def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
logging.info("*** Features ***")
for name in sorted(features.keys()):
logging.info(" name = %s, shape = %s" %
(name, features[name].shape))
def reform_a_input(raw_input):
return tf.reshape(raw_input,
[dict_run_config.inner_batch_size, -1])
def reform_b_input(raw_input):
return tf.reshape(raw_input, [dict_run_config.def_per_batch, -1])
input_ids = reform_a_input(features["input_ids"]) # [batch_size, def]
input_mask = reform_a_input(features["input_mask"])
segment_ids = reform_a_input(features["segment_ids"])
d_input_ids = reform_b_input(features["d_input_ids"])
d_input_mask = reform_b_input(features["d_input_mask"])
d_location_ids = reform_a_input(features["d_location_ids"])
ab_mapping = features["ab_mapping"]
if hasattr(ssdr_config,
"blind_dictionary") and ssdr_config.blind_dictionary:
logging.info("Hide dictionary")
d_input_ids = tf.zeros_like(d_input_ids)
d_input_mask = tf.zeros_like(d_input_mask)
if dict_run_config.prediction_op == "loss":
seed = 0
else:
seed = None
if dict_run_config.prediction_op == "loss_fixed_mask" or train_config.fixed_mask:
masked_input_ids = input_ids
masked_lm_positions = reform_a_input(
features["masked_lm_positions"])
masked_lm_ids = reform_a_input(features["masked_lm_ids"])
masked_lm_weights = reform_a_input(features["masked_lm_weights"])
else:
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
= random_masking(input_ids, input_mask, train_config.max_predictions_per_seq, MASK_ID, seed)
if dict_run_config.use_d_segment_ids:
d_segment_ids = reform_b_input(features["d_segment_ids"])
else:
d_segment_ids = None
if dict_run_config.use_ab_mapping_mask:
ab_mapping_mask = reform_a_input(features["ab_mapping_mask"])
else:
ab_mapping_mask = None
if ssdr_config.compare_attrib_value_safe("consistency", True):
print("masked_input_ids", masked_input_ids.shape)
print('d_input_ids', d_input_ids.shape)
print("ab_mapping_mask", ab_mapping_mask.shape)
masked_input_ids = tf.tile(masked_input_ids, [2, 1])
input_mask = tf.tile(input_mask, [2, 1])
segment_ids = tf.tile(segment_ids, [2, 1])
dummy = tf.zeros_like(d_input_ids, tf.int32)
#d_input_ids = tf.concat([d_input_ids, dummy], axis=0)
#d_input_mask = tf.concat([d_input_mask, dummy], axis=0)
#if d_segment_ids is not None:
# d_segment_ids = tf.concat([d_segment_ids, dummy], axis=0)
d_location_ids = tf.concat(
[d_location_ids,
tf.zeros_like(d_location_ids, tf.int32)],
axis=0)
#ab_mapping = tf.concat([ab_mapping, tf.zeros_like(ab_mapping, tf.int32)], axis=0)
ab_mapping_mask = tf.concat(
[ab_mapping_mask,
tf.zeros_like(ab_mapping_mask, tf.int32)],
axis=0)
masked_lm_positions = tf.tile(masked_lm_positions, [2, 1])
masked_lm_ids = tf.tile(masked_lm_ids, [2, 1])
masked_lm_weights = tf.tile(masked_lm_weights, [2, 1])
print("masked_input_ids", masked_input_ids.shape)
print('d_input_ids', d_input_ids.shape)
print("ab_mapping_mask", ab_mapping_mask.shape)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model = model_class(
config=bert_config,
ssdr_config=ssdr_config,
is_training=is_training,
input_ids=masked_input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
d_input_ids=d_input_ids,
d_input_mask=d_input_mask,
d_segment_ids=d_segment_ids,
d_location_ids=d_location_ids,
ab_mapping=ab_mapping,
ab_mapping_mask=ab_mapping_mask,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
(masked_lm_loss, masked_lm_example_loss,
masked_lm_log_probs) = get_masked_lm_output(
bert_config, model.get_sequence_output(),
model.get_embedding_table(), masked_lm_positions, masked_lm_ids,
masked_lm_weights)
total_loss = masked_lm_loss
tvars = tf.compat.v1.trainable_variables()
init_vars = {}
scaffold_fn = None
if train_config.init_checkpoint:
if dict_run_config.is_bert_checkpoint:
map1, map2, init_vars = dict_model_fn.get_bert_assignment_map_for_dict(
tvars, train_config.init_checkpoint)
def load_fn():
tf.compat.v1.train.init_from_checkpoint(
train_config.init_checkpoint, map1)
tf.compat.v1.train.init_from_checkpoint(
train_config.init_checkpoint, map2)
else:
map1, init_vars = get_assignment_map_as_is(
tvars, train_config.init_checkpoint)
def load_fn():
tf.compat.v1.train.init_from_checkpoint(
train_config.init_checkpoint, map1)
if train_config.use_tpu:
def tpu_scaffold():
load_fn()
return tf.compat.v1.train.Scaffold()
scaffold_fn = tpu_scaffold
else:
load_fn()
logging.info("**** Trainable Variables ****")
for var in tvars:
init_string = ""
if var.name in init_vars:
init_string = ", *INIT_FROM_CKPT*"
logging.info(" name = %s, shape = %s%s", var.name, var.shape,
init_string)
logging.info("Total parameters : %d" % get_param_num())
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
if train_config.gradient_accumulation == 1:
train_op = optimization.create_optimizer_from_config(
total_loss, train_config)
else:
logging.info("Using gradient accumulation : %d" %
train_config.gradient_accumulation)
train_op = get_accumulated_optimizer_from_config(
total_loss, train_config, tvars,
train_config.gradient_accumulation)
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (metric_fn_lm, [
masked_lm_example_loss, masked_lm_log_probs, masked_lm_ids,
masked_lm_weights
])
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
if dict_run_config.prediction_op == "gradient":
logging.info("Fetching gradient")
gradient = dict_model_fn.get_gradients(
model, masked_lm_log_probs,
train_config.max_predictions_per_seq,
bert_config.vocab_size)
predictions = {
"masked_input_ids": masked_input_ids,
"d_input_ids": d_input_ids,
"masked_lm_positions": masked_lm_positions,
"gradients": gradient,
}
elif dict_run_config.prediction_op == "scores":
logging.info("Fetching input/d_input and scores")
predictions = {
"masked_input_ids": masked_input_ids,
"d_input_ids": d_input_ids,
"masked_lm_positions": masked_lm_positions,
"masked_lm_ids": masked_lm_ids,
"ab_mapping": ab_mapping,
"d_location_ids": d_location_ids,
"scores": model.scores,
}
elif dict_run_config.prediction_op == "loss" or dict_run_config.prediction_op == "loss_fixed_mask":
logging.info("Fetching loss")
predictions = {
"masked_lm_example_loss": masked_lm_example_loss,
}
else:
raise Exception("prediction target not specified")
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
log_features(features)
def reform_a_input(raw_input):
return tf.reshape(raw_input,
[dict_run_config.inner_batch_size, -1])
def reform_b_input(raw_input):
return tf.reshape(raw_input, [dict_run_config.def_per_batch, -1])
input_ids = reform_a_input(features["input_ids"])
input_mask = reform_a_input(features["input_mask"])
segment_ids = reform_a_input(features["segment_ids"])
tf_logging.info("input_ids, input_mask")
# input_ids = features["input_ids"]
# input_mask = features["input_mask"]
# segment_ids = features["segment_ids"]
if mode == tf.estimator.ModeKeys.PREDICT:
tf.random.set_seed(0)
seed = 0
else:
seed = None
# tf_logging.info("Doing dynamic masking (random)")
# masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
# = random_masking(input_ids, input_mask, train_config.max_predictions_per_seq, MASK_ID, seed)
# if dict_run_config.prediction_op == "loss_fixed_mask" or train_config.fixed_mask:
masked_input_ids = input_ids
masked_lm_positions = reform_a_input(features["masked_lm_positions"])
masked_lm_ids = reform_a_input(features["masked_lm_ids"])
masked_lm_weights = reform_a_input(features["masked_lm_weights"])
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
if model_name == "APR":
model = APR(
masked_input_ids,
input_mask,
segment_ids,
is_training,
train_config.use_one_hot_embeddings,
bert_config,
ssdr_config,
dict_run_config.def_per_batch,
dict_run_config.inner_batch_size,
dict_run_config.max_def_length,
)
elif model_name == "BERT":
model = BertModel(
config=bert_config,
is_training=is_training,
input_ids=masked_input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
else:
assert False
masked_lm_loss, masked_lm_example_loss, masked_lm_log_probs \
= get_masked_lm_output(bert_config, model.get_sequence_output(), model.get_embedding_table(),
masked_lm_positions, masked_lm_ids, masked_lm_weights)
loss = masked_lm_loss
tvars = tf.compat.v1.trainable_variables()
assignment_fn = dict_model_fn.get_bert_assignment_map_for_dict
initialized_variable_names, init_fn = align_checkpoint_twice(
tvars, train_config.init_checkpoint, assignment_fn)
scaffold_fn = get_tpu_scaffold_or_init(init_fn, train_config.use_tpu)
log_var_assignments(tvars, initialized_variable_names)
TPUEstimatorSpec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec
if mode == tf.estimator.ModeKeys.TRAIN:
if ssdr_config.compare_attrib_value_safe("use_two_lr", True):
tf_logging.info("Using two lr for each parts")
train_op = create_optimizer_with_separate_lr(
loss, train_config)
else:
tf_logging.info("Using single lr ")
train_op = optimization.create_optimizer_from_config(
loss, train_config)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
training_hooks=[OomReportingHook()],
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (metric_fn_lm, [
masked_lm_example_loss,
masked_lm_log_probs,
masked_lm_ids,
masked_lm_weights,
])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"masked_input_ids": masked_input_ids,
"masked_lm_ids": masked_lm_ids,
"masked_lm_example_loss": masked_lm_example_loss,
"masked_lm_positions": masked_lm_positions,
}
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
tf_logging.info("*** Features ***")
for name in sorted(features.keys()):
tf_logging.info(" name = %s, shape = %s" %
(name, features[name].shape))
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
instance_id = features["instance_id"]
next_sentence_labels = get_dummy_next_sentence_labels(input_ids)
tf_logging.info("Doing dynamic masking (random)")
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
= random_masking(input_ids, input_mask, train_config.max_predictions_per_seq, MASK_ID)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model = model_class(
config=model_config,
is_training=is_training,
input_ids=masked_input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
(masked_lm_loss, masked_lm_example_loss,
masked_lm_log_probs) = get_masked_lm_output(
model_config, model.get_sequence_output(),
model.get_embedding_table(), masked_lm_positions, masked_lm_ids,
masked_lm_weights)
total_loss = masked_lm_loss
tvars = tf.compat.v1.trainable_variables()
use_multiple_checkpoint = is_multiple_checkpoint(
train_config.checkpoint_type)
initialized_variable_names, initialized_variable_names2, init_fn\
= align_checkpoint_for_lm(tvars,
train_config.checkpoint_type,
train_config.init_checkpoint,
train_config.second_init_checkpoint,
use_multiple_checkpoint)
scaffold_fn = get_tpu_scaffold_or_init(init_fn, train_config.use_tpu)
log_var_assignments(tvars, initialized_variable_names,
initialized_variable_names2)
output_spec = None
if mode == tf.estimator.ModeKeys.PREDICT:
predictions = {
"input_ids": input_ids,
"masked_lm_example_loss": masked_lm_example_loss,
"instance_id": instance_id
}
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
tf_logging.info("model_fn_apr_lm")
"""The `model_fn` for TPUEstimator."""
log_features(features)
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
if mode == tf.estimator.ModeKeys.PREDICT:
tf.random.set_seed(0)
seed = 0
else:
seed = None
tf_logging.info("Doing dynamic masking (random)")
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
= random_masking(input_ids, input_mask, train_config.max_predictions_per_seq, MASK_ID, seed)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
tf_logging.info("Using masked_input_ids")
model = APR(
masked_input_ids,
input_mask,
segment_ids,
is_training,
train_config.use_one_hot_embeddings,
bert_config,
ssdr_config,
dict_run_config.def_per_batch,
dict_run_config.inner_batch_size,
dict_run_config.max_def_length,
# MainTransformer,
# SecondTransformerEmbeddingLess,
)
masked_lm_loss, masked_lm_example_loss, masked_lm_log_probs \
= get_masked_lm_output(bert_config, model.get_sequence_output(), model.get_embedding_table(),
masked_lm_positions, masked_lm_ids, masked_lm_weights)
loss = masked_lm_loss
tvars = tf.compat.v1.trainable_variables()
assignment_fn = dict_model_fn.get_bert_assignment_map_for_dict
initialized_variable_names, init_fn = align_checkpoint_twice(
tvars, train_config.init_checkpoint, assignment_fn)
scaffold_fn = get_tpu_scaffold_or_init(init_fn, train_config.use_tpu)
log_var_assignments(tvars, initialized_variable_names)
TPUEstimatorSpec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec
if mode == tf.estimator.ModeKeys.TRAIN:
if ssdr_config.compare_attrib_value_safe("use_two_lr", True):
tf_logging.info("Using two lr for each parts")
train_op = create_optimizer_with_separate_lr(
loss, train_config)
else:
tf_logging.info("Using single lr ")
train_op = optimization.create_optimizer_from_config(
loss, train_config)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
training_hooks=[OomReportingHook()],
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (metric_fn_lm, [
masked_lm_example_loss,
masked_lm_log_probs,
masked_lm_ids,
masked_lm_weights,
])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"masked_input_ids": masked_input_ids,
"masked_lm_ids": masked_lm_ids,
"masked_lm_example_loss": masked_lm_example_loss,
"masked_lm_positions": masked_lm_positions,
}
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
logging.info("*** Features ***")
for name in sorted(features.keys()):
logging.info(" name = %s, shape = %s" % (name, features[name].shape))
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
next_sentence_labels = features["next_sentence_labels"]
seed = 0
threshold = 1e-2
logging.info("Doing All Masking")
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
= random_masking(input_ids, input_mask, train_config.max_predictions_per_seq, MASK_ID, seed)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
prefix1 = "MaybeBERT"
prefix2 = "MaybeNLI"
with tf.compat.v1.variable_scope(prefix1):
model = BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
(masked_lm_loss,
masked_lm_example_loss1, masked_lm_log_probs2) = get_masked_lm_output(
bert_config, model.get_sequence_output(), model.get_embedding_table(),
masked_lm_positions, masked_lm_ids, masked_lm_weights)
all_layers1 = model.get_all_encoder_layers()
with tf.compat.v1.variable_scope(prefix2):
model = BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
all_layers2 = model.get_all_encoder_layers()
preserved_infos = []
for a_layer, b_layer in zip(all_layers1, all_layers2):
layer_diff = a_layer - b_layer
is_preserved = tf.less(tf.abs(layer_diff), threshold)
preserved_infos.append(is_preserved)
t = tf.cast(preserved_infos[1], dtype=tf.int32) #[batch_size, seq_len, dims]
layer_1_count = tf.reduce_sum(t, axis=2)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names, init_fn = get_init_fn_for_two_checkpoints(train_config,
tvars,
train_config.init_checkpoint,
prefix1,
train_config.second_init_checkpoint,
prefix2)
scaffold_fn = get_tpu_scaffold_or_init(init_fn, train_config.use_tpu)
log_var_assignments(tvars, initialized_variable_names)
output_spec = None
if mode == tf.estimator.ModeKeys.PREDICT:
predictions = {
"input_ids": input_ids,
"layer_count": layer_1_count
}
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=None,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
logging.info("*** Features ***")
for name in sorted(features.keys()):
logging.info(" name = %s, shape = %s" %
(name, features[name].shape))
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
next_sentence_labels = features["next_sentence_labels"]
n_trial = 25
logging.info("Doing All Masking")
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
= planned_masking(input_ids, input_mask, train_config.max_predictions_per_seq, MASK_ID, n_trial)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
repeat_input_mask = tf.tile(input_mask, [n_trial, 1])
repeat_segment_ids = tf.tile(segment_ids, [n_trial, 1])
prefix1 = "MaybeBERT"
prefix2 = "MaybeBFN"
with tf.compat.v1.variable_scope(prefix1):
model = BertModel(
config=bert_config,
is_training=is_training,
input_ids=masked_input_ids,
input_mask=repeat_input_mask,
token_type_ids=repeat_segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
(masked_lm_loss, masked_lm_example_loss1,
masked_lm_log_probs2) = get_masked_lm_output(
bert_config, model.get_sequence_output(),
model.get_embedding_table(), masked_lm_positions,
masked_lm_ids, masked_lm_weights)
with tf.compat.v1.variable_scope(prefix2):
model = BertModel(
config=bert_config,
is_training=is_training,
input_ids=masked_input_ids,
input_mask=repeat_input_mask,
token_type_ids=repeat_segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
(masked_lm_loss, masked_lm_example_loss2,
masked_lm_log_probs2) = get_masked_lm_output(
bert_config, model.get_sequence_output(),
model.get_embedding_table(), masked_lm_positions,
masked_lm_ids, masked_lm_weights)
n_mask = train_config.max_predictions_per_seq
def reform(t):
t = tf.reshape(t, [n_trial, -1, n_mask])
t = tf.transpose(t, [1, 0, 2])
return t
grouped_positions = reform(masked_lm_positions)
grouped_loss1 = reform(masked_lm_example_loss1)
grouped_loss2 = reform(masked_lm_example_loss2)
tvars = tf.compat.v1.trainable_variables()
scaffold_fn = None
initialized_variable_names, init_fn = get_init_fn_for_two_checkpoints(
train_config, tvars, train_config.init_checkpoint, prefix1,
train_config.second_init_checkpoint, prefix2)
if train_config.use_tpu:
def tpu_scaffold():
init_fn()
return tf.compat.v1.train.Scaffold()
scaffold_fn = tpu_scaffold
else:
init_fn()
log_var_assignments(tvars, initialized_variable_names)
output_spec = None
if mode == tf.estimator.ModeKeys.PREDICT:
predictions = {
"input_ids": input_ids,
"input_mask": input_mask,
"segment_ids": segment_ids,
"grouped_positions": grouped_positions,
"grouped_loss1": grouped_loss1,
"grouped_loss2": grouped_loss2,
}
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=None,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
tf_logging.info("model_fn_nli_lm")
"""The `model_fn` for TPUEstimator."""
log_features(features)
input_ids = features["input_ids"] # [batch_size, seq_length]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
batch_size, seq_max = get_shape_list2(input_ids)
if "nli_input_ids" in features:
nli_input_ids = features[
"nli_input_ids"] # [batch_size, seq_length]
nli_input_mask = features["nli_input_mask"]
nli_segment_ids = features["nli_segment_ids"]
else:
nli_input_ids = input_ids
nli_input_mask = input_mask
nli_segment_ids = segment_ids
features["label_ids"] = tf.ones([batch_size], tf.int32)
if mode == tf.estimator.ModeKeys.PREDICT:
tf.random.set_seed(0)
seed = 0
else:
seed = None
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
tf_logging.info("Doing dynamic masking (random)")
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
= random_masking(input_ids, input_mask,
train_config.max_predictions_per_seq, MASK_ID, seed)
sharing_model = sharing_model_factory(
config, train_config.use_one_hot_embeddings, is_training,
masked_input_ids, input_mask, segment_ids, nli_input_ids,
nli_input_mask, nli_segment_ids)
sequence_output_lm = sharing_model.lm_sequence_output()
nli_feature = sharing_model.get_tt_feature()
masked_lm_loss, masked_lm_example_loss, masked_lm_log_probs \
= get_masked_lm_output(config, sequence_output_lm, sharing_model.get_embedding_table(),
masked_lm_positions, masked_lm_ids, masked_lm_weights)
masked_lm_log_probs = tf.reshape(masked_lm_log_probs, [batch_size, -1])
masked_lm_per_inst_loss = tf.reshape(masked_lm_example_loss,
[batch_size, -1])
task = Classification(3, features, nli_feature, is_training)
nli_loss = task.loss
task_prob = tf.nn.softmax(task.logits, axis=-1)
arg_like = task_prob[:, 1] + task_prob[:, 2]
vars = sharing_model.model.all_layer_outputs
grads_1 = tf.gradients(ys=masked_lm_loss, xs=vars) # List[ batch_szie,
grads_2 = tf.gradients(ys=arg_like, xs=vars)
l = []
for g1, g2 in zip(grads_1, grads_2):
if g1 is not None and g2 is not None:
a = tf.reshape(g1, [batch_size * 2, seq_max, -1])[:batch_size]
a = a / masked_lm_per_inst_loss
b = tf.reshape(g2, [batch_size * 2, seq_max, -1])[batch_size:]
l.append(tf.abs(a * b))
h_overlap = tf.stack(l, axis=1)
h_overlap = tf.reduce_sum(h_overlap, axis=2)
loss = combine_loss_fn(masked_lm_loss, nli_loss)
tvars = tf.compat.v1.trainable_variables()
assignment_fn = get_bert_assignment_map
initialized_variable_names, init_fn = get_init_fn(
tvars, train_config.init_checkpoint, assignment_fn)
scaffold_fn = get_tpu_scaffold_or_init(init_fn, train_config.use_tpu)
log_var_assignments(tvars, initialized_variable_names)
TPUEstimatorSpec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec
if mode == tf.estimator.ModeKeys.TRAIN:
train_op = optimization.create_optimizer_from_config(
loss, train_config)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (metric_fn_lm, [
masked_lm_example_loss,
masked_lm_log_probs,
masked_lm_ids,
masked_lm_weights,
])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"masked_input_ids": masked_input_ids,
"masked_lm_ids": masked_lm_ids,
"masked_lm_example_loss": masked_lm_example_loss,
"masked_lm_positions": masked_lm_positions,
"masked_lm_log_probs": masked_lm_log_probs,
"h_overlap": h_overlap,
}
output_spec = TPUEstimatorSpec(mode=mode,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
tf_logging.info("*** Features ***")
for name in sorted(features.keys()):
tf_logging.info(" name = %s, shape = %s" %
(name, features[name].shape))
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
if "next_sentence_labels" in features:
next_sentence_labels = features["next_sentence_labels"]
else:
next_sentence_labels = get_dummy_next_sentence_labels(input_ids)
tlm_prefix = "target_task"
with tf.compat.v1.variable_scope(tlm_prefix):
priority_score = tf.stop_gradient(priority_model(features))
priority_score = priority_score * target_model_config.amp
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights\
= biased_masking(input_ids,
input_mask,
priority_score,
target_model_config.alpha,
train_config.max_predictions_per_seq,
MASK_ID)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model = model_class(
config=bert_config,
is_training=is_training,
input_ids=masked_input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
(masked_lm_loss, masked_lm_example_loss,
masked_lm_log_probs) = get_masked_lm_output(
bert_config, model.get_sequence_output(),
model.get_embedding_table(), masked_lm_positions, masked_lm_ids,
masked_lm_weights)
(next_sentence_loss, next_sentence_example_loss,
next_sentence_log_probs) = get_next_sentence_output(
bert_config, model.get_pooled_output(), next_sentence_labels)
total_loss = masked_lm_loss + next_sentence_loss
all_vars = tf.compat.v1.all_variables()
tf_logging.info("We assume priority model is from v2")
if train_config.checkpoint_type == "v2":
assignment_map, initialized_variable_names = assignment_map_v2_to_v2(
all_vars, train_config.init_checkpoint)
assignment_map2, initialized_variable_names2 = get_assignment_map_remap_from_v2(
all_vars, tlm_prefix, train_config.second_init_checkpoint)
else:
assignment_map, assignment_map2, initialized_variable_names \
= get_tlm_assignment_map_v2(all_vars,
tlm_prefix,
train_config.init_checkpoint,
train_config.second_init_checkpoint)
initialized_variable_names2 = None
def init_fn():
if train_config.init_checkpoint:
tf.compat.v1.train.init_from_checkpoint(
train_config.init_checkpoint, assignment_map)
if train_config.second_init_checkpoint:
tf.compat.v1.train.init_from_checkpoint(
train_config.second_init_checkpoint, assignment_map2)
scaffold_fn = get_tpu_scaffold_or_init(init_fn, train_config.use_tpu)
tvars = [v for v in all_vars if not v.name.startswith(tlm_prefix)]
log_var_assignments(tvars, initialized_variable_names,
initialized_variable_names2)
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
train_op = optimization.create_optimizer_from_config(
total_loss, train_config, tvars)
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (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
])
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"masked_input_ids": masked_input_ids,
"priority_score": priority_score,
"lm_loss1": features["loss1"],
"lm_loss2": features["loss2"],
}
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
tf_logging.info("model_fn_nli_lm")
"""The `model_fn` for TPUEstimator."""
log_features(features)
input_ids = features["input_ids"] # [batch_size, seq_length]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
batch_size, _ = get_shape_list2(input_ids)
if "nli_input_ids" in features:
nli_input_ids = features[
"nli_input_ids"] # [batch_size, seq_length]
nli_input_mask = features["nli_input_mask"]
nli_segment_ids = features["nli_segment_ids"]
else:
nli_input_ids = input_ids
nli_input_mask = input_mask
nli_segment_ids = segment_ids
features["label_ids"] = tf.ones([batch_size], tf.int32)
if mode == tf.estimator.ModeKeys.PREDICT:
tf.random.set_seed(0)
seed = 0
else:
seed = None
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
tf_logging.info("Doing dynamic masking (random)")
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
= random_masking(input_ids, input_mask,
train_config.max_predictions_per_seq, MASK_ID, seed)
sharing_model = sharing_model_factory(
config, train_config.use_one_hot_embeddings, is_training,
masked_input_ids, input_mask, segment_ids, nli_input_ids,
nli_input_mask, nli_segment_ids)
sequence_output_lm = sharing_model.lm_sequence_output()
nli_feature = sharing_model.get_tt_feature()
masked_lm_loss, masked_lm_example_loss, masked_lm_log_probs \
= get_masked_lm_output(config, sequence_output_lm, sharing_model.get_embedding_table(),
masked_lm_positions, masked_lm_ids, masked_lm_weights)
masked_lm_log_probs = tf.reshape(masked_lm_log_probs, [batch_size, -1])
top_guess = masked_lm_log_probs
task = Classification(3, features, nli_feature, is_training)
nli_loss = task.loss
overlap_score = shared_gradient_fine_grained(
masked_lm_example_loss, task.logits,
train_config.max_predictions_per_seq)
loss = combine_loss_fn(masked_lm_loss, nli_loss)
tvars = tf.compat.v1.trainable_variables()
assignment_fn = get_bert_assignment_map
initialized_variable_names, init_fn = get_init_fn(
tvars, train_config.init_checkpoint, assignment_fn)
scaffold_fn = get_tpu_scaffold_or_init(init_fn, train_config.use_tpu)
log_var_assignments(tvars, initialized_variable_names)
TPUEstimatorSpec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec
if mode == tf.estimator.ModeKeys.TRAIN:
train_op = optimization.create_optimizer_from_config(
loss, train_config)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (metric_fn_lm, [
masked_lm_example_loss,
masked_lm_log_probs,
masked_lm_ids,
masked_lm_weights,
])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"masked_input_ids": masked_input_ids,
"masked_lm_ids": masked_lm_ids,
"masked_lm_example_loss": masked_lm_example_loss,
"masked_lm_positions": masked_lm_positions,
"masked_lm_log_probs": masked_lm_log_probs,
"overlap_score": overlap_score,
"top_guess": top_guess,
}
output_spec = TPUEstimatorSpec(mode=mode,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
logging.info("*** Features ***")
for name in sorted(features.keys()):
logging.info(" name = %s, shape = %s" % (name, features[name].shape))
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
d_input_ids = features["d_input_ids"]
d_input_mask = features["d_input_mask"]
d_location_ids = features["d_location_ids"]
next_sentence_labels = features["next_sentence_labels"]
if dict_run_config.prediction_op == "loss":
seed = 0
else:
seed = None
if dict_run_config.prediction_op == "loss_fixed_mask" or train_config.fixed_mask:
masked_input_ids = input_ids
masked_lm_positions = features["masked_lm_positions"]
masked_lm_ids = features["masked_lm_ids"]
masked_lm_weights = tf.ones_like(masked_lm_positions, dtype=tf.float32)
else:
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
= random_masking(input_ids, input_mask, train_config.max_predictions_per_seq, MASK_ID, seed)
if dict_run_config.use_d_segment_ids:
d_segment_ids = features["d_segment_ids"]
else:
d_segment_ids = None
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model = model_class(
config=bert_config,
d_config=dbert_config,
is_training=is_training,
input_ids=masked_input_ids,
input_mask=input_mask,
d_input_ids=d_input_ids,
d_input_mask=d_input_mask,
d_location_ids=d_location_ids,
use_target_pos_emb=dict_run_config.use_target_pos_emb,
token_type_ids=segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
d_segment_ids=d_segment_ids,
pool_dict_output=dict_run_config.pool_dict_output,
)
(masked_lm_loss,
masked_lm_example_loss, masked_lm_log_probs) = get_masked_lm_output(
bert_config, model.get_sequence_output(), model.get_embedding_table(),
masked_lm_positions, masked_lm_ids, masked_lm_weights)
(next_sentence_loss, next_sentence_example_loss,
next_sentence_log_probs) = get_next_sentence_output(
bert_config, model.get_pooled_output(), next_sentence_labels)
total_loss = masked_lm_loss
if dict_run_config.train_op == "entry_prediction":
score_label = features["useful_entry"] # [batch, 1]
score_label = tf.reshape(score_label, [-1])
entry_logits = bert_common.dense(2, bert_common.create_initializer(bert_config.initializer_range))\
(model.get_dict_pooled_output())
print("entry_logits: ", entry_logits.shape)
losses = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=entry_logits, labels=score_label)
loss = tf.reduce_mean(losses)
total_loss = loss
if dict_run_config.train_op == "lookup":
lookup_idx = features["lookup_idx"]
lookup_loss, lookup_example_loss, lookup_score = \
sequence_index_prediction(bert_config, lookup_idx, model.get_sequence_output())
total_loss += lookup_loss
tvars = tf.compat.v1.trainable_variables()
init_vars = {}
scaffold_fn = None
if train_config.init_checkpoint:
if dict_run_config.is_bert_checkpoint:
map1, map2, init_vars = get_bert_assignment_map_for_dict(tvars, train_config.init_checkpoint)
def load_fn():
tf.compat.v1.train.init_from_checkpoint(train_config.init_checkpoint, map1)
tf.compat.v1.train.init_from_checkpoint(train_config.init_checkpoint, map2)
else:
map1, init_vars = get_assignment_map_as_is(tvars, train_config.init_checkpoint)
def load_fn():
tf.compat.v1.train.init_from_checkpoint(train_config.init_checkpoint, map1)
if train_config.use_tpu:
def tpu_scaffold():
load_fn()
return tf.compat.v1.train.Scaffold()
scaffold_fn = tpu_scaffold
else:
load_fn()
logging.info("**** Trainable Variables ****")
for var in tvars:
init_string = ""
if var.name in init_vars:
init_string = ", *INIT_FROM_CKPT*"
logging.info(" name = %s, shape = %s%s", var.name, var.shape, init_string)
logging.info("Total parameters : %d" % get_param_num())
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
if train_config.gradient_accumulation == 1:
train_op = optimization.create_optimizer_from_config(total_loss, train_config)
else:
logging.info("Using gradient accumulation : %d" % train_config.gradient_accumulation)
train_op = get_accumulated_optimizer_from_config(total_loss, train_config,
tvars, train_config.gradient_accumulation)
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (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
])
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
if dict_run_config.prediction_op == "gradient":
logging.info("Fetching gradient")
gradient = get_gradients(model, masked_lm_log_probs,
train_config.max_predictions_per_seq, bert_config.vocab_size)
predictions = {
"masked_input_ids": masked_input_ids,
#"input_ids": input_ids,
"d_input_ids": d_input_ids,
"masked_lm_positions": masked_lm_positions,
"gradients": gradient,
}
elif dict_run_config.prediction_op == "loss" or dict_run_config.prediction_op == "loss_fixed_mask":
logging.info("Fetching loss")
predictions = {
"masked_lm_example_loss": masked_lm_example_loss,
}
else:
raise Exception("prediction target not specified")
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
log_features(features)
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
next_sentence_labels = features["next_sentence_labels"]
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
= random_masking(input_ids, input_mask, train_config.max_predictions_per_seq, MASK_ID)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
prefix1 = "MaybeBERT"
prefix2 = "MaybeBFN"
with tf.compat.v1.variable_scope(prefix1):
model1 = BertModel(
config=bert_config,
is_training=is_training,
input_ids=masked_input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
