def __init__(self,
model_config,
num_labels,
init_checkpoint,
load_pretrained=True,
model_io_config={},
opt_config={},
exclude_scope="",
not_storage_params=[],
target="a",
label_lst=None,
output_type="sess",
**kargs):
self.model_config = model_config
self.num_labels = num_labels
self.init_checkpoint = init_checkpoint
self.load_pretrained = load_pretrained
self.model_io_config = model_io_config
self.opt_config = opt_config
self.exclude_scope = exclude_scope
self.not_storage_params = not_storage_params
self.target = target
self.label_lst = label_lst
self.output_type = output_type
self.kargs = kargs
self.model_io_fn = model_io.ModelIO(self.model_io_config)
self.optimizer_fn = optimizer.Optimizer(self.opt_config)
def model_fn(features, labels, mode, params):
model_api = model_zoo(model_config)
model = model_api(model_config,
features,
labels,
mode,
target,
reuse=tf.AUTO_REUSE,
**kargs)
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
logits = global_discriminator_logits(model_config,
model.get_pooled_output(),
reuse=tf.AUTO_REUSE,
**kargs)
model_io_fn = model_io.ModelIO(model_io_config)
pretrained_tvars = model_io_fn.get_params(
model_config.scope, not_storage_params=not_storage_params)
global_prediction_tvars = model_io_fn.get_params(
"cls/seq_global", not_storage_params=not_storage_params)
pretrained_tvars.extend(global_prediction_tvars)
tvars = pretrained_tvars
print('==discriminator parameters==', tvars)
if load_pretrained == "yes":
use_tpu = 1 if kargs.get('use_tpu', False) else 0
scaffold_fn = model_io_fn.load_pretrained(
tvars,
init_checkpoint,
exclude_scope=exclude_scope,
use_tpu=use_tpu)
else:
scaffold_fn = None
if mode == tf.estimator.ModeKeys.PREDICT:
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
predictions={"probs": tf.nn.softmax(logits)},
export_outputs={
"output":
tf.estimator.export.PredictOutput(
{"probs": tf.nn.softmax(logits)})
})
return estimator_spec
def export_model_v2(config):
opt_config = Bunch({"init_lr":2e-5, "num_train_steps":1e30, "cycle":False})
model_io_config = Bunch({"fix_lm":False})
bert_config = json.load(open(config["config_file"], "r"))
model_config = Bunch(bert_config)
model_config.use_one_hot_embeddings = True
model_config.scope = "bert"
model_config.dropout_prob = 0.1
model_config.label_type = "single_label"
with open(config["label2id"], "r") as frobj:
label_dict = json.load(frobj)
num_classes = len(label_dict["id2label"])
max_seq_length = config["max_length"]
def serving_input_receiver_fn():
label_ids = tf.placeholder(tf.int32, [None], name='label_ids')
input_ids = tf.placeholder(tf.int32, [None, max_seq_length], name='input_ids')
input_mask = tf.placeholder(tf.int32, [None, max_seq_length], name='input_mask')
segment_ids = tf.placeholder(tf.int32, [None, max_seq_length], name='segment_ids')
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'label_ids': label_ids,
'input_ids': input_ids,
'input_mask': input_mask,
'segment_ids': segment_ids
})()
return input_fn
model_io_fn = model_io.ModelIO(model_io_config)
model_fn = bert_classifier_estimator.classifier_model_fn_builder(
model_config,
num_classes,
config["init_checkpoint"],
reuse=None,
load_pretrained=True,
model_io_fn=model_io_fn,
model_io_config=model_io_config,
opt_config=opt_config)
estimator = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=config["model_dir"])
export_dir = estimator.export_savedmodel(config["export_path"],
serving_input_receiver_fn,
checkpoint_path=config["init_checkpoint"])
print("===Succeeded in exporting saved model==={}".format(export_dir))
def model_fn(features, labels, mode):
model_api = model_zoo(model_config)
model = model_api(model_config,
features,
labels,
mode,
target,
reuse=model_reuse)
label_ids = features["label_ids"]
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
with tf.variable_scope(scope):
(loss, per_example_loss,
logits) = classifier.classifier(model_config,
model.get_pooled_output(),
num_labels, label_ids,
dropout_prob)
model_io_fn = model_io.ModelIO(model_io_config)
tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
try:
params_size = model_io_fn.count_params(model_config.scope)
print("==total params==", params_size)
except:
print("==not count params==")
if load_pretrained == "yes":
model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope='teacher')
return_dict = {
"loss": loss,
"logits": logits,
"tvars": tvars,
"model": model,
"per_example_loss": per_example_loss
}
return return_dict
def model_fn(features, labels, mode, params):
model_api = model_zoo(model_config)
model = model_api(model_config, features, labels,
mode, target, reuse=tf.AUTO_REUSE,
**kargs)
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
logits = global_discriminator_logits(model_config,
model.get_pooled_output(),
reuse=tf.AUTO_REUSE, **kargs)
model_io_fn = model_io.ModelIO(model_io_config)
pretrained_tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
global_prediction_tvars = model_io_fn.get_params("cls/seq_global",
not_storage_params=not_storage_params)
pretrained_tvars.extend(global_prediction_tvars)
tvars = pretrained_tvars
print('==discriminator parameters==', tvars)
if load_pretrained == "yes":
use_tpu = 1 if kargs.get('use_tpu', False) else 0
scaffold_fn = model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope=exclude_scope,
use_tpu=use_tpu,
restore_var_name=model_config.get('restore_var_name', []))
else:
scaffold_fn = None
return_dict = {
"logits":logits,
"tvars":tvars,
"model":model
}
return return_dict
def init_model(self):
self.graph = tf.Graph()
with self.graph.as_default():
init_checkpoint = self.config["init_checkpoint"]
bert_config = json.load(open(self.config["bert_config"], "r"))
self.model_config = Bunch(bert_config)
self.model_config.use_one_hot_embeddings = True
self.model_config.scope = "bert"
self.model_config.dropout_prob = 0.1
self.model_config.label_type = "single_label"
self.input_queue = Queue(maxsize=self.config.get("batch_size", 20))
self.output_queue = Queue(
maxsize=self.config.get("batch_size", 20))
opt_config = Bunch({
"init_lr": 2e-5,
"num_train_steps": 1e30,
"cycle": False
})
model_io_config = Bunch({"fix_lm": False})
self.num_classes = len(self.label_dict["id2label"])
self.max_seq_length = self.config["max_length"]
self.tokenizer = tokenization.FullTokenizer(
vocab_file=self.config["bert_vocab"], do_lower_case=True)
self.sess = tf.Session()
self.model_io_fn = model_io.ModelIO(model_io_config)
model_fn = bert_classifier_estimator.classifier_model_fn_builder(
self.model_config,
self.num_classes,
init_checkpoint,
reuse=None,
load_pretrained=True,
model_io_fn=self.model_io_fn,
model_io_config=model_io_config,
opt_config=opt_config)
self.estimator = tf.estimator.Estimator(
model_fn=model_fn, model_dir=self.config["model_dir"])
def model_fn(features, labels, mode):
original_loss = tf.constant(0.0)
distilled_loss = tf.constant(0.0)
st_model = st_model_fn(model_config_dict['student'],
num_labels_dict['student'],
init_checkpoint_dict['student'],
model_reuse=None,
load_pretrained=load_pretrained_dict['student'],
model_io_config=model_io_config,
opt_config=opt_config,
exclude_scope=exclude_scope_dict.get('student', ""),
not_storage_params=not_storage_params_dict.get('student', []),
target=target_dict['student'],
**kargs)
st_dict = st_model(features, labels, mode)
ta_model = ta_model_fn(model_config_dict['teacher'],
num_labels_dict['teacher'],
init_checkpoint_dict['teacher'],
model_reuse=None,
load_pretrained=load_pretrained_dict['teacher'],
model_io_config=model_io_config,
opt_config=opt_config,
exclude_scope=exclude_scope_dict.get('teacher', ""),
not_storage_params=not_storage_params_dict.get('teacher', []),
target=target_dict['teacher'],
**kargs)
ta_dict = ta_model(features, labels, mode)
studnet_logit = st_dict['logits']
teacher_logit = ta_dict['logits']
model_io_fn = model_io.ModelIO(model_io_config)
feature_flag = False
original_loss += st_dict['loss'] * (distillation_config.get('ce_loss', 1.0))
print(distillation_config.get('ce_loss', 1.0), '===ce_loss===')
sess = tf.Session()
num_train_steps = int(FLAGS.train_size / FLAGS.batch_size * FLAGS.epoch)
num_warmup_steps = int(num_train_steps * 0.1)
num_storage_steps = int(FLAGS.train_size / FLAGS.batch_size)
opt_config = Bunch({
"init_lr": 1e-5,
"num_train_steps": num_train_steps,
"num_warmup_steps": num_warmup_steps
})
model_io_config = Bunch({"fix_lm": False})
model_io_fn = model_io.ModelIO(model_io_config)
num_choice = FLAGS.num_classes
max_seq_length = FLAGS.max_length
model_train_fn = bert_classifier.multichoice_model_fn_builder(
config,
num_choice,
init_checkpoint,
reuse=None,
load_pretrained=True,
model_io_fn=model_io_fn,
model_io_config=model_io_config,
opt_config=opt_config)
model_eval_fn = bert_classifier.multichoice_model_fn_builder(
def train_eval_fn(FLAGS, worker_count, task_index, is_chief, target,
init_checkpoint, train_file, dev_file, checkpoint_dir,
is_debug, **kargs):
graph = tf.Graph()
with graph.as_default():
import json
# config = json.load(open(FLAGS.config_file, "r"))
# config = Bunch(config)
# config.use_one_hot_embeddings = True
# config.scope = "bert"
# config.dropout_prob = 0.1
# config.label_type = "single_label"
# config.model = FLAGS.model_type
config = model_config_parser(FLAGS)
if FLAGS.if_shard == "0":
train_size = FLAGS.train_size
epoch = int(FLAGS.epoch / worker_count)
elif FLAGS.if_shard == "1":
print("==number of gpus==", kargs.get('num_gpus', 1))
train_size = int(FLAGS.train_size / worker_count /
kargs.get('num_gpus', 1))
# train_size = int(FLAGS.train_size)
epoch = FLAGS.epoch
else:
train_size = int(FLAGS.train_size / worker_count)
epoch = FLAGS.epoch
init_lr = FLAGS.init_lr
distillation_dict = json.load(tf.gfile.Open(FLAGS.distillation_config))
distillation_config = Bunch(
json.load(tf.gfile.Open(FLAGS.multi_task_config)))
warmup_ratio = config.get('warmup', 0.1)
num_train_steps = int(train_size / FLAGS.batch_size * epoch)
if config.get('ln_type', 'postln') == 'postln':
num_warmup_steps = int(num_train_steps * warmup_ratio)
elif config.get('ln_type', 'preln') == 'postln':
num_warmup_steps = 0
else:
num_warmup_steps = int(num_train_steps * warmup_ratio)
print('==num warmup steps==', num_warmup_steps)
num_storage_steps = min([int(train_size / FLAGS.batch_size), 10000])
if num_storage_steps <= 100:
num_storage_steps = 500
num_eval_steps = int(FLAGS.eval_size / FLAGS.batch_size)
if is_debug == "0":
num_storage_steps = 2
num_eval_steps = 10
num_train_steps = 10
print("num_train_steps {}, num_eval_steps {}, num_storage_steps {}".
format(num_train_steps, num_eval_steps, num_storage_steps))
print(" model type {}".format(FLAGS.model_type))
print(num_train_steps, num_warmup_steps, "=============",
kargs.get('num_gpus', 1), '==number of gpus==')
if worker_count * kargs.get("num_gpus", 1) >= 2:
clip_norm_scale = 1.0
lr_scale = 0.8
else:
clip_norm_scale = 1.0
lr_scale = 1.0
lr = init_lr * worker_count * kargs.get("num_gpus", 1) * lr_scale
if lr >= 1e-3:
lr = 1e-3
print('==init lr==', lr)
opt_config = Bunch({
"init_lr": lr,
"num_train_steps": num_train_steps,
"num_warmup_steps": num_warmup_steps,
"worker_count": worker_count,
"gpu_count": worker_count * kargs.get("num_gpus", 1),
"opt_type": FLAGS.opt_type,
"is_chief": is_chief,
"train_op": kargs.get("train_op", "adam"),
"decay": kargs.get("decay", "no"),
"warmup": kargs.get("warmup", "no"),
"clip_norm": config.get("clip_norm", 1.0),
"grad_clip": config.get("grad_clip", "global_norm"),
"epoch": FLAGS.epoch,
"strategy": FLAGS.distribution_strategy
})
anneal_config = Bunch({
"initial_value": 1.0,
"num_train_steps": num_train_steps
})
model_io_config = Bunch({"fix_lm": False})
model_io_fn = model_io.ModelIO(model_io_config)
num_classes = FLAGS.num_classes
if FLAGS.opt_type == "hvd" and hvd:
checkpoint_dir = checkpoint_dir if task_index == 0 else None
elif FLAGS.opt_type == "all_reduce":
checkpoint_dir = checkpoint_dir
elif FLAGS.opt_type == "collective_reduce":
checkpoint_dir = checkpoint_dir if task_index == 0 else None
elif FLAGS.opt_type == "ps" or FLAGS.opt_type == "ps_sync":
checkpoint_dir = checkpoint_dir if task_index == 0 else None
print("==checkpoint_dir==", checkpoint_dir, is_chief)
model_config_dict = {}
num_labels_dict = {}
init_checkpoint_dict = {}
load_pretrained_dict = {}
exclude_scope_dict = {}
not_storage_params_dict = {}
target_dict = {}
for task_type in FLAGS.multi_task_type.split(","):
print("==task type==", task_type)
model_config_dict[task_type] = model_config_parser(
Bunch(distillation_config[task_type]))
print(task_type, distillation_config[task_type],
'=====task model config======')
num_labels_dict[task_type] = distillation_config[task_type][
"num_labels"]
init_checkpoint_dict[task_type] = os.path.join(
FLAGS.buckets,
distillation_config[task_type]["init_checkpoint"])
load_pretrained_dict[task_type] = distillation_config[task_type][
"load_pretrained"]
exclude_scope_dict[task_type] = distillation_config[task_type][
"exclude_scope"]
not_storage_params_dict[task_type] = distillation_config[
task_type]["not_storage_params"]
target_dict[task_type] = distillation_config[task_type]["target"]
model_fn = distillation_model_fn(
model_config_dict,
num_labels_dict,
init_checkpoint_dict,
load_pretrained_dict,
model_io_config=model_io_config,
opt_config=opt_config,
exclude_scope_dict=exclude_scope_dict,
not_storage_params_dict=not_storage_params_dict,
target_dict=target_dict,
output_type="estimator",
distillation_config=distillation_dict,
**kargs)
name_to_features = data_interface(FLAGS)
def _decode_record(record, name_to_features):
"""Decodes a record to a TensorFlow example.
"""
example = tf.parse_single_example(record, name_to_features)
