def main():
cuda = torch.cuda.is_available()
anchor_transform = transforms.Compose([
transforms.RandomAffine(degrees=90, translate=(0.25, 0.25)),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.CenterCrop(128),
transforms.Resize(IMG_SIZE),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
train_transforms = transforms.Compose([
transforms.Resize(IMG_SIZE),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# Let's use 12 while developing as it reduces the start time.
dset_train = GeoTileDataset(TILE_FILE,
transform=train_transforms,
center_transform=anchor_transform)
pd_files = dset_train.get_file_df()
weights = pd_files.frequency
train_sampler = WeightedRandomSampler(weights, len(dset_train))
# Should numworkers be 1?
kwargs = {'num_workers': 8, 'pin_memory': True} if cuda else {}
online_train_loader = DataLoader(dset_train,
batch_size=BATCH_SIZE,
sampler=train_sampler,
**kwargs)
model = Loc2Vec()
if cuda:
model.cuda()
loss_fn = OnlineTripletLoss(MARGIN, HardestNegativeTripletSelector(MARGIN),
SemihardNegativeTripletSelector(MARGIN),
RandomNegativeTripletSelector(MARGIN))
optimizer = optim.Adam(model.parameters(), lr=1e-3)
scheduler = lr_scheduler.StepLR(optimizer, 16, gamma=0.1, last_epoch=-1)
# Mixed precision training
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
# if torch.cuda.device_count() > 1:
# print("Let's use", torch.cuda.device_count(), "GPUs!")
# model = nn.DataParallel(model)
fit(online_train_loader, online_train_loader, model, loss_fn, optimizer,
scheduler, N_EPOCHS, cuda, LOG_INTERVAL)
trainLoader = torch.utils.data.DataLoader(dataset = trainDataset, batch_size=mb_size, shuffle=False, num_workers=1, sampler = sampler)
n_sampE, IE_dim = X_trainE.shape
h_dim = hdm
Z_in = h_dim
marg = mrg
lrE = lre
epoch = epch
costtr = []
auctr = []
costts = []
aucts = []
triplet_selector = RandomNegativeTripletSelector(marg)
triplet_selector2 = AllTripletSelector()
class AEE(nn.Module):
def __init__(self):
super(AEE, self).__init__()
self.EnE = torch.nn.Sequential(
nn.Linear(IE_dim, h_dim),
nn.BatchNorm1d(h_dim),
nn.ReLU(),
nn.Dropout())
def forward(self, x):
output = self.EnE(x)
return output
class OnlineTriplet(nn.Module):
train_dataset, batch_sampler=train_batch_sampler, **kwargs)
online_test_loader = torch.utils.data.DataLoader(
test_dataset, batch_sampler=test_batch_sampler, **kwargs)
# Set up the network and training parameters
from networks import EmbeddingNet
from losses import OnlineTripletLoss
from utils import AllTripletSelector, HardestNegativeTripletSelector, RandomNegativeTripletSelector, SemihardNegativeTripletSelector # Strategies for selecting triplets within a minibatch
from metrics import AverageNonzeroTripletsMetric
margin = 1.