(masked_lm_loss, masked_lm_example_loss1,
masked_lm_log_probs1) = get_masked_lm_output(
bert_config, model1.get_sequence_output(),
model1.get_embedding_table(), masked_lm_positions,
masked_lm_ids, masked_lm_weights)
masked_lm_example_loss1 = tf.reshape(masked_lm_example_loss1,
masked_lm_ids.shape)
with tf.compat.v1.variable_scope(prefix2):
model2 = BertModel(
config=bert_config,
is_training=is_training,
input_ids=masked_input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
(masked_lm_loss, masked_lm_example_loss2,
masked_lm_log_probs2) = get_masked_lm_output(
bert_config, model2.get_sequence_output(),
model2.get_embedding_table(), masked_lm_positions,
masked_lm_ids, masked_lm_weights)
print(model2.get_sequence_output().shape)
masked_lm_example_loss2 = tf.reshape(masked_lm_example_loss2,
masked_lm_ids.shape)
model = model_class(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
loss_model = IndependentLossModel(bert_config)
loss_model.train_modeling(model.get_sequence_output(),
masked_lm_positions, masked_lm_weights,
tf.stop_gradient(masked_lm_example_loss1),
tf.stop_gradient(masked_lm_example_loss2))
total_loss = loss_model.total_loss
loss1 = loss_model.loss1
loss2 = loss_model.loss2
per_example_loss1 = loss_model.per_example_loss1
per_example_loss2 = loss_model.per_example_loss2
losses1 = tf.reduce_sum(per_example_loss1, axis=1)
losses2 = tf.reduce_sum(per_example_loss2, axis=1)
prob1 = loss_model.prob1
prob2 = loss_model.prob2
checkpoint2_1, checkpoint2_2 = train_config.second_init_checkpoint.split(
",")
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names_1, init_fn_1 = get_init_fn_for_two_checkpoints(
train_config, tvars, checkpoint2_1, prefix1, checkpoint2_2,
prefix2)
assignment_fn = get_bert_assignment_map
assignment_map2, initialized_variable_names_2 = assignment_fn(
tvars, train_config.init_checkpoint)
initialized_variable_names = {}
initialized_variable_names.update(initialized_variable_names_1)
initialized_variable_names.update(initialized_variable_names_2)
def init_fn():
init_fn_1()
tf.compat.v1.train.init_from_checkpoint(
train_config.init_checkpoint, assignment_map2)
scaffold_fn = get_tpu_scaffold_or_init(init_fn, train_config.use_tpu)
log_var_assignments(tvars, initialized_variable_names)
if mode == tf.estimator.ModeKeys.TRAIN:
train_op = optimization.create_optimizer_from_config(
total_loss, train_config)
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
def metric_fn(per_example_loss1, per_example_loss2):
loss1 = tf.compat.v1.metrics.mean(values=per_example_loss1)
loss2 = tf.compat.v1.metrics.mean(values=per_example_loss2)
return {
"loss1": loss1,
"loss2": loss2,
}
eval_metrics = (metric_fn, [losses1, losses2])
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"prob1": prob1,
"prob2": prob2,
"per_example_loss1": per_example_loss1,
"per_example_loss2": per_example_loss2,
"input_ids": input_ids,
"masked_lm_positions": masked_lm_positions,
}
output_spec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
tf_logging.info("model_fn_apr_lm")
"""The `model_fn` for TPUEstimator."""
log_features(features)
raw_input_ids = features["input_ids"] # [batch_size, seq_length]
raw_input_mask = features["input_mask"]
raw_segment_ids = features["segment_ids"]
word_tokens = features["word"]
word_input_mask = tf.cast(tf.not_equal(word_tokens, 0), tf.int32)
word_segment_ids = tf.ones_like(word_tokens, tf.int32)
if mode == tf.estimator.ModeKeys.PREDICT:
tf.random.set_seed(0)
seed = 0
else:
seed = None
input_ids = tf.concat([word_tokens, raw_input_ids], axis=1)
input_mask = tf.concat([word_input_mask, raw_input_mask], axis=1)
segment_ids = tf.concat([word_segment_ids, raw_segment_ids], axis=1)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
tf_logging.info("Using masked_input_ids")
masked_input_ids, masked_lm_positions, masked_lm_ids, masked_lm_weights \
= random_masking(input_ids, input_mask, train_config.max_predictions_per_seq, MASK_ID, seed)
model = BertModel(
config=config,
is_training=is_training,
input_ids=masked_input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
(masked_lm_loss, masked_lm_example_loss,
masked_lm_log_probs) = get_masked_lm_output(
config, model.get_sequence_output(), model.get_embedding_table(),
masked_lm_positions, masked_lm_ids, masked_lm_weights)
loss = masked_lm_loss
tvars = tf.compat.v1.trainable_variables()
assignment_fn = tlm.training.assignment_map.get_bert_assignment_map
initialized_variable_names, init_fn = get_init_fn(
tvars, train_config.init_checkpoint, assignment_fn)
scaffold_fn = get_tpu_scaffold_or_init(init_fn, train_config.use_tpu)
log_var_assignments(tvars, initialized_variable_names)
TPUEstimatorSpec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec
if mode == tf.estimator.ModeKeys.TRAIN:
tf_logging.info("Using single lr ")
train_op = optimization.create_optimizer_from_config(
loss, train_config)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (metric_fn_lm, [
masked_lm_example_loss,
masked_lm_log_probs,
masked_lm_ids,
masked_lm_weights,
])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"masked_input_ids": masked_input_ids,
"masked_lm_ids": masked_lm_ids,
"masked_lm_example_loss": masked_lm_example_loss,
"masked_lm_positions": masked_lm_positions
}
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
tf_logging.info("model_fn_sero_lm")
"""The `model_fn` for TPUEstimator."""
