def train(epoch, train_adj, train_fea, idx_train, val_adj=None, val_fea=None):
if val_adj is None:
val_adj = train_adj
val_fea = train_fea
t = time.time()
model.train()
optimizer.zero_grad()
output, embeddings = model.myforward(train_fea,
train_adj,
adj_knn,
layer=1.5)
# special for reddit
if sampler.learning_type == "inductive":
loss_train = F.nll_loss(output, labels[idx_train])
acc_train = accuracy(output, labels[idx_train])
else:
loss_train = F.nll_loss(output[idx_train], labels[idx_train])
acc_train = accuracy(output[idx_train], labels[idx_train])
loss_ssl = args.lambda_ * ssl_agent.make_loss(embeddings)
loss_total = loss_train + loss_ssl
loss_total.backward()
optimizer.step()
train_t = time.time() - t
val_t = time.time()
# We can not apply the fastmode for the reddit dataset.
# if sampler.learning_type == "inductive" or not args.fastmode:
# if args.early_stopping > 0 and sampler.dataset != "reddit":
# loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
# early_stopping(loss_val, model)
if not args.fastmode and args.early_stopping > 0:
# # Evaluate validation set performance separately,
# # deactivates dropout during validation run.
model.eval()
output = model(val_fea, val_adj, adj_knn)
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
acc_val = accuracy(output[idx_val], labels[idx_val]).item()
early_stopping(acc_val, model)
else:
loss_val = 0
acc_val = 0
if args.lradjust:
scheduler.step()
val_t = time.time() - val_t
try:
return (loss_train.item(), acc_train.item(), loss_val, acc_val,
loss_ssl.item(), loss_total.item(), train_t)
except:
return (loss_train.item(), acc_train.item(), loss_val, acc_val,
loss_ssl, loss_total.item(), train_t)
def compute_metric(input, target):
metric = {
'accuracy@1': accuracy(input=input, target=target, topk=1),
'accuracy@5': accuracy(input=input, target=target, topk=5),
}
return metric
def train(epoch, train_adj, train_fea, idx_train, val_adj=None, val_fea=None):
if val_adj is None:
val_adj = train_adj
val_fea = train_fea
t = time.time()
model.train()
optimizer.zero_grad()
output = model(train_fea, train_adj)
# special for reddit
if sampler.learning_type == "inductive":
loss_train = F.nll_loss(output, labels[idx_train])
acc_train = accuracy(output, labels[idx_train])
else:
loss_train = F.nll_loss(output[idx_train], labels[idx_train])
acc_train = accuracy(output[idx_train], labels[idx_train])
loss_train.backward()
optimizer.step()
train_t = time.time() - t
val_t = time.time()
# We can not apply the fastmode for the reddit dataset.
# if sampler.learning_type == "inductive" or not args.fastmode:
grads = [
np.linalg.norm(l.grad.cpu().numpy())
for l in model.midlayer[0].model.weights
]
norms = [
np.linalg.norm(l.detach().cpu().numpy())
for l in model.midlayer[0].model.weights
]
print("Grads:", grads)
print("Norms", norms)
print(np.array(norms) / np.array(grads))
if args.early_stopping > 0 and sampler.dataset != "reddit":
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
early_stopping(loss_val, model)
if not args.fastmode:
# # Evaluate validation set performance separately,
# # deactivates dropout during validation run.
model.eval()
output = model(val_fea, val_adj)
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
acc_val = accuracy(output[idx_val], labels[idx_val]).item()
if sampler.dataset == "reddit":
early_stopping(loss_val, model)
else:
loss_val = 0
acc_val = 0
if args.lradjust:
scheduler.step()
val_t = time.time() - val_t
return (loss_train.item(), acc_train.item(), loss_val, acc_val,
get_lr(optimizer), train_t, val_t, grads, norms)
def train(epoch, train_adj, train_fea, idx_train, val_adj=None, val_fea=None):
if val_adj is None:
val_adj = train_adj
val_fea = train_fea
t = time.time()
model.train()
optimizer.zero_grad()
recovered, mu, logvar, output = model(train_fea, train_adj)
# special for reddit
if sampler.learning_type == "inductive":
#loss_train = F.nll_loss(output, labels[idx_train])
acc_train = accuracy(output, labels[idx_train])
else:
loss_nc = F.nll_loss(output[idx_train], labels[idx_train])
ae_loss = loss_function(preds=recovered,
labels=train_adj,
mu=mu,
logvar=logvar,
n_nodes=train_adj.size(0))
loss_train = loss_nc + 0.2 * ae_loss
acc_train = accuracy(output[idx_train], labels[idx_train])
loss_train.backward()
optimizer.step()
train_t = time.time() - t
val_t = time.time()
