Pytroch入坑 5. 人臉繼續實驗

原創 橘子2048 2018-08-23 08:52

0.前言

打算將centerloss移植到人臉的識別上,數據集30人，3000左右的圖片

1.實驗

1.resnet18+finetuning

Epoch 0/9
----------
train Loss 0.3178 Acc: 0.2627
val Loss 0.2674 Acc: 0.4140
Epoch 1/9
----------
train Loss 0.2228 Acc: 0.4272
val Loss 0.1230 Acc: 0.6579
Epoch 2/9
----------
train Loss 0.1787 Acc: 0.5164
val Loss 0.0783 Acc: 0.7754
Epoch 3/9
----------
train Loss 0.1484 Acc: 0.5803
val Loss 0.0725 Acc: 0.7912
Epoch 4/9
----------
train Loss 0.1384 Acc: 0.6166
val Loss 0.0774 Acc: 0.7825
Epoch 5/9
----------
train Loss 0.1282 Acc: 0.6359
val Loss 0.0802 Acc: 0.7860
Epoch 6/9
----------
train Loss 0.1198 Acc: 0.6698
val Loss 0.0466 Acc: 0.8632
Epoch 7/9
----------
train Loss 0.0838 Acc: 0.7632
val Loss 0.0256 Acc: 0.9281
Epoch 8/9
----------
train Loss 0.0751 Acc: 0.7777
val Loss 0.0218 Acc: 0.9456
Epoch 9/9
----------
train Loss 0.0661 Acc: 0.8130
val Loss 0.0224 Acc: 0.9439
Training complete in 35m 16s

Best val Acc: 0.945614


2.簡單三層mnistnet（32*7*7-->200-—》30）

出現過擬合，trainACC達到100%，test達到85%.速度很快


3.簡單三層mnistnet（32*7*7-->2-—》30）

train和test達到9%左右，看來全連接層大小影響很大.


4.centerlos+softmaxloss（128*3*3->2_-》30）

權重爲0.1,30個epoch  acc達到12%

權重爲1.5， acc在10%左右，變化不大。


5.centerlos+softmaxloss（128*3*3->10-》30）

權重0.1,30個epoch後  acc達到30%，有一定提升


6.centerlos+softmaxloss（128*3*3->50-》30）

僅採用softmax，20個epoch後， acc達到83%，lossloss達到0.05左右train的ACC接近0.98，過擬合。

Test Loss: 0.058294, Acc: 0.826316
epoch 20
Train Loss: 0.003819, Acc: 0.979069

 centerlos+softmaxloss   經過100個epoch，跑到79%的acc。180個epoc後達到82.6%,效果不大。

Predicting Epoch: 178
Validation accuracy: 83.6842105263%
Training Epoch: 179
Train Loss: 0.002816
Predicting Epoch: 179
Validation accuracy: 82.6315789474%

Training Epoch: 180

代碼： 可看   https://blog.csdn.net/sinat_37787331/article/details/80296964

import torch
from torch.autograd import Variable
import torch.utils.data as Data
import torchvision
from torchvision import transforms
import torch.nn.functional as F
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
import matplotlib.patheffects as PathEffects
from centerLoss import CenterLoss
import torch.optim.lr_scheduler as lr_scheduler
import matplotlib.cm as cm

'''
trainset = torchvision.datasets.MNIST(root='../data', train=True, transform=transforms.Compose(
    [transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]
))

testset = torchvision.datasets.MNIST(root='../data', train=False, transform=transforms.Compose(
    [transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]
))

train_loader = Data.DataLoader(dataset=trainset, batch_size=128, shuffle=True, num_workers=4)
test_loader = Data.DataLoader(dataset=testset, batch_size=128, shuffle=True, num_workers=4)
'''

train_data = torchvision.datasets.ImageFolder('/home/syj/Documents/datas/casia/train',
	                                    transform=transforms.Compose([
	                                        transforms.Resize(28),
	                                        
	                                        transforms.ToTensor(),
	                                        transforms.Normalize([0.485, 0.456, 0.406], [0.229,       0.224, 0.225])

	                                    ]))


train_loader = Data.DataLoader(dataset=train_data, batch_size=40, shuffle=True)



test_data = torchvision.datasets.ImageFolder('/home/syj/Documents/datas/casia/val',
	                                    transform=transforms.Compose([
	                                        transforms.Resize(28),
	                                      
	                                        transforms.ToTensor(),
	                                        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
	                                    ]))
test_loader = Data.DataLoader(dataset=test_data, batch_size=40, shuffle=True)

train_loss = []
test_acc = []


class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.extract = torch.nn.Sequential(
            torch.nn.Linear(784, 512),
            torch.nn.PReLU(),
            torch.nn.Linear(512, 256),
            torch.nn.PReLU(),
            torch.nn.Linear(256, 128),
            torch.nn.PReLU(),
            torch.nn.Linear(128, 64),
            torch.nn.PReLU(),
            torch.nn.Linear(64, 32),
            torch.nn.PReLU(),
            torch.nn.Linear(32, 4),
        )
        self.predict = torch.nn.Sequential(
            torch.nn.PReLU(),
            torch.nn.Linear(4, 10),
        )

    def forward(self, x):
        feature = self.extract(x.view(-1, 784))
        pred = F.log_softmax(self.predict(feature))
        return feature, pred


class ConvNet(torch.nn.Module):
    def __init__(self):
        super(ConvNet, self).__init__()
        self.extract = torch.nn.Sequential(
            torch.nn.Conv2d(3, 32, kernel_size=5, padding=2),
            torch.nn.PReLU(),
            torch.nn.Conv2d(32, 32, kernel_size=5, padding=2),
            torch.nn.PReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Conv2d(32, 64, kernel_size=5, padding=2),
            torch.nn.PReLU(),
            torch.nn.Conv2d(64, 64, kernel_size=5, padding=2),
            torch.nn.PReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Conv2d(64, 128, kernel_size=5, padding=2),
            torch.nn.PReLU(),
            torch.nn.Conv2d(128, 128, kernel_size=5, padding=2),
            torch.nn.PReLU(),
            torch.nn.MaxPool2d(2, 2),
        )
        self.feat = torch.nn.Linear(128*3*3,50)
        self.pred = torch.nn.Sequential(
            torch.nn.Linear(50, 30)
        )

    def forward(self, x):
        x = self.extract(x)
        x = x.view(-1, 128*3*3)
        feat = self.feat(x)
        pred = F.log_softmax(self.pred(feat))
        return feat, pred


#model = Net().cuda()
model = ConvNet().cuda()
optimizer4nn = torch.optim.SGD(model.parameters(), lr=1e-3, momentum=0.9, weight_decay=0.0005)
scheduler = lr_scheduler.StepLR(optimizer4nn, 20, gamma=0.8)

centerloss = CenterLoss(30, 50, 0.1).cuda()
nllloss = torch.nn.NLLLoss().cuda()
#crossentropy = torch.nn.CrossEntropyLoss().cuda()
optimizer4center = torch.optim.SGD(centerloss.parameters(), lr=0.1)


def train(train_loader, model, epoch):
    print("Training Epoch: {}".format(epoch))
    model.train()
    #running_softmax = 0.0
    running_loss = 0.0

    for step, (data, target) in enumerate(train_loader):
        data = Variable(data).cuda()
        target = Variable(target).cuda()
        feat, pred = model(data)
        loss = nllloss(pred, target) + centerloss(target, feat)
        running_loss += loss.data[0]
        optimizer4nn.zero_grad()
        optimizer4center.zero_grad()
        loss.backward()
        optimizer4nn.step()
        optimizer4center.step()
    
    print('Train Loss: {:.6f}'.format(running_loss / (len(
        train_data))))
    
    train_loss.append(running_loss / (len(train_data))) 



def test(test_loader, model, epoch):
    print("Predicting Epoch: {}".format(epoch))
    model.eval()
    test_acc = 0
    total_pred_label = []
    total_target = []
    total_feature = []
    for step, (data, target) in enumerate(test_loader):
        data = Variable(data).cuda()
        target = Variable(target).cuda()
        feature, pred = model(data)
        _, pred_label = pred.max(dim=1)
        total_pred_label.append(pred_label.data.cpu())
        total_target.append(target.data.cpu())
        total_feature.append(feature.data.cpu())

    total_pred_label = torch.cat(total_pred_label, dim=0)
    total_target = torch.cat(total_target, dim=0)
    total_feature = torch.cat(total_feature, dim=0)

    precision = torch.sum(total_pred_label == total_target) / float(total_target.shape[0])
    acc = precision*100
    print("Validation accuracy: {}%".format(precision * 100))
    #test_acc.append(acc )
    #scatter(total_feature.numpy(), total_target.numpy(), epoch)


'''
def scatter(feat, label, epoch):
    plt.ion()
    plt.clf()
    palette = np.array(sns.color_palette('hls', 10))
    ax = plt.subplot(aspect='equal')
    # sc = ax.scatter(feat[:, 0], feat[:, 1], lw=0, s=40, c=palette[label.astype(np.int)])
    for i in range(10):
        plt.plot(feat[label == i, 0], feat[label == i, 1], feat[label == i, 2],'.', c=palette[i])
    plt.legend(['0', '1', '2', '3', '4', '5', '6', '7', '8', '9'], loc='upper right')
    ax.axis('tight')
    for i in range(10):
        xtext, ytext ,ztext= np.median(feat[label == i, :], axis=0)
        txt = ax.text(xtext, ytext,str(i), fontsize=18)
        txt.set_path_effects([PathEffects.Stroke(linewidth=5, foreground="w"), PathEffects.Normal()])

    plt.draw()
    plt.savefig('./benchmark/centerloss_{}.png'.format(epoch))
    plt.pause(0.001)
'''


for epoch in range(200):
    #scheduler.step()
    
    train(train_loader, model, epoch)
    test(test_loader, model, epoch)


torch.save(model, '/home/syj/Documents/model/cen-min-50->30s.pkl')

plt.figure()
#plt.subplot(2, 1, 1)
plt.plot(train_loss, lw = 1.5, label = 'train_loss')

#plt.subplot(2, 2, 2)
#plt.plot(test_acc, lw = 1.5, label = 'test_acc')
plt.savefig("min-cen-50->30-0.1s.jpg")
plt.show()

發表評論
所有評論
還沒有人評論，想成為第一個評論的人麼? 請在上方評論欄輸入並且點擊發布.
相關文章

Pytroch自存系列-numpy常見操作

0.一些鏈接 https://blog.csdn.net/qq_33431368/article/details/86579544 https://mp.weixin.qq.com/s/b8IAf-liXvgn50-3HwT1VA 1.s

橘子2048 2020-06-30 20:06:30

Pytroch入坑 4. Centerloss+Mnist+可視化

橘子2048 2018-08-23 08:52:24

Pytroch入坑 2. 跑一個Minsit的CNN網絡

橘子2048 2018-08-23 08:52:16

Pytroch入坑 1. Deep15.5 + CUDA8.0 + pytroch環境搭建

橘子2048 2018-08-23 08:52:07