def get_test_example():
processor = processors[task_name]()
label_list = processor.get_labels()
tokenizer = tokenization.FullTokenizer(vocab_file=vocab_file,
spm_model_file=spm_model_file)
# save data to tf_record
#test_examples = processor.get_test_examples("")#测试数据目录
test_examples = processor.get_test_examples(data_dir)
features = get_test_features(test_examples, label_list, max_seq_length,
tokenizer)
return features
def create_tokenizer_from_hub_module(albert_hub_module_handle):
"""Get the vocab file and casing info from the Hub module."""
with tf.Graph().as_default():
albert_module = hub.Module(albert_hub_module_handle)
tokenization_info = albert_module(signature="tokenization_info",
as_dict=True)
with tf.Session() as sess:
vocab_file, do_lower_case = sess.run([
tokenization_info["vocab_file"],
tokenization_info["do_lower_case"]
])
return tokenization.FullTokenizer(vocab_file=vocab_file,
do_lower_case=do_lower_case,
spm_model_file=FLAGS.spm_model_file)
def transformer(path, s3_path, class_name, model='xlnet', **kwargs):
check_file(path[model], s3_path[model], **kwargs)
g = load_graph(path[model]['model'], **kwargs)
try:
with open(path[model]['setting']) as fopen:
nodes = json.load(fopen)
except:
raise Exception(
f"model corrupted due to some reasons, please run malaya.clear_cache('{class_name}/{model}/{size}') and try again"
)
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
tokenizer = sentencepiece_tokenizer_bert(path[model]['tokenizer'],
path[model]['vocab'])
if model in ['albert', 'tiny-albert']:
from albert import tokenization
tokenizer = tokenization.FullTokenizer(
vocab_file=path[model]['vocab'],
do_lower_case=False,
spm_model_file=path[model]['tokenizer'],
)
return TAGGING_BERT(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=None,
input_masks=g.get_tensor_by_name('import/Placeholder_1:0'),
logits=g.get_tensor_by_name('import/logits:0'),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
settings=nodes,
)
if model in ['xlnet', 'alxlnet']:
tokenizer = sentencepiece_tokenizer_xlnet(path[model]['tokenizer'])
return TAGGING_XLNET(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=g.get_tensor_by_name('import/Placeholder_1:0'),
input_masks=g.get_tensor_by_name('import/Placeholder_2:0'),
logits=g.get_tensor_by_name('import/logits:0'),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
settings=nodes,
)
def load(model: str = 'albert', **kwargs):
"""
Load albert model.
Parameters
----------
model : str, optional (default='base')
Model architecture supported. Allowed values:
* ``'albert'`` - base albert-bahasa released by Malaya.
* ``'albert-tiny'`` - tiny bert-bahasa released by Malaya.
Returns
-------
result : malaya.transformers.albert.Model class
"""
from malaya.path import PATH_ALBERT, S3_PATH_ALBERT
from malaya.function import check_file
model = model.lower()
check_file(PATH_ALBERT[model]['model'], S3_PATH_ALBERT[model], **kwargs)
if not os.path.exists(PATH_ALBERT[model]['directory'] + 'model.ckpt'):
import tarfile
with tarfile.open(PATH_ALBERT[model]['model']['model']) as tar:
tar.extractall(path=PATH_ALBERT[model]['path'])
from albert import tokenization
bert_checkpoint = PATH_ALBERT[model]['directory'] + 'model.ckpt'
vocab_model = PATH_ALBERT[model]['directory'] + 'sp10m.cased.v10.model'
vocab = PATH_ALBERT[model]['directory'] + 'sp10m.cased.v10.vocab'
bert_config = PATH_ALBERT[model]['directory'] + 'config.json'
tokenizer = tokenization.FullTokenizer(vocab_file=vocab,
do_lower_case=False,
spm_model_file=vocab_model)
bert_config = modeling.AlbertConfig.from_json_file(bert_config)
model = Model(bert_config, tokenizer)
model._saver.restore(model._sess, bert_checkpoint)
return model
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
tokenizer = tokenization.FullTokenizer(
vocab_file=FLAGS.vocab_file,
do_lower_case=FLAGS.do_lower_case,
spm_model_file=FLAGS.spm_model_file,
)
input_files = []
for input_pattern in FLAGS.input_file.split(","):
input_files.extend(tf.gfile.Glob(input_pattern))
tf.logging.info("*** Reading from input files ***")
for input_file in input_files:
tf.logging.info(" %s", input_file)
rng = random.Random(FLAGS.random_seed)
instances = create_training_instances(
input_files,
tokenizer,
FLAGS.max_seq_length,
FLAGS.dupe_factor,
FLAGS.short_seq_prob,
FLAGS.masked_lm_prob,
FLAGS.max_predictions_per_seq,
rng,
)
tf.logging.info("number of instances: %i", len(instances))
output_files = FLAGS.output_file.split(",")
tf.logging.info("*** Writing to output files ***")
for output_file in output_files:
tf.logging.info(" %s", output_file)
write_instance_to_example_files(
instances,
tokenizer,
FLAGS.max_seq_length,
FLAGS.max_predictions_per_seq,
output_files,
)
def sentence_to_idx(self, text):
"""
将分词后的句子转换成idx表示
:return:
"""
tokenizer = tokenization.FullTokenizer(vocab_file=self.vocab_path,
do_lower_case=True)
text = tokenization.convert_to_unicode(text)
tokens = tokenizer.tokenize(text)
tokens = ["[CLS]"] + tokens + ["[SEP]"]
input_id = tokenizer.convert_tokens_to_ids(tokens)
input_mask = [1] * len(input_id)
segment_id = [0] * len(input_id)
input_id, input_mask, segment_id = self.padding(
input_id, input_mask, segment_id)
return [input_id], [input_mask], [segment_id]
def trans_to_index(self, inputs):
"""
将输入转化为索引表示
:param inputs: 输入
:return:
"""
tokenizer = tokenization.FullTokenizer(vocab_file=self.__vocab_path,
do_lower_case=True)
input_ids = []
input_masks = []
segment_ids = []
for text in inputs:
text = tokenization.convert_to_unicode(text)
tokens = tokenizer.tokenize(text)
tokens = ["[CLS]"] + tokens + ["[SEP]"]
input_id = tokenizer.convert_tokens_to_ids(tokens)
input_ids.append(input_id)
input_masks.append([1] * len(input_id))
segment_ids.append([0] * len(input_id))
return input_ids, input_masks, segment_ids
def test_full_tokenizer(self):
vocab_tokens = [
"[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn",
"##ing", ","
]
with tempfile.NamedTemporaryFile(delete=False) as vocab_writer:
if six.PY2:
vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
else:
contents = "".join([six.ensure_str(x) + "\n" for x in vocab_tokens])
vocab_writer.write(six.ensure_binary(contents, "utf-8"))
vocab_file = vocab_writer.name
tokenizer = tokenization.FullTokenizer(vocab_file)
os.unlink(vocab_file)
tokens = tokenizer.tokenize(u"UNwant\u00E9d,running")
self.assertAllEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"])
self.assertAllEqual(
tokenizer.convert_tokens_to_ids(tokens), [7, 4, 5, 10, 8, 9])
def make_tf_record(output_dir, data_dir, vocab_file, spm_model_file):
tf.gfile.MakeDirs(output_dir) #"model/bert"
processor = processors[task_name]() #"atec"
label_list = processor.get_labels()
tokenizer = tokenization.FullTokenizer(vocab_file=vocab_file,
spm_model_file=spm_model_file)
train_file = os.path.join(output_dir, "train.tf_record")
eval_file = os.path.join(output_dir, "eval.tf_record")
# save data to tf_record
if not os.path.isfile(train_file):
train_examples = processor.get_train_examples(data_dir)
file_based_convert_examples_to_features(train_examples, label_list,
max_seq_length, tokenizer,
train_file, task_name)
del train_examples
# eval data
if not os.path.isfile(eval_file):
eval_examples = processor.get_dev_examples(data_dir)
file_based_convert_examples_to_features(eval_examples, label_list,
max_seq_length, tokenizer,
eval_file, task_name)
del eval_examples
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
albert_config = modeling.AlbertConfig.from_json_file(FLAGS.albert_config_file)
validate_flags_or_throw(albert_config)
tf.gfile.MakeDirs(FLAGS.output_dir)
tokenizer = tokenization.FullTokenizer(
vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case,
spm_model_file=FLAGS.spm_model_file)
tpu_cluster_resolver = None
if FLAGS.use_tpu and FLAGS.tpu_name:
tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver(
FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2
if FLAGS.do_train:
iterations_per_loop = int(min(FLAGS.iterations_per_loop,
FLAGS.save_checkpoints_steps))
else:
iterations_per_loop = FLAGS.iterations_per_loop
run_config = contrib_tpu.RunConfig(
cluster=tpu_cluster_resolver,
master=FLAGS.master,
model_dir=FLAGS.output_dir,
save_checkpoints_steps=FLAGS.save_checkpoints_steps,
tpu_config=contrib_tpu.TPUConfig(
iterations_per_loop=iterations_per_loop,
num_shards=FLAGS.num_tpu_cores,
per_host_input_for_training=is_per_host))
train_examples = None
num_train_steps = None
num_warmup_steps = None
train_examples = squad_utils.read_squad_examples(
input_file=FLAGS.train_file, is_training=True)
num_train_steps = int(
len(train_examples) / FLAGS.train_batch_size * FLAGS.num_train_epochs)
if FLAGS.do_train:
num_warmup_steps = int(num_train_steps * FLAGS.warmup_proportion)
# Pre-shuffle the input to avoid having to make a very large shuffle
# buffer in in the `input_fn`.
rng = random.Random(12345)
rng.shuffle(train_examples)
model_fn = squad_utils.v1_model_fn_builder(
albert_config=albert_config,
init_checkpoint=FLAGS.init_checkpoint,
learning_rate=FLAGS.learning_rate,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps,
use_tpu=FLAGS.use_tpu,
use_one_hot_embeddings=FLAGS.use_tpu)
# If TPU is not available, this will fall back to normal Estimator on CPU
# or GPU.
estimator = contrib_tpu.TPUEstimator(
use_tpu=FLAGS.use_tpu,
model_fn=model_fn,
config=run_config,
train_batch_size=FLAGS.train_batch_size,
predict_batch_size=FLAGS.predict_batch_size)
if FLAGS.do_train:
# We write to a temporary file to avoid storing very large constant tensors
# in memory.
if not tf.gfile.Exists(FLAGS.train_feature_file):
train_writer = squad_utils.FeatureWriter(
filename=os.path.join(FLAGS.train_feature_file), is_training=True)
squad_utils.convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=FLAGS.max_seq_length,
doc_stride=FLAGS.doc_stride,
max_query_length=FLAGS.max_query_length,
is_training=True,
output_fn=train_writer.process_feature,
do_lower_case=FLAGS.do_lower_case)
train_writer.close()
tf.logging.info("***** Running training *****")
tf.logging.info(" Num orig examples = %d", len(train_examples))
# tf.logging.info(" Num split examples = %d", train_writer.num_features)
tf.logging.info(" Batch size = %d", FLAGS.train_batch_size)
tf.logging.info(" Num steps = %d", num_train_steps)
del train_examples
train_input_fn = squad_utils.input_fn_builder(
input_file=FLAGS.train_feature_file,
seq_length=FLAGS.max_seq_length,
is_training=True,
drop_remainder=True,
use_tpu=FLAGS.use_tpu,
bsz=FLAGS.train_batch_size,
is_v2=False)
estimator.train(input_fn=train_input_fn, max_steps=num_train_steps)
if FLAGS.do_predict:
with tf.gfile.Open(FLAGS.predict_file) as predict_file:
prediction_json = json.load(predict_file)["data"]
eval_examples = squad_utils.read_squad_examples(
input_file=FLAGS.predict_file, is_training=False)
if (tf.gfile.Exists(FLAGS.predict_feature_file) and tf.gfile.Exists(
FLAGS.predict_feature_left_file)):
tf.logging.info("Loading eval features from {}".format(
FLAGS.predict_feature_left_file))
with tf.gfile.Open(FLAGS.predict_feature_left_file, "rb") as fin:
eval_features = pickle.load(fin)
else:
eval_writer = squad_utils.FeatureWriter(
filename=FLAGS.predict_feature_file, is_training=False)
eval_features = []
def append_feature(feature):
eval_features.append(feature)
eval_writer.process_feature(feature)
squad_utils.convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=FLAGS.max_seq_length,
doc_stride=FLAGS.doc_stride,
max_query_length=FLAGS.max_query_length,
is_training=False,
output_fn=append_feature,
do_lower_case=FLAGS.do_lower_case)
eval_writer.close()
with tf.gfile.Open(FLAGS.predict_feature_left_file, "wb") as fout:
pickle.dump(eval_features, fout)
tf.logging.info("***** Running predictions *****")
tf.logging.info(" Num orig examples = %d", len(eval_examples))
tf.logging.info(" Num split examples = %d", len(eval_features))
tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size)
predict_input_fn = squad_utils.input_fn_builder(
input_file=FLAGS.predict_feature_file,
seq_length=FLAGS.max_seq_length,
is_training=False,
drop_remainder=False,
use_tpu=FLAGS.use_tpu,
bsz=FLAGS.predict_batch_size,
is_v2=False)
def get_result(checkpoint):
"""Evaluate the checkpoint on SQuAD 1.0."""
# If running eval on the TPU, you will need to specify the number of
# steps.
reader = tf.train.NewCheckpointReader(checkpoint)
global_step = reader.get_tensor(tf.GraphKeys.GLOBAL_STEP)
all_results = []
for result in estimator.predict(
predict_input_fn, yield_single_examples=True,
checkpoint_path=checkpoint):
if len(all_results) % 1000 == 0:
tf.logging.info("Processing example: %d" % (len(all_results)))
unique_id = int(result["unique_ids"])
start_log_prob = [float(x) for x in result["start_log_prob"].flat]
end_log_prob = [float(x) for x in result["end_log_prob"].flat]
all_results.append(
squad_utils.RawResult(
unique_id=unique_id,
start_log_prob=start_log_prob,
end_log_prob=end_log_prob))
output_prediction_file = os.path.join(
FLAGS.output_dir, "predictions.json")
output_nbest_file = os.path.join(
FLAGS.output_dir, "nbest_predictions.json")
result_dict = {}
squad_utils.accumulate_predictions_v1(
result_dict, eval_examples, eval_features,
all_results, FLAGS.n_best_size, FLAGS.max_answer_length)
predictions = squad_utils.write_predictions_v1(
result_dict, eval_examples, eval_features, all_results,
FLAGS.n_best_size, FLAGS.max_answer_length,
output_prediction_file, output_nbest_file)
return squad_utils.evaluate_v1(
prediction_json, predictions), int(global_step)
def _find_valid_cands(curr_step):
filenames = tf.gfile.ListDirectory(FLAGS.output_dir)
candidates = []
for filename in filenames:
if filename.endswith(".index"):
ckpt_name = filename[:-6]
idx = ckpt_name.split("-")[-1]
if idx != "best" and int(idx) > curr_step:
candidates.append(filename)
return candidates
output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt")
checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best")
key_name = "f1"
writer = tf.gfile.GFile(output_eval_file, "w")
if tf.gfile.Exists(checkpoint_path + ".index"):
result = get_result(checkpoint_path)
best_perf = result[0][key_name]
global_step = result[1]
else:
global_step = -1
best_perf = -1
checkpoint_path = None
while global_step < num_train_steps:
steps_and_files = {}
filenames = tf.gfile.ListDirectory(FLAGS.output_dir)
for filename in filenames:
if filename.endswith(".index"):
ckpt_name = filename[:-6]
cur_filename = os.path.join(FLAGS.output_dir, ckpt_name)
if cur_filename.split("-")[-1] == "best":
continue
gstep = int(cur_filename.split("-")[-1])
if gstep not in steps_and_files:
tf.logging.info("Add {} to eval list.".format(cur_filename))
steps_and_files[gstep] = cur_filename
tf.logging.info("found {} files.".format(len(steps_and_files)))
if not steps_and_files:
tf.logging.info("found 0 file, global step: {}. Sleeping."
.format(global_step))
time.sleep(1)
else:
for ele in sorted(steps_and_files.items()):
step, checkpoint_path = ele
if global_step >= step:
if len(_find_valid_cands(step)) > 1:
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tf.logging.info("removing {}".format(src_ckpt))
tf.gfile.Remove(src_ckpt)
continue
result, global_step = get_result(checkpoint_path)
tf.logging.info("***** Eval results *****")
for key in sorted(result.keys()):
tf.logging.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
if result[key_name] > best_perf:
best_perf = result[key_name]
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tgt_ckpt = checkpoint_path.rsplit(
"-", 1)[0] + "-best.{}".format(ext)
tf.logging.info("saving {} to {}".format(src_ckpt, tgt_ckpt))
tf.gfile.Copy(src_ckpt, tgt_ckpt, overwrite=True)
writer.write("saved {} to {}\n".format(src_ckpt, tgt_ckpt))
writer.write("best {} = {}\n".format(key_name, best_perf))
tf.logging.info(" best {} = {}\n".format(key_name, best_perf))
if len(_find_valid_cands(global_step)) > 2:
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tf.logging.info("removing {}".format(src_ckpt))
tf.gfile.Remove(src_ckpt)
writer.write("=" * 50 + "\n")
checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best")
result, global_step = get_result(checkpoint_path)
tf.logging.info("***** Final Eval results *****\n")
writer.write("===== Hyperparameters =====\n")
writer.write("Training batch size: {}\n".format(FLAGS.train_batch_size))
writer.write("Max sequence length: {}\n".format(FLAGS.max_seq_length))
writer.write("Learning rate: {}\n".format(FLAGS.learning_rate))
if num_train_steps and num_warmup_steps:
writer.write("Training steps: {}\n".format(num_train_steps))
writer.write("Warmup steps: {}\n".format(num_warmup_steps))
writer.write("===== Evuations =====\n")
for key in sorted(result.keys()):
tf.logging.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
writer.write("best perf happened at step: {}".format(global_step))
# Author: dgm
# Description: 数据预处理
# Date: 2020-08-14
import math
import codecs
import random
from albert import tokenization
from utils import create_dico, create_mapping, zero_digits
tokenizer = tokenization.FullTokenizer(
vocab_file='albert_model/albert_base/vocab_chinese.txt',
do_lower_case=True)
def load_sentences(path, lower, zeros):
"""
Load sentences. A line must contain at least a word and its tag.
Sentences are separated by empty lines.
"""
sentences = []
sentence = []
num = 0
for line in codecs.open(path, 'r', 'utf8'):
num += 1
line = zero_digits(line.rstrip()) if zeros else line.rstrip()
if not line:
if len(sentence) > 0:
if 'DOCSTART' not in sentence[0][0]:
sentences.append(sentence)
sentence = []
def trans_to_features(self, examples, is_training):
"""
将输入转化为索引表示
:param examples: 输入
:param is_training:
:return:
"""
tokenizer = tokenization.FullTokenizer(vocab_file=self.__vocab_path,
do_lower_case=True)
features = []
unique_id = 1000000000
for (example_index, example) in enumerate(examples):
# 用wordpiece的方法对query进行分词处理
query_tokens = tokenizer.tokenize(example['question'])
# 给定query一个最大长度来控制query的长度
if len(query_tokens) > self.__query_length:
query_tokens = query_tokens[:self.__query_length]
# 主要是针对context构造索引，之前我们将中文，标点符号，空格，一连串的数字，英文单词分割存储在doc_tokens中
# 但在bert的分词器中会将一连串的数字，中文，英文等分割成子词，也就是说经过bert的分词之后得到的tokens和之前
# 获得的doc_tokens是不一样的，因此我们仍需要对start和end position从doc_tokens中的位置映射到当前tokens的位置
tok_to_orig_index = [] # 存储未分词的token的索引，但长度和下面的相等
orig_to_tok_index = [] # 存储分词后的token的索引，但索引不是连续的，会存在跳跃的情况
all_doc_tokens = [] # 存储分词后的token，理论上长度是要大于all_tokens的
for (i, token) in enumerate(example['doc_tokens']):
sub_tokens = tokenizer.tokenize(token)
# orig_to_tok_index的长度等于doc_tokens，里面每个值存储的是doc_tokens中的token在all_doc_tokens中的起止索引值
# 用来将在all_token中的start和end转移到all_doc_tokens中
orig_to_tok_index.append([len(all_doc_tokens)])
for sub_token in sub_tokens:
# tok_to_orig_index的长度等于all_doc_tokens, 里面会有重复的值
tok_to_orig_index.append(i)
all_doc_tokens.append(sub_token)
orig_to_tok_index[-1].append(len(all_doc_tokens) - 1)
tok_start_position = -1
tok_end_position = -1
if is_training:
# 原来token到新token的映射，这是新token的起点
tok_start_position = orig_to_tok_index[
example['start_position']][0]
tok_end_position = orig_to_tok_index[
example['end_position']][1]
tok_start_position, tok_end_position = self._improve_answer_span(
all_doc_tokens, tok_start_position, tok_end_position,
tokenizer, example['orig_answer_text'])
# The -3 accounts for [CLS], [SEP] and [SEP]
max_tokens_for_doc = self.__max_length - len(query_tokens) - 3
doc_spans = []
_DocSpan = collections.namedtuple("DocSpan", ["start", "length"])
# 在使用bert的时候，一般会将最大的序列长度控制在512，因此对于长度大于最大长度的context，我们需要将其分成多个片段
# 采用滑窗的方式，滑窗大小是小于最大长度的，因此分割的片段之间是存在重复的子片段。
start_offset = 0 # 截取的片段的起始位置
while start_offset < len(all_doc_tokens):
length = len(all_doc_tokens) - start_offset
# 当长度超标，需要使用滑窗
if length > max_tokens_for_doc:
length = max_tokens_for_doc
doc_spans.append(_DocSpan(start=start_offset, length=length))
if start_offset + length == len(
all_doc_tokens): # 当length < max_len时，该条件成立
break
start_offset += min(length, self.__doc_stride)
# 组合query和context的片段成一个序列输入到bert中
for (doc_span_index, doc_span) in enumerate(doc_spans):
tokens = []
token_to_orig_map = {}
# 因为片段之间会存在重复的子片段，但是子片段中的token在不同的片段中的重要性是不一样的，
# 在这里根据上下文的数量来决定token的重要性，在之后预测时对于出现在两个片段中的token，只取重要性高的片段
# 中的token的分数作为该token的分数
token_is_max_context = {}
segment_ids = []
tokens.append("[CLS]")
segment_ids.append(0)
for token in query_tokens:
tokens.append(token)
segment_ids.append(0)
tokens.append("[SEP]")
segment_ids.append(0)
for i in range(doc_span.length):
split_token_index = doc_span.start + i
token_to_orig_map[len(tokens)] = tok_to_orig_index[
split_token_index] # 映射当前span组成的句子对的索引到原始token的索引
# 在利用滑窗分割多个span时会存在有的词出现在两个span中，但最后统计的时候，我们只能选择一个span，因此
# 作者根据该词上下文词的数量构建了一个分数，取分数最高的那个span
is_max_context = self._check_is_max_context(
doc_spans, doc_span_index, split_token_index)
token_is_max_context[len(tokens)] = is_max_context
tokens.append(all_doc_tokens[split_token_index])
segment_ids.append(1)
tokens.append("[SEP]")
segment_ids.append(1)
input_ids = tokenizer.convert_tokens_to_ids(tokens)
# The mask has 1 for real tokens and 0 for padding tokens. Only real
# tokens are attended to.
input_mask = [1] * len(input_ids)
# Zero-pad up to the sequence length.
while len(input_ids) < self.__max_length:
input_ids.append(0)
input_mask.append(0)
segment_ids.append(0)
assert len(input_ids) == self.__max_length
assert len(input_mask) == self.__max_length
assert len(segment_ids) == self.__max_length
start_position = -1
end_position = -1
if is_training:
# For training, if our document chunk does not contain an annotation
# we throw it out, since there is nothing to predict.
if tok_start_position == -1 and tok_end_position == -1:
start_position = 0 # 问题本来没答案，0是[CLS]的位子
end_position = 0
else: # 如果原本是有答案的，那么去除没有答案的feature
out_of_span = False
doc_start = doc_span.start # 映射回原文的起点和终点
doc_end = doc_span.start + doc_span.length - 1
if not (tok_start_position >= doc_start and
tok_end_position <= doc_end): # 该划窗没答案作为无答案增强
out_of_span = True
if out_of_span:
start_position = 0
end_position = 0
else:
doc_offset = len(query_tokens) + 2
start_position = tok_start_position - doc_start + doc_offset
end_position = tok_end_position - doc_start + doc_offset
features.append({
'unique_id': unique_id,
'example_index': example_index,
'doc_span_index': doc_span_index,
'tokens': tokens,
'token_to_orig_map': token_to_orig_map,
'token_is_max_context': token_is_max_context,
'input_ids': input_ids,
'input_mask': input_mask,
'segment_ids': segment_ids,
'start_position': start_position,
'end_position': end_position
})
unique_id += 1
return features
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"xnli": XnliProcessor,
}
tokenization.validate_case_matches_checkpoint(FLAGS.do_lower_case,
FLAGS.init_checkpoint)
if not FLAGS.do_train and not FLAGS.do_eval and not FLAGS.do_predict:
raise ValueError(
"At least one of `do_train`, `do_eval` or `do_predict' must be True."
