cls loss = 1121.56, bbox loss = 345.44
Training consumes 467.38 second

导入训练的打印的log，通过逐行读取并筛选出epoch与loss。

import matplotlib.pyplot as plt
import numpy as np


class curve():
    def __init__(self, path='./log.txt'):
        self.train_losses = []
        self.avg_losses = []
        self.obj_losses = []
        self.cls_losses = []
        self.bbox_losses = []
        self.lrs = []
        self.filename = path
        self.epoch = 0

    def dataload(self):
        with open(self.filename, 'r') as f:
            lines = f.readlines()
            for line in lines:
                if line[0] == 'E':
                    a = line.split('/')[0]
                    epoch = ''.join([x for x in a if x.isdigit()])  # 判断是否为数字
                    self.epoch = int(epoch)
                    train_loss = line.split('=')[1].split(',')[0]
                    train_loss = float(train_loss)
                    self.train_losses.append(train_loss)
                    avg_loss = line.split('=')[2].split(',')[0]
                    self.avg_losses.append(avg_loss)
                    lr = line.split('=')[3]
                    lr = float(lr)
                    self.lrs.append(lr)
                    a = line[0:9]
                elif line[0:9] == 'TrainLoss':
                    objloss = line.split('=')[1].split(',')[0]
                    self.obj_losses.append(objloss)
                    classloss = line.split('=')[2].split(',')[0]
                    self.cls_losses.append(classloss)
                    bboxloss = float(line.split('=')[3].split(',')[0])
                    self.bbox_losses.append(bboxloss)

    def plot(self, title, data, ylim, color):
        epoch = self.epoch
        plt.xlim(1, epoch)
        if ylim:
            plt.ylim(ylim)
        plt.plot(np.arange(epoch), data, color=color)
        plt.xlabel('epoches')
        plt.title(title)
        plt.show()

    def show(self, tl, lr, obj, cls, bbox):
        if tl:
            title = 'train loss'
            self.plot(title, data=self.train_losses, ylim=None, color='red')
        if lr:
            title = 'learning rate'
            self.plot(title, data=self.lrs, ylim=(0.0001, 0.01), color='blue')
        if obj:
            title = 'object loss'
            self.plot(title, data=self.obj_losses, ylim=None, color='red')
        if cls:
            title = 'class loss'
            self.plot(title, data=self.cls_losses, ylim=None, color='red')
        if bbox:
            title = 'boundbox loss'
            self.plot(title,data=self.bbox_losses, ylim=None, color='red')


c = curve()
c.dataload()
c.show(tl=True, lr=True, obj=False, cls=False, bbox=True)