def get_fewshot_testdata(args):
dat_set = dataset.Omniglot(root=args.data_dir,
train=False,
transform=transforms.Compose([
transforms.Resize(32),
#transforms.Grayscale(num_output_channels=3),
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
]))
classes_per_it = args.test_way
num_samples = args.test_query+args.test_shot
sampler = PrototypicalBatchSampler(labels=dat_set.labels,
classes_per_it=classes_per_it,
num_samples=num_samples,
iterations=args.test_episodes)
dataloader = torch.utils.data.DataLoader(dat_set, batch_sampler=sampler)
return dataloader
def train(args, model, device, optimizer, tri_loss, exp_dir):
#change rho avlue accoding to training numbers
if args.trainsize > 2000 and args.trainsize <= 6000:
rho = 1.7e-3
elif args.trainsize > 6000 and args.trainsize <= 8000:
rho = 1.5e-3
elif args.trainsize > 8000 and args.trainsize <= 10000:
rho = 1.3e-3
elif args.trainsize > 10000 and args.trainsize <= 12000:
rho = 1.1e-3
elif args.trainsize > 12000 and args.trainsize <= 14000:
rho = 0.9e-3
elif args.trainsize > 14000 and args.trainsize <= 16000:
rho = 0.7e-3
else:
rho = args.rho
#start episodic training
total_NMI = np.zeros(args.iteration)
total_AMI = np.zeros(args.iteration)
total_SMI = np.zeros(args.iteration)
total_ACCU = np.zeros(args.iteration + 1)
#Tesing before self-training
print('Tesing before self-training')
accu = test(args, model, device)
total_ACCU[0] = accu
for iter_n in range(args.iteration):
#generate data loader
extraction_loader = DataLoader(
dataset.Omniglot(
root=args.data_dir,
train=True,
size=args.trainsize,
transform=transforms.Compose([
transforms.Resize(32),
#transforms.Grayscale(num_output_channels=3),
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=4,
pin_memory=True)
#extract all data features
train_features, target_labels = extract_features(
model=model, data_loader=extraction_loader, device=device)
#rerank to get the jaccard distance
rerank_dist = re_ranking(features=train_features,
MemorySave=args.memory_save)
#build the DBSCAN model
tri_mat = np.triu(rerank_dist, 1) # tri_mat.dim=2
tri_mat = tri_mat[np.nonzero(tri_mat)] # tri_mat.dim=1
tri_mat = np.sort(tri_mat, axis=None)
top_num = np.round(rho * tri_mat.size).astype(int)
eps = tri_mat[:top_num].mean()
print('eps in cluster: {:.3f}'.format(eps))
cluster = DBSCAN(eps=eps,
min_samples=4,
metric='precomputed',
n_jobs=8)
# select & cluster images as training set of this episode
print('Clustering and labeling...')
train_features = train_features.cpu().numpy()
labels = cluster.fit_predict(rerank_dist)
#calculate NMI of chosed data points of current episode
TL = target_labels
list_true = [int(TL[i].cpu().numpy()) for i in range(len(TL))]
list_pred = labels.tolist()
NMI = nmi_withGT(list_pred, list_true)
AMI = ami_withGT(list_pred, list_true)
SMI = sampling_NMI_withGT(list_pred, list_true)
total_NMI[iter_n] = NMI
total_AMI[iter_n] = AMI
total_SMI[iter_n] = SMI
num_ids = len(set(labels)) - 1
#generate new dataset
new_dataset = []
unique_labels, label_count = np.unique(labels, return_counts=True)
for i in range(len(extraction_loader.dataset.splittxt)):
idd = np.where(unique_labels == labels[i])[0][0]
if labels[i] == -1 or label_count[idd] < 6:
continue
new_dataset.append(
(extraction_loader.dataset.splittxt[i], labels[i], 0))
LL = [new_dataset[i][1] for i in range(len(new_dataset))]
print(np.unique(LL, return_counts=True))
print(
'Iteration {} have {} training ids, {} training images, NMI is {}, AMI is {}, SMI is {}'
.format(iter_n + 1, num_ids, len(new_dataset), NMI, AMI, SMI))
#triplet_proto dataloader
BS = args.batch_size * args.ims_per_id
train_loader = DataLoader(
dataset.Omniglot_clustering(root=args.data_dir,
dat_set=new_dataset,
transform=transforms.Compose([
transforms.Resize(32),
transforms.ToTensor(),
transforms.Normalize((0.1307, ),
(0.3081, ))
])),
batch_size=BS,
num_workers=4,
sampler=RandomIdentitySampler(new_dataset, args.ims_per_id),
pin_memory=True,
drop_last=True)
#training with prototipical learning methods
for ep in range(args.epochs):
# Adjust Learning Rate
adjust_lr_exp(optimizer, args.base_lr, ep + 1, args.epochs,
args.exp_decay_at_epoch)
model.train()
protoacc_meter = AverageMeter()
protoloss_meter = AverageMeter()
triloss_meter = AverageMeter()
totalloss_meter = AverageMeter()
ep_st = time.time()
for data, target in tqdm(train_loader):
#pdb.set_trace()
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
feat, _ = model(data)
protoloss, acc, support_idxs, query_idxs = loss_fn(
feat, target=target, n_support=args.train_shot)
protoloss = protoloss.to(device)
triloss, _, _, _, _, _ = global_loss_support(
tri_loss,
feat,
target,
support_idxs,
query_idxs,
normalize_feature=args.normalize_feature,
hard_mining=args.hard_mining)
total_loss = protoloss + triloss
total_loss.backward()
optimizer.step()
protoacc_meter.update(acc.item())
protoloss_meter.update(protoloss.item())
triloss_meter.update(triloss.item())
totalloss_meter.update(total_loss.item())
#Epoch log
time_log = 'Ep {}, {:.2f}s'.format(
ep,
time.time() - ep_st,
)
loss_log = (
', acc {:.2%}, protoloss {:.4f}, triloss {:.4f}, total loss {:.4f}'
.format(protoacc_meter.avg, protoloss_meter.avg,
triloss_meter.avg, totalloss_meter.avg))
final_log = time_log + loss_log
print(final_log)
#adjust learning rate back to initialized learning rate
print('Learning rate adjuested back to base learning rate {:.10f}'.
format(args.base_lr))
for g in optimizer.param_groups:
g['lr'] = args.base_lr
accu = test(args, model, device)
total_ACCU[iter_n + 1] = accu
print('total NMI value is, ', total_NMI)
print('total AMI value is, ', total_AMI)
print('total SMI value is, ', total_SMI)
print('total ACCU value is, ', total_ACCU)