)
albert_config = modeling.AlbertConfig.from_json_file(
FLAGS.albert_config_file)
if FLAGS.max_seq_length > albert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length %d because the ALBERT model "
"was only trained up to sequence length %d" %
(FLAGS.max_seq_length, albert_config.max_position_embeddings))
tf.gfile.MakeDirs(FLAGS.output_dir)
task_name = FLAGS.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
label_list = processor.get_labels()
tokenizer = tokenization.FullTokenizer(
vocab_file=FLAGS.vocab_file,
sp_cdc_file=FLAGS.cdc_spm_model_file,
do_lower_case=FLAGS.do_lower_case,
spm_model_file=FLAGS.spm_model_file)
tpu_cluster_resolver = None
if FLAGS.use_tpu and FLAGS.tpu_name:
tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver(
FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2
run_config = contrib_tpu.RunConfig(
cluster=tpu_cluster_resolver,
master=FLAGS.master,
model_dir=FLAGS.output_dir,
save_checkpoints_steps=FLAGS.save_checkpoints_steps,
keep_checkpoint_max=8,
tpu_config=contrib_tpu.TPUConfig(
iterations_per_loop=FLAGS.iterations_per_loop,
num_shards=FLAGS.num_tpu_cores,
per_host_input_for_training=is_per_host))
train_examples = None
num_train_steps = None
num_warmup_steps = None
if FLAGS.do_train:
train_examples = processor.get_train_examples(FLAGS.data_dir)
num_train_steps = int(
len(train_examples) / FLAGS.train_batch_size *
FLAGS.num_train_epochs)
num_warmup_steps = int(num_train_steps * FLAGS.warmup_proportion)
model_fn = model_fn_builder(albert_config=albert_config,
num_labels=len(label_list),
init_checkpoint=FLAGS.init_checkpoint,
learning_rate=FLAGS.learning_rate,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps,
use_tpu=FLAGS.use_tpu,
use_one_hot_embeddings=FLAGS.use_tpu)
# If TPU is not available, this will fall back to normal Estimator on CPU
# or GPU.
estimator = contrib_tpu.TPUEstimator(
use_tpu=FLAGS.use_tpu,
model_fn=model_fn,
config=run_config,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
predict_batch_size=FLAGS.predict_batch_size)
if FLAGS.do_train:
if FLAGS.data_examples:
tf.gfile.MakeDirs(FLAGS.data_examples)
train_file = os.path.join(FLAGS.data_examples, "train.tf_record")
else:
train_file = os.path.join(FLAGS.output_dir, "train.tf_record")
file_based_convert_examples_to_features(train_examples, label_list,
FLAGS.max_seq_length,
tokenizer, train_file)
tf.logging.info("***** Running training *****")
tf.logging.info(" Num examples = %d", len(train_examples))
tf.logging.info(" Batch size = %d", FLAGS.train_batch_size)
tf.logging.info(" Num steps = %d", num_train_steps)
train_input_fn = file_based_input_fn_builder(
input_file=train_file,
seq_length=FLAGS.max_seq_length,
is_training=True,
drop_remainder=True)
tf.logging.set_verbosity(tf.logging.INFO)
estimator.train(input_fn=train_input_fn, max_steps=num_train_steps)
if FLAGS.do_eval:
eval_examples = processor.get_dev_examples(FLAGS.data_dir)
num_actual_eval_examples = len(eval_examples)
if FLAGS.use_tpu:
# TPU requires a fixed batch size for all batches, therefore the number
# of examples must be a multiple of the batch size, or else examples
# will get dropped. So we pad with fake examples which are ignored
# later on. These do NOT count towards the metric (all tf.metrics
# support a per-instance weight, and these get a weight of 0.0).
while len(eval_examples) % FLAGS.eval_batch_size != 0:
eval_examples.append(PaddingInputExample())
if FLAGS.data_examples:
tf.gfile.MakeDirs(FLAGS.data_examples)
eval_file = os.path.join(FLAGS.data_examples, "eval.tf_record")
else:
eval_file = os.path.join(FLAGS.output_dir, "eval.tf_record")
file_based_convert_examples_to_features(eval_examples, label_list,
FLAGS.max_seq_length,
tokenizer, eval_file)
tf.logging.info("***** Running evaluation *****")
tf.logging.info(" Num examples = %d (%d actual, %d padding)",
len(eval_examples), num_actual_eval_examples,
len(eval_examples) - num_actual_eval_examples)
tf.logging.info(" Batch size = %d", FLAGS.eval_batch_size)
# This tells the estimator to run through the entire set.
eval_steps = None
# However, if running eval on the TPU, you will need to specify the
# number of steps.
if FLAGS.use_tpu:
assert len(eval_examples) % FLAGS.eval_batch_size == 0
eval_steps = int(len(eval_examples) // FLAGS.eval_batch_size)
eval_drop_remainder = True if FLAGS.use_tpu else False
eval_input_fn = file_based_input_fn_builder(
input_file=eval_file,
seq_length=FLAGS.max_seq_length,
is_training=False,
drop_remainder=eval_drop_remainder)
result = estimator.evaluate(input_fn=eval_input_fn, steps=eval_steps)
output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt")
with tf.gfile.GFile(output_eval_file, "w") as writer:
tf.logging.info("***** Eval results *****")
for key in sorted(result.keys()):
tf.logging.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
if FLAGS.do_predict:
predict_examples = processor.get_test_examples(FLAGS.data_dir)
num_actual_predict_examples = len(predict_examples)
if FLAGS.use_tpu:
# TPU requires a fixed batch size for all batches, therefore the number
# of examples must be a multiple of the batch size, or else examples
# will get dropped. So we pad with fake examples which are ignored
# later on.
while len(predict_examples) % FLAGS.predict_batch_size != 0:
predict_examples.append(PaddingInputExample())
if FLAGS.data_examples:
tf.gfile.MakeDirs(FLAGS.data_examples)
predict_file = os.path.join(FLAGS.data_examples,
"predict.tf_record")
else:
predict_file = os.path.join(FLAGS.output_dir, "predict.tf_record")
file_based_convert_examples_to_features(predict_examples, label_list,
FLAGS.max_seq_length,
tokenizer, predict_file)
tf.logging.info("***** Running prediction*****")
tf.logging.info(" Num examples = %d (%d actual, %d padding)",
len(predict_examples), num_actual_predict_examples,
len(predict_examples) - num_actual_predict_examples)
tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size)
predict_drop_remainder = True if FLAGS.use_tpu else False
predict_input_fn = file_based_input_fn_builder(
input_file=predict_file,
seq_length=FLAGS.max_seq_length,
is_training=False,
drop_remainder=predict_drop_remainder)
result = estimator.predict(input_fn=predict_input_fn)
output_predict_file = os.path.join(FLAGS.output_dir,
"test_results.tsv")
output_submit_file = os.path.join(FLAGS.output_dir,
"submit_results.tsv")
with tf.gfile.GFile(output_predict_file, "w") as pred_writer,\
tf.gfile.GFile(output_submit_file, "w") as sub_writer:
num_written_lines = 0
tf.logging.info("***** Predict results *****")
for (i, (example, prediction)) in\
enumerate(zip(predict_examples, result)):
probabilities = prediction["probabilities"]
if i >= num_actual_predict_examples:
break
output_line = "\t".join(
str(class_probability)
for class_probability in probabilities) + "\n"
pred_writer.write(output_line)
actual_label = label_list[int(prediction["predictions"])]
sub_writer.write(
six.ensure_str(example.guid) + "\t" + actual_label + "\n")
num_written_lines += 1
assert num_written_lines == num_actual_predict_examples
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
processors = CommonsenseQAProcessor
if not FLAGS.do_train and not FLAGS.do_eval and not FLAGS.do_predict:
raise ValueError(
"At least one of `do_train`, `do_eval` or `do_predict' must be True.")
if not FLAGS.albert_config_file:
raise ValueError("At least one of `--albert_config_file`must be set")
if FLAGS.albert_config_file:
albert_config = modeling.AlbertConfig.from_json_file(FLAGS.albert_config_file)
if FLAGS.max_seq_length > albert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length %d because the ALBERT model "
"was only trained up to sequence length %d" %
(FLAGS.max_seq_length, albert_config.max_position_embeddings))
else:
albert_config = None # Get the config from TF-Hub.
tf.gfile.MakeDirs(FLAGS.output_dir)
processor = processors(
use_spm=True if FLAGS.spm_model_file else False,
do_lower_case=FLAGS.do_lower_case)
label_list = processor.get_labels()
tokenizer = tokenization.FullTokenizer(
vocab_file=FLAGS.vocab_file,
do_lower_case=FLAGS.do_lower_case,
spm_model_file=FLAGS.spm_model_file)
tpu_cluster_resolver = None
if FLAGS.use_tpu and FLAGS.tpu_name:
tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver(
FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2
if FLAGS.do_train:
iterations_per_loop = int(min(FLAGS.iterations_per_loop, FLAGS.save_checkpoints_steps))
else:
iterations_per_loop = FLAGS.iterations_per_loop
run_config = contrib_tpu.RunConfig(
cluster=tpu_cluster_resolver,
master=FLAGS.master,
model_dir=FLAGS.output_dir,
save_checkpoints_steps=int(FLAGS.save_checkpoints_steps),
keep_checkpoint_max=0,
tpu_config=contrib_tpu.TPUConfig(
iterations_per_loop=iterations_per_loop,
num_shards=FLAGS.num_tpu_cores,
per_host_input_for_training=is_per_host))
train_examples = processor.get_train_examples(FLAGS.data_dir)
num_train_steps = int(len(train_examples) / FLAGS.train_batch_size * FLAGS.num_train_epochs)
num_warmup_steps = int(num_train_steps * FLAGS.warmup_proportion)
model_fn = model_fn_builder(
albert_config=albert_config,
num_labels=len(label_list),
init_checkpoint=FLAGS.init_checkpoint,
learning_rate=FLAGS.learning_rate,
num_train_steps= num_train_steps,
num_warmup_steps=num_warmup_steps,
use_tpu=FLAGS.use_tpu,
use_one_hot_embeddings=FLAGS.use_tpu,
optimizer=FLAGS.optimizer)
# If TPU is not available, this will fall back to normal Estimator on CPU
# or GPU.
estimator = contrib_tpu.TPUEstimator(
use_tpu=FLAGS.use_tpu,
model_fn=model_fn,
config=run_config,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
predict_batch_size=FLAGS.predict_batch_size)
if FLAGS.do_train:
cached_dir = FLAGS.cached_dir
if not cached_dir:
cached_dir = FLAGS.output_dir
train_file = os.path.join(cached_dir, "train.tf_record")
if not tf.gfile.Exists(train_file):
file_based_convert_examples_to_features(
train_examples, label_list, FLAGS.max_seq_length, tokenizer, train_file)
tf.logging.info("***** Running training *****")
tf.logging.info(" Num examples = %d", len(train_examples))
tf.logging.info(" Batch size = %d", FLAGS.train_batch_size)
tf.logging.info(" Num steps = %d", num_train_steps)
train_input_fn = file_based_input_fn_builder(
input_file=train_file,
seq_length=FLAGS.max_seq_length,
is_training=True,
drop_remainder=True)
estimator.train(input_fn=train_input_fn, max_steps=num_train_steps)
if FLAGS.do_eval:
eval_examples = processor.get_dev_examples(FLAGS.data_dir)
num_actual_eval_examples = len(eval_examples)
if FLAGS.use_tpu:
while len(eval_examples) % FLAGS.eval_batch_size != 0:
eval_examples.append(classifier_utils.PaddingInputExample())
cached_dir = FLAGS.cached_dir
if not cached_dir:
cached_dir = FLAGS.output_dir
eval_file = os.path.join(cached_dir, "eval.tf_record")
if not tf.gfile.Exists(eval_file):
file_based_convert_examples_to_features(
eval_examples, label_list, FLAGS.max_seq_length, tokenizer, eval_file)
tf.logging.info("***** Running evaluation *****")
tf.logging.info(" Num examples = %d (%d actual, %d padding)",
len(eval_examples), num_actual_eval_examples,
len(eval_examples) - num_actual_eval_examples)
tf.logging.info(" Batch size = %d", FLAGS.eval_batch_size)
# This tells the estimator to run through the entire set.
eval_steps = None
# However, if running eval on the TPU, you will need to specify the
# number of steps.
if FLAGS.use_tpu:
assert len(eval_examples) % FLAGS.eval_batch_size == 0
eval_steps = int(len(eval_examples) // FLAGS.eval_batch_size)
eval_drop_remainder = True if FLAGS.use_tpu else False
eval_input_fn = file_based_input_fn_builder(
input_file=eval_file,
seq_length=FLAGS.max_seq_length,
is_training=False,
drop_remainder=eval_drop_remainder)
best_trial_info_file = os.path.join(FLAGS.output_dir, "best_trial.txt")
def _best_trial_info():
"""Returns information about which checkpoints have been evaled so far."""
if tf.gfile.Exists(best_trial_info_file):
with tf.gfile.GFile(best_trial_info_file, "r") as best_info:
global_step, best_metric_global_step, metric_value = (
best_info.read().split(":"))
global_step = int(global_step)
best_metric_global_step = int(best_metric_global_step)
metric_value = float(metric_value)
else:
metric_value = -1
best_metric_global_step = -1
global_step = -1
tf.logging.info(
"Best trial info: Step: %s, Best Value Step: %s, "
"Best Value: %s", global_step, best_metric_global_step, metric_value)
return global_step, best_metric_global_step, metric_value
def _remove_checkpoint(checkpoint_path):
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tf.logging.info("removing {}".format(src_ckpt))
tf.gfile.Remove(src_ckpt)
def _find_valid_cands(curr_step):
filenames = tf.gfile.ListDirectory(FLAGS.output_dir)
candidates = []
for filename in filenames:
if filename.endswith(".index"):
ckpt_name = filename[:-6]
idx = ckpt_name.split("-")[-1]
if int(idx) > curr_step:
candidates.append(filename)
return candidates
output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt")
global_step, best_perf_global_step, best_perf = _best_trial_info()
writer = tf.gfile.GFile(output_eval_file, "w")
while global_step < num_train_steps:
steps_and_files = {}
filenames = tf.gfile.ListDirectory(FLAGS.output_dir)
for filename in filenames:
if filename.endswith(".index"):
ckpt_name = filename[:-6]
cur_filename = os.path.join(FLAGS.output_dir, ckpt_name)
if cur_filename.split("-")[-1] == "best":
continue
gstep = int(cur_filename.split("-")[-1])
if gstep not in steps_and_files:
tf.logging.info("Add {} to eval list.".format(cur_filename))
steps_and_files[gstep] = cur_filename
tf.logging.info("found {} files.".format(len(steps_and_files)))
if not steps_and_files:
tf.logging.info("found 0 file, global step: {}. Sleeping."
.format(global_step))
time.sleep(60)
else:
for checkpoint in sorted(steps_and_files.items()):
step, checkpoint_path = checkpoint
if global_step >= step:
if (best_perf_global_step != step and
len(_find_valid_cands(step)) > 1):
_remove_checkpoint(checkpoint_path)
continue
result = estimator.evaluate(
input_fn=eval_input_fn,
steps=eval_steps,
checkpoint_path=checkpoint_path)
global_step = result["global_step"]
tf.logging.info("***** Eval results *****")
for key in sorted(result.keys()):
tf.logging.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
writer.write("best = {}\n".format(best_perf))
if result["eval_accuracy"] > best_perf:
best_perf = result["eval_accuracy"]
best_perf_global_step = global_step
elif len(_find_valid_cands(global_step)) > 1:
_remove_checkpoint(checkpoint_path)
writer.write("=" * 50 + "\n")
writer.flush()
with tf.gfile.GFile(best_trial_info_file, "w") as best_info:
best_info.write("{}:{}:{}".format(
global_step, best_perf_global_step, best_perf))
writer.close()
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = "model.ckpt-{}.{}".format(best_perf_global_step, ext)
tgt_ckpt = "model.ckpt-best.{}".format(ext)
tf.logging.info("saving {} to {}".format(src_ckpt, tgt_ckpt))
tf.io.gfile.rename(
os.path.join(FLAGS.output_dir, src_ckpt),
os.path.join(FLAGS.output_dir, tgt_ckpt),
overwrite=True)
if FLAGS.do_predict:
predict_examples = processor.get_test_examples(FLAGS.data_dir)
num_actual_predict_examples = len(predict_examples)
if FLAGS.use_tpu:
while len(predict_examples) % FLAGS.predict_batch_size != 0:
predict_examples.append(classifier_utils.PaddingInputExample())
predict_file = os.path.join(FLAGS.output_dir, "predict.tf_record")
file_based_convert_examples_to_features(
predict_examples, label_list,
FLAGS.max_seq_length, tokenizer,
predict_file)
tf.logging.info("***** Running prediction*****")
tf.logging.info(" Num examples = %d (%d actual, %d padding)",
len(predict_examples), num_actual_predict_examples,
len(predict_examples) - num_actual_predict_examples)
tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size)
predict_drop_remainder = True if FLAGS.use_tpu else False
predict_input_fn = file_based_input_fn_builder(
input_file=predict_file,
seq_length=FLAGS.max_seq_length,
is_training=False,
drop_remainder=predict_drop_remainder)
checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best")
result = estimator.predict(
input_fn=predict_input_fn,
checkpoint_path=checkpoint_path)
output_predict_file = os.path.join(FLAGS.output_dir, "test_results.csv")
output_submit_file = os.path.join(FLAGS.output_dir, "submit_results.csv")
with tf.gfile.GFile(output_predict_file, "w") as pred_writer,\
tf.gfile.GFile(output_submit_file, "w") as sub_writer:
sub_writer.write("index" + "\t" + "prediction\n")
num_written_lines = 0
tf.logging.info("***** Predict results *****")
for (i, (example, prediction)) in\
enumerate(zip(predict_examples, result)):
probabilities = prediction["probabilities"]
if i >= num_actual_predict_examples:
break
output_line = "\t".join(
str(class_probability)
for class_probability in probabilities) + "\n"
pred_writer.write(output_line)
actual_label = label_list[int(prediction["predictions"])]
sub_writer.write(example.guid + "\t" + actual_label + "\n")
num_written_lines += 1
assert num_written_lines == num_actual_predict_examples
def transformer(model: str = 'xlnet', quantized: bool = False, **kwargs):
"""
Load Transformer toxicity model.
Parameters
----------
model : str, optional (default='bert')
Model architecture supported. Allowed values:
* ``'bert'`` - Google BERT BASE parameters.
* ``'tiny-bert'`` - Google BERT TINY parameters.
* ``'albert'`` - Google ALBERT BASE parameters.
* ``'tiny-albert'`` - Google ALBERT TINY parameters.
* ``'xlnet'`` - Google XLNET BASE parameters.
* ``'alxlnet'`` - Malaya ALXLNET BASE parameters.
quantized : bool, optional (default=False)
if True, will load 8-bit quantized model.
Quantized model not necessary faster, totally depends on the machine.
Returns
-------
result : malaya.model.bert.SIGMOID_BERT class
"""
model = model.lower()
if model not in _transformer_availability:
raise Exception(
'model not supported, please check supported models from `malaya.toxicity.available_transformer()`.'
)
check_file(
PATH_TOXIC[model], S3_PATH_TOXIC[model], quantized = quantized, **kwargs
)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(PATH_TOXIC[model][model_path], **kwargs)
path = PATH_TOXIC
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
from malaya.transformers.bert import (
_extract_attention_weights_import,
)
from malaya.transformers.bert import bert_num_layers
tokenizer = sentencepiece_tokenizer_bert(
path[model]['tokenizer'], path[model]['vocab']
)
if model in ['albert', 'tiny-albert']:
from malaya.transformers.albert import (
_extract_attention_weights_import,
)
from malaya.transformers.albert import bert_num_layers
from albert import tokenization
tokenizer = tokenization.FullTokenizer(
vocab_file = path[model]['vocab'],
do_lower_case = False,
spm_model_file = path[model]['tokenizer'],
)
return SIGMOID_BERT(
X = g.get_tensor_by_name('import/Placeholder:0'),
segment_ids = None,
input_masks = g.get_tensor_by_name('import/Placeholder_1:0'),
logits = g.get_tensor_by_name('import/logits:0'),
logits_seq = g.get_tensor_by_name('import/logits_seq:0'),
vectorizer = g.get_tensor_by_name('import/dense/BiasAdd:0'),
sess = generate_session(graph = g, **kwargs),
tokenizer = tokenizer,
label = label,
attns = _extract_attention_weights_import(
bert_num_layers[model], g
),
class_name = 'toxic',
)
if model in ['xlnet', 'alxlnet']:
if model in ['xlnet']:
from malaya.transformers.xlnet import (
_extract_attention_weights_import,
)
if model in ['alxlnet']:
from malaya.transformers.alxlnet import (
_extract_attention_weights_import,
)
tokenizer = sentencepiece_tokenizer_xlnet(path[model]['tokenizer'])
return SIGMOID_XLNET(
X = g.get_tensor_by_name('import/Placeholder:0'),
segment_ids = g.get_tensor_by_name('import/Placeholder_1:0'),
input_masks = g.get_tensor_by_name('import/Placeholder_2:0'),
logits = g.get_tensor_by_name('import/logits:0'),
logits_seq = g.get_tensor_by_name('import/logits_seq:0'),
vectorizer = g.get_tensor_by_name('import/transpose_3:0'),
sess = generate_session(graph = g, **kwargs),
tokenizer = tokenizer,
label = label,
attns = _extract_attention_weights_import(g),
class_name = 'toxic',
)
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
processors = {
"cola": classifier_utils.ColaProcessor,
"mnli": classifier_utils.MnliProcessor,
"mismnli": classifier_utils.MisMnliProcessor,
"mrpc": classifier_utils.MrpcProcessor,
"rte": classifier_utils.RteProcessor,
"sst-2": classifier_utils.Sst2Processor,
"sts-b": classifier_utils.StsbProcessor,
"qqp": classifier_utils.QqpProcessor,
"qnli": classifier_utils.QnliProcessor,
"wnli": classifier_utils.WnliProcessor,
}
tokenization.validate_case_matches_checkpoint(FLAGS.do_lower_case,
FLAGS.init_checkpoint)
if not FLAGS.do_train and not FLAGS.do_eval and not FLAGS.do_predict:
raise ValueError(
"At least one of `do_train`, `do_eval` or `do_predict' must be True."
)
albert_config = modeling.AlbertConfig.from_json_file(
FLAGS.albert_config_file)
if FLAGS.max_seq_length > albert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length %d because the ALBERT model "
"was only trained up to sequence length %d" %
(FLAGS.max_seq_length, albert_config.max_position_embeddings))
tf.gfile.MakeDirs(FLAGS.output_dir)
task_name = FLAGS.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name](
use_spm=True if FLAGS.spm_model_file else False,
do_lower_case=FLAGS.do_lower_case)
label_list = processor.get_labels()
tokenizer = tokenization.FullTokenizer(vocab_file=FLAGS.vocab_file,
do_lower_case=FLAGS.do_lower_case,
spm_model_file=FLAGS.spm_model_file)
tpu_cluster_resolver = None
if FLAGS.use_tpu and FLAGS.tpu_name:
tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver(
FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2
if FLAGS.do_train:
iterations_per_loop = int(
min(FLAGS.iterations_per_loop, FLAGS.save_checkpoints_steps))
else:
iterations_per_loop = FLAGS.iterations_per_loop
run_config = contrib_tpu.RunConfig(
cluster=tpu_cluster_resolver,
master=FLAGS.master,
model_dir=FLAGS.output_dir,
save_checkpoints_steps=int(FLAGS.save_checkpoints_steps),
keep_checkpoint_max=0,
tpu_config=contrib_tpu.TPUConfig(
iterations_per_loop=iterations_per_loop,
num_shards=FLAGS.num_tpu_cores,
per_host_input_for_training=is_per_host))
train_examples = None
if FLAGS.do_train:
train_examples = processor.get_train_examples(FLAGS.data_dir)
model_fn = classifier_utils.model_fn_builder(
albert_config=albert_config,
num_labels=len(label_list),
init_checkpoint=FLAGS.init_checkpoint,
learning_rate=FLAGS.learning_rate,
num_train_steps=FLAGS.train_step,
num_warmup_steps=FLAGS.warmup_step,
use_tpu=FLAGS.use_tpu,
use_one_hot_embeddings=FLAGS.use_tpu,
task_name=task_name,
optimizer=FLAGS.optimizer)
# If TPU is not available, this will fall back to normal Estimator on CPU
# or GPU.
estimator = contrib_tpu.TPUEstimator(
use_tpu=FLAGS.use_tpu,
model_fn=model_fn,
config=run_config,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
predict_batch_size=FLAGS.predict_batch_size)
if FLAGS.do_train:
cached_dir = FLAGS.cached_dir
if not cached_dir:
cached_dir = FLAGS.output_dir
train_file = os.path.join(cached_dir, task_name + "_train.tf_record")
if not tf.gfile.Exists(train_file):
classifier_utils.file_based_convert_examples_to_features(
train_examples, label_list, FLAGS.max_seq_length, tokenizer,
train_file, task_name)
tf.logging.info("***** Running training *****")
tf.logging.info(" Num examples = %d", len(train_examples))
tf.logging.info(" Batch size = %d", FLAGS.train_batch_size)
tf.logging.info(" Num steps = %d", FLAGS.train_step)
train_input_fn = classifier_utils.file_based_input_fn_builder(
input_file=train_file,
seq_length=FLAGS.max_seq_length,
is_training=True,
drop_remainder=True,
task_name=task_name,
use_tpu=FLAGS.use_tpu,
bsz=FLAGS.train_batch_size)
estimator.train(input_fn=train_input_fn, max_steps=FLAGS.train_step)
if FLAGS.do_eval:
eval_examples = processor.get_dev_examples(FLAGS.data_dir)
num_actual_eval_examples = len(eval_examples)
if FLAGS.use_tpu:
# TPU requires a fixed batch size for all batches, therefore the number
# of examples must be a multiple of the batch size, or else examples
# will get dropped. So we pad with fake examples which are ignored
# later on. These do NOT count towards the metric (all tf.metrics
# support a per-instance weight, and these get a weight of 0.0).
while len(eval_examples) % FLAGS.eval_batch_size != 0:
eval_examples.append(classifier_utils.PaddingInputExample())
cached_dir = FLAGS.cached_dir
if not cached_dir:
cached_dir = FLAGS.output_dir
eval_file = os.path.join(cached_dir, task_name + "_eval.tf_record")
if not tf.gfile.Exists(eval_file):
classifier_utils.file_based_convert_examples_to_features(
eval_examples, label_list, FLAGS.max_seq_length, tokenizer,
eval_file, task_name)
tf.logging.info("***** Running evaluation *****")
tf.logging.info(" Num examples = %d (%d actual, %d padding)",
len(eval_examples), num_actual_eval_examples,
len(eval_examples) - num_actual_eval_examples)
tf.logging.info(" Batch size = %d", FLAGS.eval_batch_size)
