modeling¶
Modeling classes for ALBERT model.
-
class
AlbertPretrainedModel
(*args, **kwargs)[source]¶ Bases:
paddlenlp.transformers.model_utils.PretrainedModel
An abstract class for pretrained ALBERT models. It provides ALBERT related
model_config_file
,pretrained_init_configuration
,resource_files_names
,pretrained_resource_files_map
,base_model_prefix
for downloading and loading pretrained models. SeePretrainedModel
for more details.-
base_model_class
¶
-
-
class
AlbertModel
(vocab_size=30000, embedding_size=128, hidden_size=768, num_hidden_layers=12, num_hidden_groups=1, num_attention_heads=12, intermediate_size=3072, inner_group_num=1, hidden_act='gelu', hidden_dropout_prob=0, attention_probs_dropout_prob=0, max_position_embeddings=512, type_vocab_size=2, initializer_range=0.02, layer_norm_eps=1e-12, pad_token_id=0, bos_token_id=2, eos_token_id=3, add_pooling_layer=True)[source]¶ Bases:
paddlenlp.transformers.albert.modeling.AlbertPretrainedModel
The bare Albert Model transformer outputting raw hidden-states.
This model inherits from
PretrainedModel
. Refer to the superclass documentation for the generic methods.This model is also a Paddle paddle.nn.Layer subclass. Use it as a regular Paddle Layer and refer to the Paddle documentation for all matter related to general usage and behavior.
- Parameters
vocab_size (int, optional) – Vocabulary size of
inputs_ids
inAlbertModel
. Also is the vocab size of token embedding matrix. Defines the number of different tokens that can be represented by theinputs_ids
passed when callingAlbertModel
. Defaults to30000
.embedding_size (int, optional) – Dimensionality of the embedding layer. Defaults to
128
.hidden_size (int, optional) – Dimensionality of the encoder layer and pooler layer. Defaults to
768
.num_hidden_layers (int, optional) – Number of hidden layers in the Transformer encoder. Defaults to
12
.inner_group_num (int, optional) – Number of hidden groups in the Transformer encoder. Defaults to
1
.num_attention_heads (int, optional) – Number of attention heads for each attention layer in the Transformer encoder. Defaults to
12
.intermediate_size (int, optional) – Dimensionality of the feed-forward (ff) layer in the encoder. Input tensors to ff layers are firstly projected from
hidden_size
tointermediate_size
, and then projected back tohidden_size
. Typicallyintermediate_size
is larger thanhidden_size
.inner_group_num – Number of inner groups in a hidden group. Default to
1
.hidden_act (str, optional) – The non-linear activation function in the feed-forward layer.
"gelu"
,"relu"
and any other paddle supported activation functions are supported.hidden_dropout_prob (float, optional) – The dropout probability for all fully connected layers in the embeddings and encoder. Defaults to
0
.attention_probs_dropout_prob (float, optional) – The dropout probability used in MultiHeadAttention in all encoder layers to drop some attention target. Defaults to
0
.max_position_embeddings (int, optional) – The maximum value of the dimensionality of position encoding, which dictates the maximum supported length of an input sequence. Defaults to
512
.type_vocab_size (int, optional) – The vocabulary size of
token_type_ids
. Defaults to12
.initializer_range (float, optional) –
The standard deviation of the normal initializer. Defaults to
0.02
.Note
A normal_initializer initializes weight matrices as normal distributions. See
BertPretrainedModel.init_weights()
for how weights are initialized inElectraModel
.layer_norm_eps (float, optional) – The
epsilon
parameter used inpaddle.nn.LayerNorm
for initializing layer normalization layers. A small value to the variance added to the normalization layer to prevent division by zero. Default to1e-12
.pad_token_id (int, optional) – The index of padding token in the token vocabulary. Defaults to
0
.add_pooling_layer (bool, optional) – Whether or not to add the pooling layer. Default to
False
.
-
get_input_embeddings
()[source]¶ get input embedding of model
- Returns
embedding of model
- Return type
nn.Embedding
-
set_input_embeddings
(value)[source]¶ set new input embedding for model
- Parameters
value (Embedding) – the new embedding of model
- Raises
NotImplementedError – Model has not implement
set_input_embeddings
method
-
forward
(input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, output_hidden_states=False, output_attentions=False, return_dict=False)[source]¶ The AlbertModel forward method, overrides the
__call__()
special method.- Parameters
input_ids (Tensor) – Indices of input sequence tokens in the vocabulary. They are numerical representations of tokens that build the input sequence. Its data type should be
int64
and it has a shape of [batch_size, sequence_length].attention_mask (Tensor, optional) – Mask used in multi-head attention to avoid performing attention on to some unwanted positions, usually the paddings or the subsequent positions. Its data type can be int, float and bool. When the data type is bool, the
masked
tokens haveFalse
values and the others haveTrue
values. When the data type is int, themasked
tokens have0
values and the others have1
values. When the data type is float, themasked
tokens have-INF
values and the others have0
values. It is a tensor with shape broadcasted to[batch_size, num_attention_heads, sequence_length, sequence_length]
. Defaults toNone
, which means nothing needed to be prevented attention to.token_type_ids (Tensor, optional) –
Segment token indices to indicate different portions of the inputs. Selected in the range
[0, type_vocab_size - 1]
. Iftype_vocab_size
is 2, which means the inputs have two portions. Indices can either be 0 or 1:0 corresponds to a sentence A token,
1 corresponds to a sentence B token.
Its data type should be
int64
and it has a shape of [batch_size, sequence_length]. Defaults toNone
, which means we don’t add segment embeddings.position_ids (Tensor, optional) – Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
[0, max_position_embeddings - 1]
. Shape as(batch_size, num_tokens)
and dtype as int64. Defaults toNone
.head_mask (Tensor, optional) –
Mask to nullify selected heads of the self-attention modules. Masks values can either be 0 or 1:
1 indicates the head is not masked,
0 indicated the head is masked.
inputs_embeds – If you want to control how to convert
inputs_ids
indices into associated vectors, you can pass an embedded representation directly instead of passinginputs_ids
.
- Returns
Returns tuple (
sequence_output
,pooled_output
) or a dict withlast_hidden_state
,pooled_output
,all_hidden_states
,all_attentions
fields.With the fields:
sequence_output
(Tensor):Sequence of hidden-states at the last layer of the model. It’s data type should be float32 and has a shape of [
batch_size, sequence_length, hidden_size
].
pooled_output
(Tensor):The output of first token (
[CLS]
) in sequence. We “pool” the model by simply taking the hidden state corresponding to the first token. Its data type should be float32 and has a shape of [batch_size, hidden_size].
last_hidden_state
(Tensor):The output of the last encoder layer, it is also the
sequence_output
. It’s data type should be float32 and has a shape of [batch_size, sequence_length, hidden_size].
all_hidden_states
(Tensor):Hidden_states of all layers in the Transformer encoder. The length of
all_hidden_states
isnum_hidden_layers + 1
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, sequence_length, hidden_size
].
all_attentions
(Tensor):Attentions of all layers of in the Transformer encoder. The length of
all_attentions
isnum_hidden_layers
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, num_attention_heads, sequence_length, sequence_length
].
- Return type
tuple or Dict
Example
import paddle from paddlenlp.transformers import AlbertModel, AlbertTokenizer tokenizer = AlbertTokenizer.from_pretrained('albert-base-v1') model = AlbertModel.from_pretrained('albert-base-v1') inputs = tokenizer("Welcome to use PaddlePaddle and PaddleNLP!") inputs = {k:paddle.to_tensor([v]) for (k, v) in inputs.items()} output = model(**inputs)
-
class
AlbertForPretraining
(albert, lm_head, sop_head, vocab_size)[source]¶ Bases:
paddlenlp.transformers.albert.modeling.AlbertPretrainedModel
Albert Model with a
masked language modeling
head and asentence order prediction
head on top.- Parameters
albert (
AlbertModel
) – An instance ofAlbertModel
.lm_head (
AlbertMLMHead
) – An instance ofAlbertSOPHead
.sop_head (
AlbertSOPHead
) – An instance ofAlbertSOPHead
.vocab_size (int) – See
AlbertModel
.
-
get_output_embeddings
()[source]¶ To be overwrited for models with output embeddings
- Returns
the otuput embedding of model
- Return type
Optional[Embedding]
-
get_input_embeddings
()[source]¶ get input embedding of model
- Returns
embedding of model
- Return type
nn.Embedding
-
forward
(input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, sentence_order_label=None, labels=None, output_attentions=False, output_hidden_states=False, return_dict=False)[source]¶ The AlbertForPretraining forward method, overrides the __call__() special method.
- Parameters
input_ids (Tensor) – See
AlbertModel
.attention_mask (list, optional) – See
AlbertModel
.token_type_ids (Tensor, optional) – See
AlbertModel
.position_ids (Tensor, optional) – See
AlbertModel
.head_mask (Tensor, optional) – See
AlbertModel
.inputs_embeds (Tensor, optional) – See
AlbertModel
.sentence_order_label (Tensor, optional) – Labels of the next sequence prediction. Input should be a sequence pair Indices should be 0 or 1.
0
indicates original order (sequence A, then sequence B), and1
indicates switched order (sequence B, then sequence A). Defaults toNone
.output_hidden_states (bool, optional) – Whether to return the hidden states of all layers. Defaults to
False
.output_attentions (bool, optional) – Whether to return the attentions tensors of all attention layers. Defaults to
False
.return_dict (bool, optional) – Whether to return a
ModelOutput
object. IfFalse
, the output will be a tuple of tensors. Defaults toFalse
.
- Returns
Returns tuple (
prediction_scores
,sop_scores
) or a dict withprediction_logits
,sop_logits
,pooled_output
,hidden_states
,attentions
fields.With the fields:
prediction_scores
(Tensor):The scores of masked token prediction. Its data type should be float32. and its shape is [batch_size, sequence_length, vocab_size].
sop_scores
(Tensor):The scores of sentence order prediction. Its data type should be float32 and its shape is [batch_size, 2].
prediction_logits
(Tensor):The scores of masked token prediction. Its data type should be float32. and its shape is [batch_size, sequence_length, vocab_size].
sop_logits
(Tensor):The scores of sentence order prediction. Its data type should be float32 and its shape is [batch_size, 2].
hidden_states
(Tensor):Hidden_states of all layers in the Transformer encoder. The length of
hidden_states
isnum_hidden_layers + 1
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, sequence_length, hidden_size
].
attentions
(Tensor):Attentions of all layers of in the Transformer encoder. The length of
attentions
isnum_hidden_layers
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, num_attention_heads, sequence_length, sequence_length
].
- Return type
tuple or Dict
-
class
AlbertForMaskedLM
(albert)[source]¶ Bases:
paddlenlp.transformers.albert.modeling.AlbertPretrainedModel
Albert Model with a
masked language modeling
head on top.- Parameters
albert (
AlbertModel
) – An instance ofAlbertModel
.
-
get_output_embeddings
()[source]¶ To be overwrited for models with output embeddings
- Returns
the otuput embedding of model
- Return type
Optional[Embedding]
-
get_input_embeddings
()[source]¶ get input embedding of model
- Returns
embedding of model
- Return type
nn.Embedding
-
forward
(input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, labels=None, output_hidden_states=False, output_attentions=False, return_dict=False)[source]¶ The AlbertForPretraining forward method, overrides the __call__() special method.
- Parameters
input_ids (Tensor) – See
AlbertModel
.attention_mask (list, optional) – See
AlbertModel
.token_type_ids (Tensor, optional) – See
AlbertModel
.position_ids (Tensor, optional) – See
AlbertModel
.head_mask (Tensor, optional) – See
AlbertModel
.inputs_embeds (Tensor, optional) – See
AlbertModel
.output_hidden_states (bool, optional) – Whether to return the hidden states of all layers. Defaults to
False
.output_attentions (bool, optional) – Whether to return the attentions tensors of all attention layers. Defaults to
False
.return_dict (bool, optional) – Whether to return a
ModelOutput
object. IfFalse
, the output will be a tuple of tensors. Defaults toFalse
.
