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))
label_ids = features["label_ids"]
label_ids = tf.reshape(label_ids, [-1])
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"], dtype=tf.float32)
else:
is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model = model_class(
config=model_config,
is_training=is_training,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
features=features,
)
logits = model.get_logits()
loss = model.get_loss(label_ids)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(train_config, tvars)
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
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
tvars = None
train_op = optimization.create_optimizer_from_config(loss, train_config, tvars)
output_spec = TPUEstimatorSpec(mode=mode, loss=loss, train_op=train_op, scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (classification_metric_fn, [
logits, label_ids, is_real_example
])
output_spec = TPUEstimatorSpec(mode=mode, loss=loss, eval_metrics=eval_metrics, scaffold_fn=scaffold_fn)
else:
predictions = {
"label_ids": label_ids,
"logits": logits,
}
if "data_id" in features:
predictions['data_id'] = features['data_id']
output_spec = tf.compat.v1.estimator.tpu.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."""
tf_logging.info("*** Features ***")
for name in sorted(features.keys()):
tf_logging.info(" name = %s, shape = %s" %
(name, features[name].shape))
query = features["query"]
doc = features["doc"]
doc_mask = features["doc_mask"]
data_ids = features["data_id"]
segment_len = max_seq_length - query_len - 3
step_size = model_config.step_size
input_ids, input_mask, segment_ids, n_segments = \
iterate_over(query, doc, doc_mask, total_doc_len, segment_len, step_size)
if mode == tf.estimator.ModeKeys.PREDICT:
label_ids = tf.ones([input_ids.shape[0]], dtype=tf.int32)
else:
label_ids = features["label_ids"]
label_ids = tf.reshape(label_ids, [-1])
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"],
dtype=tf.float32)
else:
is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
if "feed_features" in special_flags:
model = model_class(
config=model_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,
features=features,
)
else:
model = model_class(
config=model_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,
)
if "new_pooling" in special_flags:
pooled = mimic_pooling(model.get_sequence_output(),
model_config.hidden_size,
model_config.initializer_range)
else:
pooled = model.get_pooled_output()
if train_config.checkpoint_type != "bert_nli" and train_config.use_old_logits:
tf_logging.info("Use old version of logistic regression")
if is_training:
pooled = dropout(pooled, 0.1)
logits = tf.keras.layers.Dense(train_config.num_classes,
name="cls_dense")(pooled)
else:
tf_logging.info("Use fixed version of logistic regression")
output_weights = tf.compat.v1.get_variable(
"output_weights",
[train_config.num_classes, model_config.hidden_size],
initializer=tf.compat.v1.truncated_normal_initializer(
stddev=0.02))
output_bias = tf.compat.v1.get_variable(
"output_bias", [train_config.num_classes],
initializer=tf.compat.v1.zeros_initializer())
if is_training:
pooled = dropout(pooled, 0.1)
logits = tf.matmul(pooled, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
loss_arr = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=logits, labels=label_ids)
if "bias_loss" in special_flags:
tf_logging.info("Using special_flags : bias_loss")
loss_arr = reweight_zero(label_ids, loss_arr)
loss = tf.reduce_mean(input_tensor=loss_arr)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(
train_config, tvars)
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
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
if "simple_optimizer" in special_flags:
tf_logging.info("using simple optimizer")
train_op = create_simple_optimizer(loss,
train_config.learning_rate,
train_config.use_tpu)
else:
if "ask_tvar" in special_flags:
tvars = model.get_trainable_vars()
else:
tvars = None
train_op = optimization.create_optimizer_from_config(
loss, train_config, tvars)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (classification_metric_fn,
[logits, label_ids, is_real_example])
output_spec = TPUEstimatorSpec(mode=model,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"logits": logits,
"doc": doc,
"data_ids": data_ids,
}
useful_inputs = ["data_id", "input_ids2", "data_ids"]
for input_name in useful_inputs:
if input_name in features:
predictions[input_name] = features[input_name]
if override_prediction_fn is not None:
predictions = override_prediction_fn(predictions, model)
output_spec = tf.compat.v1.estimator.tpu.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."""