# This tells the estimator to run through the entire set.
eval_steps = None
# However, if running eval on the TPU, you will need to specify the
# number of steps.
if FLAGS.use_tpu:
assert len(eval_examples) % FLAGS.eval_batch_size == 0
eval_steps = int(len(eval_examples) // FLAGS.eval_batch_size)
eval_drop_remainder = True if FLAGS.use_tpu else False
eval_input_fn = classifier_utils.file_based_input_fn_builder(
input_file=eval_file,
seq_length=FLAGS.max_seq_length,
is_training=False,
drop_remainder=eval_drop_remainder,
task_name=task_name,
use_tpu=FLAGS.use_tpu,
bsz=FLAGS.eval_batch_size)
def _find_valid_cands(curr_step):
filenames = tf.gfile.ListDirectory(FLAGS.output_dir)
candidates = []
for filename in filenames:
if filename.endswith(".index"):
ckpt_name = filename[:-6]
idx = ckpt_name.split("-")[-1]
if idx != "best" and int(idx) > curr_step:
candidates.append(filename)
return candidates
output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt")
checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best")
if task_name == "sts-b":
key_name = "pearson"
elif task_name == "cola":
key_name = "matthew_corr"
else:
key_name = "eval_accuracy"
if tf.gfile.Exists(checkpoint_path + ".index"):
result = estimator.evaluate(input_fn=eval_input_fn,
steps=eval_steps,
checkpoint_path=checkpoint_path)
best_perf = result[key_name]
global_step = result["global_step"]
else:
global_step = -1
best_perf = -1
checkpoint_path = None
writer = tf.gfile.GFile(output_eval_file, "w")
while global_step < FLAGS.train_step:
steps_and_files = {}
filenames = tf.gfile.ListDirectory(FLAGS.output_dir)
for filename in filenames:
if filename.endswith(".index"):
ckpt_name = filename[:-6]
cur_filename = os.path.join(FLAGS.output_dir, ckpt_name)
if cur_filename.split("-")[-1] == "best":
continue
gstep = int(cur_filename.split("-")[-1])
if gstep not in steps_and_files:
tf.logging.info(
"Add {} to eval list.".format(cur_filename))
steps_and_files[gstep] = cur_filename
tf.logging.info("found {} files.".format(len(steps_and_files)))
if not steps_and_files:
tf.logging.info(
"found 0 file, global step: {}. Sleeping.".format(
global_step))
time.sleep(1)
else:
for checkpoint in sorted(steps_and_files.items()):
step, checkpoint_path = checkpoint
if global_step >= step:
if len(_find_valid_cands(step)) > 1:
for ext in [
"meta", "data-00000-of-00001", "index"
]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tf.logging.info("removing {}".format(src_ckpt))
tf.gfile.Remove(src_ckpt)
continue
result = estimator.evaluate(
input_fn=eval_input_fn,
steps=eval_steps,
checkpoint_path=checkpoint_path)
global_step = result["global_step"]
tf.logging.info("***** Eval results *****")
for key in sorted(result.keys()):
tf.logging.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
writer.write("best = {}\n".format(best_perf))
if result[key_name] > best_perf:
best_perf = result[key_name]
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tgt_ckpt = checkpoint_path.rsplit(
"-", 1)[0] + "-best.{}".format(ext)
tf.logging.info("saving {} to {}".format(
src_ckpt, tgt_ckpt))
tf.gfile.Copy(src_ckpt, tgt_ckpt, overwrite=True)
writer.write("saved {} to {}\n".format(
src_ckpt, tgt_ckpt))
if len(_find_valid_cands(global_step)) > 1:
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tf.logging.info("removing {}".format(src_ckpt))
tf.gfile.Remove(src_ckpt)
writer.write("=" * 50 + "\n")
writer.close()
if FLAGS.do_predict:
predict_examples = processor.get_test_examples(FLAGS.data_dir)
num_actual_predict_examples = len(predict_examples)
if FLAGS.use_tpu:
# TPU requires a fixed batch size for all batches, therefore the number
# of examples must be a multiple of the batch size, or else examples
# will get dropped. So we pad with fake examples which are ignored
# later on.
while len(predict_examples) % FLAGS.predict_batch_size != 0:
predict_examples.append(classifier_utils.PaddingInputExample())
predict_file = os.path.join(FLAGS.output_dir, "predict.tf_record")
classifier_utils.file_based_convert_examples_to_features(
predict_examples, label_list, FLAGS.max_seq_length, tokenizer,
predict_file, task_name)
tf.logging.info("***** Running prediction*****")
tf.logging.info(" Num examples = %d (%d actual, %d padding)",
len(predict_examples), num_actual_predict_examples,
len(predict_examples) - num_actual_predict_examples)
tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size)
predict_drop_remainder = True if FLAGS.use_tpu else False
predict_input_fn = classifier_utils.file_based_input_fn_builder(
input_file=predict_file,
seq_length=FLAGS.max_seq_length,
is_training=False,
drop_remainder=predict_drop_remainder,
task_name=task_name,
use_tpu=FLAGS.use_tpu,
bsz=FLAGS.predict_batch_size)
checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best")
result = estimator.predict(input_fn=predict_input_fn,
checkpoint_path=checkpoint_path)
output_predict_file = os.path.join(FLAGS.output_dir,
"test_results.tsv")
output_submit_file = os.path.join(FLAGS.output_dir,
"submit_results.tsv")
with tf.gfile.GFile(output_predict_file, "w") as pred_writer,\
tf.gfile.GFile(output_submit_file, "w") as sub_writer:
sub_writer.write("index" + "\t" + "prediction\n")
num_written_lines = 0
tf.logging.info("***** Predict results *****")
for (i, (example, prediction)) in\
enumerate(zip(predict_examples, result)):
probabilities = prediction["probabilities"]
if i >= num_actual_predict_examples:
break
output_line = "\t".join(
str(class_probability)
for class_probability in probabilities) + "\n"
pred_writer.write(output_line)
if task_name != "sts-b":
actual_label = label_list[int(prediction["predictions"])]
else:
actual_label = str(prediction["predictions"])
sub_writer.write(example.guid + "\t" + actual_label + "\n")
num_written_lines += 1
assert num_written_lines == num_actual_predict_examples
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
processors = {"race": race_utils.RaceProcessor}
tokenization.validate_case_matches_checkpoint(FLAGS.do_lower_case,
FLAGS.init_checkpoint)
if not FLAGS.do_train and not FLAGS.do_eval and not FLAGS.do_predict:
raise ValueError(
"At least one of `do_train`, `do_eval` or `do_predict' must be True."
)
albert_config = modeling.AlbertConfig.from_json_file(
FLAGS.albert_config_file)
albert_config.hidden_dropout_prob = FLAGS.albert_dropout_prob
albert_config.attention_probs_dropout_prob = FLAGS.albert_dropout_prob
if FLAGS.max_seq_length > albert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length %d because the ALBERT model "
"was only trained up to sequence length %d" %
(FLAGS.max_seq_length, albert_config.max_position_embeddings))
tf.gfile.MakeDirs(FLAGS.output_dir)
task_name = FLAGS.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name](
use_spm=True if FLAGS.spm_model_file else False,
do_lower_case=FLAGS.do_lower_case,
high_only=FLAGS.high_only,
middle_only=FLAGS.middle_only)
label_list = processor.get_labels()
tokenizer = tokenization.FullTokenizer(vocab_file=FLAGS.vocab_file,
do_lower_case=FLAGS.do_lower_case,
spm_model_file=FLAGS.spm_model_file)
# multiple gpus
NUM_GPUS = FLAGS.num_gpu_cores if FLAGS.strategy_type == 'mirror' else 1
using_customized_optimizer = None
if NUM_GPUS > 1 and FLAGS.strategy_type == "mirror":
os.environ["CUDA_VISIBLE_DEVICES"] = ",".join(
[str(i) for i in list(range(NUM_GPUS))])
# https://github.com/tensorflow/tensorflow/issues/21470#issuecomment-422506263
strategy = tf.contrib.distribute.MirroredStrategy(
num_gpus=NUM_GPUS,
cross_device_ops=AllReduceCrossDeviceOps('nccl',
num_packs=NUM_GPUS),
)
using_customized_optimizer = True
tf.logging.info('Use MirroredStrategy with %d devices.',
strategy.num_replicas_in_sync)
else:
strategy = tf.distribute.OneDeviceStrategy("GPU:0")
using_customized_optimizer = False
tf.logging.info('Single device mode.')
tpu_cluster_resolver = None
if FLAGS.use_tpu and FLAGS.tpu_name:
tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver(
FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2
if FLAGS.do_train:
iterations_per_loop = int(
min(FLAGS.iterations_per_loop, FLAGS.save_checkpoints_steps))
else:
iterations_per_loop = FLAGS.iterations_per_loop
run_config = contrib_tpu.RunConfig(
cluster=tpu_cluster_resolver,
master=FLAGS.master,
model_dir=FLAGS.output_dir,
save_checkpoints_steps=int(FLAGS.save_checkpoints_steps),
keep_checkpoint_max=0,
tpu_config=contrib_tpu.TPUConfig(
iterations_per_loop=iterations_per_loop,
num_shards=FLAGS.num_tpu_cores,
per_host_input_for_training=is_per_host),
train_distribute=strategy,
eval_distribute=strategy, #get error during evaluation
)
train_examples = None
total_time = None
if FLAGS.do_train:
train_examples = processor.get_train_examples(FLAGS.data_dir)
model_fn = race_utils.model_fn_builder(
albert_config=albert_config,
num_labels=len(label_list),
init_checkpoint=FLAGS.init_checkpoint,
learning_rate=FLAGS.learning_rate,
num_train_steps=FLAGS.train_step,
num_warmup_steps=FLAGS.warmup_step,
use_tpu=FLAGS.use_tpu,
use_one_hot_embeddings=FLAGS.use_tpu,
max_seq_length=FLAGS.max_seq_length,
dropout_prob=FLAGS.dropout_prob,
customized=using_customized_optimizer)
# If TPU is not available, this will fall back to normal Estimator on CPU
# or GPU.
if FLAGS.use_tpu and FLAGS.tpu_name:
tf.logging.info("Use TPUEstimator")
estimator = contrib_tpu.TPUEstimator(
use_tpu=FLAGS.use_tpu,
model_fn=model_fn,
config=run_config,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
predict_batch_size=FLAGS.predict_batch_size)
else:
tf.logging.info("Use normal Estimator")
estimator = Estimator(
model_fn=model_fn,
params={},
config=run_config,
)
if FLAGS.do_train:
if not tf.gfile.Exists(FLAGS.train_feature_file):
race_utils.file_based_convert_examples_to_features(
train_examples, label_list, FLAGS.max_seq_length, tokenizer,
FLAGS.train_feature_file, FLAGS.max_qa_length)
tf.logging.info("***** Running training *****")
tf.logging.info(" Num examples = %d", len(train_examples))
tf.logging.info(" Batch size = %d", FLAGS.train_batch_size)
tf.logging.info(" Num steps = %d", FLAGS.train_step)
train_input_fn = classifier_utils.file_based_input_fn_builder(
input_file=FLAGS.train_feature_file,
seq_length=FLAGS.max_seq_length,
is_training=True,
drop_remainder=True,
task_name=task_name,
use_tpu=FLAGS.use_tpu,
bsz=FLAGS.train_batch_size,
multiple=len(label_list))
time_hist = TimeHistory()
estimator.train(input_fn=train_input_fn, max_steps=FLAGS.train_step)
total_time = sum(time_hist.times)
if FLAGS.do_eval:
eval_examples = processor.get_dev_examples(FLAGS.data_dir)
num_actual_eval_examples = len(eval_examples)
if FLAGS.use_tpu:
# TPU requires a fixed batch size for all batches, therefore the number
# of examples must be a multiple of the batch size, or else examples
# will get dropped. So we pad with fake examples which are ignored
# later on. These do NOT count towards the metric (all tf.metrics
# support a per-instance weight, and these get a weight of 0.0).
while len(eval_examples) % FLAGS.eval_batch_size != 0:
eval_examples.append(classifier_utils.PaddingInputExample())
if not tf.gfile.Exists(FLAGS.eval_feature_file):
race_utils.file_based_convert_examples_to_features(
eval_examples, label_list, FLAGS.max_seq_length, tokenizer,
FLAGS.eval_feature_file, FLAGS.max_qa_length)
tf.logging.info("***** Running evaluation *****")
tf.logging.info(" Num examples = %d (%d actual, %d padding)",
len(eval_examples), num_actual_eval_examples,
len(eval_examples) - num_actual_eval_examples)
tf.logging.info(" Batch size = %d", FLAGS.eval_batch_size)
