def main(train_file,
dev_file,
vocab_file,
target_dir,
max_length=50,
hidden_size=300,
dropout=0.2,
num_classes=2,
epochs=1,
batch_size=256,
lr=0.0005,
patience=5,
max_grad_norm=10.0,
gpu_index=0,
checkpoint=None):
#device = torch.device("cuda:{}".format(gpu_index) if torch.cuda.is_available() else "cpu")
device = torch.device("cpu")
print(20 * "=", " Preparing for training ", 20 * "=")
# 保存模型的路径
if not os.path.exists(target_dir):
os.makedirs(target_dir)
# -------------------- Data loading ------------------- #
print("\t* Loading training data...")
train_data = LCQMC_Dataset(train_file, vocab_file, max_length)
train_loader = DataLoader(train_data, shuffle=True, batch_size=batch_size)
print("\t* Loading validation data...")
dev_data = LCQMC_Dataset(dev_file, vocab_file, max_length)
dev_loader = DataLoader(dev_data, shuffle=True, batch_size=batch_size)
# -------------------- Model definition ------------------- #
print("\t* Building model...")
# embeddings = load_embeddings(embeddings_file)
model = ESIM(hidden_size,
dropout=dropout,
num_labels=num_classes,
device=device).to(device)
# -------------------- Preparation for training ------------------- #
print('a')
criterion = nn.CrossEntropyLoss()
# 过滤出需要梯度更新的参数
parameters = filter(lambda p: p.requires_grad, model.parameters())
print('b')
# optimizer = optim.Adadelta(parameters, params["LEARNING_RATE"])
optimizer = torch.optim.Adam(parameters, lr=lr)
# optimizer = torch.optim.Adam(model.parameters(), lr=lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode="max",
factor=0.85,
patience=0)
best_score = 0.0
start_epoch = 1
# Data for loss curves plot
epochs_count = []
train_losses = []
valid_losses = []
# Continuing training from a checkpoint if one was given as argument
if checkpoint:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint["epoch"] + 1
best_score = checkpoint["best_score"]
print("\t* Training will continue on existing model from epoch {}...".
format(start_epoch))
model.load_state_dict(checkpoint["model"])
optimizer.load_state_dict(checkpoint["optimizer"])
epochs_count = checkpoint["epochs_count"]
train_losses = checkpoint["train_losses"]
valid_losses = checkpoint["valid_losses"]
# Compute loss and accuracy before starting (or resuming) training.
_, valid_loss, valid_accuracy, auc = validate(model, dev_loader, criterion)
print(
"\t* Validation loss before training: {:.4f}, accuracy: {:.4f}%, auc: {:.4f}"
.format(valid_loss, (valid_accuracy * 100), auc))
# -------------------- Training epochs ------------------- #
print("\n", 20 * "=", "Training ESIM model on device: {}".format(device),
20 * "=")
patience_counter = 0
for epoch in range(start_epoch, epochs + 1):
epochs_count.append(epoch)
print("* Training epoch {}:".format(epoch))
epoch_time, epoch_loss, epoch_accuracy = train(model, train_loader,
optimizer, criterion,
epoch, max_grad_norm)
train_losses.append(epoch_loss)
print("-> Training time: {:.4f}s, loss = {:.4f}, accuracy: {:.4f}%".
format(epoch_time, epoch_loss, (epoch_accuracy * 100)))
print("* Validation for epoch {}:".format(epoch))
epoch_time, epoch_loss, epoch_accuracy, epoch_auc = validate(
model, dev_loader, criterion)
valid_losses.append(epoch_loss)
print(
"-> Valid. time: {:.4f}s, loss: {:.4f}, accuracy: {:.4f}%, auc: {:.4f}\n"
.format(epoch_time, epoch_loss, (epoch_accuracy * 100), epoch_auc))
# Update the optimizer's learning rate with the scheduler.
scheduler.step(epoch_accuracy)
# Early stopping on validation accuracy.
if epoch_accuracy < best_score:
patience_counter += 1
else:
best_score = epoch_accuracy
patience_counter = 0
torch.save(
{
"epoch": epoch,
"model": model.state_dict(),
"best_score": best_score,
"epochs_count": epochs_count,
"train_losses": train_losses,
"valid_losses": valid_losses
}, os.path.join(target_dir, "best.pth.tar"))
# Save the model at each epoch.
torch.save(
{
"epoch": epoch,
"model": model.state_dict(),
"best_score": best_score,
"optimizer": optimizer.state_dict(),
"epochs_count": epochs_count,
"train_losses": train_losses,
"valid_losses": valid_losses
}, os.path.join(target_dir, "esim_{}.pth.tar".format(epoch)))
if patience_counter >= patience:
print("-> Early stopping: patience limit reached, stopping...")
