def train(self):
with tf.Session() as sess:
tvars = tf.trainable_variables()
(assignment_map, initialized_variable_names) = modeling.get_assignment_map_from_checkpoint(
tvars, self.__bert_checkpoint_path)
print("init bert model params")
tf.train.init_from_checkpoint(self.__bert_checkpoint_path, assignment_map)
print("init bert model params done")
sess.run(tf.variables_initializer(tf.global_variables()))
current_step = 0
for epoch in range(self.config["epochs"]):
print("----- Epoch {}/{} -----".format(epoch + 1, self.config["epochs"]))
for batch in self.data_obj.next_batch(self.t_in_ids_a, self.t_in_masks_a, self.t_seg_ids_a,
self.t_in_ids_b, self.t_in_masks_b, self.t_seg_ids_b,
self.t_lab_ids):
loss, predictions = self.model.train(sess, batch)
acc, recall, prec, f_beta = get_multi_metrics(pred_y=predictions, true_y=batch["label_ids"],
labels=self.label_list)
print("train: step: {}, loss: {}, acc: {}, recall: {}, precision: {}, f_beta: {}".format(
current_step, loss, acc, recall, prec, f_beta))
current_step += 1
if self.data_obj and current_step % self.config["checkpoint_every"] == 0:
eval_losses = []
eval_accs = []
eval_aucs = []
eval_recalls = []
eval_precs = []
eval_f_betas = []
for eval_batch in self.data_obj.next_batch(self.e_in_ids_a, self.e_in_masks_a, self.e_seg_ids_a,
self.e_in_ids_b, self.e_in_masks_b, self.e_seg_ids_b,
self.e_lab_ids):
eval_loss, eval_predictions = self.model.eval(sess, eval_batch)
eval_losses.append(eval_loss)
acc, recall, prec, f_beta = get_multi_metrics(pred_y=eval_predictions,
true_y=eval_batch["label_ids"],
labels=self.label_list)
eval_accs.append(acc)
eval_recalls.append(recall)
eval_precs.append(prec)
eval_f_betas.append(f_beta)
print("\n")
print("eval: loss: {}, acc: {}, auc: {}, recall: {}, precision: {}, f_beta: {}".format(
mean(eval_losses), mean(eval_accs), mean(eval_aucs), mean(eval_recalls),
mean(eval_precs), mean(eval_f_betas)))
print("\n")
if self.config["ckpt_model_path"]:
save_path = self.config["ckpt_model_path"]
if not os.path.exists(save_path):
os.makedirs(save_path)
model_save_path = os.path.join(save_path, self.config["model_name"])
self.model.saver.save(sess, model_save_path, global_step=current_step)
def train(self):
"""
train model
:return:
"""
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.9, allow_growth=True)
sess_config = tf.ConfigProto(log_device_placement=False, allow_soft_placement=True, gpu_options=gpu_options)
with tf.Session(config=sess_config) as sess:
# init all variable in graph
sess.run(tf.global_variables_initializer())
current_step = 0
for epoch in range(self.config["epochs"]):
print("----- Epoch {}/{} -----".format(epoch + 1, self.config["epochs"]))
for batch in self.train_data_obj.next_batch(self.train_data):
loss, predictions = self.model.train(sess, batch, self.config["keep_prob"])
label_list = list(set(batch["labels"]))
acc, recall, prec, f_beta = get_multi_metrics(pred_y=predictions, true_y=batch["labels"],
labels=label_list)
print("train: step: {}, loss: {}, acc: {}, recall: {}, precision: {}, f_beta: {}".format(
current_step, loss, acc, recall, prec, f_beta))
current_step += 1
if current_step % self.config["checkpoint_every"] == 0:
eval_losses = []
eval_accs = []
eval_recalls = []
eval_precs = []
eval_f_betas = []
for eval_batch in self.eval_data_obj.next_batch(self.eval_data):
eval_loss, eval_predictions = self.model.eval(sess, eval_batch)
eval_losses.append(eval_loss)
eval_label_list = list(set(eval_batch["labels"]))
acc, recall, prec, f_beta = get_multi_metrics(pred_y=eval_predictions,
true_y=eval_batch["labels"],
labels=eval_label_list)
eval_accs.append(acc)
eval_recalls.append(recall)
eval_precs.append(prec)
eval_f_betas.append(f_beta)
print("\n")
print("eval: loss: {}, acc: {}, recall: {}, precision: {}, f_beta: {}".format(
mean(eval_losses), mean(eval_accs), mean(eval_recalls),
mean(eval_precs), mean(eval_f_betas)))
print("\n")
if self.config["ckpt_model_path"]:
save_path = os.path.join(os.path.abspath(os.getcwd()),
self.config["ckpt_model_path"])
if not os.path.exists(save_path):
os.makedirs(save_path)
model_save_path = os.path.join(save_path, self.config["model_name"])
self.model.saver.save(sess, model_save_path, global_step=current_step)
def trainStep(batchX, batchY):
"""
训练函数
"""
feed_dict = {
transformer.inputX: batchX,
transformer.inputY: batchY,
transformer.dropoutKeepProb: config.model.dropoutKeepProb,
transformer.embeddedPosition: embeddedPosition
}
_, summary, step, loss, predictions = sess.run([
trainOp, summaryOp, globalStep, transformer.loss,
transformer.predictions
], feed_dict)
if config.numClasses == 1:
acc, recall, prec, f_beta = get_binary_metrics(
pred_y=predictions, true_y=batchY)
elif config.numClasses > 1:
acc, recall, prec, f_beta = get_multi_metrics(
pred_y=predictions, true_y=batchY, labels=labelList)
trainSummaryWriter.add_summary(summary, step)
return loss, acc, prec, recall, f_beta
def train(self):
"""
train model
:return:
"""
current_step = 0
loss_function = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(self.model.parameters(),
lr=self.config["learning_rate"])
for epoch in range(self.config["epochs"]):
print("----- Epoch {}/{} -----".format(epoch + 1,
self.config["epochs"]))
for batch in self.train_data_obj.next_batch(self.train_tasks):
predictions = self.model(batch)
# print(predictions)
optimizer.zero_grad()
labels = torch.LongTensor(batch["labels"])
# labels = self.one_hot(labels)
# print(labels)
loss = loss_function(predictions, labels)
loss.backward()
optimizer.step()
# print(predictions)
label_list = list(set(batch["labels"]))
predictions_max = torch.argmax(predictions, dim=1)
acc, recall, prec, f_beta = get_multi_metrics(
pred_y=predictions_max,
true_y=batch["labels"],
labels=label_list)
current_step += 1
print(
"train: step: {}, loss: {}, acc: {}, recall: {}, precision: {}, f_beta: {}"
.format(current_step, loss, acc, recall, prec, f_beta))
if current_step % self.config["checkpoint_every"] == 0:
eval_losses = []
eval_accs = []
eval_recalls = []
eval_precs = []
eval_f_betas = []
for eval_batch in self.train_data_obj.next_batch(
self.eval_tasks):
eval_predictions = self.model(eval_batch)
eval_labels = torch.LongTensor(eval_batch["labels"])
eval_loss = loss_function(eval_predictions,
eval_labels)
eval_losses.append(eval_loss)
eval_label_list = list(set(eval_batch["labels"]))
eval_predictions_max = torch.argmax(eval_predictions,
dim=1)
acc, recall, prec, f_beta = get_multi_metrics(
pred_y=eval_predictions_max,
true_y=eval_batch["labels"],
labels=eval_label_list)
eval_accs.append(acc)
eval_recalls.append(recall)
eval_precs.append(prec)
eval_f_betas.append(f_beta)
print("\n")
print(
"eval: loss: {}, acc: {}, recall: {}, precision: {}, f_beta: {}"
.format(mean(eval_losses), mean(eval_accs),
mean(eval_recalls), mean(eval_precs),
mean(eval_f_betas)))
print("\n")
def train(self):
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.9,
allow_growth=True)
sess_config = tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True,
gpu_options=gpu_options)
with tf.Session(config=sess_config) as sess:
sess.run(tf.global_variables_initializer()) # 初始化变量
current_step = 0
# 创建Train/Eval的summar路径和写入对象
train_summary_path = os.path.join(
self.config.BASE_DIR, self.config.summary_path + "/train")
if not os.path.exists(train_summary_path):
os.makedirs(train_summary_path)
train_summary_writer = tf.summary.FileWriter(
train_summary_path, sess.graph)
eval_summary_path = os.path.join(
self.config.BASE_DIR, self.config.summary_path + "/eval")
if not os.path.exists(eval_summary_path):
os.makedirs(eval_summary_path)
eval_summary_writer = tf.summary.FileWriter(
eval_summary_path, sess.graph)
# Train & Eval Process
for epoch in range(self.config.epochs):
print(f"----- Epoch {epoch + 1}/{self.config.epochs} -----")
for batch in self.train_data_loader.next_batch(
self.train_inputs, self.train_labels,
self.config.batch_size):
summary, loss, predictions = self.model.train(
sess, batch, self.config.keep_prob)
train_summary_writer.add_summary(summary)
if self.config.num_classes == 1:
acc = get_binary_metrics(pred_y=predictions.tolist(),
true_y=batch['y'])
print("Train step: {}, acc: {:.3f}".format(
current_step, acc))
elif self.config.num_classes > 1:
acc = get_multi_metrics(pred_y=predictions.tolist(),
true_y=batch['y'])
print("Train step: {}, acc: {:.3f}".format(
current_step, acc))
current_step += 1
if self.eval_data_loader and current_step % self.config.ckeckpoint_every == 0:
eval_losses = []
eval_accs = []
for eval_batch in self.eval_data_loader.next_batch(
self.eval_inputs, self.eval_labels,
self.config.batch_size):
eval_summary, eval_loss, eval_predictions = self.model.eval(
sess, eval_batch)
eval_summary_writer.add_summary(eval_summary)
eval_losses.append(eval_loss)
if self.config.num_classes == 1:
acc = get_binary_metrics(
pred_y=eval_predictions.tolist(),
true_y=batch['y'])
eval_accs.append(acc)
elif self.config.num_classes > 1:
acc = get_multi_metrics(
pred_y=eval_predictions.tolist(),
true_y=batch['y'])
eval_accs.append(acc)
print(
f"Eval \tloss: {list_mean(eval_losses)}, acc: {list_mean(eval_accs)}"
)
if self.config.ckpt_model_path:
save_path = os.path.join(
self.config.BASE_DIR,
self.config.ckpt_model_path)
if not os.path.exists(save_path):
os.makedirs(save_path)
model_save_path = os.path.join(
save_path, self.config.model_name)
self.model.saver.save(sess,
model_save_path,
global_step=current_step)