graph = tf.Graph()
with graph.as_default():
name_to_features = {
"input_ids": tf.FixedLenFeature([128], tf.int64),
"input_mask": tf.FixedLenFeature([128], tf.int64),
"segment_ids": tf.FixedLenFeature([128], tf.int64),
"label_ids": tf.FixedLenFeature([], tf.int64),
}
params = Bunch({})
params.epoch = epoch
params.batch_size = 32
jd_test = "/data/xuht/jd_comment/train.tfrecords"
print(params["batch_size"], "===batch size===")
input_fn = tf_data_utils.train_input_fn(jd_test,
tf_data_utils._decode_record,
name_to_features, params)
sess = tf.Session(config=sess_config)
init_op = tf.group(tf.local_variables_initializer())
sess.run(init_op)
sess.run(hvd.broadcast_global_variables(0))
i = 0
cnt = 0
while True:
try:
features = sess.run(input_fn)
i += 1
def main(_):
hvd.init()
sess_config = tf.ConfigProto()
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
print("== hvd local rank == {}".format(hvd.local_rank()))
graph = tf.Graph()
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
with graph.as_default():
import json
# config = json.load(open("/data/xuht/bert/chinese_L-12_H-768_A-12/bert_config.json", "r"))
config = json.load(open(FLAGS.config_file, "r"))
init_checkpoint = FLAGS.init_checkpoint
print("===init checkoutpoint==={}".format(init_checkpoint))
# init_checkpoint = "/data/xuht/bert/chinese_L-12_H-768_A-12/bert_model.ckpt"
# init_checkpoint = "/data/xuht/concat/model_1/oqmrc.ckpt"
config = Bunch(config)
config.use_one_hot_embeddings = True
config.scope = "bert"
config.dropout_prob = 0.1
config.label_type = "single_label"
import json
label_dict = json.load(open(FLAGS.label_id))
# label_tensor = np.asarray(label_dict["class_ratio"]).astype(np.float32)
label_tensor = None
# config.loss = "focal_loss"
# os.environ["CUDA_VISIBLE_DEVICES"] = FLAGS.gpu_id
sess = tf.Session(config=sess_config)
train_size = int(FLAGS.train_size / hvd.size())
num_train_steps = int(train_size / FLAGS.batch_size * FLAGS.epoch)
num_warmup_steps = int(num_train_steps * 0.01)
num_storage_steps = int(train_size / FLAGS.batch_size)
print(num_train_steps, num_warmup_steps, "=============")
opt_config = Bunch({
"init_lr": (1e-5 / hvd.size()),
"num_train_steps": num_train_steps,
"num_warmup_steps": num_warmup_steps
})
model_io_config = Bunch({"fix_lm": False})
model_io_fn = model_io.ModelIO(model_io_config)
num_choice = FLAGS.num_classes
max_seq_length = FLAGS.max_length
if FLAGS.model_type == "original":
model_function = bert_order_classifier.classifier_model_fn_builder
elif FLAGS.model_type == "attn":
model_function = bert_order_classifier.classifier_attn_model_fn_builder
elif FLAGS.model_type == "orignal_nonlinear":
model_function = bert_order_classifier.classifier_model_fn_builder_v1
model_train_fn = model_function(config,
num_choice,
init_checkpoint,
model_reuse=None,
load_pretrained=True,
model_io_fn=model_io_fn,
model_io_config=model_io_config,
opt_config=opt_config,
input_name=["a", "b"],
label_tensor=label_tensor)
model_eval_fn = model_function(config,
num_choice,
init_checkpoint,
model_reuse=True,
load_pretrained=True,
model_io_fn=model_io_fn,
model_io_config=model_io_config,
opt_config=opt_config,
input_name=["a", "b"],
label_tensor=label_tensor)
def metric_fn(features, logits, loss):
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.nn.softmax(logits)
accuracy = correct = tf.equal(
tf.cast(pred_label, tf.int32),
tf.cast(features["label_ids"], tf.int32))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
return {
"accuracy": accuracy,
"loss": loss,
"pred_label": pred_label,
"label_ids": features["label_ids"]
}
name_to_features = {
"input_ids_a": tf.FixedLenFeature([max_seq_length], tf.int64),
"input_mask_a": tf.FixedLenFeature([max_seq_length], tf.int64),
"segment_ids_a": tf.FixedLenFeature([max_seq_length], tf.int64),
"input_ids_b": tf.FixedLenFeature([max_seq_length], tf.int64),
"input_mask_b": tf.FixedLenFeature([max_seq_length], tf.int64),
"segment_ids_b": tf.FixedLenFeature([max_seq_length], tf.int64),
"label_ids": tf.FixedLenFeature([], tf.int64),
}
def _decode_record(record, name_to_features):
"""Decodes a record to a TensorFlow example.
"""
example = tf.parse_single_example(record, name_to_features)
# tf.Example only supports tf.int64, but the TPU only supports tf.int32.
# So cast all int64 to int32.
for name in list(example.keys()):
t = example[name]
if t.dtype == tf.int64:
t = tf.to_int32(t)
example[name] = t
return example
params = Bunch({})
params.epoch = FLAGS.epoch
params.batch_size = FLAGS.batch_size
# train_features = tf_data_utils.train_input_fn("/data/xuht/wsdm19/data/train.tfrecords",
# _decode_record, name_to_features, params)
# eval_features = tf_data_utils.eval_input_fn("/data/xuht/wsdm19/data/dev.tfrecords",
# _decode_record, name_to_features, params)
train_features = tf_data_utils.train_input_fn(FLAGS.train_file,
_decode_record,
name_to_features, params)
eval_features = tf_data_utils.eval_input_fn(FLAGS.dev_file,
_decode_record,
name_to_features, params)
[train_op, train_loss, train_per_example_loss,
train_logits] = model_train_fn(train_features, [],
tf.estimator.ModeKeys.TRAIN)
[_, eval_loss, eval_per_example_loss,
eval_logits] = model_eval_fn(eval_features, [],
tf.estimator.ModeKeys.EVAL)
result = metric_fn(eval_features, eval_logits, eval_loss)
model_io_fn.set_saver()
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
sess.run(hvd.broadcast_global_variables(0))
def eval_fn(result):
i = 0
total_accuracy = 0
label, label_id = [], []
label_weight = []
while True:
try:
eval_result = sess.run(result)
total_accuracy += eval_result["accuracy"]
label_id.extend(eval_result["label_ids"])
label.extend(eval_result["pred_label"])
for item in eval_result["label_ids"]:
label_weight.append(label_tensor[item])
i += 1
except tf.errors.OutOfRangeError:
print("End of dataset")
break
f1 = f1_score(label_id,
label,
average="macro",
sample_weight=label_weight)
accuracy = accuracy_score(label_id,
label,
sample_weight=label_weight)
print("test accuracy accuracy {} {} f1 {}".format(
total_accuracy / i, accuracy, f1))
return total_accuracy / i, f1
def train_fn(op, loss):
i = 0
cnt = 0
total_loss = 0.0
while True:
try:
[_, train_loss] = sess.run([op, loss])
total_loss += train_loss
i += 1
cnt += 1
if np.mod(i, num_storage_steps) == 0:
print(total_loss / cnt)
# model_io_fn.save_model(sess, "/data/xuht/wsdm19/data/model_11_15_focal_loss/oqmrc_{}.ckpt".format(int(i/8000)))
if hvd.rank() == 0:
model_io_fn.save_model(
sess,
FLAGS.model_output + "/oqmrc_{}.ckpt".format(
int(i / num_storage_steps)))
print("==successful storing model=={}".format(
int(i / num_storage_steps)))
total_loss = 0
cnt = 0
except tf.errors.OutOfRangeError:
break
print("===========begin to train============")
train_fn(train_op, train_loss)
# model_io_fn.save_model(sess, "/data/xuht/wsdm19/data/model_11_15_focal_loss/oqmrc.ckpt")
if hvd.rank() == 0:
model_io_fn.save_model(sess, FLAGS.model_output + "/oqmrc.ckpt")
print("===========begin to eval============")
accuracy, f1 = eval_fn(result)
print("==accuracy {} f1 {}==".format(accuracy, f1))
def main(_):
hvd.init()
sess_config = tf.ConfigProto()
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
graph = tf.Graph()
with graph.as_default():
import json
config = json.load(open(FLAGS.config_file, "r"))
init_checkpoint = FLAGS.init_checkpoint
config = Bunch(config)
config.use_one_hot_embeddings = True
config.scope = "bert"
config.dropout_prob = 0.1
config.label_type = "single_label"
if FLAGS.if_shard == "0":
train_size = FLAGS.train_size
epoch = int(FLAGS.epoch / hvd.size())
elif FLAGS.if_shard == "1":
train_size = int(FLAGS.train_size / hvd.size())
epoch = FLAGS.epoch
init_lr = 2e-5
label_dict = json.load(open(FLAGS.label_id))
num_train_steps = int(train_size / FLAGS.batch_size * epoch)
num_warmup_steps = int(num_train_steps * 0.1)
num_storage_steps = int(train_size / FLAGS.batch_size)
print(" model type {}".format(FLAGS.model_type))
print(num_train_steps, num_warmup_steps, "=============")
opt_config = Bunch({
"init_lr": init_lr / hvd.size(),
"num_train_steps": num_train_steps,
"num_warmup_steps": num_warmup_steps
})
sess = tf.Session(config=sess_config)
model_io_config = Bunch({"fix_lm": False})
model_io_fn = model_io.ModelIO(model_io_config)
optimizer_fn = optimizer.Optimizer(opt_config)
num_classes = FLAGS.num_classes
model_train_fn = bert_classifier.classifier_model_fn_builder(
config,
num_classes,
init_checkpoint,
reuse=None,
load_pretrained=True,
model_io_fn=model_io_fn,
optimizer_fn=optimizer_fn,
model_io_config=model_io_config,
opt_config=opt_config)
model_eval_fn = bert_classifier.classifier_model_fn_builder(
config,
num_classes,
init_checkpoint,
reuse=True,
load_pretrained=True,
model_io_fn=model_io_fn,
optimizer_fn=optimizer_fn,
model_io_config=model_io_config,
opt_config=opt_config)
def metric_fn(features, logits, loss):
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.nn.softmax(logits)
accuracy = correct = tf.equal(
tf.cast(pred_label, tf.int32),
tf.cast(features["label_ids"], tf.int32))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
return {
"accuracy": accuracy,
"loss": loss,
"pred_label": pred_label,
"label_ids": features["label_ids"]
}
name_to_features = {
"input_ids": tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"input_mask": tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"segment_ids": tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"label_ids": tf.FixedLenFeature([], tf.int64),
}
def _decode_record(record, name_to_features):
"""Decodes a record to a TensorFlow example.
"""
example = tf.parse_single_example(record, name_to_features)
# tf.Example only supports tf.int64, but the TPU only supports tf.int32.
# So cast all int64 to int32.
for name in list(example.keys()):
t = example[name]
if t.dtype == tf.int64:
t = tf.to_int32(t)
example[name] = t
return example
params = Bunch({})
params.epoch = FLAGS.epoch
params.batch_size = FLAGS.batch_size
train_features = tf_data_utils.train_input_fn(FLAGS.train_file,
_decode_record,
name_to_features,
params,
if_shard=FLAGS.if_shard)
eval_features = tf_data_utils.eval_input_fn(FLAGS.dev_file,
_decode_record,
name_to_features,
params,
if_shard=FLAGS.if_shard)
[train_op, train_loss, train_per_example_loss,
train_logits] = model_train_fn(train_features, [],
tf.estimator.ModeKeys.TRAIN)
train_dict = {"train_op": train_op, "train_loss": train_loss}
[_, eval_loss, eval_per_example_loss,
eval_logits] = model_eval_fn(eval_features, [],
tf.estimator.ModeKeys.EVAL)
eval_dict = metric_fn(eval_features, eval_logits, eval_loss)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
sess.run(hvd.broadcast_global_variables(0))
model_io_fn.set_saver()
print("===horovod rank==={}".format(hvd.rank()))
def run_eval(steps):
import _pickle as pkl
# eval_features = tf_data_utils.eval_input_fn(
# FLAGS.dev_file,
# _decode_record,
# name_to_features, params)
# [_, eval_loss,
# eval_per_example_loss, eval_logits] = model_eval_fn(eval_features, [], tf.estimator.ModeKeys.EVAL)
# eval_dict = metric_fn(eval_features, eval_logits, eval_loss)
# sess.run(tf.local_variables_initializer())
eval_finial_dict = eval_fn(eval_dict)
if hvd.rank() == 0:
pkl.dump(
eval_finial_dict,
open(
