def test_model(loader, model):
"""
Help function that tests the model's performance on a dataset
@param: loader - data loader for the dataset to test against
"""
correct = 0
total = 0
model.eval()
if args.model != 'RNN':
for data, lengths, labels in loader:
data_batch, length_batch, label_batch = data.to(
device), lengths.to(device), labels.to(device)
#print(data_batch,label_batch,length_batch)
outputs = F.softmax(model(data_batch, length_batch), dim=1)
predicted = outputs.max(1, keepdim=True)[1]
total += labels.size(0)
correct += predicted.eq(
labels.view_as(predicted).to(device)).sum().item()
return (100 * correct / total)
else:
for data, lengths, unsort_idx, labels in loader:
data_batch, length_batch, unsort_batch, label_batch = data.to(
device), lengths.to(device), unsort_idx.to(device), labels.to(
device)
#print(data_batch,label_batch,length_batch)
outputs = F.softmax(model(data_batch, length_batch, unsort_batch),
dim=1)
predicted = outputs.max(1, keepdim=True)[1]
total += labels.size(0)
correct += predicted.eq(
labels.view_as(predicted).to(device)).sum().item()
print("testing complete")
return (100 * correct / total)
def test(model, test_loader, criterion, device):
model.eval()
with torch.no_grad():
print("======================================================")
print("TESTING")
print("======================================================")
total_loss = 0
total_correct = 0
total_data = 0
loss = 0
for data, labels in test_loader:
model = model.to(device)
data = data.to(device)
labels = labels.to(device)
output = model(data)
loss = criterion(output, labels)
total_data += labels.size(0)
total_loss += loss.item()
_, p = torch.max(output.data, dim=1)
total_correct += (p == labels).sum().item()
print("Testing: Loss: [{:.2f}] Accuracy [{:.2f}]".format(
total_loss / len(test_loader), total_correct * 100 / total_data))
def train(net, train_iter, valid_iter, device, num_epochs):
train_loss, valid_loss = [], []
train_batch_num = 0
best_valid_acc = 0.
for epoch in range(num_epochs):
total_loss = 0.
correct = 0
sample_num = 0
net.train()
for context in train_iter:
train_batch_num += 1
data = context.Text
data = data.to(device).long()
target = context.Label
target = target.to(device).long()
masks = (data != TEXT.vocab.stoi["<pad>"])
output = net(data, masks)
optimizer.zero_grad()
loss = net.loss(output, masks, target)
loss.backward()
optimizer.step()
total_loss += loss.item()
v_loss,precision,recall,f1 = evaluate_accuracy(valid_iter, net)
valid_loss.append(v_loss)
loss = total_loss / train_batch_num
train_loss.append(loss)
if f1 > best_valid_acc:
best_valid_acc = f1
torch.save(net.state_dict(), "./best1.pth")
print('epoch %d, train loss %.4f, valid loss %.4f, valid precision %.4f, valid recall %.4f,valid f1 %.4f'
% (epoch + 1, loss, v_loss,precision,recall,f1))
return train_loss, valid_loss
def train_model(net, train_iter, test_iter, epoch, lr, batch_size):
print("begin training")
optimizer = optim.Adam(net.parameters(), lr=lr)
criterion = nn.CrossEntropyLoss()
best_acc = 0
for i in range(epoch): # 多批次循环
net.to(device)
net.train() # 必备，将模型设置为训练模式
for batch_idx, batch in enumerate(train_iter):
# 注意target=batch.label - 1，因为数据集中的label是1，2，3，4，但是pytorch的label默认是从0开始，所以这里需要减1
data, target = batch.text, batch.label - 1
data, target = data.to(device), target.to(device)
optimizer.zero_grad() # 清除所有优化的梯度
output = net(data) # 传入数据并前向传播获取输出
loss = criterion(output, target)
loss.backward()
optimizer.step()
# 打印状态信息
logging.info("train epoch=" + str(i) + ",batch_id=" +
str(batch_idx) + ",loss=" +
str(loss.item() / batch_size))
acc = model_test(net, test_iter, batch_size)
if acc > best_acc:
torch.save({'state_dict': net.cpu().state_dict()},
"saved_model/ag_fasttext_model.pth.tar")
best_acc = acc
print('Finished Training')
def model_test(net, test_iter):
net.eval() # 必备，将模型设置为训练模式
correct = torch.zeros(CLASS_NUM)
total = torch.zeros(CLASS_NUM)
with torch.no_grad():
for i, batch in enumerate(test_iter):
# 注意target=batch.label - 1，因为数据集中的label是1，2，3，4，但是pytorch的label默认是从0开始，所以这里需要减1
data, label = batch.text, batch.label
data, label = data.to(device), label.to(device)
logging.info("test batch_id=" + str(i))
data = net.embed(data)
outputs = net(data)
# torch.max()[0]表示最大值的值，troch.max()[1]表示回最大值的每个索引
_, predicted = torch.max(outputs.data,
1) # 每个output是一行n列的数据，取一行中最大的值
predicted += 1
for i in range(1, 5):
filter = (torch.ones(label.size(0)) * i).to(device)
filtered_label = ((label == filter).int()).to(device)
total[i - 1] += filtered_label.sum().item()
correct[i - 1] += ((filtered_label *
i) == predicted).int().sum().item()
totoal_accuracy = round(correct.sum().item() / total.sum().item(), 3)
print('Accuracy of the network on test set:()', totoal_accuracy)
accuray = (correct.float() / total.float()).numpy()
accuray = np.round(accuray, 3)
accuray = dict(zip(ag_news_label.values(), accuray.tolist()))
print(accuray)
return totoal_accuracy
def batchify(data, bsz):
data = TEXT.numericalize([data.examples[0].text])
# Divide the dataset into bsz parts.
nbatch = data.size(0) // bsz
# Trim off any extra elements that wouldn't cleanly fit (remainders).
data = data.narrow(0, 0, nbatch * bsz)
# Evenly divide the data across the bsz batches.
data = data.view(bsz, -1).t().contiguous()
return data.to(device)
def train(language_model, data_source, optimizer, criterion):
epoch_loss = 0
language_model.train()
for i in tqdm(range(0, data_source.size(0) - 1, args.bptt)):
optimizer.zero_grad()
data, targets = get_batch(data_source, i, args.bptt)
data = data.to(device)
#data = [bptt, batch size]
#targets = [bptt, batch size]
if args.model == 'transformer':
mask = generate_square_subsequent_mask(data.shape[0])
mask = mask.to(device)
predictions = language_model(data, mask)
else:
predictions = language_model(data)
#predictions = [batch size, bptt, vocab size]
predictions = predictions.permute(1, 0, 2)
#predictions = [bptt, batch size, vocab size]
predictions = predictions.contiguous().view(-1, vocab_size)
#predictions = [bptt * batch size, vocab size]
targets = targets.view(-1)
#targets = [bptt * batch size]
targets = targets.to(device)
loss = criterion(predictions, targets)
loss.backward()
torch.nn.utils.clip_grad_norm_(language_model.parameters(),
args.grad_clip)
optimizer.step()
epoch_loss += loss.item()
return epoch_loss / len(data_source)
def evaluate(language_model, data_source, criterion):
epoch_loss = 0
language_model.eval()
for i in tqdm(range(0, data_source.size(0) - 1, args.bptt)):
optimizer.zero_grad()
data, targets = get_batch(data_source, i, args.bptt)
data = data.to(device)
#data = [bptt, batch size]
#targets = [bptt, batch size]
with torch.no_grad():
if args.model == 'transformer':
mask = generate_square_subsequent_mask(data.shape[0])
mask = mask.to(device)
predictions = language_model(data, mask)
else:
predictions = language_model(data)
#predictions = [batch size, bptt, vocab size]
predictions = predictions.permute(1, 0, 2)
#predictions = [bptt, batch size, vocab size]
predictions = predictions.contiguous().view(-1, vocab_size)
#predictions = [bptt * batch size, vocab size]
targets = targets.view(-1)
#targets = [bptt * batch size]
targets = targets.to(device)
loss = criterion(predictions, targets)
epoch_loss += loss.item()
return epoch_loss / len(data_source)
def train(model, train_loader, num_epochs, opt, criterion, device):
train_loss_list = []
iteration_list = []
train_accuracy_list = []
print("Beginning to Train")
for epoch in range(num_epochs):
total_loss = 0
total_correct = 0
total_data = 0
loss = 0
model.train()
for i, (data, labels) in enumerate(train_loader):
model = model.to(device)
data = data.to(device)
labels = labels.to(device)
output = model(images)
loss = criterion(output, labels)
opt.zero_grad()
loss.backward()
opt.step()
total_data += labels.size(0)
total_loss += loss.item()
_, p = torch.max(output.data, dim=1)
total_correct += (p == labels).sum().item()
print("Training: epoch: [{}/{}] Loss: [{:.2f}] Accuracy [{:.2f}] ".
