def f1_avg(y_pred, y_true):
'''
mission 1&2
:param y_pred:
:param y_true:
:return:
'''
f1_micro = f1_score(y_pred=y_pred,
y_true=y_true,
pos_label=1,
average='micro')
f1_macro = f1_score(y_pred=y_pred,
y_true=y_true,
pos_label=1,
average='macro')
return (f1_micro + f1_macro) / 2
def get_f1(self):
y_pred_top = self.predict2top()
y_pred_half = self.predict2half()
y_pred_both = self.predict2both()
if self.type == "top":
f1_micro = f1_score(y_pred=y_pred_top, y_true=self.y_true, pos_label=1, average='micro')
f1_macro = f1_score(y_pred=y_pred_top, y_true=self.y_true, pos_label=1, average='macro')
elif self.type == "half":
f1_micro = f1_score(y_pred=y_pred_half, y_true=self.y_true, pos_label=1, average='micro')
f1_macro = f1_score(y_pred=y_pred_half, y_true=self.y_true, pos_label=1, average='macro')
elif self.type == "both":
f1_micro = f1_score(y_pred=y_pred_both, y_true=self.y_true, pos_label=1, average='micro')
f1_macro = f1_score(y_pred=y_pred_both, y_true=self.y_true, pos_label=1, average='macro')
return (f1_micro, f1_macro)
def eval(tagger, data_loader, criterion):
tagger.eval()
preds_collection = []
labels_collection = []
total_loss = 0
total_samples = 0
for i, data in enumerate(data_loader):
sentences, targets, labels = data
sentences, targets, labels = sentences.cuda(), targets.cuda(
), labels.cuda()
sentences = sentence_clip(sentences)
targets = targets[:, 0:sentences.size(1)].contiguous()
labels = labels[:, 0:sentences.size(1)].contiguous()
logits = tagger(sentences, targets)
logits = logits.view(-1, 2)
preds = logits.argmax(dim=-1)
labels = labels.view(-1)
loss = criterion(logits, labels).item()
batch_size = (labels != -1).long().sum().item()
total_loss += loss * batch_size
total_samples += batch_size
preds_collection.append(preds)
labels_collection.append(labels)
preds_collection = torch.cat(preds_collection, dim=0)
labels_collection = torch.cat(labels_collection, dim=0)
mask = labels_collection != -1
preds_collection = preds_collection.masked_select(mask).cpu().numpy()
labels_collection = labels_collection.masked_select(mask).cpu().numpy()
precision = precision_score(y_true=labels_collection,
y_pred=preds_collection,
labels=[0, 1],
average='macro')
recall = recall_score(y_true=labels_collection,
y_pred=preds_collection,
labels=[0, 1],
average='macro')
f1 = f1_score(y_true=labels_collection,
y_pred=preds_collection,
labels=[0, 1],
average='macro')
loss = total_loss / total_samples
return loss, precision, recall, f1
def getresult(y_pred,test_label):
f1_micro = f1_score(y_pred=y_pred, y_true=test_label, pos_label=1, average='micro')
f1_macro = f1_score(y_pred=y_pred, y_true=test_label, pos_label=1, average='macro')
return (f1_micro,f1_macro)
def score(self, y_pred, y):
return f1_score(y.numpy(), y_pred.numpy(), average='micro')
loss = criterion(logits, labels)
loss.backward()
optimizer.step()
batch_size = (labels != -1).long().sum().item()
total_loss += batch_size * loss.item()
total_samples += batch_size
preds_collection.append(preds)
labels_collection.append(labels)
# if i % 100 == 0:
# print('[epoch %d] [step %d] [loss %.4f]' % (epoch, i, loss.item()))
preds_collection = torch.cat(preds_collection, dim=0)
labels_collection = torch.cat(labels_collection, dim=0)
mask = labels_collection != -1
preds_collection = preds_collection.masked_select(mask).cpu().numpy()
labels_collection = labels_collection.masked_select(mask).cpu().numpy()
train_precision = precision_score(y_true=labels_collection, y_pred=preds_collection, labels=[0, 1], average='macro')
train_recall = recall_score(y_true=labels_collection, y_pred=preds_collection, labels=[0, 1], average='macro')
train_f1 = f1_score(y_true=labels_collection, y_pred=preds_collection, labels=[0, 1], average='macro')
train_loss = total_loss / total_samples
print('epoch: %d\ttrain\tloss: %.4f\tprecision: %.4f\trecall: %.4f\tf1: %.4f' % (epoch, train_loss, train_precision, train_recall, train_f1))
val_loss, val_precision, val_recall, val_f1 = eval(tagger, val_loader, criterion)
print('epoch: %d\tval\tloss: %.4f\tprecision: %.4f\trecall: %.4f\tf1: %.4f\n' % (epoch, val_loss, val_precision, val_recall, val_f1))
visualizer.add(epoch, train_loss, train_precision, train_recall, train_f1, val_loss, val_precision, val_recall, val_f1)
if val_f1 > max_val_f1_score:
max_val_f1_score = val_f1
torch.save(tagger, save_path)
print('save model')
print('max val f1_score: %.4f' % max_val_f1_score)
visualizer.plot()