# We can not apply the fastmode for the reddit dataset.
# if sampler.learning_type == "inductive" or not args.fastmode:
'''if args.early_stopping > 0 and sampler.dataset != "reddit":
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
early_stopping(loss_val, model)
if not args.fastmode:
# # Evaluate validation set performance separately,
# # deactivates dropout during validation run.
model.eval()
recovered, mu, logvar,output = model(val_fea, val_adj)
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
acc_val = accuracy(output[idx_val], labels[idx_val]).item()
if sampler.dataset == "reddit":
early_stopping(loss_val, model)
else:
loss_val = 0
acc_val = 0'''
model.eval()
recovered, mu, logvar, output = model(val_fea, val_adj)
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
acc_val = accuracy(output[idx_val], labels[idx_val]).item()
early_stopping(acc_val, model)
if args.lradjust:
scheduler.step()
val_t = time.time() - val_t
return (loss_train.item(), acc_train.item(), loss_val, acc_val,
get_lr(optimizer), train_t, val_t)
def train_full(epoch, train_g, val_g, idx_val, labels):
unsupervised_model.eval()
t = time.time()
classifier_model.train()
optimizer.zero_grad()
# get features for training
feats = unsupervised_model(train_g.ndata['features'], train_g)
output = classifier_model(feats)
# special for inductive, learning type must be inductive
if sampler.learning_type=='inductive':
loss_train = F.nll_loss(output, labels[idx_train])
acc_train = accuracy(output, labels[idx_train])
else:
loss_train = F.nll_loss(output[idx_train], labels[idx_train])
acc_train = accuracy(output[idx_train], labels[idx_train])
# loss_train = F.nll_loss(output, labels[idx_train])
# acc_train = accuracy(output, labels[idx_train])
# if sampler.learning_type == "inductive":
# loss_train = F.nll_loss(output, labels[idx_train])
# acc_train = accuracy(output, labels[idx_train])
# else:
# loss_train = F.nll_loss(output[idx_train], labels[idx_train])
# acc_train = accuracy(output[idx_train], labels[idx_train])
loss_train.backward()
optimizer.step()
train_t = time.time() - t
val_t = time.time()
# We can not apply the fastmode for the coauthor_phy dataset.
# if sampler.learning_type == "inductive" or not args.fastmode:
classifier_model.eval()
if sampler.dataset in ['coauthor_phy']:
unsupervised_model.cpu()
classifier_model.cpu()
labels = labels.cpu()
# get features for validation
feats = unsupervised_model(val_g.ndata['features'], val_g)
output = classifier_model(feats)
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
acc_val = accuracy(output[idx_val], labels[idx_val]).item()
early_stopping(loss_val, classifier_model)
if sampler.dataset in ['coauthor_phy']:
unsupervised_model.cuda()
classifier_model.cuda()
labels = labels.cuda()
if args.lradjust:
scheduler.step()
val_t = time.time() - val_t
return (loss_train.item(), acc_train.item(), loss_val, acc_val, get_lr(optimizer), train_t, val_t)
def test_accuracy():
input = torch.tensor([[2, 3, 1], [5, 7, 0], [0, 9, -1]], dtype=torch.float)
target = torch.tensor([1, 0, 2], dtype=torch.long)
assert torch.equal(accuracy(input=input, target=target, topk=1),
torch.tensor([1., 0., 0.]))
assert torch.equal(accuracy(input=input, target=target, topk=2),
torch.tensor([1., 1., 0.]))
def train(epoch, train_adj, train_fea, idx_train, val_adj=None, val_fea=None):
unsupervised_model.eval()
if val_adj is None:
val_adj = train_adj
val_fea = train_fea
t = time.time()
classifier_model.train()
optimizer.zero_grad()
feats = unsupervised_model(train_fea, train_adj)
output = classifier_model(feats)
# special for reddit
if sampler.learning_type == "inductive":
loss_train = F.nll_loss(output, labels[idx_train])
acc_train = accuracy(output, labels[idx_train])
else:
loss_train = F.nll_loss(output[idx_train], labels[idx_train])
acc_train = accuracy(output[idx_train], labels[idx_train])
loss_train.backward()
optimizer.step()
train_t = time.time() - t
val_t = time.time()
# We can not apply the fastmode for the reddit dataset.
# if sampler.learning_type == "inductive" or not args.fastmode:
classifier_model.eval()
feats = unsupervised_model(val_fea, val_adj)
output = classifier_model(feats)
if args.early_stopping > 0 and sampler.dataset not in ['reddit', 'coauthor_phy']:
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
acc_val = accuracy(output[idx_val], labels[idx_val]).item()
early_stopping(loss_val, classifier_model)
if not args.fastmode:
# # Evaluate validation set performance separately,
# # deactivates dropout during validation run.
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
acc_val = accuracy(output[idx_val], labels[idx_val]).item()
if sampler.dataset in ['reddit', 'coauthor_phy']:
early_stopping(loss_val, classifier_model)
else:
loss_val = 0
acc_val = 0
if args.lradjust:
scheduler.step()
val_t = time.time() - val_t
return (loss_train.item(), acc_train.item(), loss_val, acc_val, get_lr(optimizer), train_t, val_t)
def train(epoch, train_adj, train_fea, idx_train, val_adj=None, val_fea=None):
if val_adj is None:
val_adj = train_adj
val_fea = train_fea
t = time.time()
model.train()
optimizer.zero_grad()
#with torch.no_grad():
output = model(train_fea, train_adj)
# special for reddit
if sampler.learning_type == "inductive": #yes!