# tf.Example only supports tf.int64, but the TPU only supports tf.int32.
# So cast all int64 to int32.
for name in list(example.keys()):
t = example[name]
if t.dtype == tf.int64:
t = tf.to_int32(t)
example[name] = t
return example
def _decode_batch_record(record, name_to_features):
example = tf.parse_example(record, name_to_features)
# for name in list(example.keys()):
# t = example[name]
# if t.dtype == tf.int64:
# t = tf.to_int32(t)
# example[name] = t
return example
params = Bunch({})
params.epoch = FLAGS.epoch
params.batch_size = FLAGS.batch_size
if kargs.get("run_config", None):
if kargs.get("parse_type", "parse_single") == "parse_single":
train_features = lambda: tf_data_utils.all_reduce_train_input_fn(
train_file,
_decode_record,
name_to_features,
params,
if_shard=FLAGS.if_shard,
worker_count=worker_count,
task_index=task_index)
eval_features = lambda: tf_data_utils.all_reduce_eval_input_fn(
dev_file,
_decode_record,
name_to_features,
params,
if_shard=FLAGS.if_shard,
worker_count=worker_count,
task_index=task_index)
elif kargs.get("parse_type", "parse_single") == "parse_batch":
print("==apply parse example==")
train_features = lambda: tf_data_utils.all_reduce_train_batch_input_fn(
train_file,
_decode_batch_record,
name_to_features,
params,
if_shard=FLAGS.if_shard,
worker_count=worker_count,
task_index=task_index)
eval_features = lambda: tf_data_utils.all_reduce_eval_batch_input_fn(
dev_file,
_decode_batch_record,
name_to_features,
params,
if_shard=FLAGS.if_shard,
worker_count=worker_count,
task_index=task_index)
else:
train_features = lambda: tf_data_utils.train_input_fn(
train_file,
_decode_record,
name_to_features,
params,
if_shard=FLAGS.if_shard,
worker_count=worker_count,
task_index=task_index)
eval_features = lambda: tf_data_utils.eval_input_fn(
dev_file,
_decode_record,
name_to_features,
params,
if_shard=FLAGS.if_shard,
worker_count=worker_count,
task_index=task_index)
train_hooks = []
eval_hooks = []
sess_config = tf.ConfigProto(allow_soft_placement=False,
log_device_placement=False)
if FLAGS.opt_type == "ps" or FLAGS.opt_type == "ps_sync":
print("==no need for hook==")
elif FLAGS.opt_type == "pai_soar" and pai:
print("no need for hook")
elif FLAGS.opt_type == "hvd" and hvd:
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
print("==no need fo hook==")
else:
print("==no need for hooks==")
if kargs.get("run_config", None):
run_config = kargs.get("run_config", None)
run_config = run_config.replace(
save_checkpoints_steps=num_storage_steps)
print("==run config==", run_config.save_checkpoints_steps)
else:
run_config = tf.estimator.RunConfig(
model_dir=checkpoint_dir,
save_checkpoints_steps=num_storage_steps,
session_config=sess_config)
if kargs.get("profiler", "profiler") == "profiler":
if checkpoint_dir:
hooks = tf.train.ProfilerHook(
save_steps=100,
save_secs=None,
output_dir=os.path.join(checkpoint_dir, "profiler"),
)
train_hooks.append(hooks)
print("==add profiler hooks==")
model_estimator = tf.estimator.Estimator(model_fn=model_fn,
model_dir=checkpoint_dir,
config=run_config)
train_being_time = time.time()
tf.logging.info("==training distribution_strategy=={}".format(
kargs.get("distribution_strategy", "MirroredStrategy")))
if kargs.get("distribution_strategy",
"MirroredStrategy") == "MirroredStrategy":
print("==apply single machine multi-card training==")
train_spec = tf.estimator.TrainSpec(input_fn=train_features,
max_steps=num_train_steps)
eval_spec = tf.estimator.EvalSpec(input_fn=eval_features,
steps=num_eval_steps)
model_estimator.train(input_fn=train_features,
max_steps=num_train_steps,
hooks=train_hooks)
# tf.estimator.train(model_estimator, train_spec)
train_end_time = time.time()
print("==training time==", train_end_time - train_being_time)
tf.logging.info("==training time=={}".format(train_end_time -
train_being_time))
eval_results = model_estimator.evaluate(input_fn=eval_features,
steps=num_eval_steps)
print(eval_results)
elif kargs.get("distribution_strategy", "MirroredStrategy") in [
"ParameterServerStrategy", "CollectiveAllReduceStrategy"
]:
print("==apply multi-machine machine multi-card training==")
try:
print(os.environ['TF_CONFIG'], "==tf_run_config==")
except:
print("==not tf config==")
train_spec = tf.estimator.TrainSpec(input_fn=train_features,
max_steps=num_train_steps)
eval_spec = tf.estimator.EvalSpec(input_fn=eval_features,
steps=num_eval_steps)
# tf.estimator.train(model_estimator, train_spec) # tf 1.12 doesn't need evaluate
tf.estimator.train_and_evaluate(model_estimator, train_spec,
eval_spec)
def model_fn(features, labels, mode):
original_loss = tf.constant(0.0)
distilled_loss = tf.constant(0.0)
st_model = st_model_fn(
model_config_dict['student'],
num_labels_dict['student'],
init_checkpoint_dict['student'],
model_reuse=None,
load_pretrained=load_pretrained_dict['student'],
model_io_config=model_io_config,
opt_config=opt_config,
exclude_scope=exclude_scope_dict.get('student', ""),
not_storage_params=not_storage_params_dict.get('student', []),
target=target_dict['student'],
**kargs)
st_dict = st_model(features, labels, mode)
ta_model = ta_model_fn(
model_config_dict['teacher'],
num_labels_dict['teacher'],
init_checkpoint_dict['teacher'],
model_reuse=None,
load_pretrained=load_pretrained_dict['teacher'],
model_io_config=model_io_config,
opt_config=opt_config,
exclude_scope=exclude_scope_dict.get('teacher', ""),
not_storage_params=not_storage_params_dict.get('teacher', []),
target=target_dict['teacher'],
**kargs)
ta_dict = ta_model(features, labels, mode)
studnet_logit = st_dict['logits']
teacher_logit = ta_dict['logits']
model_io_fn = model_io.ModelIO(model_io_config)
feature_flag = False
original_loss += st_dict['loss'] * (distillation_config.get(
'ce_loss', 1.0))
print(distillation_config.get('ce_loss', 1.0), '===ce_loss===')
tf.summary.scalar("ce_loss", st_dict['loss'])
if 'kl_logits' in distillation_config.get('distillation_type',
['kl_logits']):
temperature = distillation_config.get('kl_temperature', 2.0)
distilled_teacher_logit = tf.nn.log_softmax(
(teacher_logit + 1e-10) / temperature) # log_softmax logits
distilled_student_logit = tf.nn.log_softmax(
(studnet_logit + 1e-10) / temperature) # log_softmax logits
kl_distilled_loss = tf.reduce_mean(
distillation_utils.kd(distilled_teacher_logit,
distilled_student_logit))
tf.summary.scalar("kl_logits_loss", kl_distilled_loss)
tf.logging.info(
"***** with knowledge distillation %s tenperature *****",
str(temperature))
# kl_distilled_loss *= np.power(temperature, 2)
distilled_loss += kl_distilled_loss * distillation_config.get(
'kl_logits_ratio', 0.9)
print(distillation_config.get('kl_logits_ratio', 0.9),
'===kl_logits_ratio===')
if 'rkd' in distillation_config.get('distillation_type',
['kl_logits']):
source = ta_dict['model'].get_pooled_output()
target = st_dict['model'].get_pooled_output()
print("==apply rkd==")
with tf.variable_scope("distillation", reuse=tf.AUTO_REUSE):
rkd_loss = repo_distillation_utils.RKD(source,
target,
l=[25, 50])
tf.summary.scalar("rkd_loss", rkd_loss)
distilled_loss += rkd_loss * distillation_config.get(
"rkd_ratio", 0.1)
if "attention_score_uniform" in distillation_config.get(
'distillation_type', ['kl_logits']):
source_attention_score = ta_dict['model'].get_multihead_attention()
target_attention_score = st_dict['model'].get_multihead_attention()
print("==apply attention_score_uniform==")
with tf.variable_scope("distillation", reuse=tf.AUTO_REUSE):
attention_loss = uniform_mapping.attention_score_matching(
source_attention_score, target_attention_score,
features['input_mask'], 0)
tf.summary.scalar("attention_score_uniform_loss", attention_loss)
feature_flag = True
distilled_loss += attention_loss * distillation_config.get(
"attention_score_uniform", 0.1)
print(distillation_config.get('attention_score_uniform', 0.1),
'===attention_score_uniform===')
if "hidden_uniform" in distillation_config.get('distillation_type',
['kl_logits']):
source_hidden = ta_dict['model'].get_all_encoder_layers()
target_hidden = st_dict['model'].get_all_encoder_layers()
print("==apply hidden_uniform==")
with tf.variable_scope("distillation", reuse=tf.AUTO_REUSE):
hidden_loss = uniform_mapping.hidden_matching(
source_hidden, target_hidden, features['input_mask'], 0)
tf.summary.scalar("hidden_uniform_loss", hidden_loss)
distilled_loss += hidden_loss * distillation_config.get(
"hidden_uniform", 0.1)
feature_flag = True
print(distillation_config.get('hidden_uniform', 0.1),
'===hidden_uniform===')
if "hidden_cls_uniform" in distillation_config.get(
'distillation_type', ['kl_logits']):
source_hidden = ta_dict['model'].get_all_encoder_layers()
target_hidden = st_dict['model'].get_all_encoder_layers()
print("==apply hidden_cls_uniform==")
with tf.variable_scope("distillation", reuse=tf.AUTO_REUSE):
hidden_cls_loss = uniform_mapping.hidden_cls_matching(
source_hidden, target_hidden, 0)
tf.summary.scalar("hidden_cls_uniform_loss", hidden_cls_loss)
distilled_loss += hidden_cls_loss * distillation_config.get(
"hidden_uniform", 0.1)
feature_flag = True
if "mdd" in distillation_config.get('distillation_type', ['mdd']):
source = ta_dict['model'].get_pooled_output()
target = st_dict['model'].get_pooled_output()
print("==apply mdd==")
if "cpc" in distillation_config.get('distillation_type', ['mdd']):
source_hidden = ta_dict['model'].get_all_encoder_layers()
target_hidden = st_dict['model'].get_all_encoder_layers()
with tf.variable_scope("distillation", reuse=tf.AUTO_REUSE):
cpc_loss = cpc_utils.CPC_Hidden(target_hidden, source_hidden,
features['input_mask'])
tf.summary.scalar("hidden_cpc_loss", cpc_loss)
distilled_loss += cpc_loss + distillation_config.get(
"cpc_hidden", 0.1)
if "wpc" in distillation_config.get('distillation_type', ['mdd']):
source_hidden = ta_dict['model'].get_all_encoder_layers()
target_hidden = st_dict['model'].get_all_encoder_layers()
with tf.variable_scope("distillation", reuse=tf.AUTO_REUSE):
wpc_loss = cpc_utils.WPC_Hidden(target_hidden, source_hidden,
features['input_mask'])
tf.summary.scalar("hidden_wpc_loss", wpc_loss)
distilled_loss += wpc_loss + distillation_config.get(
"wpc_hidden", 0.1)
total_loss = distilled_loss + original_loss
tvars = []
tvars.extend(st_dict['tvars'])
if feature_flag:
distillation_vars = model_io_fn.get_params('distillation',
not_storage_params=[])
tvars.extend(distillation_vars)
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer_fn = optimizer.Optimizer(opt_config)
model_io_fn.print_params(tvars, string=", trainable params")
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
print("==update_ops==", update_ops)
print('==total trainable vars==', list(tvars))
with tf.control_dependencies(update_ops):
train_op = optimizer_fn.get_train_op(
total_loss, list(set(tvars)), opt_config.init_lr,
opt_config.num_train_steps, **kargs)
model_io_fn.set_saver()
if kargs.get("task_index", 1) == 0 and kargs.get(
"run_config", None):
training_hooks = []
elif kargs.get("task_index", 1) == 0:
model_io_fn.get_hooks(kargs.get("checkpoint_dir", None),
kargs.get("num_storage_steps", 1000))
training_hooks = model_io_fn.checkpoint_hook
else:
training_hooks = []
if len(optimizer_fn.distributed_hooks) >= 1:
training_hooks.extend(optimizer_fn.distributed_hooks)
print(training_hooks, "==training_hooks==", "==task_index==",
kargs.get("task_index", 1))
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op,
training_hooks=training_hooks)
if output_type == "sess":
return {
"train": {
"loss": total_loss,
"logits": studnet_logit,
"train_op": train_op
},
"hooks": training_hooks
}
elif output_type == "estimator":
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
def metric_fn(per_example_loss, logits, label_ids, model_type):
"""Computes the loss and accuracy of the model."""
sentence_log_probs = tf.reshape(logits, [-1, logits.shape[-1]])
sentence_predictions = tf.argmax(logits,
axis=-1,
output_type=tf.int32)
sentence_labels = tf.reshape(label_ids, [-1])
sentence_accuracy = tf.metrics.accuracy(
labels=label_ids, predictions=sentence_predictions)
sentence_mean_loss = tf.metrics.mean(values=per_example_loss)
sentence_f = tf_metrics.f1(label_ids,
sentence_predictions,
num_labels_dict['student'],
None,
average="macro")
eval_metric_ops = {
"{}_f1".format(model_type): sentence_f,
"{}_acc".format(model_type): sentence_accuracy
}
return eval_metric_ops
if output_type == "sess":
return {
"eval": {
"per_example_loss":
st_dict['logits']['per_example_loss'],
"logits":
studnet_logit,
"loss":
tf.reduce_mean(st_dict['logits']['per_example_loss']),
"feature":
st_dict['model'].get_pooled_output()
}
}
elif output_type == "estimator":
eval_metric_ops = metric_fn(st_dict['per_example_loss'],
studnet_logit,
features['label_ids'], "student")
teacher_eval_metric_ops = metric_fn(
ta_dict['per_example_loss'], teacher_logit,
features['label_ids'], "teacher")
eval_metric_ops.update(teacher_eval_metric_ops)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=total_loss,
eval_metric_ops=eval_metric_ops)
return estimator_spec
else:
raise NotImplementedError()
def model_fn(features, labels, mode):
model = bert_encoder(model_config,
features,
labels,
mode,
target,
reuse=model_reuse)
label_ids = features["label_ids"]
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
with tf.variable_scope(scope, reuse=model_reuse):
(loss, per_example_loss,
logits) = classifier.classifier(model_config,
model.get_pooled_output(),
num_labels, label_ids,
dropout_prob)
model_io_fn = model_io.ModelIO(model_io_config)
tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
if load_pretrained:
model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope=exclude_scope)
model_io_fn.set_saver(var_lst=tvars)
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer_fn = optimizer.Optimizer(opt_config)
model_io_fn.print_params(tvars, string=", trainable params")
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer_fn.get_train_op(
loss, tvars, opt_config.init_lr,
opt_config.num_train_steps)
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
if output_type == "sess":
return {
"train": {
"loss": loss,
"logits": logits,
"train_op": train_op
}
}
elif output_type == "estimator":
return estimator_spec
elif mode == tf.estimator.ModeKeys.PREDICT:
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.nn.softmax(logits)
max_prob = tf.reduce_max(prob, axis=-1)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
predictions={
'pred_label': pred_label,
"max_prob": max_prob
},
export_outputs={
"output":
tf.estimator.export.PredictOutput({
'pred_label': pred_label,
"max_prob": max_prob
})
})
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
def metric_fn(per_example_loss, logits, label_ids):
"""Computes the loss and accuracy of the model."""
sentence_log_probs = tf.reshape(logits, [-1, logits.shape[-1]])
sentence_predictions = tf.argmax(logits,
axis=-1,
output_type=tf.int32)
sentence_labels = tf.reshape(label_ids, [-1])
sentence_accuracy = tf.metrics.accuracy(
labels=label_ids, predictions=sentence_predictions)
sentence_mean_loss = tf.metrics.mean(values=per_example_loss)
sentence_f = tf_metrics.f1(label_ids,
sentence_predictions,
num_labels,
label_lst,
average="macro")
eval_metric_ops = {"f1": sentence_f, "acc": sentence_accuracy}
return eval_metric_ops
eval_metric_ops = metric_fn(per_example_loss, logits, label_ids)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode, loss=loss, eval_metric_ops=eval_metric_ops)
if output_type == "sess":
return {
"eval": {
"per_example_loss": per_example_loss,
"logits": logits,
"loss": tf.reduce_mean(per_example_loss)
}
}
elif output_type == "estimator":
return estimator_spec
else:
raise NotImplementedError()
def model_fn(features, labels, mode):
task_type = kargs.get("task_type", "cls")
num_task = kargs.get('num_task', 1)
temp = kargs.get('temp', 0.1)
print("==task_type==", task_type)
model_io_fn = model_io.ModelIO(model_io_config)
label_ids = tf.cast(features["{}_label_ids".format(task_type)],
dtype=tf.int32)
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
is_training = True
else:
dropout_prob = 0.0
is_training = False
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
if kargs.get("get_pooled_output", "pooled_output") == "pooled_output":
pooled_feature = model.get_pooled_output()
elif kargs.get("get_pooled_output", "task_output") == "task_output":
pooled_feature_dict = model.get_task_output()
pooled_feature = pooled_feature_dict['pooled_feature']
shape_list = bert_utils.get_shape_list(
pooled_feature_dict['feature_a'], expected_rank=[2])
batch_size = shape_list[0]
if kargs.get('apply_head_proj', False):
with tf.variable_scope(scope + "/head_proj", reuse=tf.AUTO_REUSE):
feature_a = simclr_utils.projection_head(
pooled_feature_dict['feature_a'],
is_training,
head_proj_dim=128,
num_nlh_layers=1,
head_proj_mode='nonlinear',
name='head_contrastive')
pooled_feature_dict['feature_a'] = feature_a
with tf.variable_scope(scope + "/head_proj", reuse=tf.AUTO_REUSE):
feature_b = simclr_utils.projection_head(
pooled_feature_dict['feature_b'],
is_training,
head_proj_dim=128,
num_nlh_layers=1,
head_proj_mode='nonlinear',
name='head_contrastive')
pooled_feature_dict['feature_b'] = feature_b
tf.logging.info(
"****** apply contrastive feature projection *******")
else:
feature_a = pooled_feature_dict['feature_a']
feature_b = pooled_feature_dict['feature_b']
tf.logging.info("****** not apply projection *******")
# feature_a = tf.nn.l2_normalize(feature_a, axis=-1)
# feature_b = tf.nn.l2_normalize(feature_b, axis=-1)
# [batch_size, batch_size]
if kargs.get("task_seperate_proj", False):
if task_type == 'xquad' or task_type == 'wsdm':
# for passage representation
with tf.variable_scope(
scope + "/{}/feature_output_b".format(task_type),
reuse=tf.AUTO_REUSE):
feature_b = tf.layers.dense(
feature_b,
128,
use_bias=True,
activation=tf.tanh,
kernel_initializer=tf.truncated_normal_initializer(
stddev=0.01))
tf.logging.info("****** apply passage projection *******")
if task_type == 'afqmc':
# for anchor representation
with tf.variable_scope(
scope + "/{}/feature_output_a".format(task_type),
reuse=tf.AUTO_REUSE):
feature_a = tf.layers.dense(
feature_a,
128,
use_bias=True,
activation=tf.tanh,
kernel_initializer=tf.truncated_normal_initializer(
stddev=0.01))
# for successor representation
with tf.variable_scope(
scope + "/{}/feature_output_b".format(task_type),
reuse=tf.AUTO_REUSE):
feature_b = tf.layers.dense(
feature_b,
128,
use_bias=True,
activation=tf.tanh,
kernel_initializer=tf.truncated_normal_initializer(
stddev=0.01))
tf.logging.info(
"****** apply cpc anchor, successor projection *******")
cosine_score = tf.matmul(feature_a,
tf.transpose(feature_b)) / model_config.get(
'temperature', 0.5)
print("==cosine_score shape==", cosine_score.get_shape())
loss_mask = tf.cast(features["{}_loss_multipiler".format(task_type)],
tf.float32)
if task_type == 'xquad':
neg_true_mask = tf.cast(
triplet_loss_utils._get_anchor_negative_triplet_mask(
label_ids), tf.float32)
pos_true_mask = (1.0 - neg_true_mask) * tf.expand_dims(
loss_mask, axis=-1) * tf.expand_dims(loss_mask, axis=0)
neg_true_mask = neg_true_mask * tf.expand_dims(
loss_mask, axis=-1) * tf.expand_dims(loss_mask, axis=0)
elif task_type == 'wsdm':
pos_label_mask = tf.cast(
features["{}_label_ids".format(task_type)], dtype=tf.float32)
loss_mask *= pos_label_mask
pos_label_mask = tf.expand_dims(pos_label_mask,
axis=-1) # batch x batch
score_shape = bert_utils.get_shape_list(cosine_score,
expected_rank=[2, 3])
pos_true_mask = pos_label_mask * tf.eye(score_shape[0])
neg_true_mask = tf.ones_like(cosine_score) - pos_true_mask
pos_true_mask = pos_true_mask * tf.expand_dims(
loss_mask, axis=-1) * tf.expand_dims(loss_mask, axis=0)
neg_true_mask = neg_true_mask * tf.expand_dims(
loss_mask, axis=-1) * tf.expand_dims(loss_mask, axis=0)
elif task_type == 'afqmc':