format(epoch + 1, num_epochs, total_loss / len(train_loader),
total_correct * 100 / total_data))
train_loss_list.append(total_loss / len(train_loader))
iteration_list.append(epoch)
train_accuracy_list.append(total_correct * 100 / total_data)
history = {
'train_loss': train_loss_list,
'train_acc': train_accuracy_list,
}
return history
def train(net, train_iter, valid_iter, device, num_epochs):
train_loss, valid_loss = [], []
train_batch_num = 0
best_valid_acc = 0.
for epoch in range(num_epochs):
total_loss = 0.
correct = 0
sample_num = 0
net.train()
for context in train_iter:
train_batch_num += 1
data = context.Text
data = data.to(device).long()
target = context.Label
target = target.to(device).long()
output = net(data)
optimizer.zero_grad()
loss = criterion(output, target)
loss.backward()
optimizer.step()
total_loss += loss.item()
prediction = torch.argmax(output, 1)
#print(prediction)
correct += (prediction == target).sum().item()
sample_num += len(prediction)
valid_acc, v_loss = evaluate_accuracy(valid_iter, net)
valid_loss.append(v_loss)
loss = total_loss / train_batch_num
train_loss.append(loss)
acc = correct / sample_num
if valid_acc > best_valid_acc:
best_valid_acc = valid_acc
torch.save(net.state_dict(), "./best.pth")
print(
'epoch %d, train loss %.4f, train acc %.3f, valid loss %.4f, valid acc %.3f'
% (epoch + 1, loss, acc, v_loss, valid_acc))
return train_loss, valid_loss
def train_eval(model, train_loader, val_loader, num_epochs, opt, criterion,
device):
train_loss_list = []
iteration_list = []
train_accuracy_list = []
val_loss_list = []
val_accuracy_list = []
print("Beginning to Train")
for epoch in range(num_epochs):
total_loss = 0
total_correct = 0
total_data = 0
loss = 0
model.train()
for i, (data, labels) in enumerate(train_loader):
model = model.to(device)
data = data.to(device)
labels = labels.to(device)
output = model(data)
loss = criterion(output, labels)
# compute gradients, do parameter update and compute loss
opt.zero_grad()
loss.backward()
opt.step()
total_data += labels.size(0)
total_loss += loss.item()
_, p = torch.max(output.data, dim=1)
total_correct += (p == labels).sum().item()
model.eval()
val_total_loss = 0
val_total_correct = 0
val_total_data = 0
val_loss = 0
for data, labels in val_loader:
model = model.to(device)
data = data.to(device)
labels = labels.to(device)
output = model(data)
val_loss = criterion(output, labels)
val_total_data += labels.size(0)
val_total_loss += val_loss.item()
_, p = torch.max(output.data, dim=1)
val_total_correct += (p == labels).sum().item()
print(
"Training: epoch: [{}/{}] Loss: [{:.2f}] Accuracy [{:.2f}] Eval: Loss: [{:.2f}] Accuracy[{:.2f}]"
.format(epoch + 1, num_epochs, total_loss / len(train_loader),
total_correct * 100 / total_data,
val_total_loss / len(val_loader),
val_total_correct * 100 / val_total_data))
train_loss_list.append(total_loss / len(train_loader))
iteration_list.append(epoch)
train_accuracy_list.append(total_correct * 100 / total_data)
val_loss_list.append(val_total_loss / len(val_loader))
val_accuracy_list.append(val_total_correct * 100 / val_total_data)
history = {
'train_loss': train_loss_list,
'train_acc': train_accuracy_list,
'val_loss': val_loss_list,
'val_acc': val_accuracy_list
}
return history
print("Starting training")
j = 0
time_point = []
time_string = []
val_acc_list = []
train_acc_list = []
max_acc = 0
for epoch in range(num_epochs):
#linear annealing of learning rate at every 4th epoch
if epoch % 3 == 2:
optimizer = torch.optim.Adam(model.parameters(),
lr=learning_rate * 0.5)
if args.model != 'RNN':
for i, (data, lengths, labels) in enumerate(train_loader):
model.train()
data_batch, length_batch, label_batch = data.to(
device), lengths.to(device), labels.to(device)
optimizer.zero_grad()
#for k in label_batch:
# print(k.type())
outputs = model(data_batch, length_batch).to(device)
loss = criterion(outputs, label_batch)
loss.backward()
optimizer.step()
# validate every 1000 iterations
if i > 0 and i % 100 == 0:
# validate
# import pdb; pdb.set_trace()
time_point.append(j)
j += 1
val_acc = test_model(val_loader, model)
