def train():
# 创建训练、验证 批数据
sp_input_tensor, en_target_tensor, sp_input_language_tokenizer, en_target_language_tokenizer = data_util.load_dataset(
resource_data_file)
sp_input_tensor_train, sp_input_tensor_val, en_target_tensor_train, en_target_tensor_val = data_util.train_test_split(
en_input_tensor, sp_target_tensor, test_size=0.2)
dataset = tf.data.Dataset.from_tensor_slices(
(sp_input_tensor_train,
en_target_tensor_train)).shuffle(len(sp_input_tensor_train))
dataset = dataset.batch(config['batch_size'], drop_remainder=True)
for epoch in range(config['epochs']):
start = time.time()
enc_hidden = tf.zeros(
(config["batch_size"], config["enc_hidden_size"]))
total_loss = 0
for (batch, (input_lang, target_lang)) in enumerate(dataset):
# input_lang--> [batch_size, max_input_length], target_lang-->[batch_size, max_target_length]
batch_loss = seq2seqModel.train_step(input_lang, target_lang,
en_target_language_tokenizer,
enc_hidden)
total_loss += batch_loss
if batch % 100 == 0:
print('Epoch %d, Batch %d, Loss %.4f' %
(epoch + 1, batch, batch_loss.numpy()))
# 2 个 epoch， 保存一个 checkpoint
if (epoch + 1) % 2 == 0:
seq2seqModel.checkpoint.save(file_prefix=config['model_dir'])
print("Epoch %d, Loss %.4f" %
(epoch + 1, total_loss /
(len(sp_input_tensor_train) // config['batch_size'])))
print("Time taken for one epoch %f sec" % (time.time() - start))
def train():
print("Preparing data in %s" % gConfig['train_data'])
steps_per_epoch = len(input_tensor) // gConfig['batch_size']
print(steps_per_epoch)
#enc_hidden = seq2seqModel.Encoder.initHidden()
checkpoint_dir = gConfig['model_data']
#ckpt=io.gfile.listdir(checkpoint_dir)
# if ckpt:
# print("reload pretrained model")
#seq2seqModel.checkpoint.restore(tf.train.latest_checkpoint(checkpoint_dir))
checkpoint_dir = gConfig['model_data']
checkpoint_prefix = os.path.join(checkpoint_dir, ".pt")
start_time = time.time()
encoder1 = seq2seqModel.Encoder(input_lang.n_words, hidden_size).to(device)
attn_decoder1 = seq2seqModel.AttentionDencoder(hidden_size,
target_lang.n_words,
dropout_p=0.1).to(device)
#print(target_tensor)
#print(input_tensor)
max_data = gConfig['max_train_data_size']
while True:
start_time_epoch = time.time()
total_loss = 0
for i in range(0, max_data - 1):
inp = input_tensor[i]
targ = target_tensor[i]
batch_loss = seq2seqModel.train_step(
inp, targ, encoder1, attn_decoder1,
optim.SGD(encoder1.parameters(), lr=0.001),
optim.SGD(encoder1.parameters(), lr=0.001))
total_loss += batch_loss
print('训练总步数:{} 最新每步loss {:.4f}'.format(i, batch_loss))
step_time_epoch = (time.time() - start_time_epoch) / steps_per_epoch
step_loss = total_loss / steps_per_epoch
current_steps = +steps_per_epoch
step_time_total = (time.time() - start_time) / current_steps
print('训练总步数: {} 每步耗时: {} 最新每步耗时: {} 最新每步loss {:.4f}'.format(
current_steps, step_time_total, step_time_epoch, step_loss))
torch.save(
{
'modelA_state_dict': encoder1.state_dict(),
'modelB_state_dict': attn_decoder1.state_dict()
}, checkpoint_prefix)
sys.stdout.flush()
def train():
print("Preparing data in %s" % gConfig['train_data'])
steps_per_epoch = len(input_tensor) // gConfig['batch_size']
print(steps_per_epoch)
enc_hidden = seq2seqModel.encoder.initialize_hidden_state()
checkpoint_dir = gConfig['model_data']
ckpt = tf.io.gfile.listdir(checkpoint_dir)
manager = tf.train.CheckpointManager(seq2seqModel.checkpoint,
directory=checkpoint_dir,
max_to_keep=5)
if ckpt:
print("reload pretrained model")
# seq2seqModel.checkpoint.restore(tf.train.latest_checkpoint(checkpoint_dir))
seq2seqModel.checkpoint.restore(manager.latest_checkpoint)
BUFFER_SIZE = len(input_tensor)
dataset = tf.data.Dataset.from_tensor_slices(
(input_tensor, target_tensor)).shuffle(BUFFER_SIZE)
dataset = dataset.batch(BATCH_SIZE, drop_remainder=True)
checkpoint_dir = gConfig['model_data']
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
start_time = time.time()
while True:
start_time_epoch = time.time()
total_loss = 0
for (batch, (inp, targ)) in enumerate(dataset.take(steps_per_epoch)):
batch_loss = seq2seqModel.train_step(inp, targ, target_token,
enc_hidden)
total_loss += batch_loss
print(batch_loss.numpy())
step_time_epoch = (time.time() - start_time_epoch) / steps_per_epoch
step_loss = total_loss / steps_per_epoch
current_steps = +steps_per_epoch
step_time_total = (time.time() - start_time) / current_steps
print('训练总步数: {} 每步耗时: {} 最新每步耗时: {} 最新每步loss {:.4f}'.format(
current_steps, step_time_total, step_time_epoch,
step_loss.numpy()))
# seq2seqModel.checkpoint.save(file_prefix=checkpoint_prefix)
manager.save()
sys.stdout.flush()
def train():
print("Preparing data in %s" % CONFIG['train_data'])
steps_per_epoch = len(input_tensor) // CONFIG['batch_size']
print(steps_per_epoch)
enc_hidden = seq2seqModel.encoder.initialize_hidden_state()
checkpoint_dir = CONFIG['model_data']
ckpt = tf.io.gfile.listdir(checkpoint_dir)
if ckpt:
print("reload pretrained model")
seq2seqModel.checkpoint.restore(
tf.train.latest_checkpoint(checkpoint_dir))
dataset = tf.data.Dataset.from_tensor_slices(
(input_tensor, target_tensor)).shuffle(len(input_tensor))
dataset = dataset.batch(batch_size, drop_remainder=True)
checkpoint_dir = CONFIG['model_data']
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
start_time = time.time()
while True:
start_time_epoch = time.time()
total_loss = 0
for (batch, (inp, target)) in enumerate(dataset.take(steps_per_epoch)):
batch_loss = seq2seqModel.train_step(inp, target, target_lang,
enc_hidden)
total_loss += batch_loss
print(batch_loss.numpy())
step_time_epoch = (time.time() - start_time_epoch) / steps_per_epoch
step_loss = total_loss / steps_per_epoch
current_steps = +steps_per_epoch
step_time_total = (time.time() - start_time) / current_steps
print('训练总步数: {}, 每步耗时: {}, 最新每步耗时: {}, 最新每步loss值 {:.4f}'.format(
current_steps, step_time_total, step_time_epoch,
step_loss.numpy()))
seq2seqModel.checkpoint.save(file_prefix=checkpoint_prefix)
sys.stdout.flush()