def get_test_predictions(self, test_filename, save_filename):
test_dataset = DialogueDataset(
test_filename,
self.config.sentence_len,
self.vocab,
False)
test_data_loader = torch.utils.data.DataLoader(
test_dataset, self.config.val_batch_size, shuffle=True)
with open(test_filename, 'r') as f:
data = json.load(f)
start = time.clock()
phase_metrics = dict()
epoch_loss = list()
epoch_metrics = list()
results = {"accuracy": list(), "precision": list(), "recall": list(),
"F1": list()}
average_epoch_loss = None
for i, batch in enumerate(tqdm(test_data_loader,
mininterval=2, desc='test', leave=False)):
# prepare data
src_seq, src_pos, src_seg, tgt = map(
lambda x: x.to(self.device), batch[:4])
ids = batch[4]
start_end_idx = batch[5]
# forward
pred = self.model(src_seq, src_pos, src_seg, tgt)
loss = F.cross_entropy(self.prepare_pred(pred).view(-1, 2),
tgt.view(-1))
average_loss = float(loss)
epoch_loss.append(average_loss)
average_epoch_loss = np.mean(epoch_loss)
output = torch.argmax(self.prepare_pred(pred), 3)
record_predictions(output, data, ids, start_end_idx)
get_results(tgt.view(-1).cpu(), output.view(-1).cpu(), results)
phase_metrics["avg_results"] = {key: np.mean(value) for key, value in
results.items()}
phase_metrics["loss"] = average_epoch_loss
phase_metrics["time_taken"] = time.clock() - start
string = ' {} loss: {:.3f} '.format('test', average_epoch_loss)
print(string, end='\n')
data["results"] = phase_metrics
with open(save_filename, 'w') as f:
json.dump(data, f)
return phase_metrics
def __init__(self, args):
# get the dir with pre-trained model
load_dir = os.path.join(args.experiment_dir, args.old_model_dir)
# initialize, and load vocab
self.vocab = Vocab()
vocab_filename = os.path.join(load_dir, "vocab.json")
self.vocab.load_from_dict(vocab_filename)
# load configuration
with open(os.path.join(load_dir, "config.json"), "r") as f:
config = json.load(f)
args.response_len = config["response_len"]
args.history_len = config["history_len"]
# initialize an empty dataset. used to get input features
self.dataset = DialogueDataset(None,
history_len=config["history_len"],
response_len=config["response_len"],
vocab=self.vocab,
update_vocab=False)
# set device
self.device = torch.device(args.device)
# initialize model
model = Transformer(config["vocab_size"],
config["vocab_size"],
config["history_len"],
config["response_len"],
d_word_vec=config["embedding_dim"],
d_model=config["model_dim"],
d_inner=config["inner_dim"],
n_layers=config["num_layers"],
n_head=config["num_heads"],
d_k=config["dim_k"],
d_v=config["dim_v"],
dropout=config["dropout"],
pretrained_embeddings=None).to(self.device)
# load checkpoint
checkpoint = torch.load(os.path.join(load_dir, args.old_model_name),
map_location=self.device)
model.load_state_dict(checkpoint['model'])
# create chatbot
self.chatbot = Chatbot(args, model)
self.args = args
def __init__(self, args):
# set up output directory
self.output_dir = os.path.join(args.experiment_dir, args.run_name)
if not os.path.exists(args.experiment_dir):
os.mkdir(args.experiment_dir)
if not os.path.exists(self.output_dir):
os.mkdir(self.output_dir)
if not os.path.exists(os.path.join(args.experiment_dir,"runs/")):
os.mkdir(os.path.join(args.experiment_dir,"runs/"))
# initialize tensorboard writer
self.runs_dir = os.path.join(args.experiment_dir,"runs/",args.run_name)
self.writer = SummaryWriter(self.runs_dir)
# initialize global steps
self.train_gs = 0
self.val_gs = 0
# initialize model config
self.config = ModelConfig(args)
# check if there is a model to load
if args.old_model_dir is not None:
self.use_old_model = True
self.load_dir = args.old_model_dir
self.config.load_from_file(
os.path.join(self.load_dir, "config.json"))
# create vocab
self.vocab = Vocab()
self.vocab.load_from_dict(os.path.join(self.load_dir, "vocab.json"))
self.update_vocab = False
self.config.min_count=1
else:
self.use_old_model = False
self.vocab = None
self.update_vocab = True
# create data sets
self.dataset_filename = args.dataset_filename