break
y = y.to(device)
y_hat = net(sent1.to(device), sent2.to(device), mask1.to(device),
mask2.to(device))
l = crition(y_hat, y)
optimizer.zero_grad()
l.backward()
optimizer.step()
train_l_sum += l
batch_num += 1
n += y.shape[0]
train_acc_sum += (y_hat.argmax(dim=1) == y).sum().cpu().item()
if batch_num % 500 == 0:
val_acc, val_loss = val_test(net, dev_iter, crition)
if val_acc >= best_acc:
best_acc = val_acc
torch.save(net.state_dict(), args.model_path)
print(
'batch: %d, train_loss: %.3f train_acc: %.4f val_loss: %.3f val_acc: %.4f time: %.3F'
% (batch_num, train_l_sum / batch_num, train_acc_sum / n,
val_loss, val_acc, time.time() - start))
# # 加载测试数据
# def load_test_data(args):
# text = data.Field(sequential= True, use_vocab = True, lower = True, fix_length = args.max_len, include_lengths = True,batch_first = True)
# fields = [('sentence1',text),('sentence2',text)]
# # 读取数据
# test = []
# with open('./data/snli.test','r') as f:
# for line in f.readlines():
# sents = line.strip().split('|||')
# test.append(data.Example.fromlist([sents[0],sents[1]],fields))
def main():
device = args.device
print(20 * "=", " Preparing for training ", 20 * "=")
# 保存模型的路径
if not os.path.exists(args.target_dir):
os.makedirs(args.target_dir)
# -------------------- Data loading ------------------- #
print("Loading data......")
train_loader, dev_loader, test_loader, SEN1, SEN2 = load_data(
args.batch_size, device)
embedding = SEN1.vectors
vocab_size = len(embedding)
print("vocab_size:", vocab_size)
# -------------------- Model definition ------------------- #
print("\t* Building model...")
model = ESIM(args.hidden_size,
embedding=embedding,
dropout=args.dropout,
num_labels=args.num_classes,
device=device).to(device)
# -------------------- Preparation for training ------------------- #
criterion = nn.CrossEntropyLoss()
# 过滤出需要梯度更新的参数
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = torch.optim.Adam(parameters, lr=args.lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode="min",
factor=0.1,
patience=10)
best_score = 0.0
if args.ckp:
checkpoint = torch.load(os.path.join(args.target_dir, args.ckp))
best_score = checkpoint["best_score"]
model.load_state_dict(checkpoint["model"])
optimizer.load_state_dict(checkpoint["optimizer"])
_, valid_loss, valid_accuracy = validate(model, dev_loader, criterion)
print("\t* Validation loss before training: {:.4f}, accuracy: {:.4f}%".
format(valid_loss, (valid_accuracy * 100)))
# -------------------- Training epochs ------------------- #
print("\n", 20 * "=", "Training ESIM model on device: {}".format(device),
20 * "=")
patience_counter = 0
for epoch in range(args.num_epoch):
print("* Training epoch {}:".format(epoch))
epoch_time, epoch_loss, epoch_accuracy = train(model, train_loader,
optimizer, criterion,
args.max_grad_norm,
device)
print("-> Training time: {:.4f}s, loss = {:.4f}, accuracy: {:.4f}%".
format(epoch_time, epoch_loss, (epoch_accuracy * 100)))
print("* Validation for epoch {}:".format(epoch))
epoch_time, epoch_loss, epoch_accuracy = validate(
model, dev_loader, criterion, device)
print("-> Valid. time: {:.4f}s, loss: {:.4f}, accuracy: {:.4f}%\n".