loss_train = F.nll_loss(output, labels[idx_train])
acc_train = accuracy(output, labels[idx_train])
else: #only reddit yes!
loss_train = F.nll_loss(output[idx_train], labels[idx_train])
acc_train = accuracy(output[idx_train], labels[idx_train])
loss_train.backward()
#torch.nn.utils.clip_grad_norm(model.parameters(), 0.5) #
optimizer.step()
train_t = time.time() - t
val_t = time.time()
# We can not apply the fastmode for the reddit dataset.
# if sampler.learning_type == "inductive" or not args.fastmode:
if args.early_stopping > 0 and sampler.dataset != "reddit":
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
early_stopping(loss_val, model)
if not args.fastmode:
# # Evaluate validation set performance separately,
# # deactivates dropout during validation run.
model.eval()
#with torch.no_grad():
output = model(val_fea, val_adj)
loss_val = F.nll_loss(output[idx_val], labels[idx_val]).item()
acc_val = accuracy(output[idx_val], labels[idx_val]).item()
if sampler.dataset == "reddit":
early_stopping(loss_val, model)
else:
loss_val = 0
acc_val = 0
if args.lradjust:
scheduler.step()
val_t = time.time() - val_t
return (loss_train.item(), acc_train.item(), loss_val, acc_val,
get_lr(optimizer), train_t, val_t)
def train():
global best_epoch, best_acc
if args.start_epoch:
model.load_state_dict(
torch.load(
os.path.join(args.model_path,
'model-%d.pkl' % (args.start_epoch))))
# Training
for epoch in range(args.start_epoch, args.num_epochs):
train_loss = 0
train_acc = 0
scheduler.step()
model.train()
for i, x in enumerate(train_loader):
logit = model(x[0].float())
target = train_label[i]
loss = criterion(logit, target.view(1))
model.zero_grad()
loss.backward()
optimizer.step()
train_loss += loss.item()
train_acc += accuracy(logit, target.view(1))
print('[epoch', epoch + 1, '] Train loss:', train_loss / i,
'Train Acc:', train_acc / i)
if (epoch + 1) % args.val_step == 0:
model.eval()
val_loss = 0
val_acc = 0
with torch.no_grad():
for i, x in enumerate(valid_loader):
logit = model(x[0].float())
target = valid_label[i]
val_loss += criterion(logit, target.view(1)).item()
val_acc += accuracy(logit, target.view(1))
if best_acc <= (val_acc / i):
best_epoch = epoch + 1
best_acc = (val_acc / i)
torch.save(
model.state_dict(),
os.path.join(args.model_path,
'model-%d.pkl' % (best_epoch)))
print('Val loss:', val_loss / i, 'Val Acc:', val_acc / i)
def test(test_adj, test_fea, idx_test, labels):
unsupervised_model.eval()
classifier_model.eval()
if sampler.learning_type=='inductive':
unsupervised_model.cpu()
classifier_model.cpu()
labels = labels.cpu()
# construct g from test adj
if sampler.learning_type=='inductive':
test_edges = test_adj._indices().data.numpy()
else:
test_edges = test_adj._indices().data.cpu().numpy()
test_edges = sp.coo_matrix((np.ones(test_edges.shape[1]),
(test_edges[0], test_edges[1])),
shape=(test_adj.shape[0], test_adj.shape[0]),
dtype=np.float32)
test_g = nx.from_scipy_sparse_matrix(test_edges, create_using=nx.DiGraph())
test_g = DGLGraph(test_g)
feats = unsupervised_model(test_fea, test_g)
output = classifier_model(feats)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.debug:
print("Test set results:",
"loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def infer(test_reader):
"""
do inference function
"""
total_cost = 0.0
total_count = 0
preds, labels = [], []
for data in test_reader():
avg_cost, avg_acc, batch_prediction = exe.run(
inference_program,
feed=feeder.feed(data),
fetch_list=fetch_list,
return_numpy=True)
total_cost += avg_cost * len(data)
total_count += len(data)
preds.append(batch_prediction)
labels.append(np.asarray([x[-1] for x in data], dtype=np.int64))
y_pred = np.concatenate(preds)
y_label = np.concatenate(labels)
metric_res = []
for metric_name in metric_type:
if metric_name == 'accuracy_with_threshold':
metric_res.append(
(metric_name,
metric.accuracy_with_threshold(y_pred,
y_label,
threshold=0.3)))
elif metric_name == 'accuracy':
metric_res.append(
(metric_name, metric.accuracy(y_pred, y_label)))
else:
print("Unknown metric type: ", metric_name)
exit()
return total_cost / (total_count * 1.0), metric_res
def train(args, model, data, label, train_idx, feature_mask, optimizer, epoch):
model.train()
optimizer.zero_grad()
# criterion = nn.CrossEntropyLoss()
rec_vec, output, semantic = model(feature_mask)
label = label.long().view(-1, )
# classification loss
cls_loss = F.nll_loss(output[train_idx], label[train_idx])
args.logger.warning("Classfication loss " + str(cls_loss.item()))