# train
self.train_dataset = DialogueDataset(
os.path.join(self.dataset_filename, "train_data.json"),
self.config.sentence_len,
self.vocab,
self.update_vocab)
self.data_loader_train = torch.utils.data.DataLoader(
self.train_dataset, self.config.train_batch_size, shuffle=True)
self.config.train_len = len(self.train_dataset)
self.vocab = self.train_dataset.vocab
# eval
self.val_dataset = DialogueDataset(
os.path.join(self.dataset_filename, "val_data.json"),
self.config.sentence_len,
self.vocab,
self.update_vocab)
self.data_loader_val = torch.utils.data.DataLoader(
self.val_dataset, self.config.val_batch_size, shuffle=True)
self.config.val_len = len(self.val_dataset)
# update, and save vocab
self.vocab = self.val_dataset.vocab
self.train_dataset.vocab = self.vocab
if (self.config.min_count > 1):
self.config.old_vocab_size = len(self.vocab)
self.vocab.prune_vocab(self.config.min_count)
self.vocab.save_to_dict(os.path.join(self.output_dir, "vocab.json"))
self.vocab_size = len(self.vocab)
self.config.vocab_size = self.vocab_size
# load embeddings
if self.config.pretrained_embeddings_dir is None:
pretrained_embeddings = get_pretrained_embeddings(self.config.pretrained_embeddings_dir , self.vocab)
else:
pretrained_embeddings = None
# print and save the config file
self.config.print_config(self.writer)
self.config.save_config(os.path.join(self.output_dir, "config.json"))
# set device
self.device = torch.device('cuda')
# create model
self.model = Transformer(
self.config.vocab_size,
self.config.label_len,
self.config.sentence_len,
d_word_vec=self.config.embedding_dim,
d_model=self.config.model_dim,
d_inner=self.config.inner_dim,
n_layers=self.config.num_layers,
n_head=self.config.num_heads,
d_k=self.config.dim_k,
d_v=self.config.dim_v,
dropout=self.config.dropout,
pretrained_embeddings=pretrained_embeddings
).to(self.device)
# create optimizer
self.optimizer = torch.optim.Adam(
filter(lambda x: x.requires_grad, self.model.parameters()),
betas=(0.9, 0.98), eps=1e-09)
# load old model, optimizer if there is one
if self.use_old_model:
self.model, self.optimizer = load_checkpoint(
os.path.join(self.load_dir, "model.bin"),
self.model, self.optimizer, self.device)
# create a sceduled optimizer object
self.optimizer = ScheduledOptim(
self.optimizer, self.config.model_dim, self.config.warmup_steps)
def run():
logger.info("using device: {}".format(config.DEVICE))
train_data = process_raw_data()
train_list, test_list = train_test_split(train_data,
test_size=0.2,
random_state=34)
# 加载GPT2模型
model, n_ctx = create_model(True)
model.to(config.DEVICE)
# 是否使用多块GPU进行并行运算: 可以选择要使用哪几块显卡来进行训练
multi_gpu = False
if torch.cuda.is_available() and torch.cuda.device_count() > 1:
logger.info("Using more than one GPUs to train...")
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = config.DEVICE_NUM
model = DataParallel(
model, device_ids=[int(i) for i in config.DEVICE_NUM.split(",")])
multi_gpu = True
# 记录模型参数数量
num_parameters = sum(
[parameter.numel() for parameter in model.parameters()])
logger.info("number of model parameters: {}".format(num_parameters))
# 加载数据
logger.info("loading training data")
train_dataset = DialogueDataset(train_list, n_ctx)
batch_num = len(train_dataset) // config.BATCH_SIZE
test_dataset = DialogueDataset(test_list, n_ctx)
test_batch_num = len(test_dataset) // config.BATCH_SIZE
train_data_loader = DataLoader(train_dataset,
batch_size=config.BATCH_SIZE,
shuffle=True,
num_workers=4,
collate_fn=collate_fn)
test_data_loader = DataLoader(test_dataset,
batch_size=config.BATCH_SIZE,
shuffle=True,
num_workers=1,
collate_fn=collate_fn)
# 计算所有epoch进行参数优化的总步数total_steps
total_steps = int(
len(train_data_loader) * config.EPOCHS / config.BATCH_SIZE /
config.GRADIENT_ACCUMULATION)
logger.info('total training steps = {}'.format(total_steps))
# 设置优化器，并且在初始训练时，使用warmup策略
optimizer = AdamW(model.parameters(),
lr=config.LEARNING_RATE,
correct_bias=True)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=config.WARM_STEPS,
num_training_steps=total_steps)
logger.info("start training...")
best_loss = 100
best_accuracy = 0
for epoch in range(config.EPOCHS):
train_fn(model, train_data_loader, optimizer, scheduler, epoch,
batch_num, multi_gpu)
loss, accuracy = eval_fn(model, test_data_loader, test_batch_num,
multi_gpu)
if loss < best_loss or accuracy > best_accuracy:
logger.info('saving model for epoch {}, best loss: {}'.format(
epoch + 1, loss))
model_to_save = model.module if hasattr(model, 'module') else model
model_to_save.save_pretrained(config.MODEL_PATH)
best_loss = loss
best_accuracy = accuracy
def run():
logger.info("using device: {}".format(config.DEVICE))
if config.TRAIN_MMI:
train_data = process_mmi_raw_data()
else:
train_data = process_raw_data()
train_list, test_list = train_test_split(train_data,
test_size=0.2,
random_state=34)
# 加载GPT2模型
model, n_ctx = create_model(mmi=config.TRAIN_MMI)
model.to(config.DEVICE)
# 是否使用多块GPU进行并行运算: 可以选择要使用哪几块显卡来进行训练
multi_gpu = False
if torch.cuda.is_available() and torch.cuda.device_count() > 1:
logger.info("Using GPU to train...")
model = DataParallel(
model, device_ids=[int(i) for i in config.DEVICE_NUM.split(",")])
multi_gpu = True
else:
logger.info("Using cpu to train...")
# 记录模型参数数量
num_parameters = sum(
[parameter.numel() for parameter in model.parameters()])
logger.info("number of model parameters: {}".format(num_parameters))
# 加载数据
logger.info("loading training data")
train_dataset = DialogueDataset(train_list, n_ctx)
batch_num = len(train_dataset) // config.BATCH_SIZE
test_dataset = DialogueDataset(test_list, n_ctx)
test_batch_num = len(test_dataset) // config.BATCH_SIZE
train_data_loader = DataLoader(train_dataset,
batch_size=config.BATCH_SIZE,
shuffle=True,
num_workers=config.TRAIN_NUM_WORKERS,
collate_fn=collate_fn)
test_data_loader = DataLoader(test_dataset,
batch_size=config.BATCH_SIZE,
shuffle=True,
num_workers=config.TEST_NUM_WORKERS,
collate_fn=collate_fn)
# 计算所有epoch进行参数优化的总步数total_steps
total_steps = int(
len(train_data_loader) * config.EPOCHS / config.BATCH_SIZE /
config.GRADIENT_ACCUMULATION)
logger.info('total training steps = {}'.format(total_steps))
# 设置优化器，并且在初始训练时，使用warmup策略
optimizer = AdamW(model.parameters(),
lr=config.LEARNING_RATE,
correct_bias=True)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=config.WARM_STEPS,
num_training_steps=total_steps)
logger.info("start training...")
best_accuracy = 0
best_loss = 100
for epoch in range(config.EPOCHS):
epoch_start_time = datetime.now()
train_fn(model, train_data_loader, optimizer, scheduler, epoch,
batch_num, multi_gpu)
logger.info("time for epoch {}: {}".format(
epoch + 1,
datetime.now() - epoch_start_time))
loss, accuracy = eval_fn(model, test_data_loader, test_batch_num,
multi_gpu)
if accuracy > best_accuracy or loss < best_loss:
logger.info(
'saving model for epoch {}, best accuracy is {}'.format(
epoch + 1, accuracy))
if config.TRAIN_MMI: # 当前训练MMI模型
model_path = config.MMI_MODEL_PATH
else:
model_path = config.DIALOGUE_MODEL_PATH
if not os.path.exists(model_path):
os.mkdir(model_path)
model_to_save = model.module if hasattr(model, 'module') else model
model_to_save.save_pretrained(model_path)
best_accuracy = accuracy
best_loss = loss