format(epoch_time, epoch_loss, (epoch_accuracy * 100)))
# Update the optimizer's learning rate with the scheduler.
scheduler.step(epoch_accuracy)
# Early stopping on validation accuracy.
if epoch_accuracy < best_score:
patience_counter += 1
else:
print("save model！！！！")
best_score = epoch_accuracy
patience_counter = 0
torch.save(
{
"model": model.state_dict(),
"best_score": best_score,
"optimizer": optimizer.state_dict(),
}, os.path.join(args.target_dir, "best.pth.tar"))
if patience_counter >= 5:
print("-> Early stopping: patience limit reached, stopping...")
break
# ##-------------------- Testing epochs ------------------- #
# print(20 * "=", " Testing ", 20 * "=")
# if args.ckp:
# checkpoint = torch.load(os.path.join(args.target_dir, args.ckp))
# best_score = checkpoint["best_score"]
# model.load_state_dict(checkpoint["model"])
# optimizer.load_state_dict(checkpoint["optimizer"])
#
# print("best_score:", best_score)
# all_labels = test(model, test_loader, device)
# print(all_labels[:10])
# target_label = [id2label[id] for id in all_labels]
# print(target_label[:10])
# with open(os.path.join(args.target_dir, 'result.txt'), 'w+') as f:
# for label in target_label:
# f.write(label + '\n')
del train_loader
del dev_loader
del test_loader
del SEN1
del SEN2
del embedding
(epoch + 1, epoch_mins, epoch_secs))
f_log.write('\tTrain Loss: %.3f | Train Acc: %.2f %%\n' %
(train_loss, train_acc * 100))
f_log.write('\t Val. Loss: %.3f | Val. Acc: %.2f %%\n' %
(valid_loss, valid_acc * 100))
print('Epoch: %d | Epoch Time: %dm %ds' %
(epoch + 1, epoch_mins, epoch_secs))
print('\tTrain Loss: %.3f | Train Acc: %.2f %%' %
(train_loss, train_acc * 100))
print('\t Val. Loss: %.3f | Val. Acc: %.2f %%' %
(valid_loss, valid_acc * 100))
if valid_loss < best_valid_loss:
best_valid_loss = valid_loss
best_valid_acc = valid_acc
torch.save(model.state_dict(), './saved_model/esim.pt')
print("New model saved!")
f_log.write("New model saved!\n")
f_log.flush()
f_log.close()
model.load_state_dict(torch.load('./saved_model/esim.pt'))
model.eval()
f_valid = open("data/test-set.data", "r", encoding='utf-8')
f_res = open('prediction.txt', 'w')
for i, rowlist in enumerate(f_valid):
rowlist = rowlist[:-1].split('\t')
input_sent = []
for sent in rowlist[:2]:
def model_train_validate_test(train_df,
dev_df,
test_df,
embeddings_file,
vocab_file,
target_dir,
mode,
num_labels=2,
max_length=50,
hidden_size=200,
dropout=0.2,
epochs=50,
batch_size=256,
lr=0.0005,
patience=5,
max_grad_norm=10.0,
gpu_index=0,
if_save_model=False,
checkpoint=None):
device = torch.device(
"cuda:{}".format(gpu_index) if torch.cuda.is_available() else "cpu")
print(20 * "=", " Preparing for training ", 20 * "=")
# 保存模型的路径
if not os.path.exists(target_dir):
os.makedirs(target_dir)
# -------------------- Data loading ------------------- #
print("\t* Loading training data...")
train_data = My_Dataset(train_df, vocab_file, max_length, mode)
train_loader = DataLoader(train_data, shuffle=True, batch_size=batch_size)
print("\t* Loading validation data...")
dev_data = My_Dataset(dev_df, vocab_file, max_length, mode)
dev_loader = DataLoader(dev_data, shuffle=True, batch_size=batch_size)
print("\t* Loading test data...")
test_data = My_Dataset(test_df, vocab_file, max_length, mode)
test_loader = DataLoader(test_data, shuffle=False, batch_size=batch_size)
# -------------------- Model definition ------------------- #
print("\t* Building model...")
if (embeddings_file is not None):
embeddings = load_embeddings(embeddings_file)
else:
embeddings = None
model = ESIM(hidden_size,
embeddings=embeddings,
dropout=dropout,
num_labels=num_labels,
device=device).to(device)
total_params = sum(p.numel() for p in model.parameters())
print(f'{total_params:,} total parameters.')
total_trainable_params = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print(f'{total_trainable_params:,} training parameters.')
# -------------------- Preparation for training ------------------- #
criterion = nn.CrossEntropyLoss()
# 过滤出需要梯度更新的参数
parameters = filter(lambda p: p.requires_grad, model.parameters())
# optimizer = optim.Adadelta(parameters, params["LEARNING_RATE"])
optimizer = torch.optim.Adam(parameters, lr=lr)
# optimizer = torch.optim.Adam(model.parameters(), lr=lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode="max",
factor=0.85,
patience=0)
best_score = 0.0
start_epoch = 1
# Data for loss curves plot
epochs_count = []
train_losses = []
valid_losses = []
# Continuing training from a checkpoint if one was given as argument
if checkpoint:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint["epoch"] + 1
best_score = checkpoint["best_score"]
print("\t* Training will continue on existing model from epoch {}...".
format(start_epoch))
model.load_state_dict(checkpoint["model"])
optimizer.load_state_dict(checkpoint["optimizer"])
epochs_count = checkpoint["epochs_count"]
train_losses = checkpoint["train_losses"]
valid_losses = checkpoint["valid_losses"]
# Compute loss and accuracy before starting (or resuming) training.
_, valid_loss, valid_accuracy, _, = validate(model, dev_loader, criterion)
print("\t* Validation loss before training: {:.4f}, accuracy: {:.4f}%".
format(valid_loss, (valid_accuracy * 100)))
# -------------------- Training epochs ------------------- #
print("\n", 20 * "=", "Training ESIM model on device: {}".format(device),
20 * "=")
patience_counter = 0
for epoch in range(start_epoch, epochs + 1):
epochs_count.append(epoch)
print("* Training epoch {}:".format(epoch))
epoch_time, epoch_loss, epoch_accuracy = train(model, train_loader,
optimizer, criterion,
epoch, max_grad_norm)
train_losses.append(epoch_loss)
print("-> Training time: {:.4f}s, loss = {:.4f}, accuracy: {:.4f}%".
format(epoch_time, epoch_loss, (epoch_accuracy * 100)))
print("* Validation for epoch {}:".format(epoch))
epoch_time, epoch_loss, epoch_accuracy, _, = validate(
model, dev_loader, criterion)
valid_losses.append(epoch_loss)
print("-> Valid. time: {:.4f}s, loss: {:.4f}, accuracy: {:.4f}%\n".
format(epoch_time, epoch_loss, (epoch_accuracy * 100)))
# Update the optimizer's learning rate with the scheduler.
scheduler.step(epoch_accuracy)
# Early stopping on validation accuracy.
if epoch_accuracy < best_score:
patience_counter += 1
else:
best_score = epoch_accuracy
patience_counter = 0
if (if_save_model):
torch.save(
{
"epoch": epoch,
"model": model.state_dict(),
"best_score": best_score,
"epochs_count": epochs_count,
"train_losses": train_losses,
"valid_losses": valid_losses
}, os.path.join(target_dir, "best.pth.tar"))
print("save model succesfully!\n")
print("* Test for epoch {}:".format(epoch))
_, _, test_accuracy, predictions = validate(
model, test_loader, criterion)
print("Test accuracy: {:.4f}%\n".format(test_accuracy))
test_prediction = pd.DataFrame({'prediction': predictions})
test_prediction.to_csv(os.path.join(target_dir,
"test_prediction.csv"),
index=False)
if patience_counter >= patience:
print("-> Early stopping: patience limit reached, stopping...")
break
def main(args):
print(20 * "=", " Preparing for training ", 20 * "=")
# 保存模型的路径
if not os.path.exists(args.target_dir):
os.makedirs(args.target_dir)
# -------------------- Data loading ------------------- #
print("\t* Loading training data...")
# train_data = LCQMC_dataset(args.train_file, args.vocab_file, args.max_length, test_flag=False)
train_data = LCQMC_dataset(args.train_file, args.vocab_file, args.max_length, test_flag=False)
train_loader = DataLoader(train_data, batch_size=args.batch_size, shuffle=True)
print("\t* Loading valid data...")
dev_data = LCQMC_dataset(args.dev_file, args.vocab_file, args.max_length, test_flag=False)
dev_loader = DataLoader(dev_data, batch_size=args.batch_size, shuffle=True)
# -------------------- Model definition ------------------- #
print("\t* Building model...")
embeddings = load_embeddings(args.embed_file)
model = ESIM(args, embeddings=embeddings).to(args.device)
# -------------------- Preparation for training ------------------- #
criterion = nn.CrossEntropyLoss() # 交叉熵损失函数
# 过滤出需要梯度更新的参数
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = torch.optim.Adam(parameters, lr=args.lr) # 优化器
# 学习计划
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max',
factor=0.85, patience=0)
best_score = 0.0
start_epoch = 1
epochs_count = []
train_losses = []
valid_losses = []
# Continuing training from a checkpoint if one was given as argument
if args.checkpoint:
# 从文件中加载checkpoint数据, 从而继续训练模型
checkpoints = torch.load(args.checkpoint)
start_epoch = checkpoints["epoch"] + 1
best_score = checkpoints["best_score"]
print("\t* Training will continue on existing model from epoch {}...".format(start_epoch))
model.load_state_dict(checkpoints["model"]) # 模型部分
optimizer.load_state_dict(checkpoints["optimizer"])
epochs_count = checkpoints["epochs_count"]
train_losses = checkpoints["train_losses"]
valid_losses = checkpoints["valid_losses"]
# 这里改为只有从以前加载的checkpoint中才进行计算 valid, Compute loss and accuracy before starting (or resuming) training.
_, valid_loss, valid_accuracy, auc = validate(model, dev_loader, criterion)
print("\t* Validation loss before training: {:.4f}, accuracy: {:.4f}%, auc: {:.4f}"
.format(valid_loss, (valid_accuracy * 100), auc))
# -------------------- Training epochs ------------------- #
print("\n", 20 * "=", "Training ESIM model on device: {}".format(args.device), 20 * "=")
patience_counter = 0
for epoch in range(start_epoch, args.epochs + 1):
epochs_count.append(epoch)
print("* Training epoch {}:".format(epoch))
epoch_time, epoch_loss, epoch_accuracy = train(model, train_loader, optimizer,
criterion, epoch, args.max_grad_norm)
train_losses.append(epoch_loss)
print("-> Training time: {:.4f}s, loss = {:.4f}, accuracy: {:.4f}%"
.format(epoch_time, epoch_loss, (epoch_accuracy * 100)))
print("* Validation for epoch {}:".format(epoch))
epoch_time, epoch_loss, epoch_accuracy, epoch_auc = validate(model, train_loader, criterion)
valid_losses.append(epoch_loss)
print("-> Valid. time: {:.4f}s, loss: {:.4f}, accuracy: {:.4f}%, auc: {:.4f}\n"
.format(epoch_time, epoch_loss, (epoch_accuracy * 100), epoch_auc))
# Update the optimizer's learning rate with the scheduler.
scheduler.step(epoch_accuracy)
# Early stopping on validation accuracy.
if epoch_accuracy < best_score:
patience_counter += 1
else:
best_score = epoch_accuracy
patience_counter = 0
# 保存最好的结果，需要保存的参数，这些参数在checkpoint中都能找到
torch.save(
{
"epoch": epoch,
"model": model.state_dict(),
"best_score": best_score,
"epochs_count": epochs_count,
"train_losses": train_losses,
"valid_losses": valid_losses},
os.path.join(args.target_dir, "new_best.pth.tar"))
# 保存每个epoch的结果 Save the model at each epoch.(这里可要可不要)
torch.save(
{
"epoch": epoch,
"model": model.state_dict(),
"best_score": best_score,
"optimizer": optimizer.state_dict(),
"epochs_count": epochs_count,
"train_losses": train_losses,
"valid_losses": valid_losses},
os.path.join(args.target_dir, "new_esim_{}.pth.tar".format(epoch)))
if patience_counter >= args.patience:
print("-> Early stopping: patience limit reached, stopping...")
break
def main(test_file, pretrained_file, embeddings_file, batch_size=1):
"""
Test the ESIM model with pretrained weights on some dataset.
Args:
test_file: The path to a file containing preprocessed NLI data.
pretrained_file: The path to a checkpoint produced by the
'train_model' script.
vocab_size: The number of words in the vocabulary of the model
being tested.
embedding_dim: The size of the embeddings in the model.
hidden_size: The size of the hidden layers in the model. Must match
the size used during training. Defaults to 300.
num_classes: The number of classes in the output of the model. Must
match the value used during training. Defaults to 3.
batch_size: The size of the batches used for testing. Defaults to 32.
"""
debug_file = open('test_debug.txt', 'w')
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
#device = torch.device("cpu")
print(20 * "=", " Preparing for generating representations ", 20 * "=")
checkpoint = torch.load(pretrained_file)
# Retrieving model parameters from checkpoint.
vocab_size = checkpoint["model"]["_word_embedding.weight"].size(0)
embedding_dim = checkpoint["model"]['_word_embedding.weight'].size(1)
hidden_size = checkpoint["model"]["_projection.0.weight"].size(0)
num_classes = checkpoint["model"]["_classification.4.weight"].size(0)
print("\t* Loading the data...")
with open(test_file, "rb") as pkl:
test_data = NLIDataset(pickle.load(pkl))
print(test_data, file=debug_file)
test_loader = DataLoader(test_data, shuffle=False, batch_size=batch_size)
print("\t* Building model...")
# loading the embedding weights separately
# with open(embeddings_file, "rb") as pkl:
pkl = open(embeddings_file, "rb")
embeddings = torch.tensor(pickle.load(pkl), dtype=torch.float)\
.to(device)
pkl.close()
# model = ESIM(vocab_size,
# embedding_dim,
# hidden_size,
# num_classes=num_classes,
# device=device).to(device)
model = ESIM(embeddings.shape[0],
embeddings.shape[1],
hidden_size,
embeddings=embeddings,
num_classes=num_classes,
device=device).to(device)
# Writing custom load_state_dict
pretrained_dict = checkpoint["model"]
own_state = model.state_dict()
for i, (name, param) in enumerate(pretrained_dict.items()):
#print(name, type(name))
# if name is "_word_embedding.weight":
# print(name)
# continue
if i == 0:
continue
if isinstance(param, Parameter):
# backwards compatibility for serialized parameters
param = param.data
own_state[name].copy_(param)
#model.load_state_dict(checkpoint["model"])
print(
20 * "=",
" Loading the representations from ESIM model on device: {} ".format(
device), 20 * "=")
batch_time, total_time, save_rep = test(model, test_loader)
print("-> Average batch processing time: {:.4f}s, total test time:\
{:.4f}s,%".format(batch_time, total_time))
file_debug = open('test_save_rep_details.txt', 'w')
print('len of save_rep is' + str(len(save_rep)), file=file_debug)
try:
print('save_rep sample key is' + str(list(save_rep.keys())[0]),
file=file_debug)
print('save_rep sample value is' + str(list(save_rep.values())[0]),
file=file_debug)
except:
pass
# Dump save_rep as a pickle file
with open('test_nv_repr.pickle', 'wb') as handle:
pickle.dump(save_rep, handle, protocol=pickle.HIGHEST_PROTOCOL)
def main(train_q1_file,
train_q2_file,
train_labels_file,
dev_q1_file,
dev_q2_file,
dev_labels_file,
embeddings_file,
target_dir,
hidden_size=128,
dropout=0.5,
num_classes=2,
epochs=15,
batch_size=64,
lr=0.001,
patience=5,
max_grad_norm=10.0,
gpu_index=0,
checkpoint=None):
device = torch.device(
"cuda:{}".format(gpu_index) if torch.cuda.is_available() else "cpu")
print(20 * "=", " Preparing for training ", 20 * "=")
# 保存模型的路径
if not os.path.exists(target_dir):
os.makedirs(target_dir)
# -------------------- Data loading ------------------- #
print("\t* Loading training data...")
train_q1 = np.load(train_q1_file)
train_q2 = np.load(train_q2_file)
train_labels = np.load(train_labels_file)
# train_labels = label_transformer(train_labels)
train_data = {"q1": train_q1, "q2": train_q2, "labels": train_labels}
train_data = QQPDataset(train_data)
train_loader = DataLoader(train_data, shuffle=True, batch_size=batch_size)
print("\t* Loading validation data...")
dev_q1 = np.load(dev_q1_file)
dev_q2 = np.load(dev_q2_file)
dev_labels = np.load(dev_labels_file)
# dev_labels = label_transformer(dev_labels)
dev_data = {"q1": dev_q1, "q2": dev_q2, "labels": dev_labels}
dev_data = QQPDataset(dev_data)
dev_loader = DataLoader(dev_data, shuffle=True, batch_size=batch_size)
# -------------------- Model definition ------------------- #
print("\t* Building model...")
embeddings = torch.tensor(np.load(embeddings_file),
dtype=torch.float).to(device)
model = ESIM(embeddings.shape[0],
embeddings.shape[1],
hidden_size,
embeddings=embeddings,
dropout=dropout,
num_classes=num_classes,
device=device).to(device)
# -------------------- Preparation for training ------------------- #
criterion = nn.CrossEntropyLoss()
# 过滤出需要梯度更新的参数
parameters = filter(lambda p: p.requires_grad, model.parameters())
# optimizer = optim.Adadelta(parameters, params["LEARNING_RATE"])
optimizer = torch.optim.Adam(parameters, lr=lr)
# optimizer = torch.optim.Adam(model.parameters(), lr=lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode="max",
factor=0.85,
patience=0)
best_score = 0.0
start_epoch = 1
# Data for loss curves plot
epochs_count = []
train_losses = []
valid_losses = []
# Continuing training from a checkpoint if one was given as argument
if checkpoint:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint["epoch"] + 1
best_score = checkpoint["best_score"]
print("\t* Training will continue on existing model from epoch {}...".
format(start_epoch))
model.load_state_dict(checkpoint["model"])
optimizer.load_state_dict(checkpoint["optimizer"])
epochs_count = checkpoint["epochs_count"]
train_losses = checkpoint["train_losses"]
valid_losses = checkpoint["valid_losses"]
# Compute loss and accuracy before starting (or resuming) training.
_, valid_loss, valid_accuracy = validate(model, dev_loader, criterion)
print("\t* Validation loss before training: {:.4f}, accuracy: {:.4f}%".
format(valid_loss, (valid_accuracy * 100)))
# -------------------- Training epochs ------------------- #
print("\n", 20 * "=", "Training ESIM model on device: {}".format(device),
20 * "=")
patience_counter = 0
for epoch in range(start_epoch, epochs + 1):
epochs_count.append(epoch)
print("* Training epoch {}:".format(epoch))
epoch_time, epoch_loss, epoch_accuracy = train(model, train_loader,
optimizer, criterion,
epoch, max_grad_norm)
train_losses.append(epoch_loss)
print("-> Training time: {:.4f}s, loss = {:.4f}, accuracy: {:.4f}%".
format(epoch_time, epoch_loss, (epoch_accuracy * 100)))
print("* Validation for epoch {}:".format(epoch))
epoch_time, epoch_loss, epoch_accuracy = validate(
model, dev_loader, criterion)
valid_losses.append(epoch_loss)
print("-> Valid. time: {:.4f}s, loss: {:.4f}, accuracy: {:.4f}%\n".
format(epoch_time, epoch_loss, (epoch_accuracy * 100)))
# Update the optimizer's learning rate with the scheduler.
scheduler.step(epoch_accuracy)
# Early stopping on validation accuracy.
if epoch_accuracy < best_score:
patience_counter += 1
else:
best_score = epoch_accuracy
patience_counter = 0
# Save the best model. The optimizer is not saved to avoid having
# a checkpoint file that is too heavy to be shared. To resume
# training from the best model, use the 'esim_*.pth.tar'
# checkpoints instead.
torch.save(
{
"epoch": epoch,
"model": model.state_dict(),
"best_score": best_score,
"epochs_count": epochs_count,
"train_losses": train_losses,
"valid_losses": valid_losses
}, os.path.join(target_dir, "best.pth.tar"))
# Save the model at each epoch.
torch.save(
{
"epoch": epoch,
"model": model.state_dict(),
"best_score": best_score,
"optimizer": optimizer.state_dict(),
"epochs_count": epochs_count,
"train_losses": train_losses,
"valid_losses": valid_losses
}, os.path.join(target_dir, "esim_{}.pth.tar".format(epoch)))
if patience_counter >= patience:
print("-> Early stopping: patience limit reached, stopping...")
break