score_shape = bert_utils.get_shape_list(cosine_score,
expected_rank=[2, 3])
not_equal_mask = get_labels_of_similarity(features['input_ids_a'],
features['input_ids_b'])
pos_true_mask = tf.expand_dims(loss_mask, axis=-1) * tf.eye(
score_shape[0])
neg_true_mask = not_equal_mask * tf.expand_dims(
loss_mask, axis=-1) * tf.expand_dims(loss_mask, axis=0)
cosine_score_neg = neg_true_mask * cosine_score
cosine_score_pos = -pos_true_mask * cosine_score
y_pred_neg = cosine_score_neg - (1 - neg_true_mask) * 1e12
y_pred_pos = cosine_score_pos - (1 - pos_true_mask) * 1e12
# add circle-loss without margin and scale-factor
joint_neg_loss = tf.reduce_logsumexp(y_pred_neg, axis=-1)
joint_pos_loss = tf.reduce_logsumexp(y_pred_pos, axis=-1)
logits = tf.nn.softplus(joint_neg_loss + joint_pos_loss)
loss = tf.reduce_sum(
logits * loss_mask) / (1e-10 + tf.reduce_sum(loss_mask))
task_loss = loss
params_size = model_io_fn.count_params(model_config.scope)
print("==total encoder params==", params_size)
if kargs.get("feature_distillation", True):
universal_feature_a = features.get("input_ids_a_features", None)
universal_feature_b = features.get("input_ids_b_features", None)
if universal_feature_a is None or universal_feature_b is None:
tf.logging.info(
"****** not apply feature distillation *******")
feature_loss = tf.constant(0.0)
else:
feature_a = pooled_feature_dict['feature_a']
feature_a_shape = bert_utils.get_shape_list(
feature_a, expected_rank=[2, 3])
pretrain_feature_a_shape = bert_utils.get_shape_list(
universal_feature_a, expected_rank=[2, 3])
if feature_a_shape[-1] != pretrain_feature_a_shape[-1]:
with tf.variable_scope(scope + "/feature_proj",
reuse=tf.AUTO_REUSE):
proj_feature_a = tf.layers.dense(
feature_a, pretrain_feature_a_shape[-1])
# with tf.variable_scope(scope+"/feature_rec", reuse=tf.AUTO_REUSE):
# proj_feature_a_rec = tf.layers.dense(proj_feature_a, feature_a_shape[-1])
# loss += tf.reduce_mean(tf.reduce_sum(tf.square(proj_feature_a_rec-feature_a), axis=-1))/float(num_task)
tf.logging.info(
"****** apply auto-encoder for feature compression *******"
)
else:
proj_feature_a = feature_a
feature_a_norm = tf.stop_gradient(
tf.sqrt(
tf.reduce_sum(tf.pow(proj_feature_a, 2),
axis=-1,
keepdims=True)) + 1e-20)
proj_feature_a /= feature_a_norm
feature_b = pooled_feature_dict['feature_b']
if feature_a_shape[-1] != pretrain_feature_a_shape[-1]:
with tf.variable_scope(scope + "/feature_proj",
reuse=tf.AUTO_REUSE):
proj_feature_b = tf.layers.dense(
feature_b, pretrain_feature_a_shape[-1])
# with tf.variable_scope(scope+"/feature_rec", reuse=tf.AUTO_REUSE):
# proj_feature_b_rec = tf.layers.dense(proj_feature_b, feature_a_shape[-1])
# loss += tf.reduce_mean(tf.reduce_sum(tf.square(proj_feature_b_rec-feature_b), axis=-1))/float(num_task)
tf.logging.info(
"****** apply auto-encoder for feature compression *******"
)
else:
proj_feature_b = feature_b
feature_b_norm = tf.stop_gradient(
tf.sqrt(
tf.reduce_sum(tf.pow(proj_feature_b, 2),
axis=-1,
keepdims=True)) + 1e-20)
proj_feature_b /= feature_b_norm
feature_a_distillation = tf.reduce_mean(
tf.square(universal_feature_a - proj_feature_a), axis=-1)
feature_b_distillation = tf.reduce_mean(
tf.square(universal_feature_b - proj_feature_b), axis=-1)
feature_loss = tf.reduce_mean(
(feature_a_distillation + feature_b_distillation) /
2.0) / float(num_task)
loss += feature_loss
tf.logging.info(
"****** apply prertained feature distillation *******")
if kargs.get("embedding_distillation", True):
word_embed = model.emb_mat
random_embed_shape = bert_utils.get_shape_list(
word_embed, expected_rank=[2, 3])
print("==random_embed_shape==", random_embed_shape)
pretrained_embed = kargs.get('pretrained_embed', None)
if pretrained_embed is None:
tf.logging.info(
"****** not apply prertained feature distillation *******")
embed_loss = tf.constant(0.0)
else:
pretrain_embed_shape = bert_utils.get_shape_list(
pretrained_embed, expected_rank=[2, 3])
print("==pretrain_embed_shape==", pretrain_embed_shape)
if random_embed_shape[-1] != pretrain_embed_shape[-1]:
with tf.variable_scope(scope + "/embedding_proj",
reuse=tf.AUTO_REUSE):
proj_embed = tf.layers.dense(word_embed,
pretrain_embed_shape[-1])
else:
proj_embed = word_embed
embed_loss = tf.reduce_mean(
tf.reduce_mean(tf.square(proj_embed - pretrained_embed),
axis=-1)) / float(num_task)
loss += embed_loss
tf.logging.info(
"****** apply prertained feature distillation *******")
if mode == tf.estimator.ModeKeys.TRAIN:
multi_task_config = kargs.get("multi_task_config", {})
if multi_task_config.get(task_type,
{}).get("lm_augumentation", False):
print("==apply lm_augumentation==")
masked_lm_positions = features["masked_lm_positions"]
masked_lm_ids = features["masked_lm_ids"]
masked_lm_weights = features["masked_lm_weights"]
(masked_lm_loss, masked_lm_example_loss,
masked_lm_log_probs) = pretrain.get_masked_lm_output(
model_config,
model.get_sequence_output(),
model.get_embedding_table(),
masked_lm_positions,
masked_lm_ids,
masked_lm_weights,
reuse=model_reuse)
masked_lm_loss_mask = tf.expand_dims(loss_mask, -1) * tf.ones(
(1,
multi_task_config[task_type]["max_predictions_per_seq"]))
masked_lm_loss_mask = tf.reshape(masked_lm_loss_mask, (-1, ))
masked_lm_label_weights = tf.reshape(masked_lm_weights, [-1])
masked_lm_loss_mask *= tf.cast(masked_lm_label_weights,
tf.float32)
masked_lm_example_loss *= masked_lm_loss_mask # multiply task_mask
masked_lm_loss = tf.reduce_sum(masked_lm_example_loss) / (
1e-10 + tf.reduce_sum(masked_lm_loss_mask))
loss += multi_task_config[task_type][
"masked_lm_loss_ratio"] * masked_lm_loss
masked_lm_label_ids = tf.reshape(masked_lm_ids, [-1])
print(masked_lm_log_probs.get_shape(),
"===masked lm log probs===")
print(masked_lm_label_ids.get_shape(), "===masked lm ids===")
print(masked_lm_label_weights.get_shape(),
"===masked lm mask===")
lm_acc = build_accuracy(masked_lm_log_probs,
masked_lm_label_ids,
masked_lm_loss_mask)
if kargs.get("task_invariant", "no") == "yes":
print("==apply task adversarial training==")
with tf.variable_scope(scope + "/dann_task_invariant",
reuse=model_reuse):
(_, task_example_loss,
task_logits) = distillation_utils.feature_distillation(
model.get_pooled_output(), 1.0, features["task_id"],
kargs.get("num_task", 7), dropout_prob, True)
masked_task_example_loss = loss_mask * task_example_loss
masked_task_loss = tf.reduce_sum(masked_task_example_loss) / (
1e-10 + tf.reduce_sum(loss_mask))
loss += kargs.get("task_adversarial", 1e-2) * masked_task_loss
tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
if mode == tf.estimator.ModeKeys.TRAIN:
multi_task_config = kargs.get("multi_task_config", {})
if multi_task_config.get(task_type,
{}).get("lm_augumentation", False):
print("==apply lm_augumentation==")
masked_lm_pretrain_tvars = model_io_fn.get_params(
"cls/predictions", not_storage_params=not_storage_params)
tvars.extend(masked_lm_pretrain_tvars)
try:
params_size = model_io_fn.count_params(model_config.scope)
print("==total params==", params_size)
except:
print("==not count params==")
# print(tvars)
if load_pretrained == "yes":
model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope=exclude_scope)
if mode == tf.estimator.ModeKeys.TRAIN:
# acc = build_accuracy(logits,
# label_ids,
# loss_mask,
# loss_type=kargs.get('loss', 'contrastive_loss'))
return_dict = {
"loss": loss,
"logits": logits,
"task_num": tf.reduce_sum(loss_mask),
"{}_pos_num".format(task_type): tf.reduce_sum(pos_true_mask),
"{}_neg_num".format(task_type): tf.reduce_sum(neg_true_mask),
"tvars": tvars
}
# return_dict["{}_acc".format(task_type)] = acc
if kargs.get("task_invariant", "no") == "yes":
return_dict["{}_task_loss".format(
task_type)] = masked_task_loss
task_acc = build_accuracy(task_logits, features["task_id"],
loss_mask)
return_dict["{}_task_acc".format(task_type)] = task_acc
if multi_task_config.get(task_type,
{}).get("lm_augumentation", False):
return_dict["{}_masked_lm_loss".format(
task_type)] = masked_lm_loss
return_dict["{}_masked_lm_acc".format(task_type)] = lm_acc
if kargs.get("embedding_distillation", True):
return_dict["embed_loss"] = embed_loss * float(num_task)
else:
return_dict["embed_loss"] = task_loss
if kargs.get("feature_distillation", True):
return_dict["feature_loss"] = feature_loss * float(num_task)
else:
return_dict["feature_loss"] = task_loss
return_dict["task_loss"] = task_loss
return return_dict
elif mode == tf.estimator.ModeKeys.EVAL:
eval_dict = {
"loss": loss,
"logits": logits,
"feature": model.get_pooled_output()
}
if kargs.get("adversarial", "no") == "adversarial":
eval_dict["task_logits"] = task_logits
return eval_dict
def model_fn(features, labels, mode, params):
model_api = model_zoo(model_config)
if kargs.get('random_generator', '1') == '1':
if mode in [
tf.estimator.ModeKeys.TRAIN, tf.estimator.ModeKeys.TRAIN
]:
input_ori_ids = features['input_ori_ids']
# [output_ids,
# sampled_binary_mask] = random_input_ids_generation(model_config,
# features['input_ori_ids'],
# features['input_mask'],
# mask_probability=0.2,
# replace_probability=0.1,
# original_probability=0.1,
# **kargs)
[output_ids, sampled_binary_mask] = hmm_input_ids_generation(
model_config,
features['input_ori_ids'],
features['input_mask'], [
tf.cast(tf.constant(hmm_tran_prob), tf.float32)
for hmm_tran_prob in hmm_tran_prob_list
],
mask_probability=0.2,
replace_probability=0.0,
original_probability=0.0,
mask_prior=tf.constant(mask_prior, tf.float32),
**kargs)
features['input_ids'] = output_ids
tf.logging.info("****** do random generator *******")
else:
sampled_binary_mask = None
else:
sampled_binary_mask = None
model = model_api(model_config,
features,
labels,
mode,
target,
reuse=tf.AUTO_REUSE,
**kargs)
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
(nsp_loss, nsp_per_example_loss,
nsp_log_prob) = pretrain.get_next_sentence_output(
model_config,
model.get_pooled_output(),
features['next_sentence_labels'],
reuse=tf.AUTO_REUSE,
scope=generator_scope_prefix)
masked_lm_positions = features["masked_lm_positions"]
masked_lm_ids = features["masked_lm_ids"]
masked_lm_weights = features["masked_lm_weights"]
if model_config.model_type == 'bert':
masked_lm_fn = pretrain.get_masked_lm_output
seq_masked_lm_fn = pretrain.seq_mask_masked_lm_output
print("==apply bert masked lm==")
elif model_config.model_type == 'albert':
masked_lm_fn = pretrain_albert.get_masked_lm_output
seq_masked_lm_fn = pretrain_albert.seq_mask_masked_lm_output
print("==apply albert masked lm==")
else:
masked_lm_fn = pretrain.get_masked_lm_output
seq_masked_lm_fn = pretrain_albert.seq_mask_masked_lm_output
print("==apply bert masked lm==")
if sampled_binary_mask is not None:
(masked_lm_loss, masked_lm_example_loss, masked_lm_log_probs,
masked_lm_mask) = seq_masked_lm_fn(
model_config,
model.get_sequence_output(),
model.get_embedding_table(),
features['input_mask'],
features['input_ori_ids'],
features['input_ids'],
sampled_binary_mask,
reuse=tf.AUTO_REUSE,
embedding_projection=model.get_embedding_projection_table(),
scope=generator_scope_prefix)
masked_lm_ids = features['input_ori_ids']
else:
(masked_lm_loss, masked_lm_example_loss, masked_lm_log_probs,
masked_lm_mask) = masked_lm_fn(
model_config,
model.get_sequence_output(),
model.get_embedding_table(),
masked_lm_positions,
masked_lm_ids,
masked_lm_weights,
reuse=tf.AUTO_REUSE,
embedding_projection=model.get_embedding_projection_table(),
scope=generator_scope_prefix)
print(model_config.lm_ratio, '==mlm lm_ratio==')
loss = model_config.lm_ratio * masked_lm_loss + 0.0 * nsp_loss
if kargs.get("resample_discriminator", False):
input_ori_ids = features['input_ori_ids']
[output_ids, sampled_binary_mask
] = random_input_ids_generation(model_config,
features['input_ori_ids'],
features['input_mask'],
mask_probability=0.2,
replace_probability=0.1,
original_probability=0.1)
resample_features = {}
for key in features:
resample_features[key] = features[key]
resample_features['input_ids'] = tf.identity(output_ids)
model_resample = model_api(model_config,
resample_features,
labels,
mode,
target,
reuse=tf.AUTO_REUSE,
**kargs)
tf.logging.info("**** apply discriminator resample **** ")
else:
model_resample = model
resample_features = features
tf.logging.info("**** not apply discriminator resample **** ")
sampled_ids = token_generator(model_config,
model_resample.get_sequence_output(),
model_resample.get_embedding_table(),
resample_features['input_ids'],
resample_features['input_ori_ids'],
resample_features['input_mask'],
embedding_projection=model_resample.
get_embedding_projection_table(),
scope=generator_scope_prefix,
mask_method='only_mask',
use_tpu=kargs.get('use_tpu', True))
if model_config.get('gen_sample', 1) == 1:
input_ids = features['input_ori_ids']
input_mask = features['input_mask']
segment_ids = features['segment_ids']
else:
input_ids = tf.expand_dims(features['input_ori_ids'], axis=-1)
# batch x seq_length x 1
input_ids = tf.einsum(
'abc,cd->abd', input_ids,
tf.ones((1, model_config.get('gen_sample', 1))))
input_ids = tf.cast(input_ids, tf.int32)
input_shape_list = bert_utils.get_shape_list(input_ids,
expected_rank=3)
batch_size = input_shape_list[0]
seq_length = input_shape_list[1]
gen_sample = input_shape_list[2]
sampled_ids = tf.reshape(sampled_ids,
[batch * gen_sample, seq_length])
input_ids = tf.reshape(input_ids, [batch * gen_sample, seq_length])
input_mask = tf.expand_dims(features['input_mask'], axis=-1)
input_mask = tf.einsum(
'abc,cd->abd', input_mask,
tf.ones((1, model_config.get('gen_sample', 1))))
input_mask = tf.cast(input_mask, tf.int32)
segment_ids = tf.expand_dims(features['segmnet_ids'], axis=-1)
segment_ids = tf.einsum(
'abc,cd->abd', segment_ids,
tf.ones((1, model_config.get('gen_sample', 1))))
segment_ids = tf.cast(segment_ids, tf.int32)
segment_ids = tf.reshape(segment_ids,
[batch * gen_sample, seq_length])
input_mask = tf.reshape(input_mask,
[batch * gen_sample, seq_length])
model_io_fn = model_io.ModelIO(model_io_config)
pretrained_tvars = model_io_fn.get_params(
model_config.scope, not_storage_params=not_storage_params)
if generator_scope_prefix:
"""
"generator/cls/predictions"
"""
lm_pretrain_tvars = model_io_fn.get_params(
generator_scope_prefix + "/cls/predictions",
not_storage_params=not_storage_params)
nsp_pretrain_vars = model_io_fn.get_params(
generator_scope_prefix + "/cls/seq_relationship",
not_storage_params=not_storage_params)
else:
lm_pretrain_tvars = model_io_fn.get_params(
"cls/predictions", not_storage_params=not_storage_params)
nsp_pretrain_vars = model_io_fn.get_params(
"cls/seq_relationship", not_storage_params=not_storage_params)
if model_config.get('embedding_scope', None) is not None:
embedding_tvars = model_io_fn.get_params(
model_config.get('embedding_scope', 'bert') + "/embeddings",
not_storage_params=not_storage_params)
pretrained_tvars.extend(embedding_tvars)
pretrained_tvars.extend(lm_pretrain_tvars)
pretrained_tvars.extend(nsp_pretrain_vars)
tvars = pretrained_tvars
print('==generator parameters==', tvars)
if load_pretrained == "yes":
use_tpu = 1 if kargs.get('use_tpu', False) else 0
scaffold_fn = model_io_fn.load_pretrained(
tvars,
init_checkpoint,
exclude_scope=exclude_scope,
use_tpu=use_tpu,
restore_var_name=model_config.get('restore_var_name', []))
else:
scaffold_fn = None
tf.add_to_collection("generator_loss", loss)
return_dict = {
"loss": loss,
"tvars": tvars,
"model": model,
"sampled_ids": sampled_ids, # batch x gen_sample, seg_length
"sampled_input_ids": input_ids, # batch x gen_sample, seg_length,
"sampled_input_mask": input_mask,
"sampled_segment_ids": segment_ids,
"masked_lm_ids": masked_lm_ids,
"masked_lm_weights": masked_lm_mask,
"masked_lm_log_probs": masked_lm_log_probs,
"masked_lm_example_loss": masked_lm_example_loss,
"next_sentence_example_loss": nsp_per_example_loss,
"next_sentence_log_probs": nsp_log_prob,
"next_sentence_labels": features['next_sentence_labels'],
"sampled_binary_mask": sampled_binary_mask
}
return return_dict
def model_fn(features, labels, mode, params):
train_op_type = kargs.get('train_op_type', 'joint')
gen_disc_type = kargs.get('gen_disc_type', 'all_disc')
mask_method = kargs.get('mask_method', 'only_mask')
use_tpu = 1 if kargs.get('use_tpu', False) else 0
print(train_op_type, "===train op type===", gen_disc_type,
"===generator loss type===")
if mask_method == 'only_mask':
tf.logging.info(
"****** generator token generation mask type:%s with only masked token *******",
mask_method)
elif mask_method == 'all_mask':
tf.logging.info(
"****** generator token generation mask type:%s with all token *******",
mask_method)
else:
mask_method = 'only_mask'
tf.logging.info(
"****** generator token generation mask type:%s with only masked token *******",
mask_method)
if kargs.get('optimization_type', 'grl') == 'grl':
if_flip_grad = True
train_op_type = 'joint'
elif kargs.get('optimization_type', 'grl') == 'minmax':
if_flip_grad = False
else:
if_flip_grad = True
train_op_type = 'joint'
generator_fn = generator(
model_config_dict['generator'],
num_labels_dict['generator'],
init_checkpoint_dict['generator'],
model_reuse=None,
load_pretrained=load_pretrained_dict['generator'],
model_io_config=model_io_config,
opt_config=opt_config,
exclude_scope=exclude_scope_dict.get('generator', ""),
not_storage_params=not_storage_params_dict.get('generator', []),
target=target_dict['generator'],
if_flip_grad=if_flip_grad,
# mask_method='only_mask',
**kargs)
tf.logging.info("****** train_op_type:%s *******", train_op_type)
tf.logging.info("****** optimization_type:%s *******",
kargs.get('optimization_type', 'grl'))
generator_dict = generator_fn(features, labels, mode, params)
discriminator_fn = discriminator_generator(
model_config_dict['discriminator'],
num_labels_dict['discriminator'],
init_checkpoint_dict['discriminator'],
model_reuse=None,
load_pretrained=load_pretrained_dict['discriminator'],
model_io_config=model_io_config,
opt_config=opt_config,
exclude_scope=exclude_scope_dict.get('discriminator', ""),
not_storage_params=not_storage_params_dict.get(
'discriminator', []),
target=target_dict['discriminator'],
loss='cross_entropy',
**kargs)
discriminator_features = {}
# minmax_mode in ['masked', 'corrupted']
minmax_mode = kargs.get('minmax_mode', 'corrupted')
tf.logging.info("****** minmax mode for discriminator: %s *******",
minmax_mode)
if minmax_mode == 'corrupted':
tf.logging.info("****** gumbel 3-D sampled_ids *******")
elif minmax_mode == 'masked':
discriminator_features['ori_sampled_ids'] = generator_dict[
'output_ids']
discriminator_features['sampled_binary_mask'] = generator_dict[
'sampled_binary_mask']
tf.logging.info("****** conditional sampled_ids *******")
discriminator_features['input_ids'] = generator_dict['sampled_ids']
discriminator_features['input_mask'] = generator_dict[
'sampled_input_mask']
discriminator_features['segment_ids'] = generator_dict[
'sampled_segment_ids']
discriminator_features['input_ori_ids'] = generator_dict[
'sampled_input_ids']
discriminator_features['next_sentence_labels'] = features[
'next_sentence_labels']
discriminator_features['ori_input_ids'] = generator_dict['sampled_ids']
discriminator_dict = discriminator_fn(discriminator_features, labels,
mode, params)
[disc_loss, disc_logits, disc_per_example_loss
] = optimal_discriminator(model_config_dict['discriminator'],
generator_dict,
features,
discriminator_dict,
discriminator_features,
use_tpu=use_tpu)
[
equal_per_example_loss, equal_loss_all, equal_loss_self,
not_equal_per_example_loss, not_equal_loss_all, not_equal_loss_self
] = modified_loss(disc_per_example_loss,
disc_logits,
discriminator_features['input_ori_ids'],
discriminator_features['ori_input_ids'],
discriminator_features['input_mask'],
sampled_binary_mask=discriminator_features.get(
'sampled_binary_mask', None),
**kargs)
output_dict = {}
