In [13]:
mkdir
%matplotlib inline
In [2]:
!ls data/
In [3]:
import os
current_dir = os.getcwd()
data_dir = os.path.join(current_dir, 'data', 'dogscats')
train_dir = os.path.join(data_dir, 'train')
valid_dir = os.path.join(data_dir, 'valid')
test_dir = os.path.join(data_dir, 'test')
In [20]:
!mkdir $data_dir
In [ ]:
!unzip train.zip -d $data_dir
In [ ]:
!unzip test.zip -d $data_dir
In [33]:
!ls -1 $train_dir | wc -l
In [34]:
!ls -1 $test_dir | wc -l
訓練データからランダムに選んだ2000画像をvalidationデータとする
In [26]:
!mkdir $valid_dir
In [ ]:
%cd $train_dir
import os
from glob import glob
import numpy as np
g = glob('*.jpg')
shuf = np.random.permutation(g)
for i in range(2000):
os.rename(shuf[i], os.path.join(valid_dir, shuf[i]))
In [36]:
!ls -1 $valid_dir | wc -l
In [ ]:
# train
%cd $train_dir
%mkdir cats dogs
%mv cat.*.jpg cats/
%mv dog.*.jpg dogs/
In [ ]:
# valid
%cd $valid_dir
%mkdir cats dogs
%mv cat.*.jpg cats/
%mv dog.*.jpg dogs/
In [ ]:
# test
%cd $test_dir
%mkdir unknown
%mv *.jpg unknown
In [4]:
vgg16 = models.vgg16(pretrained=True)
vgg16.eval() # eval mode!
Out[4]:
# 最後のfc層のみ2クラス分類できるように置き換える
num_features = vgg16.classifier[6].in_features
vgg16.classifier[6] = nn.Linear(num_features, 2) # <= この代入はできない！
In [5]:
# 全層のパラメータを固定
for param in vgg16.parameters():
param.requires_grad = False
vgg16.classifier = nn.Sequential(
nn.Linear(25088, 4096),
nn.ReLU(inplace=True),
nn.Dropout(0.5),
nn.Linear(4096, 4096),
nn.ReLU(inplace=True),
nn.Dropout(0.5),
nn.Linear(4096, 2)
)
use_gpu = torch.cuda.is_available()
if use_gpu:
vgg16 = vgg16.cuda()
print(vgg16)
Resize(224) は動作しない
In [6]:
train_preprocess = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
test_preprocess = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
In [7]:
train_dataset = datasets.ImageFolder(train_dir, train_preprocess)
valid_dataset = datasets.ImageFolder(valid_dir, test_preprocess)
test_dataset = datasets.ImageFolder(test_dir, test_preprocess)
In [8]:
# DataSetのlenはサンプル数
print(len(train_dataset))
print(len(valid_dataset))
print(len(test_dataset))
In [9]:
classes = train_dataset.classes
print(train_dataset.classes)
print(valid_dataset.classes)
print(test_dataset.classes)
In [10]:
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=128,
shuffle=True)
valid_loader = torch.utils.data.DataLoader(valid_dataset,
batch_size=128,
shuffle=False)
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=128,
shuffle=False)
In [11]:
# DataLoaderのlenはミニバッチ数
print(len(train_loader))
print(len(valid_loader))
print(len(test_loader))
In [14]:
def imshow(images, title=None):
images = images.numpy().transpose((1, 2, 0)) # (h, w, c)
# denormalize
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])
images = std * images + mean
images = np.clip(images, 0, 1)
plt.imshow(images)
if title is not None:
plt.title(title)
images, classes = next(iter(train_loader))
print(images.size(), classes.size())
images = torchvision.utils.make_grid(images[:25], nrow=5)
imshow(images)
requires_grad = False している vgg16.parameters() を指定するとエラーになる
In [15]:
if use_gpu:
vgg16 = vgg16.cuda()
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(vgg16.classifier.parameters(), lr=0.001, momentum=0.9)
In [16]:
def train(model, criterion, optimizer, train_loader):
model.train()
running_loss = 0
for batch_idx, (images, labels) in enumerate(train_loader):
if use_gpu:
images = Variable(images.cuda())
labels = Variable(labels.cuda())
else:
images = Variable(images)
labels = Variable(labels)
optimizer.zero_grad()
outputs = model(images)
loss = criterion(outputs, labels)
running_loss += loss.data[0]
loss.backward()
optimizer.step()
train_loss = running_loss / len(train_loader)
return train_loss
def valid(model, criterion, valid_loader):
model.eval()
running_loss = 0
correct = 0
total = 0
for batch_idx, (images, labels) in enumerate(valid_loader):
if use_gpu:
images = Variable(images.cuda())
labels = Variable(labels.cuda())
else:
images = Variable(images)
labels = Variable(labels)
outputs = model(images)
loss = criterion(outputs, labels)
running_loss += loss.data[0]
_, predicted = torch.max(outputs.data, 1)
correct += (predicted == labels.data).sum()
total += labels.size(0)
val_loss = running_loss / len(valid_loader)
val_acc = correct / total
return val_loss, val_acc
In [18]:
%mkdir logs
In [ ]:
num_epochs = 5
log_dir = './logs'
best_acc = 0
loss_list = []
val_loss_list = []
val_acc_list = []
for epoch in range(num_epochs):
loss = train(vgg16, criterion, optimizer, train_loader)
val_loss, val_acc = valid(vgg16, criterion, valid_loader)
print('epoch %d, loss: %.4f val_loss: %.4f val_acc: %.4f'
% (epoch, loss, val_loss, val_acc))
if val_acc > best_acc:
print('val_acc improved from %.5f to %.5f!' % (best_acc, val_acc))
best_acc = val_acc
model_file = 'epoch%03d-%.3f-%.3f.pth' % (epoch, val_loss, val_acc)
torch.save(vgg16.state_dict(), os.path.join(log_dir, model_file))
# logging
loss_list.append(loss)
val_loss_list.append(val_loss)
val_acc_list.append(val_acc)
認識率：99.3%
epoch 0, loss: 0.1068 val_loss: 0.0346 val_acc: 0.9905
val_acc improved from 0.00000 to 0.99050!
epoch 1, loss: 0.0351 val_loss: 0.0316 val_acc: 0.9925
val_acc improved from 0.99050 to 0.99250!
epoch 2, loss: 0.0267 val_loss: 0.0315 val_acc: 0.9930
val_acc improved from 0.99250 to 0.99300!
epoch 3, loss: 0.0224 val_loss: 0.0317 val_acc: 0.9930
epoch 4, loss: 0.0175 val_loss: 0.0311 val_acc: 0.9920
In [ ]: