def init_prob_kmeans(model, eval_loader, args):
torch.manual_seed(1)
model = model.to(device)
# cluster parameter initiate
model.eval()
targets = np.zeros(len(eval_loader.dataset))
feats = np.zeros((len(eval_loader.dataset), 1024))
for _, (x, _, label, idx) in enumerate(eval_loader):
x = x.to(device)
_, feat = model(x)
feat = feat.view(x.size(0), -1)
idx = idx.data.cpu().numpy()
feats[idx, :] = feat.data.cpu().numpy()
targets[idx] = label.data.cpu().numpy()
# evaluate clustering performance
pca = PCA(n_components=args.n_clusters)
feats = pca.fit_transform(feats)
kmeans = KMeans(n_clusters=args.n_clusters, n_init=20)
y_pred = kmeans.fit_predict(feats)
acc, nmi, ari = cluster_acc(targets,
y_pred), nmi_score(targets, y_pred), ari_score(
targets, y_pred)
print('Init acc {:.4f}, nmi {:.4f}, ari {:.4f}'.format(acc, nmi, ari))
probs = feat2prob(torch.from_numpy(feats),
torch.from_numpy(kmeans.cluster_centers_))
return kmeans.cluster_centers_, probs
def test(model, test_loader, args):
model.eval()
preds=np.array([])
targets=np.array([])
for batch_idx, (x, label, _) in enumerate(tqdm(test_loader)):
x, label = x.to(device), label.to(device)
output1, output2, _ = model(x)
if args.head=='head1':
output = output1
else:
output = output2
_, pred = output.max(1)
targets=np.append(targets, label.cpu().numpy())
preds=np.append(preds, pred.cpu().numpy())
acc, nmi, ari = cluster_acc(targets.astype(int), preds.astype(int)), nmi_score(targets, preds), ari_score(targets, preds)
print('Test acc {:.4f}, nmi {:.4f}, ari {:.4f}'.format(acc, nmi, ari))
def test(model, test_loader, args, epoch='test'):
model.eval()
preds=np.array([])
targets=np.array([])
feats = np.zeros((len(test_loader.dataset), args.n_clusters))
probs= np.zeros((len(test_loader.dataset), args.n_clusters))
for batch_idx, (x, label, idx) in enumerate(tqdm(test_loader)):
x, label = x.to(device), label.to(device)
_, feat = model(x)
prob = feat2prob(feat, model.center)
_, pred = prob.max(1)
targets=np.append(targets, label.cpu().numpy())
preds=np.append(preds, pred.cpu().numpy())
idx = idx.data.cpu().numpy()
feats[idx, :] = feat.cpu().detach().numpy()
probs[idx, :] = prob.cpu().detach().numpy()
acc, nmi, ari = cluster_acc(targets.astype(int), preds.astype(int)), nmi_score(targets, preds), ari_score(targets, preds)
print('Test acc {:.4f}, nmi {:.4f}, ari {:.4f}'.format(acc, nmi, ari))
probs = torch.from_numpy(probs)
return acc, nmi, ari, probs
def test(model, eval_loader, args):
model.eval()
targets = np.zeros(len(eval_loader.dataset))
y_pred = np.zeros(len(eval_loader.dataset))
probs= np.zeros((len(eval_loader.dataset), args.n_clusters))
for _, (x, _, label, idx) in enumerate(eval_loader):
x = x.to(device)
_, feat = model(x)
prob = feat2prob(feat, model.center)
# prob = F.softmax(logit, dim=1)
idx = idx.data.cpu().numpy()
y_pred[idx] = prob.data.cpu().detach().numpy().argmax(1)
targets[idx] = label.data.cpu().numpy()
probs[idx, :] = prob.cpu().detach().numpy()
# evaluate clustering performance
y_pred = y_pred.astype(np.int64)
acc, nmi, ari = cluster_acc(targets, y_pred), nmi_score(targets, y_pred), ari_score(targets, y_pred)
print('Test acc {:.4f}, nmi {:.4f}, ari {:.4f}'.format(acc, nmi, ari))
probs = torch.from_numpy(probs)
return acc, nmi, ari, probs
def estimate_k(model, unlabeled_loader, labeled_loaders, args):
u_num = len(unlabeled_loader.dataset)
u_targets = np.zeros(u_num)
u_feats = np.zeros((u_num, 1024))
print('extracting features for unlabeld data')
for _, (x, _, label, idx) in enumerate(unlabeled_loader):
x = x.to(device)
_, feat = model(x)
feat = feat.view(x.size(0), -1)
idx = idx.data.cpu().numpy()
u_feats[idx, :] = feat.data.cpu().numpy()
u_targets[idx] = label.data.cpu().numpy()
cand_k = np.arange(args.max_cand_k)
#get acc for labeled data with short listed k
best_ks = np.zeros(len(omniglot_background_val_alphabets))
print('extracting features for labeld data')
for alphabetStr in omniglot_background_val_alphabets:
labeled_loader = labeled_loaders[alphabetStr]
args.num_val_cls = labeled_loader.num_classes
l_num = len(labeled_loader.dataset)
l_targets = np.zeros(l_num)
l_feats = np.zeros((l_num, 1024))
for _, (x, _, label, idx) in enumerate(labeled_loader):
x = x.to(device)
_, feat = model(x)
feat = feat.view(x.size(0), -1)
idx = idx.data.cpu().numpy()
l_feats[idx, :] = feat.data.cpu().numpy()
l_targets[idx] = label.data.cpu().numpy()
l_classes = set(l_targets)
num_lt_cls = int(round(len(l_classes) * args.split_ratio))
lt_classes = set(random.sample(l_classes, num_lt_cls))
lv_classes = l_classes - lt_classes
lt_feats = np.empty((0, l_feats.shape[1]))
lt_targets = np.empty(0)
for c in lt_classes:
lt_feats = np.vstack((lt_feats, l_feats[l_targets == c]))
lt_targets = np.append(lt_targets, l_targets[l_targets == c])
lv_feats = np.empty((0, l_feats.shape[1]))
lv_targets = np.empty(0)
for c in lv_classes:
lv_feats = np.vstack((lv_feats, l_feats[l_targets == c]))
lv_targets = np.append(lv_targets, l_targets[l_targets == c])
cvi_list = np.zeros(len(cand_k))
acc_list = np.zeros(len(cand_k))
cat_pred_list = np.zeros([len(cand_k), u_num + l_num])
print('estimating K ...')
for i in range(len(cand_k)):
cvi_list[i], cat_pred_i = labeled_val_fun(
np.concatenate((lv_feats, u_feats)), lt_feats, lt_targets,
cand_k[i] + args.num_val_cls)
cat_pred_list[i, :] = cat_pred_i
acc_list[i] = cluster_acc(
lv_targets,
cat_pred_i[len(lt_targets):len(lt_targets) + len(lv_targets)])
idx_cvi = np.max(np.argwhere(cvi_list == np.max(cvi_list)))
idx_acc = np.max(np.argwhere(acc_list == np.max(acc_list)))
idx_best = int(math.ceil((idx_cvi + idx_acc) * 1.0 / 2))
cat_pred = cat_pred_list[idx_best, :]
cnt_cat = Counter(cat_pred.tolist())
cnt_l = Counter(cat_pred[:l_num].tolist())
cnt_ul = Counter(cat_pred[l_num:].tolist())
bin_cat = [x[1] for x in sorted(cnt_cat.items())]
bin_l = [x[1] for x in sorted(cnt_l.items())]
bin_ul = [x[1] for x in sorted(cnt_ul.items())]
expectation = u_num * 1.0 / (cand_k[idx_best] + args.num_val_cls)
best_k = np.sum(