FLAGS.model_output + "/eval_dict_{}.pkl".format(steps),
"wb"))
return eval_finial_dict
def eval_fn(result):
i = 0
total_accuracy = 0
eval_total_dict = {}
while True:
try:
eval_result = sess.run(result)
for key in eval_result:
if key not in eval_total_dict:
if key in ["pred_label", "label_ids"]:
eval_total_dict[key] = []
eval_total_dict[key].extend(eval_result[key])
if key in ["accuracy", "loss"]:
eval_total_dict[key] = 0.0
eval_total_dict[key] += eval_result[key]
else:
if key in ["pred_label", "label_ids"]:
eval_total_dict[key].extend(eval_result[key])
if key in ["accuracy", "loss"]:
eval_total_dict[key] += eval_result[key]
i += 1
# if i == 100:
# break
except tf.errors.OutOfRangeError:
print("End of dataset")
break
label_id = eval_total_dict["label_ids"]
pred_label = eval_total_dict["pred_label"]
result = classification_report(label_id,
pred_label,
target_names=list(
label_dict["label2id"].keys()))
print(result)
eval_total_dict["classification_report"] = result
return eval_total_dict
def train_fn(op_dict):
i = 0
cnt = 0
loss_dict = {}
monitoring_train = []
monitoring_eval = []
while True:
try:
train_result = sess.run(op_dict)
for key in train_result:
if key == "train_op":
continue
else:
if np.isnan(train_result[key]):
print(train_loss, "get nan loss")
break
else:
if key in loss_dict:
loss_dict[key] += train_result[key]
else:
loss_dict[key] = train_result[key]
i += 1
cnt += 1
if np.mod(i, num_storage_steps) == 0:
string = ""
for key in loss_dict:
tmp = key + " " + str(loss_dict[key] / cnt) + "\t"
string += tmp
print(string)
monitoring_train.append(loss_dict)
if hvd.rank() == 0:
model_io_fn.save_model(
sess,
FLAGS.model_output + "/model_{}.ckpt".format(
int(i / num_storage_steps)))
print("==successful storing model=={}".format(
int(i / num_storage_steps)))
cnt = 0
# eval_finial_dict = run_eval(int(i/num_storage_steps))
# monitoring_eval.append(eval_finial_dict)
for key in loss_dict:
loss_dict[key] = 0.0
except tf.errors.OutOfRangeError:
if hvd.rank() == 0:
import _pickle as pkl
pkl.dump(
{
"train": monitoring_train,
"eval": monitoring_eval
},
open(FLAGS.model_output + "/monitoring.pkl", "wb"))
break
print("===========begin to train============")
train_fn(train_dict)
if hvd.rank() == 0:
model_io_fn.save_model(sess, FLAGS.model_output + "/model.ckpt")
print("===========begin to eval============")
eval_finial_dict = run_eval("final")
def main(_):
hvd.init()
sess_config = tf.ConfigProto()
# sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
sess_config.gpu_options.visible_device_list = \
'%d,%d' % (hvd.local_rank() * 2, hvd.local_rank() * 2 + 1)
graph = tf.Graph()
with graph.as_default():
import json
config = json.load(
open("/data/xuht/bert/chinese_L-12_H-768_A-12/bert_config.json",
"r"))
init_checkpoint = "/data/xuht/bert/chinese_L-12_H-768_A-12/bert_model.ckpt"
config = Bunch(config)
config.use_one_hot_embeddings = True
config.scope = "bert"
config.dropout_prob = 0.1
config.label_type = "single_label"
config.loss = "focal_loss"
# config.num_hidden_layers =
# os.environ["CUDA_VISIBLE_DEVICES"] = "0"
num_train = int(33056 / hvd.size())
batch_size = 32
valid_step = int(num_train / batch_size)
epoch = 2
num_train_steps = int(num_train / (batch_size) * epoch)
decay_train_steps = num_train_steps
# decay_train_steps = int(
# 33056 / batch_size * epoch)
num_warmup_steps = int(num_train_steps * 0.01)
sess = tf.Session(config=sess_config)
opt_config = Bunch({
"init_lr": float(1e-5 / hvd.size()),
"num_train_steps": decay_train_steps,
"cycle": False,
"num_warmup_steps": num_warmup_steps,