# This tells the estimator to run through the entire set.
eval_steps = None
# However, if running eval on the TPU, you will need to specify the
# number of steps.
if FLAGS.use_tpu:
assert len(eval_examples) % FLAGS.eval_batch_size == 0
eval_steps = int(len(eval_examples) // FLAGS.eval_batch_size)
eval_drop_remainder = True if FLAGS.use_tpu else False
eval_input_fn = classifier_utils.file_based_input_fn_builder(
input_file=FLAGS.eval_feature_file,
seq_length=FLAGS.max_seq_length,
is_training=False,
drop_remainder=eval_drop_remainder,
task_name=task_name,
use_tpu=FLAGS.use_tpu,
bsz=FLAGS.eval_batch_size,
multiple=len(label_list))
def _find_valid_cands(curr_step):
filenames = tf.gfile.ListDirectory(FLAGS.output_dir)
candidates = []
for filename in filenames:
if filename.endswith(".index"):
ckpt_name = filename[:-6]
idx = ckpt_name.split("-")[-1]
if idx != "best" and int(idx) > curr_step:
candidates.append(filename)
return candidates
output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt")
checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best")
key_name = "eval_accuracy"
if tf.gfile.Exists(checkpoint_path + ".index"):
result = estimator.evaluate(input_fn=eval_input_fn,
steps=eval_steps,
checkpoint_path=checkpoint_path)
best_perf = result[key_name]
global_step = result["global_step"]
else:
global_step = -1
best_perf = -1
checkpoint_path = None
writer = tf.gfile.GFile(output_eval_file, "w")
avg_time_per_batch = np.mean(time_hist.times)
writer.write("===== Hyperparameters =====\n")
writer.write("Training batch size: {}\n".format(
FLAGS.train_batch_size))
writer.write("Max sequence length: {}\n".format(FLAGS.max_seq_length))
writer.write("Max qa length: {}\n".format(FLAGS.max_qa_length))
writer.write("Learning rate: {}\n".format(FLAGS.learning_rate))
writer.write("Num of GPU cores: {}\n".format(NUM_GPUS))
if FLAGS.do_train:
avg_time_per_batch = np.mean(time_hist.times)
writer.write("Total time: {}\n".format(total_time))
writer.write("Speed: {}\n".format(FLAGS.train_batch_size *
NUM_GPUS / avg_time_per_batch))
if FLAGS.train_step and FLAGS.warmup_step:
writer.write("Training steps: {}\n".format(FLAGS.train_step))
writer.write("Warmup steps: {}\n".format(FLAGS.warmup_step))
while global_step < FLAGS.train_step:
steps_and_files = {}
filenames = tf.gfile.ListDirectory(FLAGS.output_dir)
for filename in filenames:
if filename.endswith(".index"):
ckpt_name = filename[:-6]
cur_filename = os.path.join(FLAGS.output_dir, ckpt_name)
if cur_filename.split("-")[-1] == "best":
continue
gstep = int(cur_filename.split("-")[-1])
if gstep not in steps_and_files:
tf.logging.info(
"Add {} to eval list.".format(cur_filename))
steps_and_files[gstep] = cur_filename
tf.logging.info("found {} files.".format(len(steps_and_files)))
# steps_and_files = sorted(steps_and_files, key=lambda x: x[0])
if not steps_and_files:
tf.logging.info(
"found 0 file, global step: {}. Sleeping.".format(
global_step))
time.sleep(1)
else:
for ele in sorted(steps_and_files.items()):
step, checkpoint_path = ele
if global_step >= step:
if len(_find_valid_cands(step)) > 1:
for ext in [
"meta", "data-00000-of-00001", "index"
]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tf.logging.info("removing {}".format(src_ckpt))
tf.gfile.Remove(src_ckpt)
continue
result = estimator.evaluate(
input_fn=eval_input_fn,
steps=eval_steps,
checkpoint_path=checkpoint_path)
global_step = result["global_step"]
tf.logging.info("***** Eval results *****")
for key in sorted(result.keys()):
tf.logging.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
writer.write("best = {}\n".format(best_perf))
if result[key_name] > best_perf:
best_perf = result[key_name]
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tgt_ckpt = checkpoint_path.rsplit(
"-", 1)[0] + "-best.{}".format(ext)
tf.logging.info("saving {} to {}".format(
src_ckpt, tgt_ckpt))
tf.gfile.Copy(src_ckpt, tgt_ckpt, overwrite=True)
writer.write("saved {} to {}\n".format(
src_ckpt, tgt_ckpt))
if len(_find_valid_cands(global_step)) > 1:
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tf.logging.info("removing {}".format(src_ckpt))
tf.gfile.Remove(src_ckpt)
writer.write("=" * 50 + "\n")
writer.close()
if FLAGS.do_predict:
predict_examples = processor.get_test_examples(FLAGS.data_dir)
num_actual_predict_examples = len(predict_examples)
if FLAGS.use_tpu:
# TPU requires a fixed batch size for all batches, therefore the number
# of examples must be a multiple of the batch size, or else examples
# will get dropped. So we pad with fake examples which are ignored
# later on.
while len(predict_examples) % FLAGS.predict_batch_size != 0:
predict_examples.append(classifier_utils.PaddingInputExample())
assert len(predict_examples) % FLAGS.predict_batch_size == 0
predict_steps = int(
len(predict_examples) // FLAGS.predict_batch_size)
predict_file = os.path.join(FLAGS.output_dir, "predict.tf_record")
race_utils.file_based_convert_examples_to_features(
predict_examples, label_list, FLAGS.max_seq_length, tokenizer,
predict_file, FLAGS.max_qa_length)
tf.logging.info("***** Running prediction*****")
tf.logging.info(" Num examples = %d (%d actual, %d padding)",
len(predict_examples), num_actual_predict_examples,
len(predict_examples) - num_actual_predict_examples)
tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size)
predict_drop_remainder = True if FLAGS.use_tpu else False
predict_input_fn = classifier_utils.file_based_input_fn_builder(
input_file=predict_file,
seq_length=FLAGS.max_seq_length,
is_training=False,
drop_remainder=predict_drop_remainder,
task_name=task_name,
use_tpu=FLAGS.use_tpu,
bsz=FLAGS.predict_batch_size,
multiple=len(label_list))
checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best")
result = estimator.evaluate(input_fn=predict_input_fn,
steps=predict_steps,
checkpoint_path=checkpoint_path)
output_predict_file = os.path.join(FLAGS.output_dir,
"predict_results.txt")
with tf.gfile.GFile(output_predict_file, "w") as pred_writer:
# num_written_lines = 0
tf.logging.info("***** Predict results *****")
pred_writer.write("***** Predict results *****\n")
for key in sorted(result.keys()):
tf.logging.info(" %s = %s", key, str(result[key]))
pred_writer.write("%s = %s\n" % (key, str(result[key])))
pred_writer.write("best = {}\n".format(best_perf))
def _transformer(model, bert_class, xlnet_class, **kwargs):
model = model.lower()
if model not in _availability:
raise Exception(
'model not supported, please check supported models from malaya.similarity.available_transformer()'
)
check_file(PATH_SIMILARITY[model], S3_PATH_SIMILARITY[model], **kwargs)
g = load_graph(PATH_SIMILARITY[model]['model'], **kwargs)
path = PATH_SIMILARITY
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
from malaya.transformers.bert import (
_extract_attention_weights_import, )
from malaya.transformers.bert import bert_num_layers
tokenizer = sentencepiece_tokenizer_bert(path[model]['tokenizer'],
path[model]['vocab'])
if model in ['albert', 'tiny-albert']:
from malaya.transformers.albert import (
_extract_attention_weights_import, )
from malaya.transformers.albert import bert_num_layers
from albert import tokenization
tokenizer = tokenization.FullTokenizer(
vocab_file=path[model]['vocab'],
do_lower_case=False,
spm_model_file=path[model]['tokenizer'],
)
return bert_class(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=g.get_tensor_by_name('import/Placeholder_1:0'),
input_masks=g.get_tensor_by_name('import/Placeholder_2:0'),
logits=g.get_tensor_by_name('import/logits:0'),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
label=['not similar', 'similar'],
)
if model in ['xlnet', 'alxlnet']:
if model in ['xlnet']:
from malaya.transformers.xlnet import (
_extract_attention_weights_import, )
if model in ['alxlnet']:
from malaya.transformers.alxlnet import (
_extract_attention_weights_import, )
tokenizer = sentencepiece_tokenizer_xlnet(path[model]['tokenizer'])
return xlnet_class(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=g.get_tensor_by_name('import/Placeholder_1:0'),
input_masks=g.get_tensor_by_name('import/Placeholder_2:0'),
logits=g.get_tensor_by_name('import/logits:0'),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
label=['not similar', 'similar'],
)
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
albert_config = modeling.AlbertConfig.from_json_file(
FLAGS.albert_config_file)
validate_flags_or_throw(albert_config)
tf.gfile.MakeDirs(FLAGS.output_dir)
tokenizer = tokenization.FullTokenizer(vocab_file=FLAGS.vocab_file,
do_lower_case=FLAGS.do_lower_case,
spm_model_file=FLAGS.spm_model_file)
# multiple gpus
NUM_GPUS = FLAGS.num_gpu_cores if FLAGS.strategy_type == 'mirror' else 1
using_customized_optimizer = None
if NUM_GPUS > 1 and FLAGS.strategy_type == "mirror":
os.environ["CUDA_VISIBLE_DEVICES"] = ",".join(
[str(i) for i in list(range(NUM_GPUS))])
# https://github.com/tensorflow/tensorflow/issues/21470#issuecomment-422506263
strategy = tf.contrib.distribute.MirroredStrategy(
num_gpus=NUM_GPUS,
cross_device_ops=AllReduceCrossDeviceOps('nccl',
num_packs=NUM_GPUS),
)
using_customized_optimizer = True
tf.logging.info('Use MirroredStrategy with %d devices.',
strategy.num_replicas_in_sync)
else:
strategy = tf.distribute.OneDeviceStrategy("GPU:0")
using_customized_optimizer = False
tf.logging.info('Single device mode.')
tpu_cluster_resolver = None
if FLAGS.use_tpu and FLAGS.tpu_name:
tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver(
FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2
if FLAGS.do_train:
iterations_per_loop = int(
min(FLAGS.iterations_per_loop, FLAGS.save_checkpoints_steps))
else:
iterations_per_loop = FLAGS.iterations_per_loop
run_config = contrib_tpu.RunConfig(
cluster=tpu_cluster_resolver,
master=FLAGS.master,
model_dir=FLAGS.output_dir,
save_checkpoints_steps=FLAGS.save_checkpoints_steps,
tpu_config=contrib_tpu.TPUConfig(
iterations_per_loop=iterations_per_loop,
num_shards=FLAGS.num_tpu_cores,
per_host_input_for_training=is_per_host),
train_distribute=strategy,
eval_distribute=strategy, #get error during evaluation
)
train_examples = None
num_train_steps = None
num_warmup_steps = None
train_examples = squad_utils.read_squad_examples(
input_file=FLAGS.train_file, is_training=True)
num_train_steps = int(
len(train_examples) / FLAGS.train_batch_size * FLAGS.num_train_epochs)
if FLAGS.do_train:
num_warmup_steps = int(num_train_steps * FLAGS.warmup_proportion)
# Pre-shuffle the input to avoid having to make a very large shuffle
# buffer in in the `input_fn`.
rng = random.Random(12345)
rng.shuffle(train_examples)
model_fn = squad_utils.v2_model_fn_builder(
albert_config=albert_config,
init_checkpoint=FLAGS.init_checkpoint,
learning_rate=FLAGS.learning_rate,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps,
use_tpu=FLAGS.use_tpu,
use_one_hot_embeddings=FLAGS.use_tpu,
max_seq_length=FLAGS.max_seq_length,
start_n_top=FLAGS.start_n_top,
end_n_top=FLAGS.end_n_top,
dropout_prob=FLAGS.dropout_prob,
customized=using_customized_optimizer,
optimizer=FLAGS.optimizer)
# If TPU is not available, this will fall back to normal Estimator on CPU
# or GPU.
if FLAGS.use_tpu and FLAGS.tpu_name:
tf.logging.info("Use TPUEstimator")
estimator = contrib_tpu.TPUEstimator(
use_tpu=FLAGS.use_tpu,
model_fn=model_fn,
config=run_config,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
predict_batch_size=FLAGS.predict_batch_size)
else:
tf.logging.info("Use normal Estimator")
estimator = Estimator(
model_fn=model_fn,
params={},
config=run_config,
)
if FLAGS.do_train:
# We write to a temporary file to avoid storing very large constant tensors
# in memory.
if not tf.gfile.Exists(FLAGS.train_feature_file):
train_writer = squad_utils.FeatureWriter(filename=os.path.join(
FLAGS.train_feature_file),
is_training=True)
squad_utils.convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=FLAGS.max_seq_length,
doc_stride=FLAGS.doc_stride,
max_query_length=FLAGS.max_query_length,
is_training=True,
output_fn=train_writer.process_feature,
do_lower_case=FLAGS.do_lower_case)
train_writer.close()
tf.logging.info("***** Running training *****")
tf.logging.info(" Num orig examples = %d", len(train_examples))
# tf.logging.info(" Num split examples = %d", train_writer.num_features)
tf.logging.info(
f" Batch size = {FLAGS.train_batch_size} * {NUM_GPUS}")
tf.logging.info(" Num steps = %d", num_train_steps)
del train_examples
train_input_fn = squad_utils.input_fn_builder(
input_file=FLAGS.train_feature_file,
seq_length=FLAGS.max_seq_length,
is_training=True,
drop_remainder=True,
use_tpu=FLAGS.use_tpu,
bsz=FLAGS.train_batch_size,
is_v2=True)
time_hist = TimeHistory()
estimator.train(input_fn=train_input_fn, max_steps=num_train_steps)
total_time = sum(time_hist.times)
if FLAGS.do_predict:
with tf.gfile.Open(FLAGS.predict_file) as predict_file:
prediction_json = json.load(predict_file)["data"]
eval_examples = squad_utils.read_squad_examples(
input_file=FLAGS.predict_file, is_training=False)
if (tf.gfile.Exists(FLAGS.predict_feature_file)
and tf.gfile.Exists(FLAGS.predict_feature_left_file)):
tf.logging.info("Loading eval features from {}".format(
FLAGS.predict_feature_left_file))
with tf.gfile.Open(FLAGS.predict_feature_left_file, "rb") as fin:
eval_features = pickle.load(fin)
else:
eval_writer = squad_utils.FeatureWriter(
filename=FLAGS.predict_feature_file, is_training=False)
eval_features = []
def append_feature(feature):
eval_features.append(feature)
eval_writer.process_feature(feature)
squad_utils.convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=FLAGS.max_seq_length,
doc_stride=FLAGS.doc_stride,
max_query_length=FLAGS.max_query_length,
is_training=False,
output_fn=append_feature,
do_lower_case=FLAGS.do_lower_case)
eval_writer.close()
with tf.gfile.Open(FLAGS.predict_feature_left_file, "wb") as fout:
pickle.dump(eval_features, fout)
tf.logging.info("***** Running predictions *****")
tf.logging.info(" Num orig examples = %d", len(eval_examples))
tf.logging.info(" Num split examples = %d", len(eval_features))
tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size)
predict_input_fn = squad_utils.input_fn_builder(
input_file=FLAGS.predict_feature_file,
seq_length=FLAGS.max_seq_length,
is_training=False,
drop_remainder=False,
use_tpu=FLAGS.use_tpu,
bsz=FLAGS.predict_batch_size,
is_v2=True)
def get_result(checkpoint):
"""Evaluate the checkpoint on SQuAD v2.0."""
