def fit(args, model, device, optimizer, loss_fn, dataset, labels_list, task_id):
# Dataloader
train_loader = trainer.get_loader(mnist.getTrain(dataset), args, device, 'train')
val_loader = trainer.get_loader(mnist.getVal(dataset), args, device, 'val')
# Log Best Accuracy
best_val_loss = 0
# Early Stopping
early_stop = 0
# Training loop
for epoch in range(1, args.epochs + 1):
# Prepare model for current task
model.set_task_id(labels_list[task_id])
trainer.train(args, model, device, train_loader, optimizer, epoch, loss_fn)
val_loss, _ = trainer.test(args, model, device, val_loader, loss_fn, val=True)
if val_loss > best_val_loss:
best_val_loss = val_loss
best_state = model.state_dict()
early_stop = 0
else:
early_stop += 1
if early_stop >= args.early_stop_after:
break
return best_state
def get_dataset(args, task_id, split):
trans = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])
if args.dataset=='splitMNIST10':
Dataset = mnist.SplitMNIST10
if split == 'train':
trainval = Dataset('data', task_id, train=True, download=True, transform=trans)
dataset = mnist.getTrain(trainval)
elif split == 'val':
trainval = Dataset('data', task_id, train=True, download=True, transform=trans)
dataset = mnist.getVal(trainval)
elif split == 'trainval':
dataset = Dataset('data', task_id, train=True, download=True, transform=trans)
else:
dataset = Dataset('data', task_id, train=False, download=True, transform=trans)
return dataset
def fit(args, model, device, optimizer, loss_fn, coresets, dataset,
labels_list, task_id):
# Dataloader
train_loader = trainer.get_loader(mnist.getTrain(dataset), args, device,
'train')
val_loader = trainer.get_loader(mnist.getVal(dataset), args, device, 'val')
# Log Best Accuracy
best_val_acc = 0
all_test_accs = []
# Early Stopping
early_stop = 0
# Training loop
for epoch in range(1, args.epochs + 1):
# Prepare model for current task
model.set_range(labels_list[task_id])
trainer.train(args, model, device, train_loader, optimizer, epoch,
loss_fn)
_, val_acc = trainer.test(args,
model,
device,
val_loader,
loss_fn,
val=True)
if val_acc > best_val_acc:
best_val_acc = val_acc
best_state = model.state_dict()
early_stop = 0
else:
early_stop += 1
if early_stop >= args.early_stop_after:
break
# Evaluate all tasks on test set
test_accs = test_all_tasks(coresets, args, model, device, loss_fn,
labels_list, 'test')
all_test_accs.append(test_accs)
return best_state, all_test_accs
skew = m['mu11']/m['mu02']
M = np.float32([[1, skew, -0.5*20*skew], [0, 1, 0]])
img = cv2.warpAffine(img,M,(20, 20),flags=affine_flags)
return img
# HOGDescriptor를 위한 파라미터 설정 및 생성---②
winSize = (20,20)
blockSize = (10,10)
blockStride = (5,5)
cellSize = (5,5)
nbins = 9
hogDesc = cv2.HOGDescriptor(winSize,blockSize,blockStride,cellSize,nbins)
if __name__ =='__main__':
# MNIST 이미지에서 학습용 이미지와 테스트용 이미지 가져오기 ---③
train_data, train_label = mnist.getTrain(reshape=False)
test_data, test_label = mnist.getTest(reshape=False)
# 학습 이미지 글씨 바로 세우기 ---④
deskewed = [list(map(deskew,row)) for row in train_data]
# 학습 이미지 HOG 계산 ---⑤
hogdata = [list(map(hogDesc.compute,row)) for row in deskewed]
train_data = np.float32(hogdata)
print('SVM training started...train data:', train_data.shape)
# 학습용 HOG 데이타 재배열 ---⑥
train_data = train_data.reshape(-1,train_data.shape[2])
# SVM 알고리즘 객체 생성 및 훈련 ---⑦
svm = cv2.ml.SVM_create()
startT = time.time()
svm.trainAuto(train_data, cv2.ml.ROW_SAMPLE, train_label)
endT = time.time() - startT
print('SVM training complete. %.2f Min'%(endT/60))
import numpy as np, cv2
import mnist
# 훈련 데이타와 테스트 데이타 가져오기 ---①
train, train_labels = mnist.getTrain()
test, test_labels = mnist.getTest()
# kNN 객체 생성 및 훈련 ---②
knn = cv2.ml.KNearest_create()
knn.train(train, cv2.ml.ROW_SAMPLE, train_labels)
# k값을 1~10까지 변경하면서 예측 ---③
for k in range(1, 11):
# 결과 예측 ---④
ret, result, neighbors, distance = knn.findNearest(test, k=k)
# 정확도 계산 및 출력 ---⑤
correct = np.sum(result == test_labels)
accuracy = correct / result.size * 100.0
print("K:%d, Accuracy :%.2f%%(%d/%d)" %
(k, accuracy, correct, result.size))