- Returns
Returns tensor
prediction_scores
or a dict withlogits
,hidden_states
,attentions
fields.With the fields:
prediction_scores
(Tensor):The scores of masked token prediction. Its data type should be float32. and its shape is [batch_size, sequence_length, vocab_size].
logits
(Tensor):The scores of masked token prediction. Its data type should be float32. and its shape is [batch_size, sequence_length, vocab_size].
hidden_states
(Tensor):Hidden_states of all layers in the Transformer encoder. The length of
hidden_states
isnum_hidden_layers + 1
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, sequence_length, hidden_size
].
attentions
(Tensor):Attentions of all layers of in the Transformer encoder. The length of
attentions
isnum_hidden_layers
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, num_attention_heads, sequence_length, sequence_length
].
- Return type
Tensor or Dict
-
class
AlbertForSequenceClassification
(albert, classifier_dropout_prob=0, num_classes=2)[source]¶ Bases:
paddlenlp.transformers.albert.modeling.AlbertPretrainedModel
Albert Model with a linear layer on top of the output layer, designed for sequence classification/regression tasks like GLUE tasks.
- Parameters
albert (
AlbertModel
) – An instance of AlbertModel.classifier_dropput_prob (float, optional) – The dropout probability for the classifier. Defaults to
0
.num_classes (int, optional) – The number of classes. Defaults to
2
.
-
forward
(input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, labels=None, output_hidden_states=False, output_attentions=False, return_dict=False)[source]¶ The AlbertForSequenceClassification forward method, overrides the __call__() special method.
- Parameters
input_ids (Tensor) – See
AlbertModel
.attention_mask (list, optional) – See
AlbertModel
.token_type_ids (Tensor, optional) – See
AlbertModel
.position_ids (Tensor, optional) – See
AlbertModel
.head_mask (Tensor, optional) – See
AlbertModel
.inputs_embeds (Tensor, optional) – See
AlbertModel
.labels (Tensor of shape
(batch_size,)
, optional) – Labels for computing the sequence classification/regression loss. Indices should be in[0, ..., num_classes - 1]
. Ifnum_classes == 1
a regression loss is computed (Mean-Square loss), Ifnum_classes > 1
a classification loss is computed (Cross-Entropy).output_hidden_states (bool, optional) – Whether to return the hidden states of all layers. Defaults to
False
.output_attentions (bool, optional) – Whether to return the attentions tensors of all attention layers. Defaults to
False
.return_dict (bool, optional) – Whether to return a
SequenceClassifierOutput
object. IfFalse
, the output will be a tuple of tensors. Defaults toFalse
.
- Returns
Returns tensor
logits
, or a dict withlogits
,hidden_states
,attentions
fields.With the fields:
logits
(Tensor):A tensor of the input text classification logits. Shape as
[batch_size, num_classes]
and dtype as float32.
hidden_states
(Tensor):Hidden_states of all layers in the Transformer encoder. The length of
hidden_states
isnum_hidden_layers + 1
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, sequence_length, hidden_size
].
attentions
(Tensor):Attentions of all layers of in the Transformer encoder. The length of
attentions
isnum_hidden_layers
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, num_attention_heads, sequence_length, sequence_length
].
- Return type
Tensor or Dict
Example
import paddle from paddlenlp.transformers import AlbertForSequenceClassification, AlbertTokenizer tokenizer = AlbertTokenizer.from_pretrained('albert-base-v1') model = AlbertForSequenceClassification.from_pretrained('albert-base-v1') inputs = tokenizer("Welcome to use PaddlePaddle and PaddleNLP!") inputs = {k:paddle.to_tensor([v]) for (k, v) in inputs.items()} outputs = model(**inputs) logits = outputs[0]
-
class
AlbertForTokenClassification
(albert, num_classes=2)[source]¶ Bases:
paddlenlp.transformers.albert.modeling.AlbertPretrainedModel
Albert Model with a linear layer on top of the hidden-states output layer, designed for token classification tasks like NER tasks.