output_dict['logits'] = disc_logits
output_dict['per_example_loss'] = disc_per_example_loss
output_dict['loss'] = disc_loss + 0.0 * discriminator_dict["loss"]
output_dict["equal_per_example_loss"] = equal_per_example_loss,
output_dict["equal_loss_all"] = equal_loss_all,
output_dict["equal_loss_self"] = equal_loss_self,
output_dict["not_equal_per_example_loss"] = not_equal_per_example_loss,
output_dict["not_equal_loss_all"] = not_equal_loss_all,
output_dict["not_equal_loss_self"] = not_equal_loss_self
output_dict['tvars'] = discriminator_dict['tvars']
model_io_fn = model_io.ModelIO(model_io_config)
tvars = []
loss = kargs.get('dis_loss', 1.0) * output_dict['loss']
tvars.extend(discriminator_dict['tvars'])
if kargs.get('joint_train', '1') == '1':
tf.logging.info(
"****** joint generator and discriminator training *******")
tvars.extend(generator_dict['tvars'])
loss += generator_dict['loss']
tvars = list(set(tvars))
var_checkpoint_dict_list = []
for key in init_checkpoint_dict:
if load_pretrained_dict[key] == "yes":
if key == 'generator':
tmp = {
"tvars":
generator_dict['tvars'],
"init_checkpoint":
init_checkpoint_dict['generator'],
"exclude_scope":
exclude_scope_dict[key],
"restore_var_name":
model_config_dict['generator'].get(
'restore_var_name', [])
}
if kargs.get("sharing_mode", "none") != "none":
tmp['exclude_scope'] = ''
var_checkpoint_dict_list.append(tmp)
elif key == 'discriminator':
tmp = {
"tvars":
discriminator_dict['tvars'],
"init_checkpoint":
init_checkpoint_dict['discriminator'],
"exclude_scope":
exclude_scope_dict[key],
"restore_var_name":
model_config_dict['discriminator'].get(
'restore_var_name', [])
}
var_checkpoint_dict_list.append(tmp)
use_tpu = 1 if kargs.get('use_tpu', False) else 0
if len(var_checkpoint_dict_list) >= 1:
scaffold_fn = model_io_fn.load_multi_pretrained(
var_checkpoint_dict_list, use_tpu=use_tpu)
else:
scaffold_fn = None
if mode == tf.estimator.ModeKeys.TRAIN:
if not kargs.get('use_tpu', False):
metric_dict = discriminator_metric_train(
output_dict['per_example_loss'], output_dict['logits'],
generator_dict['sampled_input_ids'],
generator_dict['sampled_ids'],
generator_dict['sampled_input_mask'])
for key in metric_dict:
tf.summary.scalar(key, metric_dict[key])
tf.summary.scalar("generator_loss", generator_dict['loss'])
tf.summary.scalar("discriminator_loss",
discriminator_dict['loss'])
if kargs.get('use_tpu', False):
optimizer_fn = optimizer.Optimizer(opt_config)
use_tpu = 1
else:
optimizer_fn = distributed_optimizer.Optimizer(opt_config)
use_tpu = 0
model_io_fn.print_params(tvars, string=", trainable params")
train_op = get_train_op(generator_dict,
output_dict,
optimizer_fn,
opt_config,
model_config_dict['generator'],
model_config_dict['discriminator'],
use_tpu=use_tpu,
train_op_type=train_op_type,
gen_disc_type=gen_disc_type)
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# with tf.control_dependencies(update_ops):
# train_op = optimizer_fn.get_train_op(loss, list(set(tvars)),
# opt_config.init_lr,
# opt_config.num_train_steps,
# use_tpu=use_tpu)
if kargs.get('use_tpu', False):
estimator_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn
# training_hooks=[logging_hook]
)
else:
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
if kargs.get('joint_train', '0') == '1':
def joint_metric(masked_lm_example_loss, masked_lm_log_probs,
masked_lm_ids, masked_lm_weights,
next_sentence_example_loss,
next_sentence_log_probs, next_sentence_labels,
per_example_loss, logits, input_ori_ids,
input_ids, input_mask):
generator_metric = generator_metric_fn_eval(
masked_lm_example_loss, masked_lm_log_probs,
masked_lm_ids, masked_lm_weights,
next_sentence_example_loss, next_sentence_log_probs,
next_sentence_labels)
discriminator_metric = discriminator_metric_eval(
per_example_loss, logits, input_ori_ids, input_ids,
input_mask)
generator_metric.update(discriminator_metric)
return generator_metric
tpu_eval_metrics = (joint_metric, [
generator_dict['masked_lm_example_loss'],
generator_dict['masked_lm_log_probs'],
generator_dict['masked_lm_ids'],
generator_dict['masked_lm_weights'],
generator_dict.get('next_sentence_example_loss', None),
generator_dict.get('next_sentence_log_probs', None),
generator_dict.get('next_sentence_labels', None),
discriminator_dict['per_example_loss'],
discriminator_dict['logits'],
generator_dict['sampled_input_ids'],
generator_dict['sampled_ids'],
generator_dict['sampled_input_mask']
])
gpu_eval_metrics = joint_metric(
generator_dict['masked_lm_example_loss'],
generator_dict['masked_lm_log_probs'],
generator_dict['masked_lm_ids'],
generator_dict['masked_lm_weights'],
generator_dict.get('next_sentence_example_loss', None),
generator_dict.get('next_sentence_log_probs', None),
generator_dict.get('next_sentence_labels', None),
discriminator_dict['per_example_loss'],
discriminator_dict['logits'],
generator_dict['sampled_input_ids'],
generator_dict['sampled_ids'],
generator_dict['sampled_input_mask'])
else:
gpu_eval_metrics = discriminator_metric_eval(
discriminator_dict['per_example_loss'],
discriminator_dict['logits'],
generator_dict['sampled_input_ids'],
generator_dict['sampled_ids'],
generator_dict['sampled_input_mask'])
tpu_eval_metrics = (discriminator_metric_eval, [
discriminator_dict['per_example_loss'],
discriminator_dict['logits'],
generator_dict['sampled_input_ids'],
generator_dict['sampled_ids'],
generator_dict['sampled_input_mask']
])
if kargs.get('use_tpu', False):
estimator_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=loss,
eval_metrics=tpu_eval_metrics,
scaffold_fn=scaffold_fn)
else:
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode, loss=loss, eval_metric_ops=gpu_eval_metrics)
return estimator_spec
else:
raise NotImplementedError()
def model_fn(features, labels, mode):
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
label_ids = features["label_ids"]
model_lst = []
for index, name in enumerate(target):
if index > 0:
reuse = True
else:
reuse = model_reuse
model_lst.append(bert_encoding(model_config, features, labels,
mode, name,
scope, dropout_rate,
reuse=reuse))
[input_mask_a, repres_a] = model_lst[0]
[input_mask_b, repres_b] = model_lst[1]
output_a, output_b = alignment_aggerate(model_config,
repres_a, repres_b,
input_mask_a,
input_mask_b,
scope,
reuse=model_reuse)
if model_config.pooling == "ave_max_pooling":
pooling_fn = ave_max_pooling
elif model_config.pooling == "multihead_pooling":
pooling_fn = multihead_pooling
repres_a = pooling_fn(model_config, output_a,
input_mask_a,
scope,
dropout_prob,
reuse=model_reuse)
repres_b = pooling_fn(model_config, output_b,
input_mask_b,
scope,
dropout_prob,
reuse=True)
pair_repres = tf.concat([repres_a, repres_b,
tf.abs(repres_a-repres_b),
repres_b*repres_a], axis=-1)
with tf.variable_scope(scope, reuse=model_reuse):
(loss,
per_example_loss,
logits) = classifier.classifier(model_config,
pair_repres,
num_labels,
label_ids,
dropout_prob)
model_io_fn = model_io.ModelIO(model_io_config)
tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
if load_pretrained:
model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope=exclude_scope)
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer_fn = optimizer.Optimizer(opt_config)
model_io_fn.print_params(tvars, string=", trainable params")
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer_fn.get_train_op(loss, tvars,
opt_config.init_lr,
opt_config.num_train_steps)
model_io_fn.set_saver()
if kargs.get("task_index", 1) == 0:
model_io_fn.get_hooks(kargs.get("checkpoint_dir", None),
kargs.get("num_storage_steps", 1000))
training_hooks = model_io_fn.checkpoint_hook
else:
training_hooks = []
if len(optimizer_fn.distributed_hooks) >= 1:
training_hooks.extend(optimizer_fn.distributed_hooks)
print(training_hooks)
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=loss, train_op=train_op,
training_hooks=training_hooks)
if output_type == "sess":
return {
"train":{
"loss":loss,
"logits":logits,
"train_op":train_op
},
"hooks":training_hooks
}
elif output_type == "estimator":
return estimator_spec
elif mode == tf.estimator.ModeKeys.PREDICT:
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.nn.softmax(logits)
max_prob = tf.reduce_max(prob, axis=-1)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
predictions={
'pred_label':pred_label,
"max_prob":max_prob
},
export_outputs={
"output":tf.estimator.export.PredictOutput(
{
'pred_label':pred_label,
"max_prob":max_prob
}
)
}
)
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
def metric_fn(per_example_loss,
logits,
label_ids):
"""Computes the loss and accuracy of the model."""
sentence_log_probs = tf.reshape(
logits, [-1, logits.shape[-1]])
sentence_predictions = tf.argmax(
logits, axis=-1, output_type=tf.int32)
sentence_labels = tf.reshape(label_ids, [-1])
sentence_accuracy = tf.metrics.accuracy(
labels=label_ids, predictions=sentence_predictions)
sentence_mean_loss = tf.metrics.mean(
values=per_example_loss)
sentence_f = tf_metrics.f1(label_ids,
sentence_predictions,
num_labels,
label_lst, average="macro")
eval_metric_ops = {
"f1": sentence_f,
"loss": sentence_mean_loss,
"acc":sentence_accuracy
}
return eval_metric_ops
eval_metric_ops = metric_fn(
per_example_loss,
logits,
label_ids)
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
eval_metric_ops=eval_metric_ops)
if output_type == "sess":
return {
"eval":{
"per_example_loss":per_example_loss,
"logits":logits,
"loss":tf.reduce_mean(per_example_loss)
}
}
elif output_type == "estimator":
return estimator_spec
else:
raise NotImplementedError()
def model_fn(features, labels, mode):
train_ops = []
train_hooks = []
logits_dict = {}
losses_dict = {}
features_dict = {}
tvars = []
task_num_dict = {}
multi_task_config = kargs.get('multi_task_config', {})
total_loss = tf.constant(0.0)
task_num = 0
encoder = {}
hook_dict = {}
print(task_type_dict.keys(), "==task type dict==")
num_task = len(task_type_dict)
from data_generator import load_w2v
flags = kargs.get('flags', Bunch({}))
print(flags.pretrained_w2v_path, "===pretrain vocab path===")
w2v_path = os.path.join(flags.buckets, flags.pretrained_w2v_path)
vocab_path = os.path.join(flags.buckets, flags.vocab_file)
# [w2v_embed, token2id,
# id2token, is_extral_symbol, use_pretrained] = load_w2v.load_pretrained_w2v(vocab_path, w2v_path)
# pretrained_embed = tf.cast(tf.constant(w2v_embed), tf.float32)
pretrained_embed = None
for index, task_type in enumerate(task_type_dict.keys()):
if model_config_dict[task_type].model_type in model_type_lst:
reuse = True
else:
reuse = None
model_type_lst.append(model_config_dict[task_type].model_type)
if model_config_dict[task_type].model_type not in encoder:
model_api = model_zoo(model_config_dict[task_type])
model = model_api(model_config_dict[task_type],
features,
labels,
mode,
target_dict[task_type],
reuse=reuse,
cnn_type=model_config_dict[task_type].get(
'cnn_type', 'bi_dgcnn'))
encoder[model_config_dict[task_type].model_type] = model
# vae_kl_model = vae_model_fn(encoder[model_config_dict[task_type].model_type],
# model_config_dict[task_type],
# num_labels_dict[task_type],
# init_checkpoint_dict[task_type],
# reuse,
# load_pretrained_dict[task_type],
# model_io_config,
# opt_config,
# exclude_scope=exclude_scope_dict[task_type],
# not_storage_params=not_storage_params_dict[task_type],
# target=target_dict[task_type],
# label_lst=None,
# output_type=output_type,
# task_layer_reuse=task_layer_reuse,
# task_type=task_type,
# num_task=num_task,
# task_adversarial=1e-2,
# get_pooled_output='task_output',
# feature_distillation=False,
# embedding_distillation=False,
# pretrained_embed=pretrained_embed,
# **kargs)
# vae_result_dict = vae_kl_model(features, labels, mode)
# tvars.extend(vae_result_dict['tvars'])
# total_loss += vae_result_dict["loss"]
# for key in vae_result_dict:
# if key in ['perplexity', 'token_acc', 'kl_div']:
# hook_dict[key] = vae_result_dict[key]
print(encoder, "==encode==")
if task_type_dict[task_type] == "cls_task":
task_model_fn = cls_model_fn(
encoder[model_config_dict[task_type].model_type],
model_config_dict[task_type],
num_labels_dict[task_type],
init_checkpoint_dict[task_type],
reuse,
load_pretrained_dict[task_type],
model_io_config,
opt_config,
exclude_scope=exclude_scope_dict[task_type],
not_storage_params=not_storage_params_dict[task_type],
target=target_dict[task_type],
label_lst=None,
output_type=output_type,
task_layer_reuse=task_layer_reuse,
task_type=task_type,
num_task=num_task,
task_adversarial=1e-2,
get_pooled_output='task_output',
feature_distillation=False,
embedding_distillation=False,
pretrained_embed=pretrained_embed,
**kargs)
result_dict = task_model_fn(features, labels, mode)
tf.logging.info("****** task: *******",
task_type_dict[task_type], task_type)
elif task_type_dict[task_type] == "embed_task":
task_model_fn = embed_model_fn(
encoder[model_config_dict[task_type].model_type],
model_config_dict[task_type],
num_labels_dict[task_type],
init_checkpoint_dict[task_type],
reuse,
load_pretrained_dict[task_type],
model_io_config,
opt_config,
exclude_scope=exclude_scope_dict[task_type],
not_storage_params=not_storage_params_dict[task_type],
target=target_dict[task_type],
label_lst=None,
output_type=output_type,
task_layer_reuse=task_layer_reuse,
task_type=task_type,
num_task=num_task,
task_adversarial=1e-2,
get_pooled_output='task_output',
feature_distillation=False,
embedding_distillation=False,
pretrained_embed=pretrained_embed,
loss='contrastive_loss',
apply_head_proj=False,
**kargs)
result_dict = task_model_fn(features, labels, mode)
tf.logging.info("****** task: *******",
task_type_dict[task_type], task_type)
# cpc_model_fn = embed_cpc_model_fn(encoder[model_config_dict[task_type].model_type],
# model_config_dict[task_type],
# num_labels_dict[task_type],
# init_checkpoint_dict[task_type],
# reuse,
# load_pretrained_dict[task_type],
# model_io_config,
# opt_config,
# exclude_scope=exclude_scope_dict[task_type],
# not_storage_params=not_storage_params_dict[task_type],
# target=target_dict[task_type],
# label_lst=None,
# output_type=output_type,
# task_layer_reuse=task_layer_reuse,
# task_type=task_type,
# num_task=num_task,
# task_adversarial=1e-2,
# get_pooled_output='task_output',
# feature_distillation=False,
# embedding_distillation=False,
# pretrained_embed=pretrained_embed,
# loss='contrastive_loss',
# apply_head_proj=False,
# **kargs)
# cpc_result_dict = cpc_model_fn(features, labels, mode)
# result_dict['loss'] += cpc_result_dict['loss']
# result_dict['tvars'].extend(cpc_result_dict['tvars'])
# hook_dict["{}_all_neg_loss".format(task_type)] = cpc_result_dict['loss']
# hook_dict["{}_all_neg_num".format(task_type)] = cpc_result_dict['task_num']
elif task_type_dict[task_type] == "cpc_task":
task_model_fn = embed_cpc_v1_model_fn(
encoder[model_config_dict[task_type].model_type],
model_config_dict[task_type],
num_labels_dict[task_type],
init_checkpoint_dict[task_type],
reuse,
load_pretrained_dict[task_type],
model_io_config,
opt_config,
exclude_scope=exclude_scope_dict[task_type],
not_storage_params=not_storage_params_dict[task_type],
target=target_dict[task_type],
label_lst=None,
output_type=output_type,
task_layer_reuse=task_layer_reuse,
task_type=task_type,
num_task=num_task,
task_adversarial=1e-2,
get_pooled_output='task_output',
feature_distillation=False,
embedding_distillation=False,
pretrained_embed=pretrained_embed,
loss='contrastive_loss',
apply_head_proj=False,
task_seperate_proj=True,
**kargs)
result_dict = task_model_fn(features, labels, mode)
tf.logging.info("****** task: *******",
task_type_dict[task_type], task_type)
elif task_type_dict[task_type] == "regression_task":
task_model_fn = regression_model_fn(
encoder[model_config_dict[task_type].model_type],
model_config_dict[task_type],
num_labels_dict[task_type],
init_checkpoint_dict[task_type],
reuse,
load_pretrained_dict[task_type],
model_io_config,
opt_config,
exclude_scope=exclude_scope_dict[task_type],
not_storage_params=not_storage_params_dict[task_type],
target=target_dict[task_type],
label_lst=None,
output_type=output_type,
task_layer_reuse=task_layer_reuse,
task_type=task_type,
num_task=num_task,
task_adversarial=1e-2,
get_pooled_output='task_output',
feature_distillation=False,
embedding_distillation=False,
pretrained_embed=pretrained_embed,
loss='contrastive_loss',
apply_head_proj=False,
**kargs)
result_dict = task_model_fn(features, labels, mode)
tf.logging.info("****** task: *******",
task_type_dict[task_type], task_type)
else:
continue
print("==SUCCEEDED IN LODING==", task_type)
# result_dict = task_model_fn(features, labels, mode)
logits_dict[task_type] = result_dict["logits"]
losses_dict[task_type] = result_dict["loss"] # task loss
for key in [
"pos_num", "neg_num", "masked_lm_loss", "task_loss", "acc",
"task_acc", "masked_lm_acc"
]:
name = "{}_{}".format(task_type, key)
if name in result_dict:
hook_dict[name] = result_dict[name]
hook_dict["{}_loss".format(task_type)] = result_dict["loss"]
hook_dict["{}_num".format(task_type)] = result_dict["task_num"]
print("==loss ratio==", task_type,
multi_task_config[task_type].get('loss_ratio', 1.0))
total_loss += result_dict["loss"] * multi_task_config[
task_type].get('loss_ratio', 1.0)
hook_dict['embed_loss'] = result_dict["embed_loss"]
hook_dict['feature_loss'] = result_dict["feature_loss"]
hook_dict["{}_task_loss".format(
task_type)] = result_dict["task_loss"]
if 'positive_label' in result_dict:
hook_dict["{}_task_positive_label".format(
task_type)] = result_dict["positive_label"]
if mode == tf.estimator.ModeKeys.TRAIN:
tvars.extend(result_dict["tvars"])
task_num += result_dict["task_num"]
task_num_dict[task_type] = result_dict["task_num"]
elif mode == tf.estimator.ModeKeys.EVAL:
features[task_type] = result_dict["feature"]
hook_dict["total_loss"] = total_loss
if mode == tf.estimator.ModeKeys.TRAIN:
model_io_fn = model_io.ModelIO(model_io_config)
optimizer_fn = optimizer.Optimizer(opt_config)
model_io_fn.print_params(list(set(tvars)),
string=", trainable params")
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
print("==update_ops==", update_ops)
with tf.control_dependencies(update_ops):
train_op = optimizer_fn.get_train_op(
total_loss, list(set(tvars)), opt_config.init_lr,
opt_config.num_train_steps, **kargs)
model_io_fn.set_saver(optimizer_fn.opt)
if kargs.get("task_index", 1) == 1 and kargs.get(
"run_config", None):
model_io_fn.get_hooks(kargs.get("checkpoint_dir", None),
kargs.get("num_storage_steps", 1000))
training_hooks = model_io_fn.checkpoint_hook
elif kargs.get("task_index", 1) == 1:
training_hooks = []
else:
training_hooks = []
if len(optimizer_fn.distributed_hooks) >= 1:
training_hooks.extend(optimizer_fn.distributed_hooks)
print(training_hooks, "==training_hooks==", "==task_index==",
kargs.get("task_index", 1))
if output_type == "sess":
return {
"train": {
"total_loss": total_loss,
"loss": losses_dict,
"logits": logits_dict,
"train_op": train_op,
"task_num_dict": task_num_dict
},
"hooks": train_hooks
}
elif output_type == "estimator":
hook_dict['learning_rate'] = optimizer_fn.learning_rate
logging_hook = tf.train.LoggingTensorHook(hook_dict,
every_n_iter=100)
training_hooks.append(logging_hook)
print("==hook_dict==")
print(hook_dict)
for key in hook_dict:
tf.summary.scalar(key, hook_dict[key])
for index, task_type in enumerate(task_type_dict.keys()):
tmp = "{}_loss".format(task_type)
if tmp == key:
tf.summary.scalar(
"loss_gap_{}".format(task_type),
hook_dict["total_loss"] - hook_dict[key])
for key in task_num_dict:
tf.summary.scalar(key + "_task_num", task_num_dict[key])
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op,
training_hooks=training_hooks)
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL: # eval execute for each class solo
def metric_fn(logits, label_ids):
"""Computes the loss and accuracy of the model."""
