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)
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, FLAGS.output_file,
rng)
elif six.PY2:
if isinstance(text, str):
return text
elif isinstance(text, six.text_type):
return six.ensure_binary(text, "utf-8")
else:
raise ValueError("Unsupported string type: %s" % (type(text)))
else:
raise ValueError("Not running on Python2 or Python 3?")
if __name__ == "__main__":
text_ = "我喜欢篮球"
bpe_tokenizer = tokenization.FullTokenizer(
vocab_file="data_proc/tokenizers/sentencepiece/char_segmented_lower-21128-clean.vocab",
do_lower_case=True,
spm_model_file="data_proc/tokenizers/sentencepiece/char_segmented_lower-21128-clean.model")
text_seg = char2char_single_sent(
text_,
tokenizer=bpe_tokenizer
)
print("vocab: ", len(bpe_tokenizer.vocab))
print(text_seg)
print(" ".join(text_seg))
bpe_tokenizer = tokenization.FullTokenizer(
vocab_file="data_proc/tokenizers/sentencepiece/char_segmented_lower-5282-clean.vocab",
do_lower_case=True,
spm_model_file="data_proc/tokenizers/sentencepiece/char_segmented_lower-5282-clean.model")
text_seg = char2char_single_sent(
text_,
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
processors = {
"chn": data_processors.ChnSentiCorpDataProcessor,
"lcqmc": data_processors.LCQMCProcessor,
"xnli": data_processors.XnliProcessor,
"book_review": data_processors.BookReviewProcessor,
"shopping": data_processors.ShoppingProcessor,
"weibo": data_processors.WeiboProcessor,
"law_qa": data_processors.LawQAProcessor,
"nlpcc_dbqa": data_processors.NlpccDbqaProcessor,
}
# 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."
)
if not FLAGS.albert_config_file and not FLAGS.albert_hub_module_handle:
raise ValueError("At least one of `--albert_config_file` and "
"`--albert_hub_module_handle` 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)
task_name = FLAGS.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % task_name)
processor = processors[task_name](FLAGS)
label_list = processor.get_labels()
bpe_tokenizer = tokenization.FullTokenizer(
vocab_file=FLAGS.vocab_file,
do_lower_case=FLAGS.do_lower_case,
spm_model_file=FLAGS.spm_model_file)
dict_char2comp = json.load(
open(FLAGS.dict_char2comp_dir, "r", encoding="utf-8"))
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,
hub_module=FLAGS.albert_hub_module_handle,
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,
bpe_tokenizer,
train_file,
task_name,
dict_char2comp=dict_char2comp)
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(data_processors.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,
bpe_tokenizer,
eval_file,
task_name,
dict_char2comp=dict_char2comp)
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)
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")
if task_name == "sts-b":
key_name = "pearson"
elif task_name == "cola":
key_name = "matthew_corr"
elif task_name == "nlpcc_dbqa":
key_name = "f1_score"
else:
key_name = "eval_accuracy"
global_step, best_perf_global_step, best_perf = _best_trial_info()
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(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[key_name] > best_perf:
best_perf = result[key_name]
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()
# output_eval_file_local = os.path.join("./tmp/", "_".join(output_eval_file.split("/")[4:]))
# tf.gfile.Copy(output_eval_file, output_eval_file_local, overwrite=True)
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:
# 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(data_processors.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,
bpe_tokenizer,
predict_file,
task_name,
dict_char2comp=dict_char2comp)
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
truth_json_dir = os.path.join(FLAGS.data_dir, "test.json")
pred_tsv_dir = output_submit_file
accuracy, results = cal_metrics(pred_tsv_dir, truth_json_dir)
tf.logging.info("***** Predict metrics *****")
tf.logging.info("***** Predict metrics *****")
tf.logging.info("accuracy: %f" % accuracy)
tf.logging.info("results: ")
tf.logging.info(json.dumps(results, ensure_ascii=False, indent=2))
tf.logging.info("***** Predict metrics *****")
tf.logging.info("***** Predict metrics *****")