"lr_decay": "polynomial_decay"
})
model_io_config = Bunch({"fix_lm": False})
model_io_fn = model_io.ModelIO(model_io_config)
optimizer_fn = optimizer.Optimizer(opt_config)
num_calsses = 2
model_train_fn = bert_classifier.classifier_model_fn_builder(
config,
num_calsses,
init_checkpoint,
reuse=None,
load_pretrained=True,
model_io_fn=model_io_fn,
optimizer_fn=optimizer_fn,
model_io_config=model_io_config,
opt_config=opt_config,
gpu_id=0,
gpu_nums=2)
# model_eval_fn = bert_classifier.classifier_model_fn_builder(config, num_calsses, init_checkpoint,
# reuse=True,
# load_pretrained=True,
# model_io_fn=model_io_fn,
# optimizer_fn=optimizer_fn,
# model_io_config=model_io_config,
# opt_config=opt_config,
# gpu_id=0,
# gpu_nums=2)
def metric_fn(features, logits, loss):
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.nn.softmax(logits)
accuracy = correct = tf.equal(
tf.cast(pred_label, tf.int32),
tf.cast(features["label_ids"], tf.int32))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
return {
"accuracy": accuracy,
"loss": loss,
"pred_label": pred_label,
"label_ids": features["label_ids"]
}
name_to_features = {
"input_ids": tf.FixedLenFeature([128], tf.int64),
"input_mask": tf.FixedLenFeature([128], tf.int64),
"segment_ids": tf.FixedLenFeature([128], tf.int64),
"label_ids": tf.FixedLenFeature([], tf.int64),
}
params = Bunch({})
params.epoch = epoch
params.batch_size = 32
train_file = "/data/xuht/eventy_detection/event/model/train.tfrecords"
train_file1 = "/data/xuht/eventy_detection/sentiment/model/sentiment_11_14/train.tfrecords"
title_sentiment = "/data/xuht/eventy_detection/sentiment/model/test/train.tfrecords"
sentiment = "/data/xuht/eventy_detection/sentiment/model/bert/train_11_15.tfrecords"
jd_train = "/data/xuht/jd_comment/train.tfrecords"
train_features = tf_data_utils.train_input_fn(
jd_train, tf_data_utils._decode_record, name_to_features, params)
test_file = [
"/data/xuht/eventy_detection/sentiment/model/sentiment_11_14/test.tfrecords"
]
test_file1_1 = [
"/data/xuht/eventy_detection/sentiment/model/test/train.tfrecords",
"/data/xuht/eventy_detection/sentiment/model/test/test.tfrecords"
]
test_file2 = "/data/xuht/eventy_detection/event/model/test.tfrecords"
title_test = "/data/xuht/eventy_detection/sentiment/model/test/test.tfrecords"
jd_test = "/data/xuht/jd_comment/test.tfrecords"
sentiment_test = "/data/xuht/eventy_detection/sentiment/model/bert/test_11_15.tfrecords"
eval_features = tf_data_utils.eval_input_fn(
jd_test, tf_data_utils._decode_record, name_to_features, params)
[train_op, train_loss, train_per_example_loss,
train_logits] = model_train_fn(train_features, [],
tf.estimator.ModeKeys.TRAIN)
# [_, eval_loss, eval_per_example_loss, eval_logits] = model_eval_fn(eval_features, [], tf.estimator.ModeKeys.EVAL)
# result = metric_fn(eval_features, eval_logits, eval_loss)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
sess.run(hvd.broadcast_global_variables(0))
model_io_fn.set_saver()
print("===horovod rank==={}".format(hvd.rank()))
def eval_fn(result):
i = 0
total_accuracy = 0
label, label_id = [], []
while True:
try:
eval_result = sess.run(result)
total_accuracy += eval_result["accuracy"]
label_id.extend(eval_result["label_ids"])
label.extend(eval_result["pred_label"])
i += 1
except tf.errors.OutOfRangeError:
print("End of dataset")
break
macro_f1 = f1_score(label_id, label, average="macro")
micro_f1 = f1_score(label_id, label, average="micro")
accuracy = accuracy_score(label_id, label)
print("test accuracy {} macro_f1 score {} micro_f1 {} accuracy {}".