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 mode == tf.estimator.ModeKeys.PREDICT:
label_ids = tf.ones([input_ids.shape[0]], dtype=tf.int32)
else:
label_ids = features["label_ids"]
label_ids = tf.reshape(label_ids, [-1])
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"],
dtype=tf.float32)
else:
is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)
domain_ids = features["domain_ids"]
domain_ids = tf.reshape(domain_ids, [-1])
is_valid_label = features["is_valid_label"]
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model_1 = BertModel(
config=model_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,
)
pooled = model_1.get_pooled_output()
if is_training:
pooled = dropout(pooled, 0.1)
logits = tf.keras.layers.Dense(train_config.num_classes,
name="cls_dense")(pooled)
pred_losses = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=logits, labels=label_ids)
num_domain = 2
pooled_for_domain = grad_reverse(pooled)
domain_logits = tf.keras.layers.Dense(
num_domain, name="domain_dense")(pooled_for_domain)
domain_losses = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=domain_logits, labels=domain_ids)
pred_loss = tf.reduce_mean(pred_losses *
tf.cast(is_valid_label, tf.float32))
domain_loss = tf.reduce_mean(domain_losses)
tf.compat.v1.summary.scalar('domain_loss', domain_loss)
tf.compat.v1.summary.scalar('pred_loss', pred_loss)
alpha = model_config.alpha
loss = pred_loss + alpha * domain_loss
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(
train_config, tvars)
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
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
tvars = None
train_op = optimization.create_optimizer_from_config(
loss, train_config, tvars)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (classification_metric_fn,
[logits, label_ids, is_real_example])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"logits": logits,
}
if "data_id" in features:
predictions['data_id'] = features['data_id']
output_spec = tf.compat.v1.estimator.tpu.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."""
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 mode == tf.estimator.ModeKeys.PREDICT:
label_ids = tf.ones([input_ids.shape[0]], dtype=tf.float32)
else:
label_ids = features["label_ids"]
label_ids = tf.reshape(label_ids, [-1])
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"], dtype=tf.float32)
else:
is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model = BertModel(
config=model_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,
)
pooled = model.get_pooled_output()
if is_training:
pooled = dropout(pooled, 0.1)
logits = tf.keras.layers.Dense(train_config.num_classes, name="cls_dense")(pooled)
scale = model_config.scale
label_ids = scale * label_ids
weight = tf.abs(label_ids)
loss_arr = tf.keras.losses.MAE(y_true=label_ids, y_pred=logits)
loss_arr = loss_arr * weight
loss = tf.reduce_mean(input_tensor=loss_arr)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(train_config, tvars)
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
def metric_fn(logits, label, is_real_example):
mae = tf.compat.v1.metrics.mean_absolute_error(
labels=label, predictions=logits, weights=is_real_example)
return {
"mae": mae
}
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
tvars = None
train_op = optimization.create_optimizer_from_config(loss, train_config, tvars)
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, [
logits, label_ids, is_real_example
])
output_spec = TPUEstimatorSpec(mode=mode, loss=loss, eval_metrics=eval_metrics, scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"logits": logits,
}
if "data_id" in features:
predictions['data_id'] = features['data_id']
output_spec = tf.compat.v1.estimator.tpu.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."""