embedding_net = EmbeddingNet()
model = embedding_net
if cuda:
model.cuda()
loss_fn = OnlineTripletLoss(margin, RandomNegativeTripletSelector(margin))
lr = 1e-3
optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=1e-4)
scheduler = lr_scheduler.StepLR(optimizer, 8, gamma=0.1, last_epoch=-1)
n_epochs = 20
log_interval = 50
fit(online_train_loader,
online_test_loader,
model,
loss_fn,
optimizer,
scheduler,
n_epochs,
cuda,
log_interval,
def main():
torch.manual_seed(args.seed) #设置随机种子
train_dataset = SpeakerTrainDataset() #设置训练集读取
n_classes = train_dataset.n_classes #说话人数
batch_sampler = BalancedBatchSampler(train_dataset.labels,
train_dataset.count, 10, 10)
print('Num of classes: {}'.format(n_classes))
model = ResNet(layers=[1, 1, 1],
embedding_size=args.embedding_size,
n_classes=n_classes)
model.to(device)
if args.optimizer == 'sgd': #优化器使用sgd
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
dampening=args.dampening,
weight_decay=args.wd)
elif args.optimizer == 'adagrad': #优化器使用adagrad
optimizer = torch.optim.Adagrad(model.parameters(),
lr=args.lr,
lr_decay=args.lr_decay,
weight_decay=args.wd)
else: #优化器使用adam
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.wd)
selector = RandomNegativeTripletSelector(args.m)
criterion = OnlineTripletLoss(args.m, selector)
start = 1
if args.resume: #是否从之前保存的模型开始
if os.path.isfile(args.resume):
print('=> loading checkpoint {}'.format(args.resume))
checkpoint = torch.load(args.resume)
if args.start is not None:
start = start
else:
start = checkpoint['epoch'] + 1
if args.load_optimizer:
optimizer.load_state_dict(checkpoint['optimizer'])
model.load_state_dict(checkpoint['state_dict'])
else:
print('=> no checkpoint found at {}'.format(args.resume))
train_loader = DataLoader(train_dataset,
batch_sampler=batch_sampler,
num_workers=8,
pin_memory=True)
test_dataset = SpeakerTestDataset(transform=args.transform)
test_loader = DataLoader(test_dataset,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=8,
pin_memory=True)
for epoch in range(start, args.epochs + 1):
train(epoch, model, criterion, optimizer, train_loader)
test(model, test_loader) #测试
task = pd.read_csv('task/task.csv',
header=None,
delimiter='[ ]',
engine='python')
pred = pd.read_csv(args.final_dir + 'pred.csv', engine='python')
y_true = np.array(task.iloc[:, 0])
y_pred = np.array(pred.iloc[:, -1])
eer, thresh = cal_eer(y_true, y_pred)
print('\nEER : {:.3%}'.format(eer))
print('Threshold: {:.5f}'.format(thresh))
def main():
torch.manual_seed(seed) # 设置随机种子
train_dataset = SpeakerTrainDataset() # 设置训练集读取
n_classes = train_dataset.n_classes # 说话人数
batch_sampler = BalancedBatchSampler(train_dataset.labels,
train_dataset.count, spks_per_batch,
utts_per_spk)
print('Num of classes: {}'.format(n_classes))
encoder = Encoder(expansion, blocks, embedding_dim).to(device)
if optimizer == 'sgd': # 优化器使用sgd
encoder_optimizer = optim.SGD(encoder.parameters(),
lr=lr,
momentum=momentum,
weight_decay=weight_decay)
elif optimizer == 'adagrad': # 优化器使用adagrad
encoder_optimizer = optim.Adagrad(encoder.parameters(),
lr=lr,
weight_decay=weight_decay)
else: # 优化器使用adam
encoder_optimizer = optim.Adam(encoder.parameters(),
lr=lr,
weight_decay=weight_decay)
selector = RandomNegativeTripletSelector(margin)
triplet_criterion = OnlineTripletLoss(margin, selector).to(device)
start = 1
metric = AverageNonzeroTripletsMetric()
if resume: # 是否从之前保存的模型开始
if os.path.isfile(os.path.join(final_dir, "net.pth")):
print('=> loading checkpoint {}'.format(
os.path.join(final_dir, "net.pth")))
checkpoint = torch.load(os.path.join(final_dir, "net.pth"))
start = checkpoint['epoch'] + 1
if load_optimizer:
encoder_optimizer.load_state_dict(
checkpoint['encoder_optimizer'])
encoder.load_state_dict(checkpoint['encoder_state_dict'])
else:
print('=> no checkpoint found at {}'.format(
os.path.join(final_dir, "net.pth")))
train_loader = DataLoader(train_dataset,
batch_sampler=batch_sampler,
num_workers=8,
pin_memory=True)
test_dataset = SpeakerTestDataset()
test_loader = DataLoader(test_dataset,
batch_size=1,
shuffle=False,
num_workers=8,
pin_memory=True)
for epoch in range(start, epochs + 1):
fake_fbank = train(epoch, encoder, triplet_criterion,
encoder_optimizer, train_loader, metric)
test(encoder, test_loader) # 测试
task = pd.read_csv(TRIAL_FILE,
header=None,
delimiter='[,]',
engine='python')
pred = pd.read_csv(final_dir + prediction, engine='python')
y_true = np.array(task.iloc[:, 0])
y_pred = np.array(pred.iloc[:, -1])
eer, thresh = cal_eer(y_true, y_pred)
print('EER : {:.3%}'.format(eer))
print('Threshold: {:.5f}'.format(thresh))
logger.log_value('eer', eer)
if cuda:
model.cuda()
if args.classify:
loss_fn = torch.nn.CrossEntropyLoss()
else:
margin = args.margin
if args.triplets == 'batch-hard':
loss_fn = OnlineTripletLoss(margin,
HardestNegativeTripletSelector(margin))
if args.triplets == 'semi-hard':
loss_fn = OnlineTripletLoss(
margin, SemihardNegativeTripletSelector(margin))
if args.triplets == 'random-negative':
loss_fn = OnlineTripletLoss(margin,
RandomNegativeTripletSelector(margin))
optimizer = optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.decay)
scheduler = lr_scheduler.StepLR(optimizer, 8, gamma=0.1, last_epoch=-1)
n_epochs = args.epochs
log_interval = args.log
if args.checkpoint:
checkpoint = torch.load('checkpoint.pth.tar')
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
epoch = checkpoint['epoch']
print("Restarting training from checkpoint...")
if args.classify:
def main():
torch.manual_seed(seed) # 设置随机种子
train_dataset = SpeakerTrainDataset() # 设置训练集读取
n_classes = train_dataset.n_classes # 说话人数
batch_sampler = BalancedBatchSampler(train_dataset.labels,
train_dataset.count, spks_per_batch,
utts_per_spk)
print('Num of classes: {}'.format(n_classes))
encoder = Encoder(expansion, blocks, embedding_dim).to(device)
generator = Generator(expansion, blocks, embedding_dim, FEATURE_LEN,
latent_dim).to(device)
discriminator = Discriminator(expansion, blocks, embedding_dim).to(device)
classifier = Classifier(expansion, blocks, n_classes).to(device)
if optimizer == 'sgd': # 优化器使用sgd
generator_optimizer = optim.SGD([{
'params': generator.parameters()
}, {
'params': encoder.parameters()
}, {
'params': discriminator.parameters()
}],
lr=lr,
momentum=momentum,
weight_decay=weight_decay)
discriminator_optimizer = optim.SGD(
[{
'params': discriminator.parameters()
}, {
'params': classifier.parameters()
}],
lr=lr,
momentum=momentum,
weight_decay=weight_decay)
elif optimizer == 'adagrad': # 优化器使用adagrad
generator_optimizer = optim.Adagrad(
[{
'params': generator.parameters()
}, {
'params': encoder.parameters()
}, {
'params': discriminator.parameters()
}],
lr=lr,
weight_decay=weight_decay)
discriminator_optimizer = optim.Adagrad(
[{
'params': discriminator.parameters()
}, {
'params': classifier.parameters()
}],
lr=lr,
weight_decay=weight_decay)
else: # 优化器使用adam
generator_optimizer = optim.Adam([{
'params': generator.parameters()
}, {
'params': encoder.parameters()
}, {
'params': discriminator.parameters()
}],
lr=lr,
weight_decay=weight_decay)
discriminator_optimizer = optim.Adam(
[{
'params': discriminator.parameters()
}, {
'params': classifier.parameters()
}],
lr=lr,
weight_decay=weight_decay)
gen_scheduler = lr_scheduler.StepLR(generator_optimizer,
200,
gamma=0.1,
last_epoch=-1)
dis_scheduler = lr_scheduler.StepLR(discriminator_optimizer,
200,
gamma=0.1,
last_epoch=-1)