- Parameters
albert (
AlbertModel
) – An instance of AlbertModel.num_classes (int, optional) – The number of classes. Defaults to
2
.
-
forward
(input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, labels=None, output_hidden_states=False, output_attentions=False, return_dict=False)[source]¶ The AlbertForTokenClassification forward method, overrides the __call__() special method.
- Parameters
input_ids (Tensor) – See
AlbertModel
.attention_mask (list, optional) – See
AlbertModel
.token_type_ids (Tensor, optional) – See
AlbertModel
.position_ids (Tensor, optional) – See
AlbertModel
.head_mask (Tensor, optional) – See
AlbertModel
.inputs_embeds (Tensor, optional) – See
AlbertModel
.labels (Tensor of shape
(batch_size, sequence_length)
, optional) – Labels for computing the token classification loss. Indices should be in[0, ..., num_classes - 1]
.output_hidden_states (bool, optional) – Whether to return the hidden states of all layers. Defaults to
False
.output_attentions (bool, optional) – Whether to return the attentions tensors of all attention layers. Defaults to
False
.return_dict (bool, optional) – Whether to return a
TokenClassifierOutput
object. IfFalse
, the output will be a tuple of tensors. Defaults toFalse
.
- Returns
Returns tensor
logits
, or a dict withlogits
,hidden_states
,attentions
fields.With the fields:
logits
(Tensor):A tensor of the input token classification logits. Shape as
[batch_size, sequence_length, num_classes]
and dtype asfloat32
.
hidden_states
(Tensor):Hidden_states of all layers in the Transformer encoder. The length of
hidden_states
isnum_hidden_layers + 1
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, sequence_length, hidden_size
].
attentions
(Tensor):Attentions of all layers of in the Transformer encoder. The length of
attentions
isnum_hidden_layers
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, num_attention_heads, sequence_length, sequence_length
].
- Return type
Tensor or Dict
Example
import paddle from paddlenlp.transformers import AlbertForTokenClassification, AlbertTokenizer tokenizer = AlbertTokenizer.from_pretrained('albert-base-v1') model = AlbertForTokenClassification.from_pretrained('albert-base-v1') inputs = tokenizer("Welcome to use PaddlePaddle and PaddleNLP!") inputs = {k:paddle.to_tensor([v]) for (k, v) in inputs.items()} outputs = model(**inputs) logits = outputs[0]
-
class
AlbertForQuestionAnswering
(albert, num_labels=2)[source]¶ Bases:
paddlenlp.transformers.albert.modeling.AlbertPretrainedModel
Albert Model with a linear layer on top of the hidden-states output to compute
span_start_logits
andspan_end_logits
, designed for question-answering tasks like SQuAD.- Parameters
albert (
AlbertModel
) – An instance of AlbertModel.num_classes (int) – The number of classes.
-
forward
(input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, start_positions=None, end_positions=None, output_hidden_states=False, output_attentions=False, return_dict=False)[source]¶ The AlbertForQuestionAnswering forward method, overrides the __call__() special method.
- Parameters
input_ids (Tensor) – See
AlbertModel
.attention_mask (list, optional) – See
AlbertModel
.token_type_ids (Tensor, optional) – See
AlbertModel
.position_ids (Tensor, optional) – See
AlbertModel
.head_mask (Tensor, optional) – See
AlbertModel
.inputs_embeds (Tensor, optional) – See
AlbertModel
.start_positions (Tensor of shape
(batch_size,)
, optional) – Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (sequence_length
). Position outside of the sequence are not taken into account for computing the loss.end_positions (Tensor of shape
(batch_size,)
, optional) – Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (sequence_length
). Position outside of the sequence are not taken into account for computing the loss.output_hidden_states (bool, optional) – Whether to return the hidden states of all layers. Defaults to
False
.output_attentions (bool, optional) – Whether to return the attentions tensors of all attention layers. Defaults to
False
.return_dict (bool, optional) – Whether to return a
QuestionAnsweringModelOutput
object. IfFalse
, the output will be a tuple of tensors. Defaults toFalse
.
- Returns
Returns tuple (
start_logits, end_logits
)or a dict withstart_logits
,end_logits
,hidden_states
,attentions
fields.With the fields:
start_logits
(Tensor):A tensor of the input token classification logits, indicates the start position of the labelled span. Its data type should be float32 and its shape is [batch_size, sequence_length].
end_logits
(Tensor):A tensor of the input token classification logits, indicates the end position of the labelled span. Its data type should be float32 and its shape is [batch_size, sequence_length].
hidden_states
(Tensor):Hidden_states of all layers in the Transformer encoder. The length of
hidden_states
isnum_hidden_layers + 1
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, sequence_length, hidden_size
].
attentions
(Tensor):Attentions of all layers of in the Transformer encoder. The length of
attentions
isnum_hidden_layers
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, num_attention_heads, sequence_length, sequence_length
].
- Return type
tuple or Dict
Example
import paddle from paddlenlp.transformers import AlbertForQuestionAnswering, AlbertTokenizer tokenizer = AlbertTokenizer.from_pretrained('albert-base-v1') model = AlbertForQuestionAnswering.from_pretrained('albert-base-v1') inputs = tokenizer("Welcome to use PaddlePaddle and PaddleNLP!") inputs = {k:paddle.to_tensor([v]) for (k, v) in inputs.items()} outputs = model(**inputs) logits = outputs[0]
-
class
AlbertForMultipleChoice
(albert)[source]¶ Bases:
paddlenlp.transformers.albert.modeling.AlbertPretrainedModel
Albert Model with a linear layer on top of the hidden-states output layer, designed for multiple choice tasks like SWAG tasks .
- Parameters
albert (
AlbertModel
) – An instance of AlbertModel.
-
forward
(input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, labels=None, output_hidden_states=False, output_attentions=False, return_dict=False)[source]¶ The AlbertForQuestionAnswering forward method, overrides the __call__() special method.
- Parameters
input_ids (Tensor) – See
AlbertModel
.attention_mask (list, optional) – See
AlbertModel
.token_type_ids (Tensor, optional) – See
AlbertModel
.position_ids (Tensor, optional) – See
AlbertModel
.head_mask (Tensor, optional) – See
AlbertModel
.inputs_embeds (Tensor, optional) – See
AlbertModel
.labels (Tensor of shape
(batch_size, )
, optional) – Labels for computing the multiple choice classification loss. Indices should be in[0, ..., num_choices-1]
wherenum_choices
is the size of the second dimension of the input tensors. (Seeinput_ids
above)output_hidden_states (bool, optional) – Whether to return the hidden states of all layers. Defaults to
False
.output_attentions (bool, optional) – Whether to return the attentions tensors of all attention layers. Defaults to
False
.return_dict (bool, optional) – Whether to return a
MultipleChoiceModelOutput
object. IfFalse
, the output will be a tuple of tensors. Defaults toFalse
.
- Returns
Returns tensor
reshaped_logits
or a dict withreshaped_logits
,hidden_states
,attentions
fields.With the fields:
reshaped_logits
(Tensor):A tensor of the input multiple choice classification logits. Shape as
[batch_size, num_classes]
and dtype asfloat32
.
hidden_states
(Tensor):Hidden_states of all layers in the Transformer encoder. The length of
hidden_states
isnum_hidden_layers + 1
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, sequence_length, hidden_size
].
attentions
(Tensor):Attentions of all layers of in the Transformer encoder. The length of
attentions
isnum_hidden_layers
. For all element in the tuple, its data type should be float32 and its shape is [batch_size, num_attention_heads, sequence_length, sequence_length
].
- Return type
Tensor or Dict