# reconstruction loss
rec_loss = 0.0
for v in range(args.view_num):
sum = torch.sum(torch.pow(torch.sub(rec_vec[v], data[v]), 2.0), 1)
fea = feature_mask[:, v].double()
loss = sum * fea
loss = torch.sum(loss)
args.logger.warning("View " + str(v) + " loss " + str(loss.item()))
rec_loss += loss
# summary loss
if epoch < 100:
loss = rec_loss
else:
loss = cls_loss + rec_loss
args.logger.warning("Total loss " + str(loss.item()))
loss.backward()
optimizer.step()
acc_train = accuracy(output[train_idx], label[train_idx]).item()
args.logger.error("Epoch : " + str(epoch) + ' train accuracy : ' +
str(acc_train))
return acc_train, semantic
def validate(self):
self.pred_labels = self.classify()
acc = accuracy(self.true_labels, self.pred_labels)
ematch = exact_match(self.true_labels, self.pred_labels)
pre_micro, rec_micro, f_micro = f_score_micro(self.true_labels, self.pred_labels)
pre_label, rec_label, f_label, prec_result, recall_result = f_score_by_label(self.true_labels, self.pred_labels, len(self.label_dict))
return acc, ematch, pre_micro, rec_micro, f_micro, pre_label, rec_label, f_label, prec_result, recall_result
def test(X):
zh, score = kmeans(X)
a = accuracy(z, zh)
print "k-means original space:", a
Y = pca_proj(X)
zh, score = kmeans(Y)
a = accuracy(z, zh)
print "k-means/1D PCA:", a
zh = kmeans_random(X, 20)
a = accuracy(z, zh)
print "k-means/random projection 1D:", a
zh = energy_random(X, 20)
a = accuracy(z, zh)
print "energy/random projection 1D", a
def train(epoch, train_adj, train_fea, val_adj=None, val_fea=None):
if val_adj is None:
val_adj = train_adj
val_fea = train_fea
t = time.time()
#adjust lr
if args.lradjust:
#scheduler.step(loss_val)
scheduler.step()
model.train()
optimizer.zero_grad()
output = model(train_fea, train_adj)
#special for reddit
if sampler.learning_type == "inductive":
loss_train = F.nll_loss(output, labels[idx_train])
acc_train = accuracy(output, labels[idx_train])
else:
loss_train = F.nll_loss(output[idx_train], labels[idx_train])
acc_train = accuracy(output[idx_train], labels[idx_train])
loss_train.backward()
optimizer.step()
#We can not apply the fastmode for the reddit dataset.
if sampler.learning_type == "inductive" or not args.fastmode:
# Evaluate validation set performance separately,
# deactivates dropout during validation run.
model.eval()
output = model(val_fea, val_adj)
loss_val = F.nll_loss(output[idx_val], labels[idx_val])
acc_val = accuracy(output[idx_val], labels[idx_val])
if args.earlystopping > 0:
early_stopping(loss_val, model)
if args.debug and epoch % 1 == 0:
print('Epoch: {:04d}'.format(epoch+1),
'loss_train: {:.4f}'.format(loss_train.item()),
'acc_train: {:.4f}'.format(acc_train.item()),
'loss_val: {:.4f}'.format(loss_val.item()),
'acc_val: {:.4f}'.format(acc_val.item()),
'time: {:.4f}s'.format(time.time() - t))
return (loss_train.item(), acc_train.item(), loss_val.item(), acc_val.item())
def test(X):
zh, score = kmeans(X)
a = accuracy(z, zh)
print "k-means original space:", a
Y = pca_proj(X)
zh, score = kmeans(Y)
a = accuracy(z, zh)
print "k-means/1D PCA:", a
zh = kmeans_random(X, 20)
a = accuracy(z, zh)
print "k-means/random projection 1D:", a
zh = energy_random(X, 20)
a = accuracy(z, zh)
print "energy/random projection 1D", a
def test(args, model, data, label, test_idx, feature_mask, epoch=0):
model.eval()
with torch.no_grad():
_, output, semantic = model(feature_mask)
loss_test = F.nll_loss(output[test_idx], label[test_idx])
acc_test = accuracy(output[test_idx], label[test_idx]).item()
args.logger.error("Epoch : " + str(epoch) + ' test accuracy : ' +
str(acc_test))
return acc_test, semantic
def test(test_adj,test_fea):
model.eval()
output = model(test_fea, test_adj)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
if args.debug:
print("Test set results:",
"loss= {:.4f}".format(loss_test.item()),
"accuracy= {:.4f}".format(acc_test.item()))
return (loss_test.item(), acc_test.item())
def test_accuracy_should_examine_each_sentence(self):
reference = [
"the dog walks on his 4 legs in the park",
"bob looked at the stars and saw jupiter"
]
translation = [
"the dog is on his 4 leg in the stadium",
"bob look at a star and saw some wood"
]
self.assertEqual(103 / 180, metric.accuracy(reference, translation))
def trainGraphConvolutionNetwork(hparam):
from model import GraphConvNetwork
from data.dataset import CoraDataSet
device = 'cuda' if hparam.gpu else 'cpu'
db = CoraDataSet(basepath=hparam.dataPath)
X, y, A, idx_train, idx_val, idx_test = db.getTorchTensor(device)
net = GraphConvNetwork(inchannel=db.featureDim,
nhidden=hparam.nHidden,
outchannel=db.numClass,
dropout=hparam.dropout).to(device).train()
from losses import LossFamily
lossfunc = LossFamily[hparam.loss]
import torch.optim as optim
optimizer = optim.Adam(net.parameters(),
lr=hparam.lr,
weight_decay=hparam.weight_decay)
from metric import accuracy
for epoch in range(hparam.epoch):
optimizer.zero_grad()
#forward path
yhat = net(X, A)
_loss = lossfunc(yhat[idx_train], y[idx_train])
# backward
_loss.backward()
optimizer.step()
#统计截断
trainLoss = _loss.item()
valLoss = lossfunc(yhat[idx_val], y[idx_val]).item()
trainAcc = accuracy(yhat[idx_train], y[idx_train])
valAcc = accuracy(yhat[idx_val], y[idx_val])
print('epoch %d,train accuracy %.2f,val accuracy %.2f' %
(trainAcc, valAcc))
torch.save(net.state_dict(), hparam.savepath)
def _forward(data,
model,
loss_fn,
window,
forecast_length,
training=True,
teacher_ratio=1):
outputs = []
label_x, feature_x, label_y, feature_y, _, _ = data
batch_size = label_x.shape[0]
model.init_hidden(batch_size)
# concat the true value of day(t-1) and the features of day(t) to forecast day(t)
inp = torch.cat([
label_x[:, :-1].reshape(batch_size, window - 1, 1), feature_x[:, 1:, :]
],
dim=2)
# no need to iterate the first day
for time_step in range(window - 1):
output = model(inp[:, time_step:time_step + 1, :])
outputs.append(output)
for idx in range(forecast_length):
if idx == 0:
inp = torch.cat([
label_x[:, -1:].reshape([-1, 1, 1]), feature_y[:,
idx:idx + 1, :]
],
dim=2)
else:
if training:
if np.random.random() < teacher_ratio:
inp = torch.cat([
label_y[:, idx - 1].reshape([-1, 1, 1]),
feature_y[:, idx:idx + 1, :]
],
dim=2)
else:
inp = torch.cat([
output.reshape([-1, 1, 1]), feature_y[:,
idx:idx + 1, :]
],
dim=2)
else:
inp = torch.cat(
[output.reshape([-1, 1, 1]), feature_y[:, idx:idx + 1, :]],
dim=2)
output = model(inp)
outputs.append(output)
outputs = torch.stack(outputs, 1)
loss = loss_fn(outputs, torch.cat([label_x[:, 1:], label_y], 1))
avg_acc = accuracy(outputs[:, -forecast_length:], label_y)
return loss, outputs, avg_acc
def test(test_adj, test_fea):
model.eval()
# output = model(test_fea, test_adj)
output, embeddings = model.myforward(test_fea, test_adj, adj_knn)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
print("Test set results:", "loss= {:.4f}".format(loss_test.item()),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def train(epoch, train_adj, train_fea, idx_train, val_adj=None, val_fea=None):
if val_adj is None:
val_adj = train_adj
val_fea = train_fea
t = time.time()
model.train()
optimizer.zero_grad()
output = model(train_fea, train_adj)
#special for reddit
if sampler.learning_type == "inductive":
loss_train = F.nll_loss(output, labels[idx_train])
acc_train = accuracy(output, labels[idx_train])
else:
loss_train = F.nll_loss(output[idx_train], labels[idx_train])
acc_train = accuracy(output[idx_train], labels[idx_train])
loss_train.backward()
optimizer.step()
train_t = time.time() - t
val_t = time.time()
#We can not apply the fastmode for the reddit dataset.
if sampler.learning_type == "inductive" or not args.fastmode:
# # Evaluate validation set performance separately,
# # deactivates dropout during validation run.
model.eval()
output = model(val_fea, val_adj)
loss_val = F.nll_loss(output[idx_val], labels[idx_val])
acc_val = accuracy(output[idx_val], labels[idx_val])
if args.lradjust:
scheduler.step()
if args.early_stopping > 0:
early_stopping(loss_val, model)
val_t = time.time() - val_t
return (loss_train.item(), acc_train.item(), loss_val.item(),
acc_val.item(), get_lr(optimizer), train_t, val_t)
def test(test_adj, test_fea):
model.eval()
recovered, mu, logvar, output = model(test_fea, test_adj)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.debug:
'''print("Test set results:",
"loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))'''
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def test_sampling(model, test_g, val_batch_size):
model.eval()
output = model.inference(test_g, test_g.ndata['features'], val_batch_size,
'cpu')
loss_test = F.nll_loss(output[idx_test.cpu()], labels[idx_test].cpu())
acc_test = accuracy(output[idx_test.cpu()], labels[idx_test].cpu())
auc_test = roc_auc_compute_fn(output[idx_test.cpu()],
labels[idx_test].cpu())
if args.debug:
print("Test set results:", "loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def validation_class(network, dataset, type='validation'):
accuracy = 0.
time_max = dataset.get_size(type='validation') // batch_size
for index in range(0, time_max, 1):
data_x, data_y = dataset.get_minbatch(batch_size, index, type=type)
pred_label = network.predict(data_x)
pred_label = pred_label.astype(np.int)
pred_label = np.equal(pred_label, data_y).astype(np.int)
data_y = np.ones_like(pred_label)
accuracy += metric.accuracy(pred_label, data_y)
accuracy = accuracy / time_max
return accuracy
def test(test_adj, test_fea):
unsupervised_model.eval()
classifier_model.eval()
feats = unsupervised_model(test_fea, test_adj)
output = classifier_model(feats)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.debug:
print("Test set results:",
"loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def main():
data = np.genfromtxt('./winequality-white.csv',
delimiter=';', dtype=float, skip_header=1)
train_feature = data[:3000, :-1]
train_label = data[:3000, -1]
test_feature = data[3000:, :-1]
test_label = data[3000:, -1]
adaboost = Adaboost(train_feature, train_label)
adaboost.train(3)
predict_label = []
for item in test_feature:
predict_label.append(adaboost.predict(item))
print("test accuracy: ", accuracy(predict_label, test_label))
def test_sampling(test_g, val_batch_size):
unsupervised_model.eval()
classifier_model.eval()
feats = unsupervised_model.inference(test_g, test_g.ndata['features'], val_batch_size, 'cpu')
output = classifier_model(feats[idx_test.cpu()].cuda())
loss_test = F.nll_loss(output, labels[idx_test])
acc_test = accuracy(output, labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.debug:
print("Test set results:",
"loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def test_full(model, test_g, idx_test, labels):
model.eval()
if sampler.dataset in ['coauthor_phy']:
model.cpu()
labels = labels.cpu()
output = model(test_g.ndata['features'], test_g)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.debug:
print("Test set results:", "loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def eval(model, eval_dataloader, device):
model.eval()
eval_accuracy = 0.
#eval_map, eval_accuracy, eval_mrr = 0., 0., 0.
nb_eval_steps, nb_eval_examples = 0, 0
preds, labels = [], []
for batch in tqdm(eval_dataloader, desc="Evaluating"):
batch = (t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids, x_input_ids, x_input_mask, x_segment_ids, y_input_ids, y_input_mask, y_segment_ids = batch
with torch.no_grad():
# tmp_eval_loss = model(input_ids, segment_ids, input_mask, label_ids)
logits = model(x_input_ids, x_input_mask, x_segment_ids,
y_input_ids, y_input_mask, y_segment_ids, input_ids,
segment_ids, input_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
preds.extend(np.argmax(logits, 1).tolist())
labels.extend(label_ids.tolist())
tmp_eval_accuracy = accuracy(logits, label_ids)
#tmp_eval_map = mean_average_precision(label_ids, logits[:,1])
#tmp_eval_mrr = mean_reciprocal_rank(label_ids, logits[:, 1])
# eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
#eval_map += tmp_eval_map
#eval_mrr += tmp_eval_mrr
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
# eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
eval_f1 = f1_score(np.array(labels), np.array(preds))
#eval_map = eval_map / nb_eval_examples
#eval_mrr = eval_mrr / nb_eval_examples
result = { # 'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'eval_f1_score': eval_f1
}
#'eval_map': eval_map,
#'eval_mrr': eval_mrr}
pprint(result)
#fname = "normal_density2.pdf" fname = "lognormal_density2.pdf" ############################################################################### t = PrettyTable(['Method', 'Accuracy']) km = KMeans(k, n_init=5) km.fit(Y) zh_kmeans = km.labels_ x1_kmeans = X[np.where(zh_kmeans==0)][:, np.newaxis] x2_kmeans = X[np.where(zh_kmeans==1)][:, np.newaxis] x1_mu_kmeans, x2_mu_kmeans = km.cluster_centers_ x1_mu_kmeans, x2_mu_kmeans = x1_mu_kmeans[0], x2_mu_kmeans[0] x1_var_kmeans, x2_var_kmeans = np.var(x1_kmeans), np.var(x2_kmeans) acc_kmeans = metric.accuracy(z, zh_kmeans) t.add_row(['k-means', acc_kmeans]) gm = GMM(k, n_init=5, init_params="kmeans") gm.fit(Y) zh_gmm = gm.predict(Y) #x1_gmm = X[np.where(zh_gmm==0)][:, np.newaxis] #x2_gmm = X[np.where(zh_gmm==1)][:, np.newaxis] x1_mu_gmm, x2_mu_gmm = gm.means_ x1_mu_gmm, x2_mu_gmm = x1_mu_gmm[0], x2_mu_gmm[0] x1_var_gmm, x2_var_gmm = gm.covariances_ x1_var_gmm, x2_var_gmm = x1_var_gmm[0][0], x2_var_gmm[0][0] acc_gmm = metric.accuracy(z, zh_gmm) t.add_row(['gmm', acc_gmm]) G = eclust.kernel_matrix(Y, lambda x, y: np.linalg.norm(x-y))
G = kernel_matrix(X, rho)
# initialization
mu0, z0 = initialization.kmeanspp(k, X, ret='both')
Z0 = ztoZ(z0)
z1 = initialization.spectral(k, G)
Z1 = ztoZ(z1)
t = BeautifulTable()
t.column_headers = ["Method", "Accuracy", "Objective", "Exec Time"]
start = timer()
zh = energy_clustering_brute(k, G, Z0)
end = timer()
Zh = ztoZ(zh)
t.append_row(["E-clustering brute", metric.accuracy(z, zh),
objective(Zh, G), end-start])
start = timer()
zh = energy_hartigan(k, G, Z0)
end = timer()
Zh = ztoZ(zh)
t.append_row(["E-H-clustering++", metric.accuracy(z, zh),
objective(Zh, G), end-start])
t.append_row(['Spectral Clustering:', metric.accuracy(z, z1),
objective(Z1,G), '-'])
start = timer()
zh = energy_hartigan(k, G, Z1)
end = timer()
import data
from metric import accuracy
import eclust
import initialization
num_experiments = 10
table = np.zeros((num_experiments, 5))
for i in range(num_experiments):
X, z = data.univariate_lognormal([0, -1.5], [0.3, 1.5], [100, 100])
#X, z = data.univariate_normal([0, 5], [1, 22], [15, 15])
Y = np.array([[x] for x in X])
k = 2
# 1D energy clustering
zh, cost = two_clusters1D(X)
table[i,0] = accuracy(z, zh)
# initialization
z0 = initialization.kmeanspp(k, Y, ret='labels')
Z0 = eclust.ztoZ(z0)
rho = lambda x, y: np.linalg.norm(x-y)
G = eclust.kernel_matrix(Y, rho)
z1 = initialization.spectral(k, G)
Z1 = eclust.ztoZ(z1)
# Hartigan's method
zh = eclust.energy_hartigan(k, G, Z0)
table[i,1] = accuracy(z, zh)
zh = eclust.energy_hartigan(k, G, Z1)
table[i,2] = accuracy(z, zh)
k = 2 n = 2000 n1, n2 = np.random.multinomial(n, [0.5, 0.5]) m1 = 0 s1 = 1.5 m2 = 1.5 s2 = 0.3 #X, z = data.univariate_normal([m1, m2], [s1, s2], [n1, n2]) X, z = data.univariate_lognormal([m1, m2], [s1, s2], [n1, n2]) Y = np.array([[x] for x in X]) ### clustering t = PrettyTable(['Method', 'Accuracy']) G = eclust.kernel_matrix(Y, lambda x, y: np.linalg.norm(x-y)) zh_kmeans = wrapper.kmeans(k, Y) t.add_row(['k-means', metric.accuracy(z, zh_kmeans)]) zh_gmm = wrapper.gmm(k, Y) t.add_row(['gmm', metric.accuracy(z, zh_gmm)]) zh_kgroups = wrapper.kernel_kgroups(k, Y, G) t.add_row(['kernel k-groups', metric.accuracy(z, zh_kgroups)]) print t ### estimated classes x1_true = X[np.where(z==0)] x2_true = X[np.where(z==1)] x1_kmeans = X[np.where(zh_kmeans==0)] x2_kmeans = X[np.where(zh_kmeans==1)] x1_gmm = X[np.where(zh_gmm==0)] x2_gmm = X[np.where(zh_gmm==1)]
n = 400
d = 10
n1, n2 = np.random.multinomial(n, [1/2, 1/2])
m1 = np.zeros(d)
m2 = 0.7*np.ones(d)
s1 = s2 = np.eye(d)
X, z = data.multivariate_normal([m1, m2], [s1, s2], [n1, n2])
G = eclust.kernel_matrix(X, lambda x, y: np.linalg.norm(x-y))
W = np.eye(n)
k = 2
t = PrettyTable(["Method", "Accuracy"])
zh = kernel_kmeans(k, X, G, W, run_times=5, ini="k-means++")
a = metric.accuracy(z, zh)
t.add_row(["Kernel k-means", a])
zh = kernel_kgroups(k, X, G, W, run_times=5, ini="k-means++")
a = metric.accuracy(z, zh)
t.add_row(["Kernel k-groups", a])
zh = spectral(k, X, G, W, run_times=5)
a = metric.accuracy(z, zh)
t.add_row(["Spectral", a])
zh = kmeans(k, X, run_times=5)
a = metric.accuracy(z, zh)
t.add_row(["k-means", a])
zh = gmm(k, X, run_times=5)
import data
import metric
#np.random.seed(12)
D = 10
m1 = np.zeros(D)
s1 = np.eye(D)
m2 = np.ones(D)
s2 = 2*np.eye(D)
X, z = data.multivariate_normal([m1, m2], [s1, s2], [100, 100])