# If running eval on the TPU, you will need to specify the number of
# steps.
reader = tf.train.NewCheckpointReader(checkpoint)
global_step = reader.get_tensor(tf.GraphKeys.GLOBAL_STEP)
all_results = []
for result in estimator.predict(predict_input_fn,
yield_single_examples=True,
checkpoint_path=checkpoint):
if len(all_results) % 1000 == 0:
tf.logging.info("Processing example: %d" %
(len(all_results)))
unique_id = int(result["unique_ids"])
start_top_log_probs = ([
float(x) for x in result["start_top_log_probs"].flat
])
start_top_index = [
int(x) for x in result["start_top_index"].flat
]
end_top_log_probs = ([
float(x) for x in result["end_top_log_probs"].flat
])
end_top_index = [int(x) for x in result["end_top_index"].flat]
cls_logits = float(result["cls_logits"].flat[0])
all_results.append(
squad_utils.RawResultV2(
unique_id=unique_id,
start_top_log_probs=start_top_log_probs,
start_top_index=start_top_index,
end_top_log_probs=end_top_log_probs,
end_top_index=end_top_index,
cls_logits=cls_logits))
output_prediction_file = os.path.join(FLAGS.output_dir,
"predictions.json")
output_nbest_file = os.path.join(FLAGS.output_dir,
"nbest_predictions.json")
output_null_log_odds_file = os.path.join(FLAGS.output_dir,
"null_odds.json")
result_dict = {}
cls_dict = {}
squad_utils.accumulate_predictions_v2(
result_dict, cls_dict, eval_examples, eval_features,
all_results, FLAGS.n_best_size, FLAGS.max_answer_length,
FLAGS.start_n_top, FLAGS.end_n_top)
return squad_utils.evaluate_v2(
result_dict, cls_dict, prediction_json, eval_examples,
eval_features, all_results, FLAGS.n_best_size,
FLAGS.max_answer_length, output_prediction_file,
output_nbest_file, output_null_log_odds_file), int(global_step)
def _find_valid_cands(curr_step):
filenames = tf.gfile.ListDirectory(FLAGS.output_dir)
candidates = []
for filename in filenames:
if filename.endswith(".index"):
ckpt_name = filename[:-6]
idx = ckpt_name.split("-")[-1]
if idx != "best" and int(idx) > curr_step:
candidates.append(filename)
return candidates
output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt")
checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best")
key_name = "f1"
writer = tf.gfile.GFile(output_eval_file, "w")
avg_time_per_batch = np.mean(time_hist.times)
writer.write("===== Hyperparameters =====\n")
writer.write("Training batch size: {}\n".format(
FLAGS.train_batch_size))
writer.write("Max sequence length: {}\n".format(FLAGS.max_seq_length))
writer.write("Learning rate: {}\n".format(FLAGS.learning_rate))
writer.write("Num of GPU cores: {}\n".format(NUM_GPUS))
if FLAGS.do_train:
avg_time_per_batch = np.mean(time_hist.times)
writer.write("Total time: {}\n".format(total_time))
writer.write("Speed: {}\n".format(FLAGS.train_batch_size *
NUM_GPUS / avg_time_per_batch))
if num_train_steps and num_warmup_steps:
writer.write("Training steps: {}\n".format(num_train_steps))
writer.write("Warmup steps: {}\n".format(num_warmup_steps))
if tf.gfile.Exists(checkpoint_path + ".index"):
result = get_result(checkpoint_path)
best_perf = result[0][key_name]
global_step = result[1]
else:
global_step = -1
best_perf = -1
checkpoint_path = None
while global_step < num_train_steps:
steps_and_files = {}
filenames = tf.gfile.ListDirectory(FLAGS.output_dir)
for filename in filenames:
if filename.endswith(".index"):
ckpt_name = filename[:-6]
cur_filename = os.path.join(FLAGS.output_dir, ckpt_name)
if cur_filename.split("-")[-1] == "best":
continue
gstep = int(cur_filename.split("-")[-1])
if gstep not in steps_and_files:
tf.logging.info(
"Add {} to eval list.".format(cur_filename))
steps_and_files[gstep] = cur_filename
tf.logging.info("found {} files.".format(len(steps_and_files)))
if not steps_and_files:
tf.logging.info(
"found 0 file, global step: {}. Sleeping.".format(
global_step))
time.sleep(1)
else:
for ele in sorted(steps_and_files.items()):
step, checkpoint_path = ele
if global_step >= step:
if len(_find_valid_cands(step)) > 1:
for ext in [
"meta", "data-00000-of-00001", "index"
]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tf.logging.info("removing {}".format(src_ckpt))
tf.gfile.Remove(src_ckpt)
continue
result, global_step = get_result(checkpoint_path)
tf.logging.info("***** Eval results *****")
for key in sorted(result.keys()):
tf.logging.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
if result[key_name] > best_perf:
best_perf = result[key_name]
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tgt_ckpt = checkpoint_path.rsplit(
"-", 1)[0] + "-best.{}".format(ext)
tf.logging.info("saving {} to {}".format(
src_ckpt, tgt_ckpt))
tf.gfile.Copy(src_ckpt, tgt_ckpt, overwrite=True)
writer.write("saved {} to {}\n".format(
src_ckpt, tgt_ckpt))
writer.write("best {} = {}\n".format(key_name, best_perf))
tf.logging.info(" best {} = {}\n".format(
key_name, best_perf))
if len(_find_valid_cands(global_step)) > 2:
for ext in ["meta", "data-00000-of-00001", "index"]:
src_ckpt = checkpoint_path + ".{}".format(ext)
tf.logging.info("removing {}".format(src_ckpt))
tf.gfile.Remove(src_ckpt)
writer.write("=" * 50 + "\n")
checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best")
result, global_step = get_result(checkpoint_path)
tf.logging.info("***** Final Eval results *****")
tf.logging.info(f"num_gpu_cores = {NUM_GPUS}")
writer.write("===== Evuations =====\n")
for key in sorted(result.keys()):
tf.logging.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
writer.write("best perf happened at step: {}".format(global_step))
def save_as_record(path, data, vocab_path, max_length, fields):
path_tag_size = [40, 10, 100]
max_ner_size = 64
events = {
'EquityFreeze': 0,
'EquityRepurchase': 1,
'EquityUnderweight': 2,
'EquityOverweight': 3,
'EquityPledge': 4
}
events_fields = {
'EquityFreeze': [
'EquityHolder', 'FrozeShares', 'LegalInstitution',
'TotalHoldingShares', 'TotalHoldingRatio', 'StartDate', 'EndDate',
'UnfrozeDate'
],
'EquityRepurchase': [
'CompanyName', 'HighestTradingPrice', 'LowestTradingPrice',
'RepurchasedShares', 'ClosingDate', 'RepurchaseAmount'
],
'EquityUnderweight': [
'EquityHolder', 'TradedShares', 'StartDate', 'EndDate',
'LaterHoldingShares', 'AveragePrice'
],
'EquityOverweight': [
'EquityHolder', 'TradedShares', 'StartDate', 'EndDate',
'LaterHoldingShares', 'AveragePrice'
],
'EquityPledge': [
'Pledger', 'PledgedShares', 'Pledgee', 'TotalHoldingShares',
'TotalHoldingRatio', 'TotalPledgedShares', 'StartDate', 'EndDate',
'ReleasedDate'
]
}
train_writer = tf.python_io.TFRecordWriter(path)
vocab = tokenization.FullTokenizer(vocab_path)
for x in tqdm(data):
# 处理原文
sentences, sentences_mask = process(x['sentences'], vocab, max_length)
# 处理ner 标签
# 空白的ner_tag
ner_tag = np.zeros(max_length, dtype=np.int32)
flag = 0
for w in x['ann_mspan2dranges'].keys():
field = w
tag = x['ann_mspan2guess_field'][field]
indexs = x['ann_mspan2dranges'][field]
for index in indexs:
if index[2] > max_length[1] - 1:
flag = 1
break
ner_tag[index[0]][index[1] + 1] = fields['%s_B' %
str(tag).upper()]
for i in range(index[1] + 1, index[2]):
ner_tag[index[0]][i + 1] = fields['%s_I' %
str(tag).upper()]
if flag == 1:
continue
# 生成路径tag
# 维度分别是路径数，字段数，候选值
path_tag = np.zeros(path_tag_size, dtype=np.int32) + (-1)
# 存储完整路径
path_entity_list = np.zeros(path_tag_size[:2], dtype=np.int32) + (-1)
path_event_type = np.zeros([path_tag_size[0]], dtype=np.int32) + (-1)
event_tag = [0 for i in range(5)]
# 辅助数据：实体index
ners = [k for k in x['ann_mspan2dranges'].keys()]
ner_index = []
ner_list_index = [0]
for k in ners:
ner_index.extend(x['ann_mspan2dranges'][k])
# ner_index[-1][1] += 1
# ner_index[-1][2] += 1
ner_list_index.append(ner_list_index[-1] +
len(x['ann_mspan2dranges'][k]))
for k in ner_index:
k[1] += 1
k[2] += 1
for i in range(max_ner_size - len(ner_index)):
ner_index.append([-1, -1, -1])
for i in range(max_ner_size - len(ner_list_index)):
ner_list_index.append(-1)
event_tree = {}
for e in x['recguid_eventname_eventdict_list']:
# 处理事件tag
event_tag[events[e[1]]] = 1
if events[e[1]] not in event_tree.keys():
event_tree[events[e[1]]] = {}
et = event_tree[events[e[1]]]
# 合并相同前缀
for f in events_fields[e[1]]:
value = e[2][f]
if value == None:
value = 'NA'
if value not in et.keys():
et[value] = {}
et = et[value]
# 创建路径tag
path_type = np.zeros([10], dtype=np.int32) + (-1)
paths = []
for index, k in enumerate(event_tree.keys()):
start = len(paths)
paths.extend(get_path(event_tree[k]))
path_type[start:len(paths)] = k
for index2, path in enumerate(paths[start:]):
cache = event_tree[k][path[0]]
# 跳过第一个节点
if ners.index(path[0]) != 'NA':
path_entity_list[start + index2,
0] = ners.index(path[0]) + 1
else:
path_entity_list[start + index2, 0] = 0
for i, p in enumerate(path[1:]):
tag = np.array(
[0 if f not in cache.keys() else 1 for f in ners],
dtype=np.int32)
tag = np.concatenate([
tag,
np.zeros([path_tag_size[-1] - tag.size],
dtype=np.int32)
],
axis=0)
path_tag[start + index2, i + 1, :] = tag
if p != 'NA':
path_entity_list[start + index2,
i + 1] = ners.index(p) + 1
else:
path_entity_list[start + index2, i + 1] = 0
cache = cache[p]
path_event_type[start + index2] = k
tag = np.array(
[0 if f not in [c[0] for c in paths] else 1 for f in ners],
dtype=np.int32)
tag = np.concatenate([
tag,
np.zeros([path_tag_size[-1] - tag.size], dtype=np.int32)
],
axis=0)
path_tag[start:len(paths), 0, :] = tag
if len(ner_list_index) != max_ner_size:
continue
if len(ner_index) != max_ner_size:
continue
# # test
# def select_path(path_tag, path_num, path_event_type, path_entity_list):
# path_index = np.random.randint(0, path_num[0], size=1, dtype=np.int32)[0]
# return path_tag[path_index], path_index, path_event_type[path_index], path_entity_list[path_index]
#
# path_tag, path_index, path_event_type, path_entity_list = select_path(path_tag, [len(paths)], path_event_type,
# path_entity_list)
#
# # 去除padding的ner_index
# def select_nert_index(path_entity_list):
# size2 = path_entity_list.argmin(axis=0)
# return path_entity_list[:size2]
#
# path_entity_list = select_nert_index(path_entity_list)
features = tf.train.Features(
feature={
'sentences': get_byte_feature(sentences), # 原始文本
'sentences_mask': get_byte_feature(sentences_mask), # 原始文本长度
'event_tag': get_byte_feature(event_tag), # 事件标签
'ner_tag': get_byte_feature(ner_tag), # 实体标签
'path_tag': get_byte_feature(path_tag), # 路径标签
'ner_list_index': get_byte_feature(ner_list_index), #
'ner_index': get_byte_feature(ner_index), #
'path_event_type': get_int_feature(path_event_type),
'path_num': get_int_feature([len(paths)]),
'path_entity_list': get_byte_feature(path_entity_list)
# 'mask1': tf.train.Feature(int64_list=tf.train.Int64List(value=mask1)),
})
example = tf.train.Example(features=features)
train_writer.write(example.SerializeToString())
train_writer.close()
def transformer(model: str = 'bert', **kwargs):
"""
Load Transformer sentiment model.