np.array(bin_ul) /
np.max(bin_ul).astype(float) > args.min_max_ratio)
print('current best K {}'.format(best_k))
i_alpha = omniglot_background_val_alphabets.index(alphabetStr)
best_ks[i_alpha] = best_k
best_k = np.ceil(np.mean(best_ks)).astype(np.int32)
kmeans = KMeans(n_clusters=best_k)
u_pred = kmeans.fit_predict(u_feats).astype(np.int32)
acc, nmi, ari = cluster_acc(u_targets, u_pred), nmi_score(
u_targets, u_pred), ari_score(u_targets, u_pred)
print('Final K {}, acc {:.4f}, nmi {:.4f}, ari {:.4f}'.format(
best_k, acc, nmi, ari))
return best_k
def estimate_k(model, unlabeled_loader, labeled_loader, args):
u_num = len(unlabeled_loader.dataset)
u_targets = np.zeros(u_num)
u_feats = np.zeros((u_num, 512))
print('extracting features for unlabeld data')
for _, (x, label, idx) in enumerate(tqdm(unlabeled_loader)):
x = x.to(device)
_, feat = model(x)
feat = feat.view(x.size(0), -1)
idx = idx.data.cpu().numpy()
u_feats[idx, :] = feat.data.cpu().numpy()
u_targets[idx] = label.data.cpu().numpy()
cand_k = np.arange(args.max_cand_k)
#get acc for labeled data with short listed k
l_num = len(labeled_loader.dataset)
l_targets = np.zeros(l_num)
l_feats = np.zeros((l_num, 512))
print('extracting features for labeld data')
for _, (x, label, idx) in enumerate(tqdm(labeled_loader)):
x = x.to(device)
_, feat = model(x)
feat = feat.view(x.size(0), -1)
idx = idx.data.cpu().numpy()
l_feats[idx, :] = feat.data.cpu().numpy()
l_targets[idx] = label.data.cpu().numpy()
l_classes = set(l_targets)
num_lt_cls = int(round(len(l_classes) * args.split_ratio))
lt_classes = set(random.sample(
l_classes,
num_lt_cls)) #random sample 5 classes from all labeled classes
lv_classes = l_classes - lt_classes
lt_feats = np.empty((0, l_feats.shape[1]))
lt_targets = np.empty(0)
for c in lt_classes:
lt_feats = np.vstack((lt_feats, l_feats[l_targets == c]))
lt_targets = np.append(lt_targets, l_targets[l_targets == c])
lv_feats = np.empty((0, l_feats.shape[1]))
lv_targets = np.empty(0)
for c in lv_classes:
lv_feats = np.vstack((lv_feats, l_feats[l_targets == c]))
lv_targets = np.append(lv_targets, l_targets[l_targets == c])
cvi_list = np.zeros(len(cand_k))
acc_list = np.zeros(len(cand_k))
cat_pred_list = np.zeros([len(cand_k), u_num + l_num])
print('estimating K ...')
for i in range(len(cand_k)):
cvi_list[i], cat_pred_i = labeled_val_fun(
np.concatenate((lv_feats, u_feats)), lt_feats, lt_targets,
cand_k[i] + args.num_val_cls)
cat_pred_list[i, :] = cat_pred_i
acc_list[i] = cluster_acc(
lv_targets,
cat_pred_i[len(lt_targets):len(lt_targets) + len(lv_targets)])
best_k = get_best_k(cvi_list[:i + 1], acc_list[:i + 1],
cat_pred_list[:i + 1], l_num)
print('current best K {}'.format(best_k))
kmeans = KMeans(n_clusters=best_k)
u_pred = kmeans.fit_predict(u_feats).astype(np.int32)
acc, nmi, ari = cluster_acc(u_targets, u_pred), nmi_score(
u_targets, u_pred), ari_score(u_targets, u_pred)
print('Final K {}, acc {:.4f}, nmi {:.4f}, ari {:.4f}'.format(
best_k, acc, nmi, ari))
return best_k
