def train(dataset, model, args, same_feat=True, val_dataset=None, test_dataset=None, writer=None,
mask_nodes = True):
writer_batch_idx = [0, 3, 6, 9]
optimizer = torch.optim.Adam(filter(lambda p : p.requires_grad, model.parameters()), lr=0.001)
iter = 0
best_val_result = {
'epoch': 0,
'loss': 0,
'acc': 0}
test_result = {
'epoch': 0,
'loss': 0,
'acc': 0}
train_accs = []
train_epochs = []
best_val_accs = []
best_val_epochs = []
test_accs = []
test_epochs = []
val_accs = []
for epoch in range(args.num_epochs):
total_time = 0
avg_loss = 0.0
model.train()
print('Epoch: ', epoch)
for batch_idx, data in enumerate(dataset):
begin_time = time.time()
model.zero_grad()
adj = Variable(data['adj'].float(), requires_grad=False).cuda()
h0 = Variable(data['feats'].float(), requires_grad=False).cuda()
label = Variable(data['label'].long()).cuda()
batch_num_nodes = data['num_nodes'].int().numpy() if mask_nodes else None
assign_input = Variable(data['assign_feats'].float(), requires_grad=False).cuda()
ypred = model(h0, adj, batch_num_nodes, assign_x=assign_input)
if not args.method == 'soft-assign' or not args.linkpred:
loss = model.loss(ypred, label)
else:
loss = model.loss(ypred, label, adj, batch_num_nodes)
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.clip)
optimizer.step()
iter += 1
avg_loss += loss
#if iter % 20 == 0:
# print('Iter: ', iter, ', loss: ', loss.data[0])
elapsed = time.time() - begin_time
total_time += elapsed
# log once per XX epochs
if epoch % 10 == 0 and batch_idx == len(dataset) // 2 and args.method == 'soft-assign' and writer is not None:
log_assignment(model.assign_tensor, writer, epoch, writer_batch_idx)
if args.log_graph:
log_graph(adj, batch_num_nodes, writer, epoch, writer_batch_idx, model.assign_tensor)
avg_loss /= batch_idx + 1
if writer is not None:
writer.add_scalar('loss/avg_loss', avg_loss, epoch)
if args.linkpred:
writer.add_scalar('loss/linkpred_loss', model.link_loss, epoch)
print('Avg loss: ', avg_loss, '; epoch time: ', total_time)
result = evaluate(dataset, model, args, name='Train', max_num_examples=100)
train_accs.append(result['acc'])
train_epochs.append(epoch)
if val_dataset is not None:
print("validation>>>>>>")
val_result = evaluate(val_dataset, model, args, name='Validation')
val_accs.append(val_result['acc'])
if val_result['acc'] > best_val_result['acc'] - 1e-7:
print("saving best results>>>>>")
best_val_result['acc'] = val_result['acc']
best_val_result['epoch'] = epoch
best_val_result['loss'] = avg_loss
if test_dataset is not None:
test_result = evaluate(test_dataset, model, args, name='Test')
test_result['epoch'] = epoch
if writer is not None:
writer.add_scalar('acc/train_acc', result['acc'], epoch)
writer.add_scalar('acc/val_acc', val_result['acc'], epoch)
writer.add_scalar('loss/best_val_loss', best_val_result['loss'], epoch)
if test_dataset is not None:
writer.add_scalar('acc/test_acc', test_result['acc'], epoch)
print('Best val result: ', best_val_result)
best_val_epochs.append(best_val_result['epoch'])
best_val_accs.append(best_val_result['acc'])
if test_dataset is not None:
print('Test result: ', test_result)
test_epochs.append(test_result['epoch'])
test_accs.append(test_result['acc'])
matplotlib.style.use('seaborn')
plt.switch_backend('agg')
plt.figure()
plt.plot(train_epochs, util.exp_moving_avg(train_accs, 0.85), '-', lw=1)
if test_dataset is not None:
plt.plot(best_val_epochs, best_val_accs, 'bo', test_epochs, test_accs, 'go')
plt.legend(['train', 'val', 'test'])
else:
plt.plot(best_val_epochs, best_val_accs, 'bo')