Parameters
----------
model : str, optional (default='bert')
Model architecture supported. Allowed values:
* ``'bert'`` - BERT architecture from google.
* ``'tiny-bert'`` - BERT architecture from google with smaller parameters.
* ``'albert'`` - ALBERT architecture from google.
* ``'tiny-albert'`` - ALBERT architecture from google with smaller parameters.
* ``'xlnet'`` - XLNET architecture from google.
* ``'alxlnet'`` - XLNET architecture from google + Malaya.
Returns
-------
BERT : malaya._models._bert_model.BINARY_BERT class
"""
model = model.lower()
if model not in _availability:
raise Exception(
'model not supported, please check supported models from malaya.similarity.available_transformer_model()'
)
check_file(PATH_SIMILARITY[model], S3_PATH_SIMILARITY[model], **kwargs)
g = load_graph(PATH_SIMILARITY[model]['model'])
path = PATH_SIMILARITY
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
from malaya.transformers.bert import (
_extract_attention_weights_import, )
from malaya.transformers.bert import bert_num_layers
tokenizer = sentencepiece_tokenizer_bert(path[model]['tokenizer'],
path[model]['vocab'])
if model in ['albert', 'tiny-albert']:
from malaya.transformers.albert import (
_extract_attention_weights_import, )
from malaya.transformers.albert import bert_num_layers
from albert import tokenization
tokenizer = tokenization.FullTokenizer(
vocab_file=path[model]['vocab'],
do_lower_case=False,
spm_model_file=path[model]['tokenizer'],
)
return SIAMESE_BERT(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=g.get_tensor_by_name('import/Placeholder_1:0'),
input_masks=g.get_tensor_by_name('import/Placeholder_2:0'),
logits=g.get_tensor_by_name('import/logits:0'),
sess=generate_session(graph=g),
tokenizer=tokenizer,
label=['not similar', 'similar'],
)
if model in ['xlnet', 'alxlnet']:
if model in ['xlnet']:
from malaya.transformers.xlnet import (
_extract_attention_weights_import, )
if model in ['alxlnet']:
from malaya.transformers.alxlnet import (
_extract_attention_weights_import, )
tokenizer = sentencepiece_tokenizer_xlnet(path[model]['tokenizer'])
return SIAMESE_XLNET(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=g.get_tensor_by_name('import/Placeholder_1:0'),
input_masks=g.get_tensor_by_name('import/Placeholder_2:0'),
logits=g.get_tensor_by_name('import/logits:0'),
sess=generate_session(graph=g),
tokenizer=tokenizer,
label=['not similar', 'similar'],
)
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
layer_indexes = [int(x) for x in FLAGS.layers.split(",")]
bert_config = modeling.AlbertConfig.from_json_file(FLAGS.bert_config_file)
tokenizer = tokenization.FullTokenizer(
spm_model_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case)
is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2
run_config = tf.contrib.tpu.RunConfig(
master=FLAGS.master,
tpu_config=tf.contrib.tpu.TPUConfig(
num_shards=FLAGS.num_tpu_cores,
per_host_input_for_training=is_per_host))
# examples = read_examples(FLAGS.input_file)
json_examples = []
for x in ['test', 'train', 'dev']:
with open(os.path.join(FLAGS.input_file, x + '.english.jsonlines')) as f:
json_examples.extend((json.loads(jsonline) for jsonline in f.readlines()))
orig_examples = []
bert_examples = []
for i, json_e in enumerate(json_examples):
e = process_example(json_e, i, should_filter_embedded_mentions=True)
orig_examples.append(e)
bert_examples.append(e.bertify(tokenizer))
model_fn = model_fn_builder(
bert_config=bert_config,
init_checkpoint=FLAGS.init_checkpoint,
layer_indexes=layer_indexes,
use_tpu=FLAGS.use_tpu,
use_one_hot_embeddings=FLAGS.use_one_hot_embeddings)
# If TPU is not available, this will fall back to normal Estimator on CPU
# or GPU.
estimator = tf.contrib.tpu.TPUEstimator(
use_tpu=FLAGS.use_tpu,
model_fn=model_fn,
config=run_config,
predict_batch_size=FLAGS.batch_size)
input_fn = input_fn_builder(
examples=bert_examples, window_size=FLAGS.window_size, stride=FLAGS.stride, tokenizer=tokenizer)
writer = h5py.File(FLAGS.output_file, 'w')
with tqdm(total=sum(len(e.tokens) for e in orig_examples)) as t:
for result in estimator.predict(input_fn, yield_single_examples=True):
document_index = int(result["unique_ids"])
bert_example = bert_examples[document_index]
orig_example = orig_examples[document_index]
file_key = bert_example.doc_key.replace('/', ':')
t.update(n=(result['extract_indices'] >= 0).sum())
for output_index, bert_token_index in enumerate(result['extract_indices']):
if bert_token_index < 0:
continue
token_index = bert_example.bert_to_orig_map[bert_token_index]
sentence_index, token_index = orig_example.unravel_token_index(token_index)
dataset_key ="{}/{}".format(file_key, sentence_index)
if dataset_key not in writer:
writer.create_dataset(dataset_key,
(len(orig_example.sentence_tokens[sentence_index]), bert_config.hidden_size, len(layer_indexes)),
dtype=np.float32)
dset = writer[dataset_key]
for j, layer_index in enumerate(layer_indexes):
layer_output = result["layer_output_%d" % j]
dset[token_index, :, j] = layer_output[output_index]
writer.close()
def transformer(
path,
s3_path,
class_name,
label,
model='bert',
quantized=False,
**kwargs,
):
check_file(path[model], s3_path[model], quantized=quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
if len(label) > 2 or class_name == 'relevancy':
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
selected_class = MULTICLASS_BERT
selected_node = 'import/dense/BiasAdd:0'
if model in ['xlnet', 'alxlnet']:
selected_class = MULTICLASS_XLNET
selected_node = 'import/transpose_3:0'
else:
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
selected_class = BINARY_BERT
selected_node = 'import/dense/BiasAdd:0'
if model in ['xlnet', 'alxlnet']:
selected_class = BINARY_XLNET
selected_node = 'import/transpose_3:0'
if model in ['albert', 'bert', 'tiny-albert', 'tiny-bert']:
if model in ['bert', 'tiny-bert']:
from malaya.transformers.bert import (
_extract_attention_weights_import, )
from malaya.transformers.bert import bert_num_layers
tokenizer = sentencepiece_tokenizer_bert(path[model]['tokenizer'],
path[model]['vocab'])
if model in ['albert', 'tiny-albert']:
from malaya.transformers.albert import (
_extract_attention_weights_import, )
from malaya.transformers.albert import bert_num_layers
from albert import tokenization
tokenizer = tokenization.FullTokenizer(
vocab_file=path[model]['vocab'],
do_lower_case=False,
spm_model_file=path[model]['tokenizer'],
)
return selected_class(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=None,
input_masks=g.get_tensor_by_name('import/Placeholder_1:0'),
logits=g.get_tensor_by_name('import/logits:0'),
logits_seq=g.get_tensor_by_name('import/logits_seq:0'),
vectorizer=g.get_tensor_by_name(selected_node),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
label=label,
attns=_extract_attention_weights_import(bert_num_layers[model], g),
class_name=class_name,
)
if model in ['xlnet', 'alxlnet']:
if model in ['xlnet']:
from malaya.transformers.xlnet import (
_extract_attention_weights_import, )
if model in ['alxlnet']:
from malaya.transformers.alxlnet import (
_extract_attention_weights_import, )
tokenizer = sentencepiece_tokenizer_xlnet(path[model]['tokenizer'])
return selected_class(
X=g.get_tensor_by_name('import/Placeholder:0'),
segment_ids=g.get_tensor_by_name('import/Placeholder_1:0'),
input_masks=g.get_tensor_by_name('import/Placeholder_2:0'),
logits=g.get_tensor_by_name('import/logits:0'),
logits_seq=g.get_tensor_by_name('import/logits_seq:0'),
vectorizer=g.get_tensor_by_name(selected_node),
sess=generate_session(graph=g, **kwargs),
tokenizer=tokenizer,
label=label,
attns=_extract_attention_weights_import(g),
class_name=class_name,
)
def trans_to_features(self, example):
"""
将输入转化为索引表示
:param example: 输入
:return:
"""
tokenizer = tokenization.FullTokenizer(vocab_file=self.vocab_path, do_lower_case=True)
features = []
unique_id = 1000000000
query_tokens = tokenizer.tokenize(example['question'])
# 给定query一个最大长度来控制query的长度
if len(query_tokens) > self.query_length:
query_tokens = query_tokens[: self.query_length]
# 主要是针对context构造索引，之前我们将中文，标点符号，空格，一连串的数字，英文单词分割存储在doc_tokens中
# 但在bert的分词器中会将一连串的数字，中文，英文等分割成子词，也就是说经过bert的分词之后得到的tokens和之前
# 获得的doc_tokens是不一样的，因此我们仍需要对start和end position从doc_tokens中的位置映射到当前tokens的位置
tok_to_orig_index = [] # 存储未分词的token的索引，但长度和下面的相等
orig_to_tok_index = [] # 存储分词后的token的索引，但索引不是连续的，会存在跳跃的情况
all_doc_tokens = [] # 存储分词后的token，理论上长度是要大于all_tokens的
for (i, token) in enumerate(example['doc_tokens']):
sub_tokens = tokenizer.tokenize(token)
# orig_to_tok_index的长度等于doc_tokens，里面每个值存储的是doc_tokens中的token在all_doc_tokens中的起止索引值
# 用来将在all_token中的start和end转移到all_doc_tokens中
orig_to_tok_index.append([len(all_doc_tokens)])
for sub_token in sub_tokens:
# tok_to_orig_index的长度等于all_doc_tokens, 里面会有重复的值
tok_to_orig_index.append(i)
all_doc_tokens.append(sub_token)
orig_to_tok_index[-1].append(len(all_doc_tokens) - 1)
# The -3 accounts for [CLS], [SEP] and [SEP]
max_tokens_for_doc = self.max_length - len(query_tokens) - 3
doc_spans = []
_DocSpan = collections.namedtuple("DocSpan", ["start", "length"])
# 在使用bert的时候，一般会将最大的序列长度控制在512，因此对于长度大于最大长度的context，我们需要将其分成多个片段
# 采用滑窗的方式，滑窗大小是小于最大长度的，因此分割的片段之间是存在重复的子片段。
start_offset = 0 # 截取的片段的起始位置
while start_offset < len(all_doc_tokens):
length = len(all_doc_tokens) - start_offset
# 当长度超标，需要使用滑窗
if length > max_tokens_for_doc:
length = max_tokens_for_doc
doc_spans.append(_DocSpan(start=start_offset, length=length))
if start_offset + length == len(all_doc_tokens): # 当length < max_len时，该条件成立
break
start_offset += min(length, self.doc_stride)
# 组合query和context的片段成一个序列输入到bert中
for (doc_span_index, doc_span) in enumerate(doc_spans):
tokens = []
token_to_orig_map = {}
# 因为片段之间会存在重复的子片段，但是子片段中的token在不同的片段中的重要性是不一样的，
# 在这里根据上下文的数量来决定token的重要性，在之后预测时对于出现在两个片段中的token，只取重要性高的片段
# 中的token的分数作为该token的分数
token_is_max_context = {}
segment_ids = []
tokens.append("[CLS]")
segment_ids.append(0)
for token in query_tokens:
tokens.append(token)
segment_ids.append(0)
tokens.append("[SEP]")
segment_ids.append(0)
for i in range(doc_span.length):
split_token_index = doc_span.start + i
token_to_orig_map[len(tokens)] = tok_to_orig_index[
split_token_index] # 映射当前span组成的句子对的索引到原始token的索引
# 在利用滑窗分割多个span时会存在有的词出现在两个span中，但最后统计的时候，我们只能选择一个span，因此
# 作者根据该词上下文词的数量构建了一个分数，取分数最高的那个span
is_max_context = self._check_is_max_context(doc_spans, doc_span_index, split_token_index)
token_is_max_context[len(tokens)] = is_max_context
tokens.append(all_doc_tokens[split_token_index])
segment_ids.append(1)
tokens.append("[SEP]")
segment_ids.append(1)
input_ids = tokenizer.convert_tokens_to_ids(tokens)
# The mask has 1 for real tokens and 0 for padding tokens. Only real
# tokens are attended to.
input_mask = [1] * len(input_ids)
# Zero-pad up to the sequence length.
while len(input_ids) < self.max_length:
input_ids.append(0)
input_mask.append(0)
segment_ids.append(0)
assert len(input_ids) == self.max_length
assert len(input_mask) == self.max_length
assert len(segment_ids) == self.max_length
features.append({'unique_id': unique_id,
'doc_span_index': doc_span_index,
'tokens': tokens,
'token_to_orig_map': token_to_orig_map,
'token_is_max_context': token_is_max_context,
'input_ids': input_ids,
'input_mask': input_mask,
'segment_ids': segment_ids,
'start_position': -1,
'end_position': -1})
unique_id += 1
return features