def train(self):
self.model.train()
best_acc = 0.0
save_model_prefix = os.path.join(self.model_path,
self.config.model_prefix)
for epoch in range(self.num_epochs):
self.logger.info("Epoch %d/%d" % (epoch + 1, self.num_epochs))
start_time = time.time()
for batch in self.train_data_loader:
output = self.model(MyDataset.to(batch, self.config.device))
self.model.zero_grad()
loss = self._calc_loss(output, batch)
loss.backward()
self.optimizer.step()
self.writer.add_scalar("scalar/loss", loss.cpu().item(), epoch)
time_diff = time.time() - start_time
self.logger.info("epoch %d time consumed: %dm%ds." %
(epoch + 1, time_diff // 60, time_diff % 60))
# evaluate model
cur_acc = self.eval_dev(self.dev_data_loader)
self.model.train()
self.logger.info("Current accuracy: %.3f" % cur_acc)
self.writer.add_scalar("scalar/accuracy", cur_acc)
if cur_acc > best_acc: # and epoch > 10:
save_filename = save_model_prefix + str(cur_acc)
torch.save(self.model.state_dict(), save_filename)
best_acc = cur_acc
def eval_dev(self, dev_data_loader):
self.model.eval()
correct_count = 0
total_count = 0
for batch in dev_data_loader:
output = self.model(MyDataset.to(batch, self.config.device))
pred = torch.argmax(output, 1)
correct_count += (pred.cpu().detach().numpy() ==
batch['answer_index'].numpy()).sum()
total_count += len(batch['query_length'])
return float(correct_count) / total_count