class TensorPredictor:
def __init__(self, checkpoint_path):
self.device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
self.model = EmbeddingNet()
self.model.load_state_dict(torch.load(checkpoint_path))
self.model.to(self.device)
self.model.eval()
def predict(self, image_list):
tensor_list = []
for image_tensor in image_list:
#image_tensor = torch.index_select(image_tensor, 2, torch.tensor([2, 1, 0], device=self.device)) # BGR -> RBG ?
image_tensor = F.interpolate(torch.unsqueeze(image_tensor,0), size=(224, 224))[0]
tensor_list.append(image_tensor)
input_tensor = torch.stack(tensor_list)
with torch.no_grad():
input_tensor = input_tensor.to(self.device)
embeddings = self.model(input_tensor)
return embeddings
class Predictor:
def __init__(self, checkpoint_path):
self.device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
self.model = EmbeddingNet()
self.model.load_state_dict(torch.load(checkpoint_path))
self.model.to(self.device)
self.model.eval()
self.transform = transforms.Compose([transforms.Lambda(lambda image: image.convert('RGB')),
transforms.Resize((224, 224)),
transforms.ToTensor()])
def _preprocess(self, image):
image = transforms.ToPILImage()(image)
image = self.transform(image)
return image
def predict(self, image_list):
image_tensor = torch.cat([self._preprocess(im).unsqueeze(0) for im in image_list], dim=0)
with torch.no_grad():
image_tensor = image_tensor.cuda()
embedings = self.model(image_tensor)
return embedings.cpu().numpy()
def main(args):
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
p = args.labels_per_batch
k = args.samples_per_label
batch_size = p * k
model = EmbeddingNet()
if args.resume:
model.load_state_dict(torch.load(args.resume))
model.to(device)
criterion = TripletMarginLoss(margin=args.margin)
optimizer = Adam(model.parameters(), lr=args.lr)
transform = transforms.Compose([
transforms.Lambda(lambda image: image.convert("RGB")),
transforms.Resize((224, 224)),
transforms.PILToTensor(),
transforms.ConvertImageDtype(torch.float),
])
# Using FMNIST to demonstrate embedding learning using triplet loss. This dataset can
# be replaced with any classification dataset.
train_dataset = FashionMNIST(args.dataset_dir,
train=True,
transform=transform,
download=True)
test_dataset = FashionMNIST(args.dataset_dir,
train=False,
transform=transform,
download=True)
# targets is a list where the i_th element corresponds to the label of i_th dataset element.
# This is required for PKSampler to randomly sample from exactly p classes. You will need to
# construct targets while building your dataset. Some datasets (such as ImageFolder) have a
# targets attribute with the same format.
targets = train_dataset.targets.tolist()
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
sampler=PKSampler(targets, p, k),
num_workers=args.workers)
test_loader = DataLoader(test_dataset,
batch_size=args.eval_batch_size,
shuffle=False,
num_workers=args.workers)
for epoch in range(1, args.epochs + 1):
print("Training...")
train_epoch(model, optimizer, criterion, train_loader, device, epoch,
args.print_freq)
print("Evaluating...")
evaluate(model, test_loader, device)
print("Saving...")
save(model, epoch, args.save_dir, "ckpt.pth")
def create_embedder(embedding_model=''):
embedder = EmbeddingNet()
if embedding_model != '':
embedder.load_state_dict(torch.load(embedding_model))
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
embedder = torch.nn.DataParallel(embedder)
embedder.cuda()
return embedder
def main(args):
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
p = args.labels_per_batch
k = args.samples_per_label
batch_size = p * k
model = EmbeddingNet(backbone=None, pretrained=args.pretrained)
print('pretrained: ', args.pretrained)
if args.resume:
model.load_state_dict(torch.load(args.resume))
model.to(device)
criterion = TripletMarginLoss(margin=args.margin)
optimizer = Adam(model.parameters(), lr=args.lr)
train_transform = transforms.Compose([transforms.Lambda(lambda image: image.convert('RGB')),
transforms.RandomHorizontalFlip(),
transforms.Resize((224, 224)),
transforms.ToTensor()])
test_transform = transforms.Compose([transforms.Lambda(lambda image: image.convert('RGB')),
transforms.Resize((224, 224)),
transforms.ToTensor()])
# Using FMNIST to demonstrate embedding learning using triplet loss. This dataset can
# be replaced with any classification dataset.
train_dataset = ImageFolder(os.path.join(args.dataset_dir, 'train'), transform=train_transform)
test_dataset = ImageFolder(os.path.join(args.dataset_dir, 'test'), transform=test_transform)
# targets is a list where the i_th element corresponds to the label of i_th dataset element.
# This is required for PKSampler to randomly sample from exactly p classes. You will need to
# construct targets while building your dataset. Some datasets (such as ImageFolder) have a
# targets attribute with the same format.
targets = train_dataset.targets
train_loader = DataLoader(train_dataset, batch_size=batch_size,
sampler=PKSampler(targets, p, k),
num_workers=args.workers)
test_loader = DataLoader(test_dataset, batch_size=args.eval_batch_size,
shuffle=False,
num_workers=args.workers)
for epoch in range(1, args.epochs + 1):
print('Training...')