tf_logging.info("*** Features ***")
for name in sorted(features.keys()):
tf_logging.info(" name = %s, shape = %s" %
(name, features[name].shape))
q_input_ids_1 = features["q_input_ids_1"]
q_input_mask_1 = features["q_input_mask_1"]
d_input_ids_1 = features["d_input_ids_1"]
d_input_mask_1 = features["d_input_mask_1"]
q_input_ids_2 = features["q_input_ids_2"]
q_input_mask_2 = features["q_input_mask_2"]
d_input_ids_2 = features["d_input_ids_2"]
d_input_mask_2 = features["d_input_mask_2"]
q_input_ids = tf.stack([q_input_ids_1, q_input_ids_2], axis=0)
q_input_mask = tf.stack([q_input_mask_1, q_input_mask_2], axis=0)
q_segment_ids = tf.zeros_like(q_input_ids, tf.int32)
d_input_ids = tf.stack([d_input_ids_1, d_input_ids_2], axis=0)
d_input_mask = tf.stack([d_input_mask_1, d_input_mask_2], axis=0)
d_segment_ids = tf.zeros_like(d_input_ids, tf.int32)
label_ids = features["label_ids"]
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"],
dtype=tf.float32)
else:
is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)
with tf.compat.v1.variable_scope("query"):
model_q = model_class(
config=model_config,
is_training=is_training,
input_ids=q_input_ids,
input_mask=q_input_mask,
token_type_ids=q_segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
with tf.compat.v1.variable_scope("document"):
model_d = model_class(
config=model_config,
is_training=is_training,
input_ids=d_input_ids,
input_mask=d_input_mask,
token_type_ids=d_segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
pooled_q = model_q.get_pooled_output()
pooled_d = model_d.get_pooled_output()
logits = tf.matmul(pooled_q, pooled_d, transpose_b=True)
y = tf.cast(label_ids, tf.float32) * 2 - 1
losses = tf.maximum(1.0 - logits * y, 0)
loss = tf.reduce_mean(losses)
pred = tf.cast(logits > 0, tf.int32)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(
train_config, tvars)
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
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
if "simple_optimizer" in special_flags:
tf_logging.info("using simple optimizer")
train_op = create_simple_optimizer(loss,
train_config.learning_rate,
train_config.use_tpu)
else:
train_op = optimization.create_optimizer_from_config(
loss, train_config, tvars)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (classification_metric_fn,
[pred, label_ids, is_real_example])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"q_input_ids": q_input_ids,
"d_input_ids": d_input_ids,
"score": logits
}
useful_inputs = ["data_id", "input_ids2", "data_ids"]
for input_name in useful_inputs:
if input_name in features:
predictions[input_name] = features[input_name]
output_spec = tf.compat.v1.estimator.tpu.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."""
tf_logging.info("*** Features ***")
for name in sorted(features.keys()):
tf_logging.info(" name = %s, shape = %s" %
(name, features[name].shape))
q_input_ids = features["q_input_ids"]
q_input_mask = features["q_input_mask"]
d_input_ids = features["d_input_ids"]
d_input_mask = features["d_input_mask"]
input_shape = get_shape_list(q_input_ids, expected_rank=2)
batch_size = input_shape[0]
doc_length = model_config.max_doc_length
num_docs = model_config.num_docs
d_input_ids_unpacked = tf.reshape(d_input_ids,
[-1, num_docs, doc_length])
d_input_mask_unpacked = tf.reshape(d_input_mask,
[-1, num_docs, doc_length])
d_input_ids_flat = tf.reshape(d_input_ids_unpacked, [-1, doc_length])
d_input_mask_flat = tf.reshape(d_input_mask_unpacked, [-1, doc_length])
q_segment_ids = tf.zeros_like(q_input_ids, tf.int32)
d_segment_ids = tf.zeros_like(d_input_ids_flat, tf.int32)
label_ids = features["label_ids"]
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"],
dtype=tf.float32)
else:
is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)
with tf.compat.v1.variable_scope(dual_model_prefix1):
q_model_config = copy.deepcopy(model_config)
q_model_config.max_seq_length = model_config.max_sent_length
model_q = model_class(
config=model_config,
is_training=is_training,
input_ids=q_input_ids,
input_mask=q_input_mask,
token_type_ids=q_segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
with tf.compat.v1.variable_scope(dual_model_prefix2):
d_model_config = copy.deepcopy(model_config)
d_model_config.max_seq_length = model_config.max_doc_length
model_d = model_class(
config=model_config,
is_training=is_training,
input_ids=d_input_ids_flat,
input_mask=d_input_mask_flat,
token_type_ids=d_segment_ids,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
pooled_q = model_q.get_pooled_output() # [batch, vector_size]
pooled_d_flat = model_d.get_pooled_output(
) # [batch, num_window, vector_size]
pooled_d = tf.reshape(pooled_d_flat, [batch_size, num_docs, -1])
pooled_q_t = tf.expand_dims(pooled_q, 1)
pooled_d_t = tf.transpose(pooled_d, [0, 2, 1])
all_logits = tf.matmul(pooled_q_t,
pooled_d_t) # [batch, 1, num_window]
if "hinge_all" in special_flags:
apply_loss_modeing = hinge_all
elif "sigmoid_all" in special_flags:
apply_loss_modeing = sigmoid_all
else:
apply_loss_modeing = hinge_max
logits, loss = apply_loss_modeing(all_logits, label_ids)
pred = tf.cast(logits > 0, tf.int32)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(
train_config, tvars)
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
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
if "simple_optimizer" in special_flags:
tf_logging.info("using simple optimizer")
train_op = create_simple_optimizer(loss,
train_config.learning_rate,
train_config.use_tpu)
else:
train_op = optimization.create_optimizer_from_config(
loss, train_config, tvars)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (classification_metric_fn,
[pred, label_ids, is_real_example])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"q_input_ids": q_input_ids,
"d_input_ids": d_input_ids,
"logits": logits
}
useful_inputs = ["data_id", "input_ids2", "data_ids"]
for input_name in useful_inputs:
if input_name in features:
predictions[input_name] = features[input_name]
output_spec = tf.compat.v1.estimator.tpu.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."""
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 mode == tf.estimator.ModeKeys.PREDICT:
label_ids = tf.ones([input_ids.shape[0]], dtype=tf.int32)
else:
label_ids = features["label_ids"]
label_ids = tf.reshape(label_ids, [-1])
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"],
dtype=tf.float32)
else:
is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
input_ids2 = features["input_ids2"]
input_mask2 = features["input_mask2"]
segment_ids2 = features["segment_ids2"]
with tf.compat.v1.variable_scope(dual_model_prefix1):
model_1 = BertModel(
config=model_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,
)
pooled = model_1.get_pooled_output()
if is_training:
pooled = dropout(pooled, 0.1)
logits = tf.keras.layers.Dense(train_config.num_classes,
name="cls_dense")(pooled)
with tf.compat.v1.variable_scope(dual_model_prefix2):
model_2 = BertModel(
config=model_config,
is_training=is_training,
input_ids=input_ids2,
input_mask=input_mask2,
token_type_ids=segment_ids2,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
pooled = model_2.get_pooled_output()
if is_training:
pooled = dropout(pooled, 0.1)
conf_probs = tf.keras.layers.Dense(
train_config.num_classes,
name="cls_dense",
activation=tf.keras.activations.softmax)(pooled)
confidence = conf_probs[:, 1]
confidence_loss = 1 - confidence
cls_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=logits, labels=label_ids)
k = model_config.k
alpha = model_config.alpha
loss_arr = cls_loss * confidence + confidence_loss * k
loss_arr = apply_weighted_loss(loss_arr, label_ids, alpha)
loss = tf.reduce_mean(input_tensor=loss_arr)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(
train_config, tvars)
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
def metric_fn(log_probs, label, is_real_example, confidence):
r = classification_metric_fn(log_probs, label, is_real_example)
r['confidence'] = tf.compat.v1.metrics.mean(confidence)
return r
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
tvars = None
train_op = optimization.create_optimizer_from_config(
loss, train_config, tvars)
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,
[logits, label_ids, is_real_example, confidence])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"logits": logits,
"confidence": confidence,
}
if "data_id" in features:
predictions['data_id'] = features['data_id']
output_spec = tf.compat.v1.estimator.tpu.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."""
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"]
label_ids = features["label_ids"]
label_ids = tf.reshape(label_ids, [-1])
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"], dtype=tf.float32)
else:
is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model = model_class(
config=model_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,
features=features,
)
probs = model.get_prob()
logits = probs
epsilon = 1e-12
# probs_ad = tf.clip_by_value(probs, epsilon, 1.0 - epsilon)
# logits1 = tf.math.log(probs_ad)
# logits0 = tf.zeros_like(logits1)
# logits = tf.stack([logits0, logits1], axis=1)
# prob2 = tf.nn.softmax(logits, axis=1)
# prob_err = prob2[:, 1] - probs
y_true = tf.cast(label_ids, tf.float32)
loss_arr = tf.keras.losses.BinaryCrossentropy()(y_true, probs)
loss = tf.reduce_mean(input_tensor=loss_arr)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(train_config, tvars)
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
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
tvars = None
train_op = optimization.create_optimizer_from_config(loss, train_config, tvars)
output_spec = TPUEstimatorSpec(mode=mode, loss=loss, train_op=train_op, scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (classification_metric_fn, [
logits, label_ids, is_real_example
])
output_spec = TPUEstimatorSpec(mode=mode, loss=loss, eval_metrics=eval_metrics, scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"label_ids": label_ids,
"logits": logits,
"score1": model.score1,
"score2": model.score2,
# "prob_err": prob_err,
}
if "data_id" in features:
predictions['data_id'] = features['data_id']
output_spec = tf.compat.v1.estimator.tpu.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."""
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"]
label_ids = features["label_ids"]
label_ids = tf.reshape(label_ids, [-1])
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"],
dtype=tf.float32)
else:
is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
if "feed_features" in special_flags:
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,
features=features,
)
else:
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,
)
if "new_pooling" in special_flags:
pooled = mimic_pooling(model.get_sequence_output(),
bert_config.hidden_size,
bert_config.initializer_range)
else:
pooled = model.get_pooled_output()
if train_config.checkpoint_type != "bert_nli" and train_config.use_old_logits:
tf_logging.info("Use old version of logistic regression")
logits = tf.keras.layers.Dense(train_config.num_classes,
name="cls_dense")(pooled)
else:
tf_logging.info("Use fixed version of logistic regression")
output_weights = tf.compat.v1.get_variable(
"output_weights", [3, bert_config.hidden_size],
initializer=tf.compat.v1.truncated_normal_initializer(
stddev=0.02))
output_bias = tf.compat.v1.get_variable(
"output_bias", [3],
initializer=tf.compat.v1.zeros_initializer())
if is_training:
pooled = dropout(pooled, 0.1)
logits = tf.matmul(pooled, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
loss_arr = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=logits, labels=label_ids)
if "bias_loss" in special_flags:
tf_logging.info("Using special_flags : bias_loss")
loss_arr = reweight_zero(label_ids, loss_arr)
loss = tf.reduce_mean(input_tensor=loss_arr)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(
train_config, tvars)
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
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
if "simple_optimizer" in special_flags:
tf_logging.info("using simple optimizer")
train_op = create_simple_optimizer(loss,
train_config.learning_rate,
train_config.use_tpu)
else:
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 = (classification_metric_fn,
[logits, label_ids, is_real_example])
output_spec = TPUEstimatorSpec(mode=model,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
probs = tf.nn.softmax(logits, axis=-1)
gradient_list = tf.gradients(probs[:, 1], model.embedding_output)
print(len(gradient_list))
gradient = gradient_list[0]
print(gradient.shape)
gradient = tf.reduce_sum(gradient, axis=2)
predictions = {
"input_ids": input_ids,
"gradient": gradient,
"labels": label_ids,
"logits": logits
}
output_spec = tf.compat.v1.estimator.tpu.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."""
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"]
label_ids = features["label_ids"]
label_ids = tf.reshape(label_ids, [-1])
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"],
dtype=tf.float32)
else:
is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model_1 = BertModel(
config=model_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,
)
pooled = model_1.get_pooled_output()
if is_training:
pooled = dropout(pooled, 0.1)
logits = tf.keras.layers.Dense(train_config.num_classes,
name="cls_dense")(pooled)
loss_arr = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=logits, labels=label_ids)
loss = tf.reduce_mean(input_tensor=loss_arr)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(
train_config, tvars)
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
global_step = tf.compat.v1.train.get_or_create_global_step()
init_lr = train_config.learning_rate
num_warmup_steps = train_config.num_warmup_steps
num_train_steps = train_config.num_train_steps
learning_rate2_const = tf.constant(value=init_lr,
shape=[],
dtype=tf.float32)
learning_rate2_decayed = tf.compat.v1.train.polynomial_decay(
learning_rate2_const,
global_step,
num_train_steps,
end_learning_rate=0.0,
power=1.0,
cycle=False)
if num_warmup_steps:
global_steps_int = tf.cast(global_step, tf.int32)
warmup_steps_int = tf.constant(num_warmup_steps, dtype=tf.int32)
global_steps_float = tf.cast(global_steps_int, tf.float32)
warmup_steps_float = tf.cast(warmup_steps_int, tf.float32)
warmup_percent_done = global_steps_float / warmup_steps_float
warmup_learning_rate = init_lr * warmup_percent_done
is_warmup = tf.cast(global_steps_int < warmup_steps_int,
tf.float32)
learning_rate = ((1.0 - is_warmup) * learning_rate2_decayed +
is_warmup * warmup_learning_rate)
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
tvars = None
train_op = optimization.create_optimizer_from_config(
loss, train_config, tvars)
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
eval_metrics = (classification_metric_fn,
[logits, label_ids, is_real_example])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
def reform_scala(t):
return tf.reshape(t, [1])
predictions = {
"input_ids": input_ids,
"label_ids": label_ids,
"logits": logits,
"learning_rate2_const": reform_scala(learning_rate2_const),
"warmup_percent_done": reform_scala(warmup_percent_done),
"warmup_learning_rate": reform_scala(warmup_learning_rate),
"learning_rate": reform_scala(learning_rate),
"learning_rate2_decayed": reform_scala(learning_rate2_decayed),
}
if "data_id" in features:
predictions['data_id'] = features['data_id']
output_spec = tf.compat.v1.estimator.tpu.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."""
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"]
label_ids = features["label_ids"]
input_shape = get_shape_list2(input_ids)
batch_size, seq_length = input_shape
if "is_real_example" in features:
is_real_example = tf.cast(features["is_real_example"],
dtype=tf.float32)
else:
is_real_example = tf.ones([batch_size, 1], dtype=tf.float32)
label_ids = tf.reshape(
label_ids, [batch_size, seq_length, train_config.num_classes])
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model = BertModel(
config=model_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,
)
seq_out = model.get_sequence_output()
if is_training:
seq_out = dropout(seq_out, 0.1)
logits = tf.keras.layers.Dense(train_config.num_classes,
name="cls_dense")(seq_out)
probs = tf.math.sigmoid(logits)
eps = 1e-10
label_logs = tf.math.log(label_ids + eps)
#scale = model_config.scale
#label_ids = scale * label_ids
is_valid_mask = tf.cast(segment_ids, tf.float32)
#loss_arr = tf.keras.losses.MAE(y_true=label_ids, y_pred=probs)
loss_arr = tf.keras.losses.MAE(y_true=label_logs, y_pred=logits)
loss_arr = loss_arr * is_valid_mask
loss = tf.reduce_mean(input_tensor=loss_arr)
tvars = tf.compat.v1.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if train_config.init_checkpoint:
initialized_variable_names, init_fn = get_init_fn(
train_config, tvars)
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
def metric_fn(probs, label, is_real_example):
cut = math.exp(-10)
pred_binary = probs > cut
label_binary_all = label > cut