sentence_log_probs = tf.reshape(logits, [-1, logits.shape[-1]])
sentence_predictions = tf.argmax(logits,
axis=-1,
output_type=tf.int32)
sentence_labels = tf.reshape(label_ids, [-1])
sentence_accuracy = tf.metrics.accuracy(
labels=label_ids, predictions=sentence_predictions)
sentence_f = tf_metrics.f1(label_ids,
sentence_predictions,
num_labels,
label_lst,
average="macro")
eval_metric_ops = {"f1": sentence_f, "acc": sentence_accuracy}
return eval_metric_ops
if output_type == "sess":
return {
"eval": {
"logits": logits_dict,
"total_loss": total_loss,
"feature": features,
"loss": losses_dict
}
}
elif output_type == "estimator":
eval_metric_ops = {}
for key in logits_dict:
eval_dict = metric_fn(logits_dict[key],
features_task_dict[key]["label_ids"])
for sub_key in eval_dict.keys():
eval_key = "{}_{}".format(key, sub_key)
eval_metric_ops[eval_key] = eval_dict[sub_key]
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=total_loss / task_num,
eval_metric_ops=eval_metric_ops)
return estimator_spec
else:
raise NotImplementedError()
def train_eval_fn(FLAGS, worker_count, task_index, is_chief, target,
init_checkpoint, train_file, dev_file, checkpoint_dir,
is_debug, **kargs):
graph = tf.Graph()
with graph.as_default():
import json
config = json.load(open(FLAGS.config_file, "r"))
config = Bunch(config)
config.use_one_hot_embeddings = True
config.scope = "bert"
config.dropout_prob = 0.1
config.label_type = "single_label"
config.model = FLAGS.model_type
config.init_lr = 1e-4
config.ln_type = FLAGS.ln_type
config.loss = 'entropy'
print('==init learning rate==', config.init_lr)
if FLAGS.if_shard == "0":
train_size = FLAGS.train_size
epoch = int(FLAGS.epoch / worker_count)
elif FLAGS.if_shard == "1":
train_size = int(FLAGS.train_size / worker_count)
epoch = FLAGS.epoch
else:
train_size = int(FLAGS.train_size / worker_count)
epoch = FLAGS.epoch
init_lr = config.init_lr
label_dict = json.load(tf.gfile.Open(FLAGS.label_id))
num_train_steps = int(train_size / FLAGS.batch_size * epoch)
num_warmup_steps = int(num_train_steps * 0.1)
num_storage_steps = int(train_size / FLAGS.batch_size)
num_eval_steps = int(FLAGS.eval_size / FLAGS.batch_size)
if is_debug == "0":
num_storage_steps = 2
num_eval_steps = 10
num_train_steps = 10
print("num_train_steps {}, num_eval_steps {}, num_storage_steps {}".
format(num_train_steps, num_eval_steps, num_storage_steps))
print(" model type {}".format(FLAGS.model_type))
print(num_train_steps, num_warmup_steps, "=============")
if worker_count * kargs.get("num_gpus", 1) >= 2:
clip_norm_scale = 1.0
lr_scale = 0.75
else:
clip_norm_scale = 1.0
lr_scale = 1.0
lr = init_lr * worker_count * kargs.get("num_gpus", 1) * lr_scale
# if lr >= 1e-3:
# lr = 1e-3
lr = config.init_lr
print('--training learning rate--', lr)
opt_config = Bunch({
"init_lr": lr,
"num_train_steps": num_train_steps,
"num_warmup_steps": num_warmup_steps,
"worker_count": worker_count,
"opt_type": FLAGS.opt_type,
"is_chief": is_chief,
"train_op": kargs.get("train_op", "adam"),
"decay": kargs.get("decay", "no"),
"warmup": kargs.get("warmup", "no"),
"clip_norm": 1
})
anneal_config = Bunch({
"initial_value": 1.0,
"num_train_steps": num_train_steps
})
model_io_config = Bunch({"fix_lm": False})
model_io_fn = model_io.ModelIO(model_io_config)
num_classes = FLAGS.num_classes
if FLAGS.opt_type == "hvd" and hvd:
checkpoint_dir = checkpoint_dir if task_index == 0 else None
elif FLAGS.opt_type == "all_reduce":
checkpoint_dir = checkpoint_dir
elif FLAGS.opt_type == "collective_reduce":
checkpoint_dir = checkpoint_dir if task_index == 0 else None
elif FLAGS.opt_type == "ps" or FLAGS.opt_type == "ps_sync":
checkpoint_dir = checkpoint_dir if task_index == 0 else None
print("==checkpoint_dir==", checkpoint_dir, is_chief)
model_fn = model_fn_builder(config,
num_classes,
init_checkpoint,
model_reuse=None,
load_pretrained=FLAGS.load_pretrained,
model_io_config=model_io_config,
opt_config=opt_config,
model_io_fn=model_io_fn,
exclude_scope="",
not_storage_params=[],
target=kargs.get("input_target", ""),
output_type="estimator",
checkpoint_dir=checkpoint_dir,
num_storage_steps=num_storage_steps,
task_index=task_index,
anneal_config=anneal_config,
**kargs)
name_to_features = {
"input_ids": tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"input_mask": tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"segment_ids": tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"label_ids": tf.FixedLenFeature([], tf.int64),
}
def _decode_record(record, name_to_features):
"""Decodes a record to a TensorFlow example.
"""
example = tf.parse_single_example(record, name_to_features)
# tf.Example only supports tf.int64, but the TPU only supports tf.int32.
# So cast all int64 to int32.
for name in list(example.keys()):
t = example[name]
if t.dtype == tf.int64:
t = tf.to_int32(t)
example[name] = t
return example
def _decode_batch_record(record, name_to_features):
example = tf.parse_example(record, name_to_features)
# for name in list(example.keys()):
# t = example[name]
# if t.dtype == tf.int64:
# t = tf.to_int32(t)
# example[name] = t
return example
params = Bunch({})
params.epoch = FLAGS.epoch
params.batch_size = FLAGS.batch_size
train_features = lambda: tf_data_utils.all_reduce_train_batch_input_fn(
train_file,
_decode_batch_record,
name_to_features,
params,
if_shard=FLAGS.if_shard,
worker_count=worker_count,
task_index=task_index)
eval_features = lambda: tf_data_utils.all_reduce_eval_batch_input_fn(
dev_file,
_decode_batch_record,
name_to_features,
params,
if_shard=FLAGS.if_shard,
worker_count=worker_count,
task_index=task_index)
sess_config = tf.ConfigProto(allow_soft_placement=False,
log_device_placement=False)
if FLAGS.opt_type == "ps" or FLAGS.opt_type == "ps_sync":
print("==no need for hook==")
elif FLAGS.opt_type == "pai_soar" and pai:
print("no need for hook")
elif FLAGS.opt_type == "hvd" and hvd:
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
print("==no need fo hook==")
else:
print("==no need for hooks==")
if kargs.get("run_config", None):
run_config = kargs.get("run_config", None)
run_config = run_config.replace(
save_checkpoints_steps=num_storage_steps)
print("==run config==", run_config.save_checkpoints_steps)
else:
run_config = tf.estimator.RunConfig(
model_dir=checkpoint_dir,
save_checkpoints_steps=num_storage_steps,
session_config=sess_config)
train_hooks = []
if kargs.get("profiler", "profiler") == "profiler":
if checkpoint_dir:
hooks = tf.train.ProfilerHook(
save_steps=100,
save_secs=None,
output_dir=os.path.join(checkpoint_dir, "profiler"),
)
train_hooks.append(hooks)
print("==add profiler hooks==")
model_estimator = tf.estimator.Estimator(model_fn=model_fn,
model_dir=checkpoint_dir,
config=run_config)
train_being_time = time.time()
tf.logging.info("==training distribution_strategy=={}".format(
kargs.get("distribution_strategy", "MirroredStrategy")))
if kargs.get("distribution_strategy",
"MirroredStrategy") == "MirroredStrategy":
print("==apply single machine multi-card training==")
# model_estimator.train(input_fn=train_features,
# max_steps=num_train_steps,
# hooks=train_hooks)
train_spec = tf.estimator.TrainSpec(input_fn=train_features,
max_steps=num_train_steps)
eval_spec = tf.estimator.EvalSpec(input_fn=eval_features,
steps=num_eval_steps)
model_estimator.train(input_fn=train_features,
max_steps=num_train_steps,
hooks=train_hooks)
# tf.estimator.train(model_estimator, train_spec)
# tf.estimator.evaluate(model_estimator, eval_spec)
train_end_time = time.time()
print("==training time==", train_end_time - train_being_time)
tf.logging.info("==training time=={}".format(train_end_time -
train_being_time))
eval_results = model_estimator.evaluate(input_fn=eval_features,
steps=num_eval_steps)
# print(eval_results)
elif kargs.get("distribution_strategy", "MirroredStrategy") in [
"ParameterServerStrategy", "CollectiveAllReduceStrategy"
]:
print("==apply multi-machine machine multi-card training==")
try:
print(os.environ['TF_CONFIG'], "==tf_run_config==")
except:
print("==not tf config==")
train_spec = tf.estimator.TrainSpec(input_fn=train_features,
max_steps=num_train_steps)
eval_spec = tf.estimator.EvalSpec(input_fn=eval_features,
steps=num_eval_steps)
tf.estimator.train_and_evaluate(model_estimator, train_spec,
eval_spec)
train_end_time = time.time()
print("==training time==", train_end_time - train_being_time)
def model_fn(features, labels, mode, params):
train_op_type = kargs.get('train_op_type', 'joint')
ebm_noise_fce = EBM_NOISE_NCE(
model_config_dict,
num_labels_dict,
init_checkpoint_dict,
load_pretrained_dict,
model_io_config=model_io_config,
opt_config=opt_config,
exclude_scope_dict=exclude_scope_dict,
not_storage_params_dict=not_storage_params_dict,
target_dict=target_dict,
**kargs)
model_io_fn = model_io.ModelIO(model_io_config)
use_tpu = 1 if kargs.get('use_tpu', False) else 0
if mode == tf.estimator.ModeKeys.TRAIN:
if kargs.get('use_tpu', False):
optimizer_fn = optimizer.Optimizer(opt_config)
use_tpu = 1
else:
optimizer_fn = distributed_optimizer.Optimizer(opt_config)
use_tpu = 0
train_op = get_train_op(ebm_noise_fce,
optimizer_fn,
opt_config,
model_config_dict['ebm_dist'],
model_config_dict['noise_dist'],
model_config_dict['generator'],
features,
labels,
mode,
params,
use_tpu=use_tpu,
train_op_type=train_op_type,
alternate_order=['ebm', 'generator'])
ebm_noise_fce.load_pretrained_model(**kargs)
var_checkpoint_dict_list = ebm_noise_fce.var_checkpoint_dict_list
loss = ebm_noise_fce.loss
tvars = ebm_noise_fce.tvars
if len(var_checkpoint_dict_list) >= 1:
scaffold_fn = model_io_fn.load_multi_pretrained(
var_checkpoint_dict_list, use_tpu=use_tpu)
else:
scaffold_fn = None
metric_dict = ebm_train_metric(
ebm_noise_fce.true_ebm_dist_dict['logits'],
ebm_noise_fce.fake_ebm_dist_dict['logits'])
if not kargs.get('use_tpu', False):
for key in metric_dict:
tf.summary.scalar(key, metric_dict[key])
tf.summary.scalar("ebm_loss",
ebm_noise_fce.ebm_opt_dict['ebm_loss'])
tf.summary.scalar("mlm_loss",
ebm_noise_fce.ebm_opt_dict['mlm_loss'])
tf.summary.scalar("all_loss",
ebm_noise_fce.ebm_opt_dict['all_loss'])
model_io_fn.print_params(tvars, string=", trainable params")
if kargs.get('use_tpu', False):
estimator_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
else:
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
ebm_noise_fce.get_loss(features, labels, mode, params, **kargs)
ebm_noise_fce.load_pretrained_model(**kargs)
var_checkpoint_dict_list = ebm_noise_fce.var_checkpoint_dict_list
loss = ebm_noise_fce.loss
if len(var_checkpoint_dict_list) >= 1:
scaffold_fn = model_io_fn.load_multi_pretrained(
var_checkpoint_dict_list, use_tpu=use_tpu)
else:
scaffold_fn = None
tpu_eval_metrics = (ebm_eval_metric, [
ebm_noise_fce.true_ebm_dist_dict['logits'],
ebm_noise_fce.fake_ebm_dist_dict['logits']
])
gpu_eval_metrics = ebm_eval_metric(
ebm_noise_fce.true_ebm_dist_dict['logits'],
ebm_noise_fce.fake_ebm_dist_dict['logits'])
if kargs.get('use_tpu', False):
estimator_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=loss,
eval_metrics=tpu_eval_metrics,
scaffold_fn=scaffold_fn)
else:
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode, loss=loss, eval_metric_ops=gpu_eval_metrics)
return estimator_spec
else:
raise NotImplementedError()
def model_fn(features, labels, mode, params):
model_api = model_zoo(model_config)
model = model_api(model_config, features, labels,
mode, target, reuse=tf.AUTO_REUSE)
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
(nsp_loss,
nsp_per_example_loss,
nsp_log_prob) = pretrain.get_next_sentence_output(model_config,
model.get_pooled_output(),
features['next_sentence_labels'],
reuse=tf.AUTO_REUSE)
masked_lm_positions = features["masked_lm_positions"]
masked_lm_ids = features["masked_lm_ids"]
masked_lm_weights = features["masked_lm_weights"]
if model_config.model_type == 'bert':
masked_lm_fn = pretrain.get_masked_lm_output
print("==apply bert masked lm==")
elif model_config.model_type == 'albert':
masked_lm_fn = pretrain_albert.get_masked_lm_output
print("==apply albert masked lm==")
else:
masked_lm_fn = pretrain.get_masked_lm_output
print("==apply bert masked lm==")
(masked_lm_loss,
masked_lm_example_loss,
masked_lm_log_probs,
masked_lm_mask) = masked_lm_fn(
model_config,
model.get_sequence_output(),
model.get_embedding_table(),
masked_lm_positions,
masked_lm_ids,
masked_lm_weights,
reuse=tf.AUTO_REUSE,
embedding_projection=model.get_embedding_projection_table())
print(model_config.lm_ratio, '==mlm lm_ratio==')
loss = model_config.lm_ratio * masked_lm_loss #+ model_config.nsp_ratio * nsp_loss
model_io_fn = model_io.ModelIO(model_io_config)
pretrained_tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
lm_pretrain_tvars = model_io_fn.get_params("cls/predictions",
not_storage_params=not_storage_params)
pretrained_tvars.extend(lm_pretrain_tvars)
if load_pretrained == "yes":
scaffold_fn = model_io_fn.load_pretrained(pretrained_tvars,
init_checkpoint,
exclude_scope=exclude_scope,
use_tpu=1)
else:
scaffold_fn = None
print("******* scaffold fn *******", scaffold_fn)
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer_fn = optimizer.Optimizer(opt_config)
tvars = pretrained_tvars
model_io_fn.print_params(tvars, string=", trainable params")
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# with tf.control_dependencies(update_ops):
print('==gpu count==', opt_config.get('gpu_count', 1))
train_op = optimizer_fn.get_train_op(loss, tvars,
opt_config.init_lr,
opt_config.num_train_steps,
use_tpu=opt_config.use_tpu)
train_metric_dict = train_metric_fn(
masked_lm_example_loss, masked_lm_log_probs,
masked_lm_ids,
masked_lm_weights,
nsp_per_example_loss,
nsp_log_prob,
features['next_sentence_labels'],
masked_lm_mask=masked_lm_mask
)
# for key in train_metric_dict:
# tf.summary.scalar(key, train_metric_dict[key])
# tf.summary.scalar('learning_rate', optimizer_fn.learning_rate)
estimator_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
def metric_fn(masked_lm_example_loss, masked_lm_log_probs, masked_lm_ids,
masked_lm_weights, next_sentence_example_loss,
next_sentence_log_probs, next_sentence_labels):
"""Computes the loss and accuracy of the model."""
masked_lm_log_probs = tf.reshape(masked_lm_log_probs,
[-1, masked_lm_log_probs.shape[-1]])
masked_lm_predictions = tf.argmax(
masked_lm_log_probs, axis=-1, output_type=tf.int32)
masked_lm_example_loss = tf.reshape(masked_lm_example_loss, [-1])
masked_lm_ids = tf.reshape(masked_lm_ids, [-1])
masked_lm_weights = tf.reshape(masked_lm_weights, [-1])
masked_lm_accuracy = tf.metrics.accuracy(
labels=masked_lm_ids,
predictions=masked_lm_predictions,
weights=masked_lm_weights)
masked_lm_mean_loss = tf.metrics.mean(
values=masked_lm_example_loss, weights=masked_lm_weights)
next_sentence_log_probs = tf.reshape(
next_sentence_log_probs, [-1, next_sentence_log_probs.shape[-1]])
next_sentence_predictions = tf.argmax(
next_sentence_log_probs, axis=-1, output_type=tf.int32)
next_sentence_labels = tf.reshape(next_sentence_labels, [-1])
next_sentence_accuracy = tf.metrics.accuracy(
labels=next_sentence_labels, predictions=next_sentence_predictions)
next_sentence_mean_loss = tf.metrics.mean(
values=next_sentence_example_loss)
return {
"masked_lm_accuracy": masked_lm_accuracy,
"masked_lm_loss": masked_lm_mean_loss,
"next_sentence_accuracy": next_sentence_accuracy,
"next_sentence_loss": next_sentence_mean_loss
}
eval_metrics = (metric_fn, [
masked_lm_example_loss, masked_lm_log_probs, masked_lm_ids,
masked_lm_weights, nsp_per_example_loss,
nsp_log_prob, features['next_sentence_labels']
])
estimator_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
return estimator_spec
else:
raise NotImplementedError()
def model_fn(features, labels, mode):
# model = bert_encoder(model_config, features, labels,
# mode, target, reuse=model_reuse)
model = albert_encoder(model_config, features, labels,
mode, target, reuse=model_reuse)
label_ids = features["label_ids"]
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
with tf.variable_scope(scope, reuse=model_reuse):
(loss,
per_example_loss,
logits) = classifier.classifier(model_config,
model.get_pooled_output(),
num_labels,
label_ids,
dropout_prob)
model_io_fn = model_io.ModelIO(model_io_config)
tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
if load_pretrained == "yes":
model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope=exclude_scope)
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer_fn = optimizer.Optimizer(opt_config)
model_io_fn.print_params(tvars, string=", trainable params")
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
print("==update_ops==", update_ops)
with tf.control_dependencies(update_ops):
train_op = optimizer_fn.get_train_op(loss, tvars,
opt_config.init_lr,
opt_config.num_train_steps,
**kargs)
# train_op, hooks = model_io_fn.get_ema_hooks(train_op,
# tvars,
# kargs.get('params_moving_average_decay', 0.99),
# scope, mode,
# first_stage_steps=opt_config.num_warmup_steps,
# two_stage=True)
model_io_fn.set_saver()
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=loss, train_op=train_op)
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
# _, hooks = model_io_fn.get_ema_hooks(None,
# None,
# kargs.get('params_moving_average_decay', 0.99),
# scope, mode)
hooks = None
def metric_fn(per_example_loss,
logits,
label_ids):
"""Computes the loss and accuracy of the model."""
sentence_log_probs = tf.reshape(
logits, [-1, logits.shape[-1]])
sentence_predictions = tf.argmax(
logits, axis=-1, output_type=tf.int32)
sentence_labels = tf.reshape(label_ids, [-1])
sentence_accuracy = tf.metrics.accuracy(
labels=label_ids, predictions=sentence_predictions)
sentence_mean_loss = tf.metrics.mean(
values=per_example_loss)
sentence_f = tf_metrics.f1(label_ids,
sentence_predictions,
num_labels,
label_lst, average="macro")
eval_metric_ops = {
"f1": sentence_f,
"acc":sentence_accuracy
}
return eval_metric_ops
eval_metric_ops = metric_fn(
per_example_loss,
logits,
label_ids)
eval_hooks = [hooks] if hooks else []
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
eval_metric_ops=eval_metric_ops,
evaluation_hooks=eval_hooks
)
if output_type == "sess":
return {
"eval":{
"per_example_loss":per_example_loss,
"logits":logits,
"loss":tf.reduce_mean(per_example_loss)
}
}
elif output_type == "estimator":
return estimator_spec
else:
raise NotImplementedError()
def model_fn(features, labels, mode):
model_api = model_zoo(model_config)
model = model_api(model_config,
features,
labels,
mode,
target,
reuse=model_reuse,
**kargs)
label_ids = features["label_ids"]
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
with tf.variable_scope(scope, reuse=model_reuse):
(loss, per_example_loss,
logits) = classifier.classifier(model_config,
model.get_pooled_output(),
num_labels, label_ids,
dropout_prob)
model_io_fn = model_io.ModelIO(model_io_config)
tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
try:
params_size = model_io_fn.count_params(model_config.scope)
print("==total params==", params_size)
except:
print("==not count params==")
print(tvars)
if load_pretrained == "yes":
model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope=exclude_scope)
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer_fn = optimizer.Optimizer(opt_config)
model_io_fn.print_params(tvars, string=", trainable params")
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
print("==update_ops==", update_ops)
with tf.control_dependencies(update_ops):
train_op = optimizer_fn.get_train_op(
loss, tvars, opt_config.init_lr,
opt_config.num_train_steps, **kargs)
model_io_fn.set_saver()
if kargs.get("task_index", 1) == 0 and kargs.get(
"run_config", None):
training_hooks = []
elif kargs.get("task_index", 1) == 0:
model_io_fn.get_hooks(kargs.get("checkpoint_dir", None),
kargs.get("num_storage_steps", 1000))
training_hooks = model_io_fn.checkpoint_hook
else:
training_hooks = []
if len(optimizer_fn.distributed_hooks) >= 1:
training_hooks.extend(optimizer_fn.distributed_hooks)
print(training_hooks, "==training_hooks==", "==task_index==",
kargs.get("task_index", 1))
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
training_hooks=training_hooks)
print(tf.global_variables(), "==global_variables==")
if output_type == "sess":
return {
"train": {
"loss": loss,
"logits": logits,
"train_op": train_op
},
"hooks": training_hooks
}
elif output_type == "estimator":
return estimator_spec
elif mode == tf.estimator.ModeKeys.PREDICT:
# if model_config.get('label_type', 'single_label') == 'single_label':
# print(logits.get_shape(), "===logits shape===")
# pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
# prob = tf.nn.softmax(logits)
# max_prob = tf.reduce_max(prob, axis=-1)
# estimator_spec = tf.estimator.EstimatorSpec(
# mode=mode,
# predictions={
# 'pred_label':pred_label,
# "max_prob":max_prob
# },
# export_outputs={
# "output":tf.estimator.export.PredictOutput(
# {
# 'pred_label':pred_label,
# "max_prob":max_prob
# }
# )
# }
# )
if model_config.get('label_type', 'single_label') == 'multi_label':
prob = tf.nn.sigmoid(logits)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
predictions={
'pred_label': prob,
"max_prob": prob
},
export_outputs={
"output":
tf.estimator.export.PredictOutput({
'pred_label': prob,
"max_prob": prob
})
})
elif model_config.get('label_type',
'single_label') == "single_label":
prob = tf.nn.softmax(logits)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
predictions={
'pred_label': prob,
"max_prob": prob
},
export_outputs={
"output":
tf.estimator.export.PredictOutput({
'pred_label': prob,
"max_prob": prob
})
})
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
def metric_fn(per_example_loss, logits, label_ids):
"""Computes the loss and accuracy of the model."""
sentence_log_probs = tf.reshape(logits, [-1, logits.shape[-1]])
sentence_predictions = tf.argmax(logits,
axis=-1,
output_type=tf.int32)
sentence_labels = tf.reshape(label_ids, [-1])
sentence_accuracy = tf.metrics.accuracy(
labels=label_ids, predictions=sentence_predictions)
sentence_mean_loss = tf.metrics.mean(values=per_example_loss)
sentence_f = tf_metrics.f1(label_ids,
sentence_predictions,
num_labels,
label_lst,
average="macro")
eval_metric_ops = {"f1": sentence_f, "acc": sentence_accuracy}
return eval_metric_ops
if output_type == "sess":
return {
"eval": {
"per_example_loss": per_example_loss,
"logits": logits,
"loss": tf.reduce_mean(per_example_loss),
"feature": model.get_pooled_output()
}
}
elif output_type == "estimator":
eval_metric_ops = metric_fn(per_example_loss, logits,
label_ids)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode, loss=loss, eval_metric_ops=eval_metric_ops)
return estimator_spec
else:
raise NotImplementedError()
def model_fn(features, labels, mode):
shape_lst_a = bert_utils.get_shape_list(features['input_ids_a'])
batch_size_a = shape_lst_a[0]
total_length_a = shape_lst_a[1]
shape_lst_b = bert_utils.get_shape_list(features['input_ids_b'])
batch_size_b = shape_lst_b[0]
total_length_b = shape_lst_b[1]
features['input_ids_a'] = tf.reshape(features['input_ids_a'],
[-1, model_config.max_length])
features['segment_ids_a'] = tf.reshape(features['segment_ids_a'],
[-1, model_config.max_length])
features['input_mask_a'] = tf.cast(
tf.not_equal(features['input_ids_a'], kargs.get('[PAD]', 0)),
tf.int64)
features['input_ids_b'] = tf.reshape(
features['input_ids_b'],
[-1, model_config.max_predictions_per_seq])
features['segment_ids_b'] = tf.reshape(
features['segment_ids_b'],
[-1, model_config.max_predictions_per_seq])
features['input_mask_b'] = tf.cast(
tf.not_equal(features['input_ids_b'], kargs.get('[PAD]', 0)),
tf.int64)
features['batch_size'] = batch_size_a
features['total_length_a'] = total_length_a
features['total_length_b'] = total_length_b
model_dict = {}
for target in ["a", "b"]:
model = bert_encoder(model_config,
features,
labels,
mode,
target,
reuse=tf.AUTO_REUSE)
model_dict[target] = model
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
with tf.variable_scope(scope, reuse=model_reuse):
(loss, per_example_loss, logits,
transition_params) = multi_position_crf_classifier(
model_config, features, model_dict, num_labels, dropout_prob)
model_io_fn = model_io.ModelIO(model_io_config)
tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
try:
params_size = model_io_fn.count_params(model_config.scope)
print("==total params==", params_size)
except:
print("==not count params==")
print(tvars)
if load_pretrained == "yes":
model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope=exclude_scope)
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer_fn = optimizer.Optimizer(opt_config)
model_io_fn.print_params(tvars, string=", trainable params")
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
print("==update_ops==", update_ops)
with tf.control_dependencies(update_ops):
train_op = optimizer_fn.get_train_op(
loss, tvars, opt_config.init_lr,
opt_config.num_train_steps, **kargs)
train_op, hooks = model_io_fn.get_ema_hooks(
train_op,
tvars,
kargs.get('params_moving_average_decay', 0.99),
scope,
mode,
first_stage_steps=opt_config.num_warmup_steps,
two_stage=True)
model_io_fn.set_saver()
if kargs.get("task_index", 1) == 0 and kargs.get(
"run_config", None):
training_hooks = []
elif kargs.get("task_index", 1) == 0:
model_io_fn.get_hooks(kargs.get("checkpoint_dir", None),
kargs.get("num_storage_steps", 1000))
training_hooks = model_io_fn.checkpoint_hook
else:
training_hooks = []
if len(optimizer_fn.distributed_hooks) >= 1:
training_hooks.extend(optimizer_fn.distributed_hooks)
print(training_hooks, "==training_hooks==", "==task_index==",
kargs.get("task_index", 1))
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
training_hooks=training_hooks)
print(tf.global_variables(), "==global_variables==")
if output_type == "sess":
return {
"train": {
"loss": loss,
"logits": logits,
"train_op": train_op
},
"hooks": training_hooks
}
elif output_type == "estimator":
return estimator_spec
elif mode == tf.estimator.ModeKeys.PREDICT:
print(logits.get_shape(), "===logits shape===")
label_weights = tf.cast(features['label_weights'], tf.int32)
label_seq_length = tf.reduce_sum(label_weights, axis=-1)
decode_tags, best_score = tf.contrib.crf.crf_decode(
logits, transition_params, label_seq_length)
_, hooks = model_io_fn.get_ema_hooks(
None, None, kargs.get('params_moving_average_decay', 0.99),
scope, mode)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
predictions={
'decode_tags': decode_tags,
"best_score": best_score,
"transition_params": transition_params,
"logits": logits
},
export_outputs={
"output":
tf.estimator.export.PredictOutput({
'decode_tags': decode_tags,
"best_score": best_score,
"transition_params": transition_params,
"logits": logits
})
},
prediction_hooks=[hooks])
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
_, hooks = model_io_fn.get_ema_hooks(
None, None, kargs.get('params_moving_average_decay', 0.99),
scope, mode)
eval_hooks = []
if output_type == "sess":
return {
"eval": {
"per_example_loss": per_example_loss,
"logits": logits,
"loss": tf.reduce_mean(per_example_loss),
"feature": model.get_pooled_output()
}
}
elif output_type == "estimator":
eval_metric_ops = eval_logtis(logits, features, num_labels,
transition_params)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
eval_metric_ops=eval_metric_ops,
evaluation_hooks=eval_hooks)
return estimator_spec
else:
raise NotImplementedError()
def model_fn(features, labels, mode):
model_api = model_zoo(model_config)
model = model_api(model_config,
features,
labels,
mode,
target,
reuse=model_reuse)
label_ids = features["label_ids"]
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
else:
dropout_prob = 0.0
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
with tf.variable_scope(scope, reuse=model_reuse):
(loss, per_example_loss,
logits) = classifier.classifier(model_config,
model.get_pooled_output(),
num_labels, label_ids,
dropout_prob)
label_loss = tf.reduce_sum(
per_example_loss * features["label_ratio"]) / (
1e-10 + tf.reduce_sum(features["label_ratio"]))
if mode == tf.estimator.ModeKeys.TRAIN:
distillation_api = distill.KnowledgeDistillation(
kargs.get(
"disitllation_config",
Bunch({
"logits_ratio_decay": "constant",
"logits_ratio": 0.5,
"logits_decay_rate": 0.999,
"distillation": ['relation_kd', 'logits'],
"feature_ratio": 0.5,
"feature_ratio_decay": "constant",
"feature_decay_rate": 0.999,
"kd_type": "kd",
"scope": scope
})))
# get teacher logits
teacher_logit = tf.log(features["label_probs"] +
1e-10) / kargs.get(
"temperature",
2.0) # log_softmax logits
student_logit = tf.nn.log_softmax(
logits / kargs.get("temperature", 2.0)) # log_softmax logits
distillation_features = {
"student_logits_tensor": student_logit,
"teacher_logits_tensor": teacher_logit,
"student_feature_tensor": model.get_pooled_output(),
"teacher_feature_tensor": features["distillation_feature"],
"student_label": tf.ones_like(label_ids, dtype=tf.int32),
"teacher_label": tf.zeros_like(label_ids, dtype=tf.int32),
"logits_ratio": kargs.get("logits_ratio", 0.5),
"feature_ratio": kargs.get("logits_ratio", 0.5),
"distillation_ratio": features["distillation_ratio"],
"src_f_logit": logits,
"tgt_f_logit": logits,
"src_tensor": model.get_pooled_output(),
"tgt_tensor": features["distillation_feature"]
}
distillation_loss = distillation_api.distillation(
distillation_features,
2,
dropout_prob,
model_reuse,
opt_config.num_train_steps,
feature_ratio=1.0,
logits_ratio_decay="constant",
feature_ratio_decay="constant",
feature_decay_rate=0.999,
logits_decay_rate=0.999,
logits_ratio=0.5,
scope=scope + "/adv_classifier",
num_classes=num_labels,
gamma=kargs.get("gamma", 4))
loss = label_loss + distillation_loss["distillation_loss"]
model_io_fn = model_io.ModelIO(model_io_config)
tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
print(tvars)
if load_pretrained == "yes":
model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope=exclude_scope)
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer_fn = optimizer.Optimizer(opt_config)
model_io_fn.print_params(tvars, string=", trainable params")
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer_fn.get_train_op(
loss, tvars, opt_config.init_lr,
opt_config.num_train_steps, **kargs)
model_io_fn.set_saver()
if kargs.get("task_index", 1) == 0 and kargs.get(
"run_config", None):
training_hooks = []
elif kargs.get("task_index", 1) == 0:
model_io_fn.get_hooks(kargs.get("checkpoint_dir", None),
kargs.get("num_storage_steps", 1000))
training_hooks = model_io_fn.checkpoint_hook
else:
training_hooks = []
if len(optimizer_fn.distributed_hooks) >= 1:
training_hooks.extend(optimizer_fn.distributed_hooks)
print(training_hooks, "==training_hooks==", "==task_index==",
kargs.get("task_index", 1))
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
training_hooks=training_hooks)
if output_type == "sess":
return {
"train": {
"loss":
loss,
"logits":
logits,
"train_op":
train_op,
"cross_entropy":
label_loss,
"distillation_loss":
distillation_loss["distillation_loss"],
"kd_num":
tf.reduce_sum(features["distillation_ratio"]),
"ce_num":
tf.reduce_sum(features["label_ratio"]),
"label_ratio":
features["label_ratio"],
"distilaltion_logits_loss":
distillation_loss["distillation_logits_loss"],
"distilaltion_feature_loss":
distillation_loss["distillation_feature_loss"],
"rkd_loss":
distillation_loss["rkd_loss"]
},
"hooks": training_hooks
}
elif output_type == "estimator":
return estimator_spec
elif mode == tf.estimator.ModeKeys.PREDICT:
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.nn.softmax(logits)
max_prob = tf.reduce_max(prob, axis=-1)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
predictions={
'pred_label': pred_label,
"max_prob": max_prob
},
export_outputs={
"output":
tf.estimator.export.PredictOutput({
'pred_label': pred_label,
"max_prob": max_prob
})
})
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
def metric_fn(per_example_loss, logits, label_ids):
"""Computes the loss and accuracy of the model."""
sentence_log_probs = tf.reshape(logits, [-1, logits.shape[-1]])
sentence_predictions = tf.argmax(logits,
axis=-1,
output_type=tf.int32)
sentence_labels = tf.reshape(label_ids, [-1])
sentence_accuracy = tf.metrics.accuracy(
labels=label_ids, predictions=sentence_predictions)
sentence_mean_loss = tf.metrics.mean(values=per_example_loss)
sentence_f = tf_metrics.f1(label_ids,
sentence_predictions,
num_labels,
label_lst,
average="macro")
eval_metric_ops = {"f1": sentence_f, "acc": sentence_accuracy}
return eval_metric_ops
eval_metric_ops = metric_fn(per_example_loss, logits, label_ids)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode, loss=loss, eval_metric_ops=eval_metric_ops)
if output_type == "sess":
return {
"eval": {
"per_example_loss": per_example_loss,
"logits": logits,
"loss": tf.reduce_mean(per_example_loss)
}
}
elif output_type == "estimator":
return estimator_spec
else:
raise NotImplementedError()
def main(_):
hvd.init()
sess_config = tf.ConfigProto()
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
graph = tf.Graph()
with graph.as_default():
import json
config = json.load(open(FLAGS.config_file, "r"))
init_checkpoint = FLAGS.init_checkpoint
config = Bunch(config)
config.use_one_hot_embeddings = True
config.scope = "bert"
config.dropout_prob = 0.1
config.label_type = "single_label"
if FLAGS.if_shard == "0":
train_size = FLAGS.train_size
epoch = int(FLAGS.epoch / hvd.size())
elif FLAGS.if_shard == "1":
train_size = int(FLAGS.train_size/hvd.size())
epoch = FLAGS.epoch
tokenizer = tokenization.FullTokenizer(
vocab_file=FLAGS.vocab_file,
do_lower_case=FLAGS.lower_case)
classifier_data_api = classifier_processor.EvaluationProcessor()
classifier_data_api.get_labels(FLAGS.label_id)
train_examples = classifier_data_api.get_train_examples(FLAGS.train_file)
write_to_tfrecords.convert_classifier_examples_to_features(train_examples,
classifier_data_api.label2id,
FLAGS.max_length,
tokenizer,
FLAGS.eval_data_file)
init_lr = 2e-5
num_train_steps = int(
train_size / FLAGS.batch_size * epoch)
num_warmup_steps = int(num_train_steps * 0.1)
num_storage_steps = int(train_size / FLAGS.batch_size)
print(" model type {}".format(FLAGS.model_type))
print(num_train_steps, num_warmup_steps, "=============")
opt_config = Bunch({"init_lr":init_lr/hvd.size(),
"num_train_steps":num_train_steps,
"num_warmup_steps":num_warmup_steps})
sess = tf.Session(config=sess_config)
model_io_config = Bunch({"fix_lm":False})
model_io_fn = model_io.ModelIO(model_io_config)
optimizer_fn = optimizer.Optimizer(opt_config)
num_classes = FLAGS.num_classes
model_eval_fn = bert_classifier.classifier_model_fn_builder(config, num_classes, init_checkpoint,
reuse=False,
load_pretrained=True,
model_io_fn=model_io_fn,
optimizer_fn=optimizer_fn,
model_io_config=model_io_config,
opt_config=opt_config)
def metric_fn(features, logits, loss):
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.nn.softmax(logits)
accuracy = correct = tf.equal(
tf.cast(pred_label, tf.int32),
tf.cast(features["label_ids"], tf.int32)
)
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
return {"accuracy":accuracy, "loss":loss, "pred_label":pred_label,
"label_ids":features["label_ids"],
"prob":prob}
name_to_features = {
"input_ids":
tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"input_mask":
tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"segment_ids":
tf.FixedLenFeature([FLAGS.max_length], tf.int64),
"label_ids":
tf.FixedLenFeature([], tf.int64),
}
def _decode_record(record, name_to_features):
"""Decodes a record to a TensorFlow example.
"""
example = tf.parse_single_example(record, name_to_features)
# tf.Example only supports tf.int64, but the TPU only supports tf.int32.
# So cast all int64 to int32.
for name in list(example.keys()):
t = example[name]
if t.dtype == tf.int64:
t = tf.to_int32(t)
example[name] = t
return example
params = Bunch({})
params.epoch = FLAGS.epoch
params.batch_size = FLAGS.batch_size
eval_features = tf_data_utils.eval_input_fn(FLAGS.eval_data_file,
_decode_record, name_to_features, params, if_shard=FLAGS.if_shard)
[_, eval_loss, eval_per_example_loss, eval_logits] = model_eval_fn(eval_features, [], tf.estimator.ModeKeys.EVAL)
result = metric_fn(eval_features, eval_logits, eval_loss)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
sess.run(hvd.broadcast_global_variables(0))
print("===horovod rank==={}".format(hvd.rank()))
def eval_fn(result):
i = 0
total_accuracy = 0
label, label_id, prob = [], [], []
while True:
try:
eval_result = sess.run(result)
total_accuracy += eval_result["accuracy"]
label_id.extend(eval_result["label_ids"])
label.extend(eval_result["pred_label"])
prob.extend(eval_result["prob"])
i += 1
except tf.errors.OutOfRangeError:
print("End of dataset")
break
macro_f1 = f1_score(label_id, label, average="macro")
micro_f1 = f1_score(label_id, label, average="micro")
macro_precision = precision_score(label_id, label, average="macro")
micro_precision = precision_score(label_id, label, average="micro")
macro_recall = recall_score(label_id, label, average="macro")
micro_recall = recall_score(label_id, label, average="micro")
accuracy = accuracy_score(label_id, label)
print("test accuracy {} macro_f1 score {} micro_f1 {} accuracy {}".format(total_accuracy/ i,
macro_f1, micro_f1, accuracy))
return total_accuracy/ i, label_id, label, prob
import time
import time
start = time.time()
acc, true_label, pred_label, prob = eval_fn(result)
end = time.time()
print("==total time {} numbers of devices {}".format(end - start, hvd.size()))
if hvd.rank() == 0:
import _pickle as pkl
pkl.dump({"true_label":true_label,
"pred_label":pred_label,
"prob":prob},
open(FLAGS.model_output+"/predict.pkl", "wb"))
def model_fn(features, labels, mode):
model = gpt_encoder(model_config, features, labels,
mode, target, reuse=tf.AUTO_REUSE)
scope = model_config.scope
if mode == tf.estimator.ModeKeys.TRAIN:
# batch x seq_length
sequence_mask = tf.to_float(tf.not_equal(features['input_ids'][:, 1:],
kargs.get('[PAD]', 0)))
# batch x seq_length
seq_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=features['input_ids'][:, 1:],
logits=model.get_sequence_output_logits()[:, :-1])
per_example_loss = tf.reduce_sum(seq_loss*sequence_mask, axis=-1) / (tf.reduce_sum(sequence_mask, axis=-1)+1e-10)
loss = tf.reduce_mean(per_example_loss)
model_io_fn = model_io.ModelIO(model_io_config)
tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
try:
params_size = model_io_fn.count_params(model_config.scope)
print("==total params==", params_size)
except:
print("==not count params==")
print(tvars)
if load_pretrained == "yes":
model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope=exclude_scope)
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer_fn = optimizer.Optimizer(opt_config)
model_io_fn.print_params(tvars, string=", trainable params")
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
print("==update_ops==", update_ops)
with tf.control_dependencies(update_ops):
train_op = optimizer_fn.get_train_op(loss, tvars,
opt_config.init_lr,
opt_config.num_train_steps,
**kargs)
model_io_fn.set_saver()
if kargs.get("task_index", 1) == 0 and kargs.get("run_config", None):
training_hooks = []
elif kargs.get("task_index", 1) == 0:
model_io_fn.get_hooks(kargs.get("checkpoint_dir", None),
kargs.get("num_storage_steps", 1000))
training_hooks = model_io_fn.checkpoint_hook
else:
training_hooks = []
if len(optimizer_fn.distributed_hooks) >= 1:
training_hooks.extend(optimizer_fn.distributed_hooks)
print(training_hooks, "==training_hooks==", "==task_index==", kargs.get("task_index", 1))
train_metric_dict = train_metric(features['input_ids'],
model.get_sequence_output_logits(),
**kargs)
for key in train_metric_dict:
tf.summary.scalar(key, train_metric_dict[key])
tf.summary.scalar('learning_rate', optimizer_fn.learning_rate)
tf.summary.scalar('seq_length', tf.reduce_mean(tf.reduce_sum(sequence_mask, axis=-1)))
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=loss, train_op=train_op,
training_hooks=training_hooks)
print(tf.global_variables(), "==global_variables==")
if output_type == "sess":
return {
"train":{
"loss":loss,
"logits":logits,
"train_op":train_op
},
"hooks":training_hooks
}
elif output_type == "estimator":
return estimator_spec
elif mode == tf.estimator.ModeKeys.PREDICT:
if kargs.get('predict_type', 'sample_sequence') == 'sample_sequence':
results = sample.sample_sequence(
gpt_encoder, hparams=model_config,
length=kargs.get('max_length', 64),
start_token=None,
batch_size=10,
context=features['input_ids'],
temperature=2,
top_k=10)
sampled_token = results['tokens'][:, 1:]
sampled_token_logits = results['logits'][:, 1:]
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
predictions={
'token':sampled_token,
"logits":sampled_token_logits
},
export_outputs={
"output":tf.estimator.export.PredictOutput(
{
'token':sampled_token,
"logits":sampled_token_logits
}
)
}
)
return estimator_spec
elif kargs.get('predict_type', 'sample_sequence') == 'infer_inputs':
sequence_mask = tf.to_float(tf.not_equal(features['input_ids'][:, 1:],
kargs.get('[PAD]', 0)))
output_logits = model.get_sequence_output_logits()[:, :-1]
# output_logits = tf.nn.log_softmax(output_logits, axis=-1)
output_id_logits = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=features['input_ids'][:, 1:],
logits=output_logits)
per_example_perplexity = tf.reduce_sum(output_id_logits * sequence_mask,
axis=-1) # batch
per_example_perplexity /= tf.reduce_sum(sequence_mask, axis=-1) # batch
perplexity = tf.exp(per_example_perplexity)
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
predictions={
'token':features['input_ids'][:, 1:],
"logits":output_id_logits,
'perplexity':perplexity
},
export_outputs={
"output":tf.estimator.export.PredictOutput(
{
'token':features['input_ids'][:,1:],
"logits":output_id_logits,
'perplexity':perplexity
}
)
}
)
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL:
def metric_fn(per_example_loss,
logits,
label_ids):
"""Computes the loss and accuracy of the model."""
sentence_log_probs = tf.reshape(
logits, [-1, logits.shape[-1]])
sentence_predictions = tf.argmax(
logits, axis=-1, output_type=tf.int32)
sentence_labels = tf.reshape(label_ids, [-1])
sentence_accuracy = tf.metrics.accuracy(
labels=label_ids, predictions=sentence_predictions)
sentence_mean_loss = tf.metrics.mean(
values=per_example_loss)
sentence_f = tf_metrics.f1(label_ids,
sentence_predictions,
num_labels,
label_lst, average="macro")
eval_metric_ops = {
"f1": sentence_f,
"acc":sentence_accuracy
}
return eval_metric_ops
if output_type == "sess":
return {
"eval":{
"per_example_loss":per_example_loss,
"logits":logits,
"loss":tf.reduce_mean(per_example_loss),
"feature":model.get_pooled_output()
}
}
elif output_type == "estimator":
eval_metric_ops = metric_fn(
per_example_loss,
logits,
label_ids)
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
eval_metric_ops=eval_metric_ops)
return estimator_spec
else:
raise NotImplementedError()
def main(_):
graph = tf.Graph()
# from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
with graph.as_default():
import json
tokenizer = tokenization.FullTokenizer(
vocab_file=FLAGS.vocab_file,
do_lower_case=True)
classifier_data_api = classifier_processor.PiarInteractionProcessor()
eval_examples = classifier_data_api.get_test_examples(FLAGS.eval_data_file,
FLAGS.lang)
print(eval_examples[0].guid)
label_tensor = None
label_id = json.load(open(FLAGS.label_id, "r"))
num_choice = 3
write_to_tfrecords.convert_interaction_classifier_examples_to_features_v1(
eval_examples,
label_id["label2id"],
FLAGS.max_length,
tokenizer,
FLAGS.output_file)
config = json.load(open(FLAGS.config_file, "r"))
init_checkpoint = FLAGS.init_checkpoint
max_seq_length = FLAGS.max_length * 2 + 3
print("===init checkoutpoint==={}".format(init_checkpoint))
config = Bunch(config)
config.use_one_hot_embeddings = True
config.scope = "esim/bert"
config.dropout_prob = 0.2
config.label_type = "single_label"
config.lstm_dim = 128
config.num_heads = 12
config.num_units = 768
# os.environ["CUDA_VISIBLE_DEVICES"] = "2"
sess = tf.Session()
opt_config = Bunch({"init_lr":(5e-5),
"num_train_steps":0,
"num_warmup_steps":0,
"train_op":"adam"})
model_io_config = Bunch({"fix_lm":False})
model_io_fn = model_io.ModelIO(model_io_config)
model_function = bert_esim.classifier_attn_model_fn_builder
model_eval_fn = model_function(
config,
num_choice,
init_checkpoint,
model_reuse=None,
load_pretrained=True,
model_io_fn=model_io_fn,
model_io_config=model_io_config,
opt_config=opt_config,
input_name=["a", "b"],
label_tensor=label_tensor,
not_storage_params=["adam", "adam_1"],
exclude_scope_dict={"task":"esim"})
# def metric_fn(features, logits):
# print(logits.get_shape(), "===logits shape===")
# pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
# return {"pred_label":pred_label, "qas_id":features["qas_id"]}
def metric_fn(features, logits):
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.exp(tf.nn.log_softmax(logits))
return {"pred_label":pred_label,
"qas_id":features["qas_id"],
"prob":prob}
name_to_features = {
"input_ids_a":
tf.FixedLenFeature([max_seq_length], tf.int64),
"input_mask_a":
tf.FixedLenFeature([max_seq_length], tf.int64),
"segment_ids_a":
tf.FixedLenFeature([max_seq_length], tf.int64),
"input_ids_b":
tf.FixedLenFeature([max_seq_length], tf.int64),
"input_mask_b":
tf.FixedLenFeature([max_seq_length], tf.int64),
"segment_ids_b":
tf.FixedLenFeature([max_seq_length], tf.int64),
"label_ids":
tf.FixedLenFeature([], tf.int64),
"qas_id":
tf.FixedLenFeature([], tf.int64),
}
def _decode_record(record, name_to_features):
"""Decodes a record to a TensorFlow example.
"""
example = tf.parse_single_example(record, name_to_features)
# tf.Example only supports tf.int64, but the TPU only supports tf.int32.
# So cast all int64 to int32.
for name in list(example.keys()):
t = example[name]
if t.dtype == tf.int64:
t = tf.to_int32(t)
example[name] = t
return example
params = Bunch({})
params.epoch = 2
params.batch_size = 32
eval_features = tf_data_utils.eval_input_fn(FLAGS.output_file,
_decode_record, name_to_features, params)
[_, eval_loss, eval_per_example_loss, eval_logits] = model_eval_fn(eval_features, [], tf.estimator.ModeKeys.EVAL)
result = metric_fn(eval_features, eval_logits)
model_io_fn.set_saver()
init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
sess.run(init_op)
model_io_fn.load_model(sess, init_checkpoint)
print(" ==succeeded in loading model== ")
def eval_fn(result):
i = 0
pred_label, qas_id, prob = [], [], []
while True:
try:
eval_result = sess.run(result)
pred_label.extend(eval_result["pred_label"].tolist())
qas_id.extend(eval_result["qas_id"].tolist())
prob.extend(eval_result["prob"].tolist())
i += 1
except tf.errors.OutOfRangeError:
print("End of dataset")
break
return pred_label, qas_id, prob
print("===========begin to eval============")
[pred_label, qas_id, prob] = eval_fn(result)
result = dict(zip(qas_id, pred_label))
print(FLAGS.result_file.split("."))
tmp_output = FLAGS.result_file.split(".")[0] + ".json"
print(tmp_output, "===temp output===")
json.dump({"id":qas_id,
"label":pred_label,
"prob":prob},
open(tmp_output, "w"))
print(len(result), "=====valid result======")
print(len(result), "=====valid result======")
import pandas as pd
df = pd.read_csv(FLAGS.eval_data_file)
output = {}
for index in range(df.shape[0]):
output[df.loc[index]["id"]] = ""
final_output = []
cnt = 0
for key in output:
if key in result:
final_output.append({"Id":key,
"Category":label_id["id2label"][str(result[key])]})
cnt += 1
else:
final_output.append({"Id":key, "Category":"unrelated"})
df_out = pd.DataFrame(final_output)
df_out.to_csv(FLAGS.result_file)
print(len(output), cnt, len(final_output), "======num of results from model==========")
def main(_):
hvd.init()
sess_config = tf.ConfigProto()
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
graph = tf.Graph()
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
with graph.as_default():
import json
# config = json.load(open("/data/xuht/bert/chinese_L-12_H-768_A-12/bert_config.json", "r"))
config = json.load(open(FLAGS.config_file, "r"))
init_checkpoint = FLAGS.init_checkpoint
print("===init checkoutpoint==={}".format(init_checkpoint))
# init_checkpoint = "/data/xuht/bert/chinese_L-12_H-768_A-12/bert_model.ckpt"
# init_checkpoint = "/data/xuht/concat/model_1/oqmrc.ckpt"
config = Bunch(config)
config.use_one_hot_embeddings = True
config.scope = "esim/bert"
config.dropout_prob = 0.1
config.label_type = "single_label"
config.lstm_dim = 128
config.num_heads = 4
import json
label_dict = json.load(open(FLAGS.label_id))
# label_tensor = np.asarray(label_dict["class_ratio"]).astype(np.float32)
label_tensor = None
# config.loss = "focal_loss"
json.dump(config, open(FLAGS.model_output + "/config.json", "w"))
# os.environ["CUDA_VISIBLE_DEVICES"] = FLAGS.gpu_id
sess = tf.Session(config=sess_config)
train_size = int(FLAGS.train_size / hvd.size())
num_train_steps = int(train_size / FLAGS.batch_size * FLAGS.epoch)
num_warmup_steps = int(num_train_steps * 0.01)
num_storage_steps = int(train_size / FLAGS.batch_size)
print(num_train_steps, num_warmup_steps, "=============")
opt_config = Bunch({
"init_lr": (5e-5 / hvd.size()),
"num_train_steps": num_train_steps,
"num_warmup_steps": num_warmup_steps,
"train_op": "adam"
})
model_io_config = Bunch({"fix_lm": True})
model_io_fn = model_io.ModelIO(model_io_config)
num_choice = FLAGS.num_classes
max_seq_length = FLAGS.max_length
if FLAGS.model_type == "original":
model_function = bert_order_classifier.classifier_model_fn_builder
elif FLAGS.model_type == "attn":
model_function = bert_order_classifier.classifier_attn_model_fn_builder
elif FLAGS.model_type == "orignal_nonlinear":
model_function = bert_order_classifier.classifier_model_fn_builder_v1
elif FLAGS.model_type == "esim_bert":
model_function = esim_bert.classifier_attn_model_fn_builder
model_eval_fn = model_function(config,
num_choice,
init_checkpoint,
model_reuse=None,
load_pretrained=True,
model_io_fn=model_io_fn,
model_io_config=model_io_config,
opt_config=opt_config,
input_name=["a", "b"],
label_tensor=label_tensor,
not_storage_params=["adam", "adam_1"],
exclude_scope_dict={"task": "esim"})
def metric_fn(features, logits, loss):
print(logits.get_shape(), "===logits shape===")
pred_label = tf.argmax(logits, axis=-1, output_type=tf.int32)
prob = tf.nn.softmax(logits)
accuracy = correct = tf.equal(
tf.cast(pred_label, tf.int32),
tf.cast(features["label_ids"], tf.int32))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
return {
"accuracy": accuracy,
"loss": loss,
"pred_label": pred_label,
"label_ids": features["label_ids"]
}
name_to_features = {
"input_ids_a": tf.FixedLenFeature([max_seq_length], tf.int64),
"input_mask_a": tf.FixedLenFeature([max_seq_length], tf.int64),
"segment_ids_a": tf.FixedLenFeature([max_seq_length], tf.int64),
"input_ids_b": tf.FixedLenFeature([max_seq_length], tf.int64),
"input_mask_b": tf.FixedLenFeature([max_seq_length], tf.int64),
"segment_ids_b": tf.FixedLenFeature([max_seq_length], tf.int64),
"label_ids": tf.FixedLenFeature([], tf.int64),
}
def _decode_record(record, name_to_features):
"""Decodes a record to a TensorFlow example.
"""
example = tf.parse_single_example(record, name_to_features)