k = 2
# scikit-learn library has a better procedure to estimate the covariance
# matrix.
g = GMM(k)
zh = g.fit_predict(X)
print "GMM class:", metric.accuracy(z, zh)
zh = gmm(k, X)
print "GMM func:", metric.accuracy(z, zh)
sg = sk_GMM(k)
sg.fit(X)
zh = sg.predict(X)
print "GMM sklearn:", metric.accuracy(z, zh)
rho = lambda x,y: np.power(np.linalg.norm(x-y), 1)
G = kernel_matrix(X, rho)
# initialization
z0, mu0 = init.kmeans_plus2(k, X)
Z0 = ztoZ(z0)
z1 = init.spectral(k, G, W)
Z1 = ztoZ(z1)
t = PrettyTable(["Method", "Accuracy", "Objective", "Exec Time"])
start = timer()
zh = kernel_kgroups(k, G, Z0, W)
end = timer()
Zh = ztoZ(zh)
t.add_row(["kernel k-groups (k-means++)", metric.accuracy(z, zh),
objective(Zh, G, W), end-start])
start = timer()
zh = kernel_kgroups(k, G, Z1, W)
end = timer()
Zh = ztoZ(zh)
t.add_row(["kernel k-groups (spectral)", metric.accuracy(z, zh),
objective(Zh, G, W), end-start])
start = timer()
zh = kernel_kmeans(k, G, Z0, W)
end = timer()
Zh = ztoZ(zh)
t.add_row(["kernel k-means (k-means++)", metric.accuracy(z, zh),
objective(Zh, G, W), end-start])
dist = np.zeros((n_samples, self.n_clusters))
self._compute_dist(K, dist, self.within_distances_,
update_within=False)
return dist.argmin(axis=1)
###############################################################################
if __name__ == '__main__':
import energy
import data
from metric import accuracy
from sklearn.cluster import KMeans
X, z = data.multivariate_normal(
[[0,0], [2,0]],
[np.eye(2), np.eye(2)],
[100, 100]
)
kernel = energy.energy_kernel
km = KernelEnergy(n_clusters=2, max_iter=100, verbose=1,
kernel_params={'alpha':.8})
zh = km.fit_predict(X)
print accuracy(z, zh)
km = KMeans(n_clusters=2)
zh = km.fit_predict(X)
print accuracy(z, zh)
G = eclust.kernel_matrix(data, rho)
#G = eclust.kernel_matrix(data, rho_gauss)
#G = eclust.kernel_matrix(data, rho_exp)
k = 3
r = []
r.append(wrapper.kmeans(k, data, run_times=5))
r.append(wrapper.gmm(k, data, run_times=5))
r.append(wrapper.spectral_clustering(k, data, G, run_times=5))
r.append(wrapper.spectral(k, data, G, run_times=5))
r.append(wrapper.kernel_kmeans(k, data, G, run_times=5, ini='random'))
#r.append(wrapper.kernel_kmeans(k, data, G, run_times=5, ini='k-means++'))
#r.append(wrapper.kernel_kmeans(k, data, G, run_times=5, ini='spectral'))
r.append(wrapper.kernel_kgroups(k,data,G,run_times=5, ini='random'))
#r.append(wrapper.kernel_kgroups(k,data,G,run_times=5, ini='k-means++'))
#r.append(wrapper.kernel_kgroups(k,data,G,run_times=5, ini='spectral'))
t = PrettyTable(['Algorithm', 'Accuracy', 'A-Rand'])
algos = ['kmeans', 'GMM', 'spectral clustering', 'spectral',
'kernel k-means', 'kernel k-groups']
for algo, zh in zip(algos, r):
t.add_row([algo,
metric.accuracy(z, zh),
sklearn.metrics.adjusted_rand_score(z, zh)
])
print t
costs.append(cost)
costs = np.array(costs)
min_index = costs.argmin()
min_cost = costs[min_index]
return min_cost, min_index
###############################################################################
if __name__ == '__main__':
import data
from metric import accuracy
m1 = np.array([0,0])
s1 = np.array([[1,0],[0,1]])
n1 = 100
m2 = np.array([3,0])
s2 = np.array([[1,0],[0,10]])
n2 = 100
X, true_labels = data.multivariate_normal([m1,m2], [s1,s2], [n1,n2])
ec = EClust(n_clusters=2, max_iter=10, init='kmeans++')
labels = ec.fit_predict(X)
print accuracy(labels, true_labels)
km = KMeans(2)
labels2 = km.fit_predict(X)
print accuracy(labels2, true_labels)