format(total_accuracy / i, macro_f1, micro_f1, accuracy))
return total_accuracy / i, label_id, label
def train_fn(op, loss):
i = 0
total_loss = 0
cnt = 0
while True:
try:
[_, train_loss] = sess.run([op, loss])
i += 1
cnt += 1
total_loss += train_loss
# print("==device id {} global step {}".format(hvd.rank(), step))
if np.mod(i, valid_step) == 0:
print(total_loss / cnt)
cnt = 0
total_loss = 0
except tf.errors.OutOfRangeError:
print("End of dataset")
break
import time
start = time.time()
train_fn(train_op, train_loss)
# acc, true_label, pred_label = eval_fn(result)
end = time.time()
print("==total time {} numbers of devices {}".format(
end - start, hvd.size()))
def main(_):
hvd.init()
sess_config = tf.ConfigProto()
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
graph = tf.Graph()
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
with graph.as_default():
import json
# config = json.load(open("/data/xuht/bert/chinese_L-12_H-768_A-12/bert_config.json", "r"))
# init_checkpoint = "/data/xuht/bert/chinese_L-12_H-768_A-12/bert_model.ckpt"
config = json.load(open(FLAGS.config_file))
init_checkpoint = FLAGS.init_checkpoint
# init_checkpoint = "/data/xuht/ai_challenge_cqmrc/bert/concat/model/oqmrc.ckpt"
config = Bunch(config)
config.use_one_hot_embeddings = True
config.scope = "bert"
config.dropout_prob = 0.1
config.label_type = "single_label"
# os.environ["CUDA_VISIBLE_DEVICES"] = FLAGS.gpu_id
sess = tf.Session(config=sess_config)
train_size = int(FLAGS.train_size/hvd.size())
print("===train size===", train_size)
num_train_steps = int(
train_size / FLAGS.batch_size * FLAGS.epoch)
decay_train_steps = num_train_steps
# decay_train_steps = int(
# FLAGS.train_size / FLAGS.batch_size * FLAGS.epoch)
num_warmup_steps = int(num_train_steps * 0.01)
# num_warmup_steps = 0
num_storage_steps = int(train_size / FLAGS.batch_size)
init_lr = 1e-5
lr = float(init_lr/hvd.size())
# lr = init_lr
# end_lr = (hvd.size()-1)/(np.power(hvd.size(), 2)) * init_lr
end_lr = 0.0
opt_config = Bunch({"init_lr":lr,
"num_train_steps":decay_train_steps,
"num_warmup_steps":num_warmup_steps,
"end_learning_rate":end_lr})
model_io_config = Bunch({"fix_lm":False})
model_io_fn = model_io.ModelIO(model_io_config)
num_choice = FLAGS.num_classes
max_seq_length = FLAGS.max_length
model_train_fn = bert_classifier.multichoice_model_fn_builder(config, num_choice, init_checkpoint,
reuse=None,
load_pretrained=True,
model_io_fn=model_io_fn,
model_io_config=model_io_config,
opt_config=opt_config)
model_eval_fn = bert_classifier.multichoice_model_fn_builder(config, num_choice, init_checkpoint,
reuse=True,
load_pretrained=True,
model_io_fn=model_io_fn,
model_io_config=model_io_config,
opt_config=opt_config)
def metric_fn(features, logits, loss):
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.nn.softmax(logits)
accuracy = correct = tf.equal(
tf.cast(pred_label, tf.int32),
tf.cast(features["label_ids"], tf.int32)
)
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
return {"accuracy":accuracy, "loss":loss, "pred_label":pred_label, "label_ids":features["label_ids"]}
name_to_features = {
"input_ids":
tf.FixedLenFeature([max_seq_length*num_choice], tf.int64),
"input_mask":
tf.FixedLenFeature([max_seq_length*num_choice], tf.int64),
"segment_ids":
tf.FixedLenFeature([max_seq_length*num_choice], tf.int64),
"label_ids":
tf.FixedLenFeature([], tf.int64),
}
def _decode_record(record, name_to_features):
"""Decodes a record to a TensorFlow example.