train_epoch(model, optimizer, criterion, train_loader, device, epoch, args.print_freq)
print('Evaluating...')
evaluate(model, test_loader, device)
print('Saving...')
save(model, epoch, args.save_dir, 'ckpt.pth')
def create_classifier(embedding_model='', model=''):
embedder = EmbeddingNet()
if embedding_model != '':
embedder.load_state_dict(torch.load(embedding_model))
classifier = FullNet(embedder)
if model != '':
classifier.load_state_dict(torch.load(model))
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
classifier = torch.nn.DataParallel(classifier)
classifier.cuda()
return classifier
def create_lcc(embedding_model='', model=''):
embedder = EmbeddingNet()
if embedding_model != '':
embedder.load_state_dict(torch.load(embedding_model))
lcc = LCCNet(embedder)
if model != '':
lcc.load_state_dict(torch.load(model))
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
lcc = torch.nn.DataParallel(lcc)
lcc.cuda()
return lcc
def create_distance_model(embedding_model='', model=''):
embedder = EmbeddingNet()
if embedding_model != '':
embedder.load_state_dict(torch.load(embedding_model))
distanceModel = DistanceNet(embedder)
if model != '':
distanceModel.load_state_dict(torch.load(model))
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
distanceModel = torch.nn.DataParallel(distanceModel)
distanceModel.cuda()
return distanceModel
def main():
# 4. dataset
mean, std = 0.1307, 0.3081
transform = tfs.Compose([tfs.Normalize((mean, ), (std, ))])
test_transform = tfs.Compose([tfs.ToTensor(),
tfs.Normalize((mean,), (std,))])
train_set = MNIST('./data/MNIST',
train=True,
download=True,
transform=None)
train_set = SEMI_MNIST(train_set,
transform=transform,
num_samples=100)
test_set = MNIST('./data/MNIST',
train=False,
download=True,
transform=test_transform)
test_set = SEMI_MNIST(test_set,
transform=transform,
num_samples=100)
# 5. data loader
train_loader = DataLoader(dataset=train_set,
shuffle=True,
batch_size=1,
num_workers=8,
pin_memory=True
)
test_loader = DataLoader(dataset=test_set,
shuffle=False,
batch_size=1,
)
# 6. model
model = EmbeddingNet().cuda()
model.load_state_dict(torch.load('./saves/state_dict.{}'.format(15)))
# 7. criterion
criterion = MetricCrossEntropy()
data = []
y = []
is_known_ = []
# for idx, (imgs, targets, samples, is_known) in enumerate(train_loader):
# model.train()
# batch_size = 1
# imgs = imgs.cuda() # [N, 1, 28, 28]
# targets = targets.cuda() # [N]
# samples = samples.cuda() # [N, 1, 32, 32]
# is_known = is_known.cuda()
#
# output = model(imgs)
# y.append(targets.cpu().detach().numpy())
# is_known_.append(is_known.cpu().detach().numpy())
#
# if idx % 100 == 0:
# print(idx)
# print(output.size())
#
# data.append(output.cpu().detach().numpy())
#
# data_numpy = np.array(data)
# y_numpy = np.array(y)
# is_known_numpy = np.array(is_known_)
#
# np.save('data', data_numpy)
# np.save('known', is_known_numpy)
# np.save('y', y)
data_numpy = np.load('data.npy')
y_numpy = np.load('y.npy')
is_known_numpy = np.load('known.npy')
print(data_numpy.shape)
print(y_numpy.shape)
data_numpy = np.squeeze(data_numpy)
y_numpy = np.squeeze(y_numpy)
is_known_numpy = np.squeeze(is_known_numpy)
print(data_numpy.shape)
print(y_numpy.shape)
from sklearn.manifold import TSNE
import matplotlib.pyplot as plt
colors = ['#476A2A', '#7851B8', '#BD3430', '#4A2D4E', '#875525',
'#A83683', '#4E655E', '#853541', '#3A3120', '#535D8E']
colors = ['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728',
'#9467bd', '#8c564b', '#e377c2', '#7f7f7f',
'#bcbd22', '#17becf', '#ada699']
# t-SNE 모델 생성 및 학습
# tsne = TSNE(random_state=0)
# digits_tsne = tsne.fit_transform(data_numpy)
# np.save('tsne', digits_tsne)
digits_tsne = np.load('tsne.npy')
print('complete t-sne')
# ------------------------------ 1 ------------------------------
plt.figure(figsize=(10, 10))
for i in range(11):
inds = np.where(y_numpy == i)[0]
known = is_known_numpy[inds]
known_idx = np.where(known == 1)
unknown_idx = np.where(known == 0)
plt.scatter(digits_tsne[inds[unknown_idx], 0], digits_tsne[inds[unknown_idx], 1], alpha=0.5, color=colors[10])
plt.scatter(digits_tsne[inds[known_idx], 0], digits_tsne[inds[known_idx], 1], alpha=0.5, color=colors[i])
# plt.scatter(digits_tsne[inds, 0], digits_tsne[inds, 1], alpha=0.5, color=colors[i])
plt.legend(['0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'unknown'])
plt.show() # 그래프 출력
def main():
# 1. argparse
parser = argparse.ArgumentParser()
parser.add_argument('--epoch', type=int, default=100)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--batch_size', type=int, default=16)
parser.add_argument('--resume', type=int, default=0)
opts = parser.parse_args()
print(opts)
# 2. device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# 3. visdom
vis = visdom.Visdom()
# 4. dataset
mean, std = 0.1307, 0.3081
transform = tfs.Compose([tfs.Normalize((mean, ), (std, ))])
test_transform = tfs.Compose(
[tfs.ToTensor(), tfs.Normalize((mean, ), (std, ))])
train_set = MNIST('./data/MNIST',
train=True,
download=True,
transform=None)
train_set = SEMI_MNIST(train_set, transform=transform, num_samples=100)
test_set = MNIST('./data/MNIST',
train=False,
download=True,
transform=test_transform)
# 5. data loader
train_loader = DataLoader(dataset=train_set,
shuffle=True,
batch_size=opts.batch_size,
num_workers=8,
pin_memory=True)
test_loader = DataLoader(
dataset=test_set,
shuffle=False,
batch_size=opts.batch_size,
)
# 6. model
model = EmbeddingNet().to(device)
# 7. criterion
criterion = MetricCrossEntropy().to(device)
# 8. optimizer
optimizer = torch.optim.SGD(params=model.parameters(),
lr=opts.lr,
momentum=0.9,
weight_decay=5e-4)
# 9. scheduler
scheduler = StepLR(optimizer=optimizer, step_size=50, gamma=1)
# 10. resume
if opts.resume:
model.load_state_dict(
torch.load('./saves/state_dict.{}'.format(opts.resume)))
print("resume from {} epoch..".format(opts.resume - 1))
else:
print("no checkpoint to resume.. train from scratch.")
# --
for epoch in range(opts.resume, opts.epoch):
# 11. trian
for idx, (imgs, targets, samples, is_known) in enumerate(train_loader):
model.train()
batch_size = opts.batch_size
imgs = imgs.to(device) # [N, 1, 28, 28]
targets = targets.to(device) # [N]
samples = samples.to(device) # [N, 1, 32, 32]
is_known = is_known.to(device)
samples = samples.view(batch_size * 10, 1, 28, 28)
out_x = model(imgs) # [N, 10]
out_z = model(samples).view(batch_size, 10,
out_x.size(-1)) # [N * 10 , 2]
loss = criterion(out_x, targets, out_z, is_known, 10, 1)
optimizer.zero_grad()
loss.backward()
optimizer.step()
for param_group in optimizer.param_groups:
lr = param_group['lr']
if idx % 100 == 0:
print('Epoch : {}\t'
'step : [{}/{}]\t'
'loss : {}\t'
'lr : {}\t'.format(epoch, idx, len(train_loader), loss,
lr))
vis.line(X=torch.ones(
(1, 1)) * idx + epoch * len(train_loader),
Y=torch.Tensor([loss]).unsqueeze(0),
update='append',
win='loss',
opts=dict(x_label='step',
y_label='loss',
title='loss',
legend=['total_loss']))
torch.save(model.state_dict(), './saves/state_dict.{}'.format(epoch))
# 12. test
correct = 0
avg_loss = 0
for idx, (img, target) in enumerate(test_loader):
model.load_state_dict(
torch.load('./saves/state_dict.{}'.format(epoch)))
model.eval()
img = img.to(device) # [N, 1, 28, 28]
target = target.to(device) # [N]
output = model(img) # [N, 10]
output = torch.softmax(output, -1)
pred, idx_ = output.max(-1)
print(idx_)
correct += torch.eq(target, idx_).sum()
#loss = criterion(output, target)
#avg_loss += loss.item()
print('Epoch {} test : '.format(epoch))
accuracy = correct.item() / len(test_set)
print("accuracy : {:.4f}%".format(accuracy * 100.))
#avg_loss = avg_loss / len(test_loader)
#print("avg_loss : {:.4f}".format(avg_loss))
vis.line(X=torch.ones((1, 1)) * epoch,
Y=torch.Tensor([accuracy]).unsqueeze(0),
update='append',
win='test',
opts=dict(x_label='epoch',
y_label='test_',
title='test_loss',
legend=['accuracy']))
scheduler.step()