selector = RandomNegativeTripletSelector(margin)
triplet_criterion = OnlineTripletLoss(margin, selector).to(device)
bce_criterion = nn.BCELoss().to(device)
softmax_criterion = nn.CrossEntropyLoss().to(device)
start = 1
metric = AverageNonzeroTripletsMetric()
if resume: # 是否从之前保存的模型开始
if os.path.isfile(os.path.join(final_dir, "net.pth")):
print('=> loading checkpoint {}'.format(
os.path.join(final_dir, "net.pth")))
checkpoint = torch.load(os.path.join(final_dir, "net.pth"))
start = checkpoint['epoch'] + 1
if load_optimizer:
generator_optimizer.load_state_dict(
checkpoint['generator_optimizer'])
discriminator_optimizer.load_state_dict(
checkpoint['discriminator_optimizer'])
generator.load_state_dict(checkpoint['generator_state_dict'])
discriminator.load_state_dict(
checkpoint['discriminator_state_dict'])
else:
print('=> no checkpoint found at {}'.format(
os.path.join(final_dir, "net.pth")))
train_loader = DataLoader(train_dataset,
batch_sampler=batch_sampler,
num_workers=8,
pin_memory=True)
test_dataset = SpeakerTestDataset()
test_loader = DataLoader(test_dataset,
batch_size=1,
shuffle=False,
num_workers=8,
pin_memory=True)
for epoch in range(start, epochs + 1):
real_data, fake_fbank = train(epoch, encoder, generator, discriminator,
classifier, triplet_criterion,
bce_criterion, softmax_criterion,
generator_optimizer,
discriminator_optimizer, train_loader,
metric)
save_image(torch.Tensor(random.sample(real_data.tolist(), 4)),
'real_fbank/{}.png'.format(epoch),
nrow=2,
normalize=True)
save_image(torch.Tensor(random.sample(fake_fbank.tolist(), 4)),
'fake_fbank/{}.png'.format(epoch),
nrow=2,
normalize=True)
test(encoder, test_loader) # 测试
gen_scheduler.step()
dis_scheduler.step()
task = pd.read_csv(TRIAL_FILE,
header=None,
delimiter='[,]',
engine='python')
pred = pd.read_csv(final_dir + prediction, engine='python')
y_true = np.array(task.iloc[:, 0])
y_pred = np.array(pred.iloc[:, -1])
eer, thresh = cal_eer(y_true, y_pred)
logger.log_value('eer', eer, epoch)
print('EER : {:.3%}'.format(eer))
print('Threshold: {:.5f}'.format(thresh))
def main():
args = get_args()
logdir = 'log/{}-emb{}-{}layers-{}resblk-lr{}-wd{}-maxlen{}-alpha10-margin{}'\
'{}class-{}sample-{}selector'\
.format(args.name,
args.embedding_size,
args.layers,
args.resblk,
args.lr,
args.wd,
args.maxlen,
args.margin,
args.n_classes,
args.n_samples,
args.selection)
if not os.path.exists(logdir):
os.makedirs(logdir)
resblock = []
for i in range(args.layers):
resblock.append(args.resblk)
if args.train:
logger = Logger(logdir)
if not os.path.exists(args.trainfeature):
os.mkdir(args.trainfeature)
extractFeature(args.training-dataset, args.trainfeature)
trainset = DeepSpkDataset(args.trainfeature, args.maxlen)
pre_loader = DataLoader(trainset, batch_size = 128, shuffle = True, num_workers = 8)
train_batch_sampler = BalancedBatchSampler(trainset.train_labels,
n_classes = args.n_classes,
n_samples = args.n_samples)
kwargs = {'num_workers' : 1, 'pin_memory' : True}
online_train_loader = torch.utils.data.DataLoader(trainset,
batch_sampler=train_batch_sampler,
**kwargs)
margin = args.margin
embedding_net = EmbeddingNet(resblock,
embedding_size = args.embedding_size,
layers = args.layers)
model = DeepSpeaker(embedding_net, trainset.get_num_class())
device = torch.device('cuda:0')
model.to(device) # 要在初始化optimizer之前把model转换到GPU上，这样初始化optimizer的时候也是在GPU上
optimizer = optim.SGD(model.embedding_net.parameters(),
lr = args.lr,
momentum = 0.99,
weight_decay = args.wd)
start_epoch = 0
if args.resume:
if os.path.isfile(args.resume):
print('=> loading checkpoint {}'.format(args.resume))
checkpoint = torch.load(args.resume)
start_epoch = checkpoint['epoch']
model.embedding_net.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
else:
print('=> no checkpoint found at {}'.format(args.resume))
pretrain_epoch = args.pretrain_epoch
if args.selection == 'randomhard':
selector = RandomNegativeTripletSelector(margin)
if args.selection == 'hardest':
selector = HardestNegativeTripletSelector(margin)
if args.selection == 'semihard':
selector = SemihardNegativeTripletSelector(margin)
if args.selection == 'all':
print('warning : select all triplet may take very long time')
selector = AllTripletSelector()
loss_fn = OnlineTripletLoss(margin, selector)
scheduler = lr_scheduler.StepLR(optimizer,
step_size = args.lr_adjust_step,
gamma = args.lr_decay,
last_epoch = -1)
n_epochs = args.n_epochs
log_interval = 50
fit(online_train_loader,
pre_loader,
model,
loss_fn,
optimizer,
scheduler,
pretrain_epoch,
n_epochs,
True,
device,
log_interval,
log_dir = logdir,
eval_path = args.evalfeature,
logger = logger,
metrics = [AverageNonzeroTripletsMetric()],
evaluatee = args.eval,
start_epoch = start_epoch)
else:
if not os.path.exists(args.testfeature):
os.mkdir(args.testfeature)
extractFeature(args.test-dataset, args.testfeature)
model = EmbeddingNet(resblock,
embedding_size = args.embedding_size,
layers = args.layers)
model.cpu()
if args.model:
if os.path.isfile(args.model):
print('=> loading checkpoint {}'.format(args.model))
checkpoint = torch.load(args.model)
model.load_state_dict(checkpoint['state_dict'])
else:
print('=> no checkpoint found at {}'.format(args.model))
thres = np.loadtxt(logdir + '/thres.txt')
acc = np.loadtxt(logdir + '/acc.txt')
idx = np.argmax(acc)
best_thres = thres[idx]
predict(model, args.testfeature, best_thres)
def main(args):
if os.path.exists('models') is False:
os.makedirs('models')
# img_list, base_path, item_dict = read_data(args.dataset, args.bbox)
img_list, base_path, item_dict = read_data("DeepFashion2", bbox_gt=False)
model_save_path = args.model_path # 'models/siames_triplet_df2.pth'
# writer = SummaryWriter('runs/fashion_mnist_experiment_1')
model = ResNetbasedNet()
if os.path.exists(model_save_path):
model.load_state_dict(torch.load(model_save_path))
cuda = torch.cuda.is_available()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if cuda:
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
model.to(device)
kwargs = {'num_workers': 8, 'pin_memory': True} if cuda else {}
if not args.phase:
train_dataset = DeepFashionDataset(img_list['train'], root=base_path)
train_batch_sampler = BalancedBatchSampler(train_dataset.labels,
train_dataset.source,
n_classes=32,
n_samples=4)
online_train_loader = torch.utils.data.DataLoader(
train_dataset, batch_sampler=train_batch_sampler, **kwargs)
test_dataset = DeepFashionDataset(img_list['validation'],
root=base_path)
test_batch_sampler = BalancedBatchSampler(test_dataset.labels,
test_dataset.source,
n_classes=32,
n_samples=4)
online_test_loader = torch.utils.data.DataLoader(
test_dataset, batch_sampler=test_batch_sampler, **kwargs)
margin = 1.
loss_fn = OnlineTripletLoss(margin,
RandomNegativeTripletSelector(margin))
lr = 1e-3
optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=1e-4)
scheduler = lr_scheduler.StepLR(optimizer, 5, gamma=0.1, last_epoch=-1)
n_epochs = 20
log_interval = 200
fit(online_train_loader,
online_test_loader,
model,
loss_fn,
optimizer,
scheduler,
n_epochs,
cuda,
log_interval,
model_save_path,
metrics=[AverageNonzeroTripletsMetric()])
else:
model.eval()
gallery_dataset = DeepFashionDataset(img_list['validation'],
root=base_path)
acc, query_cnt = get_topK_acc(gallery_dataset, item_dict['validation'],
model, cuda,
open('retrieval_result.txt', 'a'), 100)
np.savetxt('TopK_accuracy.txt',
np.concatenate((acc, np.asarray([query_cnt]))),
fmt='%1.5f',
delimiter=',')