"""
example = tf.parse_single_example(record, name_to_features)
# tf.Example only supports tf.int64, but the TPU only supports tf.int32.
# So cast all int64 to int32.
for name in list(example.keys()):
t = example[name]
if t.dtype == tf.int64:
t = tf.to_int32(t)
example[name] = t
for name in ["input_ids", "input_mask", "segment_ids"]:
example[name] = tf.reshape(example[name], [-1, max_seq_length])
return example
params = Bunch({})
params.epoch = FLAGS.epoch
params.batch_size = FLAGS.batch_size
train_features = tf_data_utils.train_input_fn(FLAGS.train_file,
_decode_record, name_to_features, params)
eval_features = tf_data_utils.eval_input_fn(FLAGS.dev_file,
_decode_record, name_to_features, params)
[train_op, train_loss, train_per_example_loss, train_logits] = model_train_fn(train_features, [], tf.estimator.ModeKeys.TRAIN)
[_, eval_loss, eval_per_example_loss, eval_logits] = model_eval_fn(eval_features, [], tf.estimator.ModeKeys.EVAL)
result = metric_fn(eval_features, eval_logits, eval_loss)
model_io_fn.set_saver()
print("====succeeded in set savering====", hvd.rank())
init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
sess.run(init_op)
print("====succeeded in initializing global and local variables====", hvd.rank())
sess.run(hvd.broadcast_global_variables(0))
print("====succeeded in initializing params====", hvd.rank())
def eval_fn(result):
i = 0
total_accuracy = 0
label, label_id = [], []
while True:
try:
eval_result = sess.run(result)
total_accuracy += eval_result["accuracy"]
label_id.extend(eval_result["label_ids"])
label.extend(eval_result["pred_label"])
i += 1
except tf.errors.OutOfRangeError:
print("End of dataset")
break
f1 = f1_score(label_id, label, average="macro")
accuracy = accuracy_score(label_id, label)
print("test accuracy accuracy {} {}, f1 {}".format(total_accuracy/i,
accuracy, f1))
return total_accuracy/ i, f1
output_dict = []
fwobj = open(FLAGS.model_output+"/log.txt", "w")
def train_fn(op, loss):
i = 0
cnt = 0
total_loss = 0.0
while True:
try:
[_, train_loss] = sess.run([op, loss])
# print(train_loss, i)
total_loss += train_loss
i += 1
cnt += 1
if np.mod(i, num_storage_steps) == 0:
if hvd.rank() == 0:
fwobj.write("loss {} rank {}\n".format(total_loss/cnt, hvd.rank()))
model_io_fn.save_model(sess, FLAGS.model_output+"/oqmrc_epoch_{}_hvd_{}.ckpt".format(int(i/num_storage_steps), hvd.rank()))
total_loss = 0
cnt = 0
except tf.errors.OutOfRangeError:
break
if hvd.rank() == 0:
model_io_fn.save_model(sess, FLAGS.model_output+"/oqmrc_hvd_{}_initial.ckpt".format(hvd.rank()))
print("===========begin to train============")
train_fn(train_op, train_loss)
if hvd.rank() == 0:
print("===========begin to eval============")
eval_fn(result)
model_io_fn.save_model(sess, FLAGS.model_output+"/oqmrc_hvd_{}.ckpt".format(hvd.rank()))
fwobj.close()
def main(_):
hvd.init()
sess_config = tf.ConfigProto()
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
graph = tf.Graph()
with graph.as_default():
import json
config = json.load(open(FLAGS.config_file, "r"))
init_checkpoint = FLAGS.init_checkpoint
config = Bunch(config)
config.use_one_hot_embeddings = True
config.scope = "bert"
config.dropout_prob = 0.1
config.label_type = "single_label"
if FLAGS.if_shard == "0":
train_size = FLAGS.train_size
epoch = int(FLAGS.epoch / hvd.size())
elif FLAGS.if_shard == "1":
train_size = int(FLAGS.train_size / hvd.size())
epoch = FLAGS.epoch
init_lr = 2e-5
num_train_steps = int(train_size / FLAGS.batch_size * epoch)
num_warmup_steps = int(num_train_steps * 0.1)
num_storage_steps = int(train_size / FLAGS.batch_size)
print(" model type {}".format(FLAGS.model_type))
print(num_train_steps, num_warmup_steps, "=============")
opt_config = Bunch({
"init_lr": init_lr / hvd.size(),
"num_train_steps": num_train_steps,
"num_warmup_steps": num_warmup_steps
})
sess = tf.Session(config=sess_config)
model_io_config = Bunch({"fix_lm": False})
model_io_fn = model_io.ModelIO(model_io_config)
optimizer_fn = optimizer.Optimizer(opt_config)
num_classes = FLAGS.num_classes
model_train_fn = bert_classifier.classifier_model_fn_builder(
config,
num_classes,
init_checkpoint,
reuse=None,
load_pretrained=True,
model_io_fn=model_io_fn,
optimizer_fn=optimizer_fn,
model_io_config=model_io_config,
opt_config=opt_config)
model_eval_fn = bert_classifier.classifier_model_fn_builder(
config,
num_classes,
init_checkpoint,
reuse=True,
load_pretrained=True,
model_io_fn=model_io_fn,
optimizer_fn=optimizer_fn,
model_io_config=model_io_config,
opt_config=opt_config)
def metric_fn(features, logits, loss):
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.nn.softmax(logits)
accuracy = correct = tf.equal(
tf.cast(pred_label, tf.int32),
tf.cast(features["label_ids"], tf.int32))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
return {
"accuracy": accuracy,
"loss": loss,
"pred_label": pred_label,
"label_ids": features["label_ids"]
}
name_to_features = {
"input_ids": tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"input_mask": tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"segment_ids": tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"label_ids": tf.FixedLenFeature([], tf.int64),
}
def _decode_record(record, name_to_features):
"""Decodes a record to a TensorFlow example.