plt.legend(['train', 'val'])
plt.savefig(gen_train_plt_name(args), dpi=600)
plt.close()
matplotlib.style.use('default')
return model, val_accs
def train(dataset,
model,
user_feature,
adjacency,
args,
same_feat=True,
val_dataset=None,
test_dataset=None,
writer=None,
mask_nodes=True):
writer_batch_idx = [0, 3, 6, 9]
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=0.001)
iter = 0
best_val_result = {'epoch': 0, 'loss': 0, 'acc': 0}
test_result = {'epoch': 0, 'loss': 0, 'acc': 0}
train_accs = []
train_epochs = []
best_val_accs = []
best_val_epochs = []
test_accs = []
test_epochs = []
val_accs = []
adj = Variable(torch.Tensor(adjacency), requires_grad=False)
h0 = Variable(torch.Tensor(user_feature), requires_grad=False)
assign_input = h0
for epoch in range(args.num_epochs):
begin_time = time.time()
avg_loss = 0.0
model.train()
print('Epoch: ', epoch)
embed_all = []
assign_all = []
for batch_idx, data in enumerate(dataset):
model.zero_grad()
train_data = (data['uid'], data['iid'])
batch_num_nodes = None
# batch_num_nodes = data['num_nodes'].int().numpy() if mask_nodes else None
# forward function invoking
# logits = model(h0, adj, batch_num_nodes, assign_x=assign_input)
# assign_x = h0
embed_all, assign_all = model(h0, adj, batch_num_nodes)
if not args.method == 'soft-assign' or not args.linkpred:
loss = model.pred_loss(
model.pred_item_based_logits(embed_all, assign_all,
train_data, args))
else:
loss = model.pred_loss(
model.pred_item_based_logits(embed_all, assign_all,
train_data, args))
loss.backward()
nn.utils.clip_grad_norm(model.parameters(), args.clip)
optimizer.step()
iter += 1
avg_loss += loss
print('Avg loss: ', avg_loss / (batch_idx + 1))
'''
if epoch % 10 == 0 and batch_idx == len(dataset) // 2 and args.method == 'soft-assign' and writer is not None:
log_assignment(model.assign_tensor, writer, epoch, writer_batch_idx)
log_graph(adj, batch_num_nodes, writer, epoch, writer_batch_idx, model.assign_tensor)
'''
# save
item_embeds = model.get_item_embeds()
save_numpy(embed_all, assign_all, item_embeds, epoch, args)
avg_loss /= batch_idx + 1
elapsed = time.time() - begin_time
if writer is not None:
writer.add_scalar('loss/avg_loss', avg_loss, epoch)
if args.linkpred:
writer.add_scalar('loss/linkpred_loss', model.link_loss, epoch)
print('Avg loss: ', avg_loss, '; epoch time: ', elapsed)
if val_dataset is not None:
val_result = evaluate(val_dataset, model, args, name='Validation')
val_accs.append(val_result['acc'])
if val_result['acc'] > best_val_result['acc'] - 1e-7:
best_val_result['acc'] = val_result['acc']
best_val_result['epoch'] = epoch
best_val_result['loss'] = avg_loss
if test_dataset is not None:
test_result = evaluate(test_dataset, model, args, name='Test')
test_result['epoch'] = epoch
print('Best val result: ', best_val_result)
best_val_epochs.append(best_val_result['epoch'])
best_val_accs.append(best_val_result['acc'])
if test_dataset is not None:
print('Test result: ', test_result)
test_epochs.append(test_result['epoch'])
test_accs.append(test_result['acc'])
matplotlib.style.use('seaborn')
plt.switch_backend('agg')
plt.figure()
plt.plot(train_epochs, util.exp_moving_avg(train_accs, 0.85), '-', lw=1)
if test_dataset is not None:
plt.plot(best_val_epochs, best_val_accs, 'bo', test_epochs, test_accs,
'go')
plt.legend(['train', 'val', 'test'])
else:
plt.plot(best_val_epochs, best_val_accs, 'bo')
plt.legend(['train', 'val'])
plt.savefig(gen_train_plt_name(args), dpi=600)
plt.close()
matplotlib.style.use('default')
return model, val_accs