pred_binary = pred_binary[:, :, 0]
label_binary_1 = label_binary_all[:, :, 1]
label_binary_0 = label_binary_all[:, :, 0]
precision = tf.compat.v1.metrics.precision(predictions=pred_binary,
labels=label_binary_0)
recall = tf.compat.v1.metrics.recall(predictions=pred_binary,
labels=label_binary_0)
true_rate_1 = tf.compat.v1.metrics.mean(label_binary_1)
true_rate_0 = tf.compat.v1.metrics.mean(label_binary_0)
mae = tf.compat.v1.metrics.mean_absolute_error(
labels=label, predictions=probs, weights=is_real_example)
return {
"mae": mae,
"precision": precision,
"recall": recall,
"true_rate_1": true_rate_1,
"true_rate_0": true_rate_0,
}
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
tvars = None
train_op = optimization.create_optimizer_from_config(
loss, train_config, tvars)
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, [probs, label_ids, is_real_example])
output_spec = TPUEstimatorSpec(mode=mode,
loss=loss,
eval_metrics=eval_metrics,
scaffold_fn=scaffold_fn)
else:
predictions = {
"input_ids": input_ids,
"logits": logits,
"label_ids": label_ids,
}
if "data_id" in features:
predictions['data_id'] = features['data_id']
output_spec = tf.compat.v1.estimator.tpu.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("model_fn_pooling_long_things")
log_features(features)
input_ids = features["input_ids"] # [batch_size, seq_length]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
label_ids = features["label_ids"]
label_ids = tf.reshape(label_ids, [-1])
batch_size, _ = get_shape_list(
input_mask) # This is not real batch_size, 2 * real_batch_size
use_context = tf.ones([batch_size, 1], tf.int32)
total_sequence_length = config.total_sequence_length
stacked_input_ids, stacked_input_mask, stacked_segment_ids, \
= split_and_append_sep2(input_ids[:, :total_sequence_length],
input_mask[:, :total_sequence_length],
segment_ids[:, :total_sequence_length],
total_sequence_length, config.window_size, CLS_ID, EOW_ID)
if "focus_mask" in features:
focus_mask = features["focus_mask"]
_, stacked_focus_mask, _, \
= split_and_append_sep2(input_ids[:, :total_sequence_length],
focus_mask[:, :total_sequence_length],
segment_ids[:, :total_sequence_length],
total_sequence_length, config.window_size, CLS_ID, EOW_ID)
features["focus_mask"] = r3to2(stacked_focus_mask)
batch_size, num_seg, seq_len = get_shape_list2(stacked_input_ids)
input_ids_2d = r3to2(stacked_input_ids)
input_mask_2d = r3to2(stacked_input_mask)
segment_ids_2d = r3to2(stacked_segment_ids)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
if "feed_features" in special_flags:
model = model_class(
config=config,
is_training=is_training,
input_ids=input_ids_2d,
input_mask=input_mask_2d,
token_type_ids=segment_ids_2d,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
features=features,
)
else:
model = model_class(
config=config,
is_training=is_training,
input_ids=input_ids_2d,
input_mask=input_mask_2d,
token_type_ids=segment_ids_2d,
use_one_hot_embeddings=train_config.use_one_hot_embeddings,
)
sequence_output_2d = model.get_sequence_output()
pooled_output = model.get_pooled_output()
if is_training:
pooled_output = dropout(pooled_output, 0.1)
pooled_output_3d = tf.reshape(pooled_output, [batch_size, num_seg, -1])
sequence_output_3d = tf.reshape(sequence_output_2d,
[batch_size, num_seg, seq_len, -1])
logits = pooling_modeling(config.option_name, train_config.num_classes,
pooled_output_3d, sequence_output_3d)
loss_arr = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=logits, labels=label_ids)
loss = tf.reduce_mean(input_tensor=loss_arr)
tvars = tf.compat.v1.trainable_variables()
if train_config.init_checkpoint:
initialized_variable_names, init_fn = classification_model_fn.get_init_fn(
train_config, tvars)
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,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
output_spec = TPUEstimatorSpec(mode=model,
loss=loss,
eval_metrics=None,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.PREDICT:
predictions = {"input_ids": input_ids, "logits": logits}
if "data_id" in features:
predictions['data_id'] = features['data_id']
output_spec = TPUEstimatorSpec(mode=model,
loss=loss,
predictions=predictions,
scaffold_fn=scaffold_fn)
return output_spec