"""
example = tf.parse_single_example(record, name_to_features)
# tf.Example only supports tf.int64, but the TPU only supports tf.int32.
# So cast all int64 to int32.
for name in list(example.keys()):
t = example[name]
if t.dtype == tf.int64:
t = tf.to_int32(t)
example[name] = t
return example
params = Bunch({})
params.epoch = FLAGS.epoch
params.batch_size = FLAGS.batch_size
train_features = tf_data_utils.train_input_fn(FLAGS.train_file,
_decode_record,
name_to_features,
params,
if_shard=FLAGS.if_shard)
eval_features = tf_data_utils.eval_input_fn(FLAGS.dev_file,
_decode_record,
name_to_features,
params,
if_shard=FLAGS.if_shard)
[train_op, train_loss, train_per_example_loss,
train_logits] = model_train_fn(train_features, [],
tf.estimator.ModeKeys.TRAIN)
[_, eval_loss, eval_per_example_loss,
eval_logits] = model_eval_fn(eval_features, [],
tf.estimator.ModeKeys.EVAL)
result = metric_fn(eval_features, eval_logits, eval_loss)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
sess.run(hvd.broadcast_global_variables(0))
model_io_fn.set_saver()
print("===horovod rank==={}".format(hvd.rank()))
def eval_fn(result):
i = 0
total_accuracy = 0
label, label_id = [], []
while True:
try:
eval_result = sess.run(result)
total_accuracy += eval_result["accuracy"]
label_id.extend(eval_result["label_ids"])
label.extend(eval_result["pred_label"])
i += 1
except tf.errors.OutOfRangeError:
print("End of dataset")
break
macro_f1 = f1_score(label_id, label, average="macro")
micro_f1 = f1_score(label_id, label, average="micro")
macro_precision = precision_score(label_id, label, average="macro")
micro_precision = precision_score(label_id, label, average="micro")
macro_recall = recall_score(label_id, label, average="macro")
micro_recall = recall_score(label_id, label, average="micro")
accuracy = accuracy_score(label_id, label)
print("test accuracy {} macro_f1 score {} micro_f1 {} accuracy {}".
format(total_accuracy / i, macro_f1, micro_f1, accuracy))
return total_accuracy / i, label_id, label
def train_fn(op, loss):
i = 0
total_loss = 0
cnt = 0
while True:
try:
[_, train_loss] = sess.run([op, loss])
i += 1
cnt += 1
total_loss += train_loss
# print("==device id {} global step {}".format(hvd.rank(), step))
if np.mod(i, num_storage_steps) == 0:
print(total_loss / cnt)
if hvd.rank() == 0:
model_io_fn.save_model(
sess,
FLAGS.model_output + "/oqmrc_{}.ckpt".format(
int(i / num_storage_steps)))
cnt = 0
total_loss = 0
except tf.errors.OutOfRangeError:
print("End of dataset")
break
import time
import time
start = time.time()
train_fn(train_op, train_loss)
acc, true_label, pred_label = eval_fn(result)
end = time.time()
print("==total time {} numbers of devices {}".format(
end - start, hvd.size()))
if hvd.rank() == 0:
model_io_fn.save_model(sess, FLAGS.model_output + "/oqmrc.ckpt")
import _pickle as pkl
pkl.dump({
"true_label": true_label,
"pred_label": pred_label
}, open(FLAGS.model_output + "/eval_result.json", "wb"))
