环境
系统 : win 10
显卡:gtx965m
cpu :i7-6700HQ
python 3.61
pytorch 0.3
包引用
import torch from torch.autograd import Variable import torch.nn.functional as F import numpy as np import visdom import time from torch import nn,optim
数据准备
use_gpu = True ones = np.ones((500,2)) x1 = torch.normal(6*torch.from_numpy(ones),2) y1 = torch.zeros(500) x2 = torch.normal(6*torch.from_numpy(ones*[-1,1]),2) y2 = y1 +1 x3 = torch.normal(-6*torch.from_numpy(ones),2) y3 = y1 +2 x4 = torch.normal(6*torch.from_numpy(ones*[1,-1]),2) y4 = y1 +3 x = torch.cat((x1, x2, x3 ,x4), 0).float() y = torch.cat((y1, y2, y3, y4), ).long()
可视化如下看一下:
visdom可视化准备
先建立需要观察的windows
viz = visdom.Visdom() colors = np.random.randint(0,255,(4,3)) #颜色随机 #线图用来观察loss 和 accuracy line = viz.line(X=np.arange(1,10,1), Y=np.arange(1,10,1)) #散点图用来观察分类变化 scatter = viz.scatter( X=x, Y=y+1, opts=dict( markercolor = colors, marksize = 5, legend=["0","1","2","3"]),) #text 窗口用来显示loss 、accuracy 、时间 text = viz.text("FOR TEST") #散点图做对比 viz.scatter( X=x, Y=y+1, opts=dict( markercolor = colors, marksize = 5, legend=["0","1","2","3"] ), )
效果如下:
逻辑回归处理
输入2,输出4
logstic = nn.Sequential( nn.Linear(2,4) )
gpu还是cpu选择:
if use_gpu: gpu_status = torch.cuda.is_available() if gpu_status: logstic = logstic.cuda() # net = net.cuda() print("###############使用gpu##############") else : print("###############使用cpu##############") else: gpu_status = False print("###############使用cpu##############")
优化器和loss函数:
loss_f = nn.CrossEntropyLoss() optimizer_l = optim.SGD(logstic.parameters(), lr=0.001)
训练2000次:
start_time = time.time() time_point, loss_point, accuracy_point = [], [], [] for t in range(2000): if gpu_status: train_x = Variable(x).cuda() train_y = Variable(y).cuda() else: train_x = Variable(x) train_y = Variable(y) # out = net(train_x) out_l = logstic(train_x) loss = loss_f(out_l,train_y) optimizer_l.zero_grad() loss.backward() optimizer_l.step()
训练过成观察及可视化:
if t % 10 == 0: prediction = torch.max(F.softmax(out_l, 1), 1)[1] pred_y = prediction.data accuracy = sum(pred_y ==train_y.data)/float(2000.0) loss_point.append(loss.data[0]) accuracy_point.append(accuracy) time_point.append(time.time()-start_time) print("[{}/{}] | accuracy : {:.3f} | loss : {:.3f} | time : {:.2f} ".format(t + 1, 2000, accuracy, loss.data[0], time.time() - start_time)) viz.line(X=np.column_stack((np.array(time_point),np.array(time_point))), Y=np.column_stack((np.array(loss_point),np.array(accuracy_point))), win=line, opts=dict(legend=["loss", "accuracy"])) #这里的数据如果用gpu跑会出错,要把数据换成cpu的数据 .cpu()即可 viz.scatter(X=train_x.cpu().data, Y=pred_y.cpu()+1, win=scatter,name="add", opts=dict(markercolor=colors,legend=["0", "1", "2", "3"])) viz.text("<h3 align='center' style='color:blue'>accuracy : {}</h3><br><h3 align='center' style='color:pink'>" "loss : {:.4f}</h3><br><h3 align ='center' style='color:green'>time : {:.1f}</h3>" .format(accuracy,loss.data[0],time.time()-start_time),win =text)
我们先用cpu运行一次,结果如下:
然后用gpu运行一下,结果如下:
发现cpu的速度比gpu快很多,但是我听说机器学习应该是gpu更快啊,百度了一下,知乎上的答案是:
我的理解就是gpu在处理图片识别大量矩阵运算等方面运算能力远高于cpu,在处理一些输入和输出都很少的,还是cpu更具优势。
添加神经层:
net = nn.Sequential( nn.Linear(2, 10), nn.ReLU(), #激活函数 nn.Linear(10, 4) )
添加一层10单元神经层,看看效果是否会有所提升:
使用cpu:
使用gpu:
比较观察,似乎并没有什么区别,看来处理简单分类问题(输入,输出少)的问题,神经层和gpu不会对机器学习加持。