# tf.Example only supports tf.int64, but the TPU only supports tf.int32.
# So cast all int64 to int32.
for name in list(example.keys()):
t = example[name]
if t.dtype == tf.int64:
t = tf.to_int32(t)
example[name] = t
return example
params = Bunch({})
params.epoch = FLAGS.epoch
params.batch_size = FLAGS.batch_size
# train_features = tf_data_utils.train_input_fn("/data/xuht/wsdm19/data/train.tfrecords",
# _decode_record, name_to_features, params)
# eval_features = tf_data_utils.eval_input_fn("/data/xuht/wsdm19/data/dev.tfrecords",
# _decode_record, name_to_features, params)
# train_features = tf_data_utils.train_input_fn(FLAGS.train_file,
# _decode_record, name_to_features, params)
eval_features = tf_data_utils.eval_input_fn(FLAGS.dev_file,
_decode_record,
name_to_features, params)
# [train_op, train_loss, train_per_example_loss, train_logits] = model_train_fn(train_features, [], tf.estimator.ModeKeys.TRAIN)
[_, eval_loss, eval_per_example_loss,
eval_logits] = model_eval_fn(eval_features, [],
tf.estimator.ModeKeys.EVAL)
result = metric_fn(eval_features, eval_logits, eval_loss)
model_io_fn.set_saver()
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
model_io_fn.load_model(sess, init_checkpoint)
print(" ==succeeded in loading model== ")
sess.run(hvd.broadcast_global_variables(0))
def eval_fn(result):
i = 0
total_accuracy = 0
label, label_id = [], []
# label_weight = []
while True:
try:
eval_result = sess.run(result)
total_accuracy += eval_result["accuracy"]
label_id.extend(eval_result["label_ids"])
label.extend(eval_result["pred_label"])
# for item in eval_result["label_ids"]:
# label_weight.append(label_tensor[item])
i += 1
except tf.errors.OutOfRangeError:
print("End of dataset")
break
# f1 = f1_score(label_id, label, average="macro", sample_weight=label_weight)
# accuracy = accuracy_score(label_id, label, sample_weight=label_weight)
f1 = f1_score(label_id, label, average="macro")
accuracy = accuracy_score(label_id, label)
print("test accuracy accuracy {} {} f1 {}".format(
total_accuracy / i, accuracy, f1))
return total_accuracy / i, f1
# print("===========begin to train============")
# train_fn(train_op, train_loss)
print("===========begin to eval============")
accuracy, f1 = eval_fn(result)
print("==accuracy {} f1 {}==".format(accuracy, f1))
def model_fn(features, labels, mode):
task_type = 'all_neg'
num_task = kargs.get('num_task', 1)
model_io_fn = model_io.ModelIO(model_io_config)
if mode == tf.estimator.ModeKeys.TRAIN:
dropout_prob = model_config.dropout_prob
is_training = True
else:
dropout_prob = 0.0
is_training = False
if model_io_config.fix_lm == True:
scope = model_config.scope + "_finetuning"
else:
scope = model_config.scope
if kargs.get("get_pooled_output", "pooled_output") == "pooled_output":
pooled_feature = model.get_pooled_output()
elif kargs.get("get_pooled_output", "task_output") == "task_output":
pooled_feature_dict = model.get_task_output()
pooled_feature = pooled_feature_dict['pooled_feature']
if kargs.get('apply_head_proj', False):
with tf.variable_scope(scope + "/head_proj", reuse=tf.AUTO_REUSE):
feature_a = simclr_utils.projection_head(
pooled_feature_dict['feature_a'],
is_training,
head_proj_dim=128,
num_nlh_layers=1,
head_proj_mode='nonlinear',
name='head_contrastive')
pooled_feature_dict['feature_a'] = feature_a
with tf.variable_scope(scope + "/head_proj", reuse=tf.AUTO_REUSE):
feature_b = simclr_utils.projection_head(
pooled_feature_dict['feature_b'],
is_training,
head_proj_dim=128,
num_nlh_layers=1,
head_proj_mode='nonlinear',
name='head_contrastive')
pooled_feature_dict['feature_b'] = feature_b
tf.logging.info(
"****** apply contrastive feature projection *******")
shape_list = bert_utils.get_shape_list(pooled_feature,
expected_rank=[2])
batch_size = shape_list[0]
random_batch = tf.random_shuffle(tf.range(batch_size))
input_a = features['input_ids_a']
input_b = tf.gather(features['input_ids_b'], random_batch)
not_equal = tf.cast(tf.not_equal(input_a, input_b), tf.int32)
not_equal = tf.reduce_sum(not_equal, axis=-1)
loss_mask = tf.cast(tf.not_equal(not_equal, tf.zeros_like(not_equal)),
tf.float32)
feat_a = pooled_feature_dict['feature_a']
feat_b = tf.gather(pooled_feature_dict['feature_b'], random_batch)
pooled_feature_dict['feature_b'] = feat_b
label_ids = tf.zeros_like(loss_mask)
loss = tf.constant(0.0)
params_size = model_io_fn.count_params(model_config.scope)
print("==total encoder params==", params_size)
if kargs.get("feature_distillation", True):
universal_feature_a = features.get("input_ids_a_features", None)
universal_feature_b = features.get("input_ids_b_features", None)
if universal_feature_a is None or universal_feature_b is None:
tf.logging.info(
"****** not apply feature distillation *******")
feature_loss = tf.constant(0.0)
else:
feature_a = pooled_feature_dict['feature_a']
feature_a_shape = bert_utils.get_shape_list(
feature_a, expected_rank=[2, 3])
pretrain_feature_a_shape = bert_utils.get_shape_list(
universal_feature_a, expected_rank=[2, 3])
if feature_a_shape[-1] != pretrain_feature_a_shape[-1]:
with tf.variable_scope(scope + "/feature_proj",
reuse=tf.AUTO_REUSE):
proj_feature_a = tf.layers.dense(
feature_a, pretrain_feature_a_shape[-1])
# with tf.variable_scope(scope+"/feature_rec", reuse=tf.AUTO_REUSE):
# proj_feature_a_rec = tf.layers.dense(proj_feature_a, feature_a_shape[-1])
# loss += tf.reduce_mean(tf.reduce_sum(tf.square(proj_feature_a_rec-feature_a), axis=-1))/float(num_task)
tf.logging.info(
"****** apply auto-encoder for feature compression *******"
)
else:
proj_feature_a = feature_a
feature_a_norm = tf.stop_gradient(
tf.sqrt(
tf.reduce_sum(tf.pow(proj_feature_a, 2),
axis=-1,
keepdims=True)) + 1e-20)
proj_feature_a /= feature_a_norm
feature_b = pooled_feature_dict['feature_b']
if feature_a_shape[-1] != pretrain_feature_a_shape[-1]:
with tf.variable_scope(scope + "/feature_proj",
reuse=tf.AUTO_REUSE):
proj_feature_b = tf.layers.dense(
feature_b, pretrain_feature_a_shape[-1])
# with tf.variable_scope(scope+"/feature_rec", reuse=tf.AUTO_REUSE):
# proj_feature_b_rec = tf.layers.dense(proj_feature_b, feature_a_shape[-1])
# loss += tf.reduce_mean(tf.reduce_sum(tf.square(proj_feature_b_rec-feature_b), axis=-1))/float(num_task)
tf.logging.info(
"****** apply auto-encoder for feature compression *******"
)
else:
proj_feature_b = feature_b
feature_b_norm = tf.stop_gradient(
tf.sqrt(
tf.reduce_sum(tf.pow(proj_feature_b, 2),
axis=-1,
keepdims=True)) + 1e-20)
proj_feature_b /= feature_b_norm
feature_a_distillation = tf.reduce_mean(
tf.square(universal_feature_a - proj_feature_a), axis=-1)
feature_b_distillation = tf.reduce_mean(
tf.square(universal_feature_b - proj_feature_b), axis=-1)
feature_loss = tf.reduce_mean(
(feature_a_distillation + feature_b_distillation) /
2.0) / float(num_task)
loss += feature_loss
tf.logging.info(
"****** apply prertained feature distillation *******")
if kargs.get("embedding_distillation", True):
word_embed = model.emb_mat
random_embed_shape = bert_utils.get_shape_list(
word_embed, expected_rank=[2, 3])
print("==random_embed_shape==", random_embed_shape)
pretrained_embed = kargs.get('pretrained_embed', None)
if pretrained_embed is None:
tf.logging.info(
"****** not apply prertained feature distillation *******")
embed_loss = tf.constant(0.0)
else:
pretrain_embed_shape = bert_utils.get_shape_list(
pretrained_embed, expected_rank=[2, 3])
print("==pretrain_embed_shape==", pretrain_embed_shape)
if random_embed_shape[-1] != pretrain_embed_shape[-1]:
with tf.variable_scope(scope + "/embedding_proj",
reuse=tf.AUTO_REUSE):
proj_embed = tf.layers.dense(word_embed,
pretrain_embed_shape[-1])
else:
proj_embed = word_embed
embed_loss = tf.reduce_mean(
tf.reduce_mean(tf.square(proj_embed - pretrained_embed),
axis=-1)) / float(num_task)
loss += embed_loss
tf.logging.info(
"****** apply prertained feature distillation *******")
if kargs.get('loss', 'contrastive_loss') == 'contrastive_loss':
feature_a = tf.nn.l2_normalize(pooled_feature_dict['feature_a'],
axis=-1)
feature_b = tf.nn.l2_normalize(pooled_feature_dict['feature_b'],
axis=-1)
per_example_loss, logits = loss_utils.contrastive_loss(
label_ids, feature_a, feature_b, kargs.get('margin', 1.0))
tf.logging.info("****** contrastive_loss *******")
elif kargs.get(
'loss',
'contrastive_loss') == 'exponent_neg_manhattan_distance_mse':
feature_a = tf.nn.l2_normalize(pooled_feature_dict['feature_a'],
axis=-1)
feature_b = tf.nn.l2_normalize(pooled_feature_dict['feature_b'],
axis=-1)
per_example_loss, logits = loss_utils.exponent_neg_manhattan_distance(
label_ids, feature_a, feature_b, 'mse')
tf.logging.info(
"****** exponent_neg_manhattan_distance_mse *******")
else:
feature_a = tf.nn.l2_normalize(pooled_feature_dict['feature_a'],
axis=-1)
feature_b = tf.nn.l2_normalize(pooled_feature_dict['feature_b'],
axis=-1)
per_example_loss, logits = loss_utils.contrastive_loss(
label_ids, feature_a, feature_b, kargs.get('margin', 1.0))
tf.logging.info("****** contrastive_loss *******")
masked_per_example_loss = per_example_loss * loss_mask
task_loss = tf.reduce_sum(masked_per_example_loss) / (
1e-10 + tf.reduce_sum(loss_mask))
loss += task_loss / float(num_task)
if mode == tf.estimator.ModeKeys.TRAIN:
multi_task_config = kargs.get("multi_task_config", {})
if multi_task_config.get(task_type,
{}).get("lm_augumentation", False):
print("==apply lm_augumentation==")
masked_lm_positions = features["masked_lm_positions"]
masked_lm_ids = features["masked_lm_ids"]
masked_lm_weights = features["masked_lm_weights"]
(masked_lm_loss, masked_lm_example_loss,
masked_lm_log_probs) = pretrain.get_masked_lm_output(
model_config,
model.get_sequence_output(),
model.get_embedding_table(),
masked_lm_positions,
masked_lm_ids,
masked_lm_weights,
reuse=model_reuse)
masked_lm_loss_mask = tf.expand_dims(loss_mask, -1) * tf.ones(
(1,
multi_task_config[task_type]["max_predictions_per_seq"]))
masked_lm_loss_mask = tf.reshape(masked_lm_loss_mask, (-1, ))
masked_lm_label_weights = tf.reshape(masked_lm_weights, [-1])
masked_lm_loss_mask *= tf.cast(masked_lm_label_weights,
tf.float32)
masked_lm_example_loss *= masked_lm_loss_mask # multiply task_mask
masked_lm_loss = tf.reduce_sum(masked_lm_example_loss) / (
1e-10 + tf.reduce_sum(masked_lm_loss_mask))
loss += multi_task_config[task_type][
"masked_lm_loss_ratio"] * masked_lm_loss
masked_lm_label_ids = tf.reshape(masked_lm_ids, [-1])
print(masked_lm_log_probs.get_shape(),
"===masked lm log probs===")
print(masked_lm_label_ids.get_shape(), "===masked lm ids===")
print(masked_lm_label_weights.get_shape(),
"===masked lm mask===")
lm_acc = build_accuracy(masked_lm_log_probs,
masked_lm_label_ids,
masked_lm_loss_mask)
if kargs.get("task_invariant", "no") == "yes":
print("==apply task adversarial training==")
with tf.variable_scope(scope + "/dann_task_invariant",
reuse=model_reuse):
(_, task_example_loss,
task_logits) = distillation_utils.feature_distillation(
model.get_pooled_output(), 1.0, features["task_id"],
kargs.get("num_task", 7), dropout_prob, True)
masked_task_example_loss = loss_mask * task_example_loss
masked_task_loss = tf.reduce_sum(masked_task_example_loss) / (
1e-10 + tf.reduce_sum(loss_mask))
loss += kargs.get("task_adversarial", 1e-2) * masked_task_loss
tvars = model_io_fn.get_params(model_config.scope,
not_storage_params=not_storage_params)
if mode == tf.estimator.ModeKeys.TRAIN:
multi_task_config = kargs.get("multi_task_config", {})
if multi_task_config.get(task_type,
{}).get("lm_augumentation", False):
print("==apply lm_augumentation==")
masked_lm_pretrain_tvars = model_io_fn.get_params(
"cls/predictions", not_storage_params=not_storage_params)
tvars.extend(masked_lm_pretrain_tvars)
try:
params_size = model_io_fn.count_params(model_config.scope)
print("==total params==", params_size)
except:
print("==not count params==")
# print(tvars)
if load_pretrained == "yes":
model_io_fn.load_pretrained(tvars,
init_checkpoint,
exclude_scope=exclude_scope)
if mode == tf.estimator.ModeKeys.TRAIN:
acc = build_accuracy(logits,
label_ids,
loss_mask,
loss_type=kargs.get('loss',
'contrastive_loss'))
return_dict = {
"loss": loss,
"logits": logits,
"task_num": tf.reduce_sum(loss_mask),
"tvars": tvars
}
return_dict["{}_acc".format(task_type)] = acc
if kargs.get("task_invariant", "no") == "yes":
return_dict["{}_task_loss".format(
task_type)] = masked_task_loss
task_acc = build_accuracy(task_logits, features["task_id"],
loss_mask)
return_dict["{}_task_acc".format(task_type)] = task_acc
if multi_task_config.get(task_type,
{}).get("lm_augumentation", False):
return_dict["{}_masked_lm_loss".format(
task_type)] = masked_lm_loss
return_dict["{}_masked_lm_acc".format(task_type)] = lm_acc
if kargs.get("embedding_distillation", True):
return_dict["embed_loss"] = embed_loss * float(num_task)
else:
return_dict["embed_loss"] = task_loss
if kargs.get("feature_distillation", True):
return_dict["feature_loss"] = feature_loss * float(num_task)
else:
return_dict["feature_loss"] = task_loss
return_dict["task_loss"] = task_loss
return return_dict
elif mode == tf.estimator.ModeKeys.EVAL:
eval_dict = {
"loss": loss,
"logits": logits,
"feature": model.get_pooled_output()
}
if kargs.get("adversarial", "no") == "adversarial":
eval_dict["task_logits"] = task_logits
return eval_dict
def export_model_v1(config):
opt_config = Bunch({"init_lr":2e-5, "num_train_steps":1e30, "cycle":False})
model_io_config = Bunch({"fix_lm":False})
bert_config = json.load(open(config["config_file"], "r"))
model_config = Bunch(bert_config)
model_config.use_one_hot_embeddings = True
model_config.scope = "bert"
model_config.dropout_prob = 0.1
model_config.label_type = "single_label"
with open(config["label2id"], "r") as frobj:
label_dict = json.load(frobj)
num_classes = len(label_dict["id2label"])
max_seq_length = config["max_length"]
def serving_input_receiver_fn():
# receive tensors
receiver_tensors = {
"input_ids":
tf.placeholder(tf.int32, [None, max_seq_length], name='input_ids'),
"input_mask":
tf.placeholder(tf.int32, [None, max_seq_length], name='input_mask'),
"segment_ids":
tf.placeholder(tf.int32, [None, max_seq_length], name='segment_ids'),
"label_ids":
tf.placeholder(tf.int32, [None], name='label_ids'),
}
# Convert give inputs to adjust to the model.
features = {}
for key in receiver_tensors:
features[key] = receiver_tensors[key]
return tf.estimator.export.ServingInputReceiver(receiver_tensors=receiver_tensors,
features=features)
# def serving_input_receiver_fn():
# receive serialized example
# serialized_tf_example = tf.placeholder(dtype=tf.string,
# shape=None,
# name='input_example_tensor')
# receiver_tensors = {'examples': serialized_tf_example}
# features = tf.parse_example(serialized_tf_example, feature_spec)
# return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)
model_io_fn = model_io.ModelIO(model_io_config)
model_fn = bert_classifier_estimator.classifier_model_fn_builder(
model_config,
num_classes,
config["init_checkpoint"],
reuse=None,
load_pretrained=True,
model_io_fn=model_io_fn,
model_io_config=model_io_config,
opt_config=opt_config)
estimator = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=config["model_dir"])
export_dir = estimator.export_savedmodel(config["export_path"],
serving_input_receiver_fn,
checkpoint_path=config["init_checkpoint"])
print("===Succeeded in exporting saved model==={}".format(export_dir))
def model_fn(features, labels, mode):
train_ops = []
train_hooks = []
logits_dict = {}
losses_dict = {}
features_dict = {}
tvars = []
task_num_dict = {}
total_loss = tf.constant(0.0)
task_num = 0
encoder = {}
hook_dict = {}
print(task_type_dict.keys(), "==task type dict==")
num_task = len(task_type_dict)
for index, task_type in enumerate(task_type_dict.keys()):
if model_config_dict[task_type].model_type in model_type_lst:
reuse = True
else:
reuse = None
model_type_lst.append(model_config_dict[task_type].model_type)
if task_type_dict[task_type] == "cls_task":
if model_config_dict[task_type].model_type not in encoder:
model_api = model_zoo(model_config_dict[task_type])
model = model_api(model_config_dict[task_type],
features,
labels,
mode,
target_dict[task_type],
reuse=reuse)
encoder[model_config_dict[task_type].model_type] = model
print(encoder, "==encode==")
task_model_fn = cls_model_fn(
encoder[model_config_dict[task_type].model_type],
model_config_dict[task_type],
num_labels_dict[task_type],
init_checkpoint_dict[task_type],
reuse,
load_pretrained_dict[task_type],
model_io_config,
opt_config,
exclude_scope=exclude_scope_dict[task_type],
not_storage_params=not_storage_params_dict[task_type],
target=target_dict[task_type],
label_lst=None,
output_type=output_type,
task_layer_reuse=task_layer_reuse,
task_type=task_type,
num_task=num_task,
task_adversarial=1e-2,
**kargs)
print("==SUCCEEDED IN LODING==", task_type)
result_dict = task_model_fn(features, labels, mode)
logits_dict[task_type] = result_dict["logits"]
losses_dict[task_type] = result_dict["loss"] # task loss
for key in [
"masked_lm_loss", "task_loss", "acc", "task_acc",
"masked_lm_acc"
]:
name = "{}_{}".format(task_type, key)
if name in result_dict:
hook_dict[name] = result_dict[name]
hook_dict["{}_loss".format(task_type)] = result_dict["loss"]
total_loss += result_dict["loss"]
if mode == tf.estimator.ModeKeys.TRAIN:
tvars.extend(result_dict["tvars"])
task_num += result_dict["task_num"]
task_num_dict[task_type] = result_dict["task_num"]
elif mode == tf.estimator.ModeKeys.EVAL:
features[task_type] = result_dict["feature"]
else:
continue
hook_dict["total_loss"] = total_loss
if mode == tf.estimator.ModeKeys.TRAIN:
model_io_fn = model_io.ModelIO(model_io_config)
optimizer_fn = optimizer.Optimizer(opt_config)
model_io_fn.print_params(list(set(tvars)),
string=", trainable params")
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
print("==update_ops==", update_ops)
with tf.control_dependencies(update_ops):
train_op = optimizer_fn.get_train_op(
total_loss, list(set(tvars)), opt_config.init_lr,
opt_config.num_train_steps, **kargs)
model_io_fn.set_saver(optimizer_fn.opt)
if kargs.get("task_index", 1) == 0 and kargs.get(
"run_config", None):
model_io_fn.get_hooks(kargs.get("checkpoint_dir", None),
kargs.get("num_storage_steps", 1000))
training_hooks = model_io_fn.checkpoint_hook
elif kargs.get("task_index", 1) == 1:
training_hooks = []
else:
training_hooks = []
if len(optimizer_fn.distributed_hooks) >= 1:
training_hooks.extend(optimizer_fn.distributed_hooks)
print(training_hooks, "==training_hooks==", "==task_index==",
kargs.get("task_index", 1))
if output_type == "sess":
return {
"train": {
"total_loss": total_loss,
"loss": losses_dict,
"logits": logits_dict,
"train_op": train_op,
"task_num_dict": task_num_dict
},
"hooks": train_hooks
}
elif output_type == "estimator":
hook_dict['learning_rate'] = optimizer_fn.learning_rate
logging_hook = tf.train.LoggingTensorHook(hook_dict,
every_n_iter=100)
training_hooks.append(logging_hook)
print("==hook_dict==")
print(hook_dict)
for key in hook_dict:
tf.summary.scalar(key, hook_dict[key])
for index, task_type in enumerate(task_type_dict.keys()):
tmp = "{}_loss".format(task_type)
if tmp == key:
tf.summary.scalar(
"loss_gap_{}".format(task_type),
hook_dict["total_loss"] - hook_dict[key])
for key in task_num_dict:
tf.summary.scalar(key + "_task_num", task_num_dict[key])
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=total_loss,
train_op=train_op)
# training_hooks=training_hooks)
return estimator_spec
elif mode == tf.estimator.ModeKeys.EVAL: # eval execute for each class solo
def metric_fn(logits, label_ids):
"""Computes the loss and accuracy of the model."""
sentence_log_probs = tf.reshape(logits, [-1, logits.shape[-1]])
sentence_predictions = tf.argmax(logits,
axis=-1,
output_type=tf.int32)
sentence_labels = tf.reshape(label_ids, [-1])
sentence_accuracy = tf.metrics.accuracy(
labels=label_ids, predictions=sentence_predictions)
sentence_f = tf_metrics.f1(label_ids,
sentence_predictions,
num_labels,
label_lst,
average="macro")
eval_metric_ops = {"f1": sentence_f, "acc": sentence_accuracy}
return eval_metric_ops
if output_type == "sess":
return {
"eval": {
"logits": logits_dict,
"total_loss": total_loss,
"feature": features,
"loss": losses_dict
}
}
elif output_type == "estimator":
eval_metric_ops = {}
for key in logits:
eval_dict = metric_fn(logits[key],
features_task_dict[key]["label_ids"])
for sub_key in eval_dict.keys():
eval_key = "{}_{}".format(key, sub_key)
eval_metric_ops[eval_key] = eval_dict[sub_key]
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=total_loss / task_num,
eval_metric_ops=eval_metric_ops)
return estimator_spec
else:
raise NotImplementedError()