log_features(features)
input_ids = features["input_ids"] # [batch_size, seq_length]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
is_sero_modeling = "sero" in modeling
if is_sero_modeling:
use_context = features["use_context"]
elif modeling == "bert":
batch_size, _ = get_shape_list(input_mask)
use_context = tf.ones([batch_size, 1], tf.int32)
else:
assert False
if mode == tf.estimator.ModeKeys.PREDICT:
tf.random.set_seed(0)
seed = 0
else:
seed = None
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
tf_logging.info("Using masked_input_ids")
if is_sero_modeling:
stacked_input_ids, stacked_input_mask, stacked_segment_ids, \
= split_and_append_sep(input_ids, input_mask, segment_ids,
config.total_sequence_length, config.window_size, CLS_ID, EOW_ID)
input_ids_2d = r3to2(stacked_input_ids)
input_mask_2d = r3to2(stacked_input_mask)
elif modeling == "bert":
stacked_input_ids, stacked_input_mask, stacked_segment_ids = input_ids, input_mask, segment_ids
input_ids_2d = stacked_input_ids
input_mask_2d = stacked_input_mask
else:
assert False
tf_logging.info("Doing dynamic masking (random)")
# TODO make stacked_input_ids 2D and recover
masked_input_ids_2d, masked_lm_positions_2d, masked_lm_ids_2d, masked_lm_weights_2d \
= random_masking(input_ids_2d, input_mask_2d,
train_config.max_predictions_per_seq, MASK_ID, seed, [EOW_ID])
if is_sero_modeling:
masked_input_ids = tf.reshape(masked_input_ids_2d,
stacked_input_ids.shape)
elif modeling == "bert":
masked_input_ids = tf.expand_dims(masked_input_ids_2d, 1)
stacked_input_mask = tf.expand_dims(stacked_input_mask, 1)
stacked_segment_ids = tf.expand_dims(stacked_segment_ids, 1)
else:
assert False
if modeling == "sero":
model_class = SeroDelta
elif modeling == "sero_epsilon":
model_class = SeroEpsilon
with tf.compat.v1.variable_scope("sero"):
model = model_class(config, is_training,
train_config.use_one_hot_embeddings)
sequence_output_3d = model.network_stacked(masked_input_ids,
stacked_input_mask,
stacked_segment_ids,
use_context)
masked_lm_loss, masked_lm_example_loss, masked_lm_log_probs \
= get_masked_lm_output(config, sequence_output_3d, model.get_embedding_table(),
masked_lm_positions_2d, masked_lm_ids_2d, masked_lm_weights_2d)
predictions = None
if prediction_op == "gradient_to_long_context":
predictions = {}
for idx, input_tensor in enumerate(model.upper_module_inputs):
g = tf.abs(tf.gradients(ys=masked_lm_loss, xs=input_tensor)[0])
main_g = g[:, :config.window_size, :]
context_g = g[:, config.window_size:, :]
main_g = tf.reduce_mean(tf.reduce_mean(main_g, axis=2), axis=1)
context_g = tf.reduce_mean(tf.reduce_mean(context_g, axis=2),
axis=1)
predictions['main_g_{}'.format(idx)] = main_g
predictions['context_g_{}'.format(idx)] = context_g
loss = masked_lm_loss #+ bert_task.masked_lm_loss
tvars = tf.compat.v1.trainable_variables()
if train_config.init_checkpoint:
assignment_fn = get_assignment_map_from_checkpoint_type(
train_config.checkpoint_type, config.lower_layers)
else:
assignment_fn = None
initialized_variable_names, init_fn = get_init_fn(
tvars, train_config.init_checkpoint, assignment_fn)
log_var_assignments(tvars, initialized_variable_names)
scaffold_fn = get_tpu_scaffold_or_init(init_fn, train_config.use_tpu)
TPUEstimatorSpec = tf.compat.v1.estimator.tpu.TPUEstimatorSpec
if mode == tf.estimator.ModeKeys.TRAIN:
train_op = optimization.create_optimizer_from_config(
loss, train_config)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
training_hooks=[OomReportingHook()],
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
output_spec = TPUEstimatorSpec(mode=model,
loss=loss,
eval_metrics=None,
scaffold_fn=scaffold_fn)
else:
if predictions is None:
predictions = {
"input_ids": input_ids,
"masked_input_ids": masked_input_ids,
"masked_lm_ids": masked_lm_ids_2d,
"masked_lm_example_loss": masked_lm_example_loss,
"masked_lm_positions": masked_lm_positions_2d,
}
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec