# coding=utf-8
# Copyright 2021 The Google Flax Team Authors and The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Callable, Optional, Tuple
import numpy as np
import flax.linen as nn
import jax
import jax.numpy as jnp
from flax.core.frozen_dict import FrozenDict
from flax.linen import dot_product_attention
from jax import lax
from jax.random import PRNGKey
from ...file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
from ...modeling_flax_outputs import (
FlaxBaseModelOutput,
FlaxBaseModelOutputWithPooling,
FlaxMaskedLMOutput,
FlaxMultipleChoiceModelOutput,
FlaxQuestionAnsweringModelOutput,
FlaxSequenceClassifierOutput,
FlaxTokenClassifierOutput,
)
from ...modeling_flax_utils import ACT2FN, FlaxPreTrainedModel, append_call_sample_docstring, overwrite_call_docstring
from ...utils import logging
from .configuration_roberta import RobertaConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "roberta-base"
_CONFIG_FOR_DOC = "RobertaConfig"
_TOKENIZER_FOR_DOC = "RobertaTokenizer"
def create_position_ids_from_input_ids(input_ids, padding_idx):
"""
Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
are ignored. This is modified from fairseq's `utils.make_positions`.
Args:
input_ids: jnp.ndarray
padding_idx: int
Returns: jnp.ndarray
"""
# The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
mask = (input_ids != padding_idx).astype("i4")
if mask.ndim > 2:
mask = mask.reshape((-1, mask.shape[-1]))
incremental_indices = jnp.cumsum(mask, axis=1).astype("i4") * mask
incremental_indices = incremental_indices.reshape(input_ids.shape)
else:
incremental_indices = jnp.cumsum(mask, axis=1).astype("i4") * mask
return incremental_indices.astype("i4") + padding_idx
ROBERTA_START_DOCSTRING = r"""
This model inherits from :class:`~transformers.FlaxPreTrainedModel`. Check the superclass documentation for the
generic methods the library implements for all its model (such as downloading, saving and converting weights from
PyTorch models)
This model is also a Flax Linen `flax.linen.Module
<https://flax.readthedocs.io/en/latest/flax.linen.html#module>`__ subclass. Use it as a regular Flax linen Module
and refer to the Flax documentation for all matter related to general usage and behavior.
Finally, this model supports inherent JAX features such as:
- `Just-In-Time (JIT) compilation <https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit>`__
- `Automatic Differentiation <https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation>`__
- `Vectorization <https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap>`__
- `Parallelization <https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap>`__
Parameters:
config (:class:`~transformers.RobertaConfig`): Model configuration class with all the parameters of the
model. Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the :meth:`~transformers.FlaxPreTrainedModel.from_pretrained` method to load the
model weights.
"""
ROBERTA_INPUTS_DOCSTRING = r"""
Args:
input_ids (:obj:`numpy.ndarray` of shape :obj:`({0})`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using :class:`~transformers.BertTokenizer`. See
:func:`transformers.PreTrainedTokenizer.encode` and :func:`transformers.PreTrainedTokenizer.__call__` for
details.
`What are input IDs? <../glossary.html#input-ids>`__
attention_mask (:obj:`numpy.ndarray` of shape :obj:`({0})`, `optional`):
Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
`What are attention masks? <../glossary.html#attention-mask>`__
token_type_ids (:obj:`numpy.ndarray` of shape :obj:`({0})`, `optional`):
Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
1]``:
- 0 corresponds to a `sentence A` token,
- 1 corresponds to a `sentence B` token.
`What are token type IDs? <../glossary.html#token-type-ids>`__
position_ids (:obj:`numpy.ndarray` of shape :obj:`({0})`, `optional`):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
config.max_position_embeddings - 1]``.
return_dict (:obj:`bool`, `optional`):
Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
"""
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertEmbeddings with Bert->Roberta
class FlaxRobertaEmbeddings(nn.Module):
"""Construct the embeddings from word, position and token_type embeddings."""
config: RobertaConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.word_embeddings = nn.Embed(
self.config.vocab_size,
self.config.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
dtype=self.dtype,
)
self.position_embeddings = nn.Embed(
self.config.max_position_embeddings,
self.config.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
dtype=self.dtype,
)
self.token_type_embeddings = nn.Embed(
self.config.type_vocab_size,
self.config.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
dtype=self.dtype,
)
self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
def __call__(self, input_ids, token_type_ids, position_ids, attention_mask, deterministic: bool = True):
# Embed
inputs_embeds = self.word_embeddings(input_ids.astype("i4"))
position_embeds = self.position_embeddings(position_ids.astype("i4"))
token_type_embeddings = self.token_type_embeddings(token_type_ids.astype("i4"))
# Sum all embeddings
hidden_states = inputs_embeds + token_type_embeddings + position_embeds
# Layer Norm
hidden_states = self.LayerNorm(hidden_states)
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
return hidden_states
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertSelfAttention with Bert->Roberta
class FlaxRobertaSelfAttention(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
if self.config.hidden_size % self.config.num_attention_heads != 0:
raise ValueError(
"`config.hidden_size`: {self.config.hidden_size} has to be a multiple of `config.num_attention_heads`: {self.config.num_attention_heads}"
)
self.query = nn.Dense(
self.config.hidden_size,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
)
self.key = nn.Dense(
self.config.hidden_size,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
)
self.value = nn.Dense(
self.config.hidden_size,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
)
def __call__(self, hidden_states, attention_mask, deterministic=True, output_attentions: bool = False):
head_dim = self.config.hidden_size // self.config.num_attention_heads
query_states = self.query(hidden_states).reshape(
hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
)
value_states = self.value(hidden_states).reshape(
hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
)
key_states = self.key(hidden_states).reshape(
hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
)
# Convert the boolean attention mask to an attention bias.
if attention_mask is not None:
# attention mask in the form of attention bias
attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
attention_bias = lax.select(
attention_mask > 0,
jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
jnp.full(attention_mask.shape, -1e10).astype(self.dtype),
)
else:
attention_bias = None
dropout_rng = None
if not deterministic and self.config.attention_probs_dropout_prob > 0.0:
dropout_rng = self.make_rng("dropout")
attn_output = dot_product_attention(
query_states,
key_states,
value_states,
bias=attention_bias,
dropout_rng=dropout_rng,
dropout_rate=self.config.attention_probs_dropout_prob,
broadcast_dropout=True,
deterministic=deterministic,
dtype=self.dtype,
precision=None,
)
outputs = (attn_output.reshape(attn_output.shape[:2] + (-1,)),)
# TODO: at the moment it's not possible to retrieve attn_weights from
# dot_product_attention, but should be in the future -> add functionality then
return outputs
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertSelfOutput with Bert->Roberta
class FlaxRobertaSelfOutput(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.dense = nn.Dense(
self.config.hidden_size,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
dtype=self.dtype,
)
self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
def __call__(self, hidden_states, input_tensor, deterministic: bool = True):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
return hidden_states
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertAttention with Bert->Roberta
class FlaxRobertaAttention(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32
def setup(self):
self.self = FlaxRobertaSelfAttention(self.config, dtype=self.dtype)
self.output = FlaxRobertaSelfOutput(self.config, dtype=self.dtype)
def __call__(self, hidden_states, attention_mask, deterministic=True, output_attentions: bool = False):
# Attention mask comes in as attention_mask.shape == (*batch_sizes, kv_length)
# FLAX expects: attention_mask.shape == (*batch_sizes, 1, 1, kv_length) such that it is broadcastable
# with attn_weights.shape == (*batch_sizes, num_heads, q_length, kv_length)
attn_outputs = self.self(
hidden_states, attention_mask, deterministic=deterministic, output_attentions=output_attentions
)
attn_output = attn_outputs[0]
hidden_states = self.output(attn_output, hidden_states, deterministic=deterministic)
outputs = (hidden_states,)
if output_attentions:
outputs += attn_outputs[1]
return outputs
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertIntermediate with Bert->Roberta
class FlaxRobertaIntermediate(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.dense = nn.Dense(
self.config.intermediate_size,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
dtype=self.dtype,
)
self.activation = ACT2FN[self.config.hidden_act]
def __call__(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = self.activation(hidden_states)
return hidden_states
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertOutput with Bert->Roberta
class FlaxRobertaOutput(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.dense = nn.Dense(
self.config.hidden_size,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
dtype=self.dtype,
)
self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
def __call__(self, hidden_states, attention_output, deterministic: bool = True):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
hidden_states = self.LayerNorm(hidden_states + attention_output)
return hidden_states
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertLayer with Bert->Roberta
class FlaxRobertaLayer(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.attention = FlaxRobertaAttention(self.config, dtype=self.dtype)
self.intermediate = FlaxRobertaIntermediate(self.config, dtype=self.dtype)
self.output = FlaxRobertaOutput(self.config, dtype=self.dtype)
def __call__(self, hidden_states, attention_mask, deterministic: bool = True, output_attentions: bool = False):
attention_outputs = self.attention(
hidden_states, attention_mask, deterministic=deterministic, output_attentions=output_attentions
)
attention_output = attention_outputs[0]
hidden_states = self.intermediate(attention_output)
hidden_states = self.output(hidden_states, attention_output, deterministic=deterministic)
outputs = (hidden_states,)
if output_attentions:
outputs += (attention_outputs[1],)
return outputs
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertLayerCollection with Bert->Roberta
class FlaxRobertaLayerCollection(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.layers = [
FlaxRobertaLayer(self.config, name=str(i), dtype=self.dtype) for i in range(self.config.num_hidden_layers)
]
def __call__(
self,
hidden_states,
attention_mask,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
all_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
for i, layer in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
layer_outputs = layer(hidden_states, attention_mask, deterministic=deterministic)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions += (layer_outputs[1],)
if output_hidden_states:
all_hidden_states += (hidden_states,)
outputs = (hidden_states,)
if not return_dict:
return tuple(v for v in outputs if v is not None)
return FlaxBaseModelOutput(
last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
)
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertEncoder with Bert->Roberta
class FlaxRobertaEncoder(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.layer = FlaxRobertaLayerCollection(self.config, dtype=self.dtype)
def __call__(
self,
hidden_states,
attention_mask,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
return self.layer(
hidden_states,
attention_mask,
deterministic=deterministic,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertPooler with Bert->Roberta
class FlaxRobertaPooler(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
self.dense = nn.Dense(
self.config.hidden_size,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
dtype=self.dtype,
)
def __call__(self, hidden_states):
cls_hidden_state = hidden_states[:, 0]
cls_hidden_state = self.dense(cls_hidden_state)
return nn.tanh(cls_hidden_state)
class FlaxRobertaLMHead(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32
bias_init: Callable[..., np.ndarray] = jax.nn.initializers.zeros
def setup(self):
self.dense = nn.Dense(
self.config.hidden_size,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
)
self.layer_norm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
self.decoder = nn.Dense(
self.config.vocab_size,
dtype=self.dtype,
use_bias=False,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
)
self.bias = self.param("bias", self.bias_init, (self.config.vocab_size,))
def __call__(self, hidden_states, shared_embedding=None):
hidden_states = self.dense(hidden_states)
hidden_states = ACT2FN["gelu"](hidden_states)
hidden_states = self.layer_norm(hidden_states)
if shared_embedding is not None:
hidden_states = self.decoder.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
else:
hidden_states = self.decoder(hidden_states)
hidden_states += self.bias
return hidden_states
class FlaxRobertaClassificationHead(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32
def setup(self):
self.dense = nn.Dense(
self.config.hidden_size,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
)
self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
self.out_proj = nn.Dense(
self.config.num_labels,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(self.config.initializer_range, self.dtype),
)
def __call__(self, hidden_states, deterministic=True):
hidden_states = hidden_states[:, 0, :] # take <s> token (equiv. to [CLS])
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
hidden_states = self.dense(hidden_states)
hidden_states = nn.tanh(hidden_states)
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
hidden_states = self.out_proj(hidden_states)
return hidden_states
class FlaxRobertaPreTrainedModel(FlaxPreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = RobertaConfig
base_model_prefix = "roberta"
module_class: nn.Module = None
def __init__(
self,
config: RobertaConfig,
input_shape: Tuple = (1, 1),
seed: int = 0,
dtype: jnp.dtype = jnp.float32,
**kwargs
):
module = self.module_class(config=config, dtype=dtype, **kwargs)
super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype)
def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
# init input tensors
input_ids = jnp.zeros(input_shape, dtype="i4")
token_type_ids = jnp.ones_like(input_ids)
position_ids = create_position_ids_from_input_ids(input_ids, self.config.pad_token_id)
attention_mask = jnp.ones_like(input_ids)
params_rng, dropout_rng = jax.random.split(rng)
rngs = {"params": params_rng, "dropout": dropout_rng}
return self.module.init(rngs, input_ids, attention_mask, token_type_ids, position_ids)["params"]
@add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
def __call__(
self,
input_ids,
attention_mask=None,
token_type_ids=None,
position_ids=None,
params: dict = None,
dropout_rng: PRNGKey = None,
train: bool = False,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
):
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.return_dict
if output_attentions:
raise NotImplementedError(
"Currently attention scores cannot be returned." "Please set `output_attentions` to False for now."
)
# init input tensors if not passed
if token_type_ids is None:
token_type_ids = jnp.ones_like(input_ids)
if position_ids is None:
position_ids = create_position_ids_from_input_ids(input_ids, self.config.pad_token_id)
if attention_mask is None:
attention_mask = jnp.ones_like(input_ids)
# Handle any PRNG if needed
rngs = {}
if dropout_rng is not None:
rngs["dropout"] = dropout_rng
return self.module.apply(
{"params": params or self.params},
jnp.array(input_ids, dtype="i4"),
jnp.array(attention_mask, dtype="i4"),
jnp.array(token_type_ids, dtype="i4"),
jnp.array(position_ids, dtype="i4"),
not train,
output_attentions,
output_hidden_states,
return_dict,
rngs=rngs,
)
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertModule with Bert->Roberta
class FlaxRobertaModule(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
add_pooling_layer: bool = True
def setup(self):
self.embeddings = FlaxRobertaEmbeddings(self.config, dtype=self.dtype)
self.encoder = FlaxRobertaEncoder(self.config, dtype=self.dtype)
self.pooler = FlaxRobertaPooler(self.config, dtype=self.dtype)
def __call__(
self,
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
hidden_states = self.embeddings(
input_ids, token_type_ids, position_ids, attention_mask, deterministic=deterministic
)
outputs = self.encoder(
hidden_states,
attention_mask,
deterministic=deterministic,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
hidden_states = outputs[0]
pooled = self.pooler(hidden_states) if self.add_pooling_layer else None
if not return_dict:
# if pooled is None, don't return it
if pooled is None:
return (hidden_states,) + outputs[1:]
return (hidden_states, pooled) + outputs[1:]
return FlaxBaseModelOutputWithPooling(
last_hidden_state=hidden_states,
pooler_output=pooled,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
[docs]@add_start_docstrings(
"The bare RoBERTa Model transformer outputting raw hidden-states without any specific head on top.",
ROBERTA_START_DOCSTRING,
)
class FlaxRobertaModel(FlaxRobertaPreTrainedModel):
module_class = FlaxRobertaModule
append_call_sample_docstring(
FlaxRobertaModel, _TOKENIZER_FOR_DOC, _CHECKPOINT_FOR_DOC, FlaxBaseModelOutputWithPooling, _CONFIG_FOR_DOC
)
class FlaxRobertaForMaskedLMModule(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32
def setup(self):
self.roberta = FlaxRobertaModule(config=self.config, add_pooling_layer=False, dtype=self.dtype)
self.lm_head = FlaxRobertaLMHead(config=self.config, dtype=self.dtype)
def __call__(
self,
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
# Model
outputs = self.roberta(
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic=deterministic,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
hidden_states = outputs[0]
if self.config.tie_word_embeddings:
shared_embedding = self.roberta.variables["params"]["embeddings"]["word_embeddings"]["embedding"]
else:
shared_embedding = None
# Compute the prediction scores
logits = self.lm_head(hidden_states, shared_embedding=shared_embedding)
if not return_dict:
return (logits,) + outputs[1:]
return FlaxMaskedLMOutput(
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
append_call_sample_docstring(
FlaxRobertaForMaskedLM,
_TOKENIZER_FOR_DOC,
_CHECKPOINT_FOR_DOC,
FlaxBaseModelOutputWithPooling,
_CONFIG_FOR_DOC,
mask="<mask>",
)
class FlaxRobertaForSequenceClassificationModule(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32
def setup(self):
self.roberta = FlaxRobertaModule(config=self.config, dtype=self.dtype, add_pooling_layer=False)
self.classifier = FlaxRobertaClassificationHead(config=self.config, dtype=self.dtype)
def __call__(
self,
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
# Model
outputs = self.roberta(
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic=deterministic,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
logits = self.classifier(sequence_output, deterministic=deterministic)
if not return_dict:
return (logits,) + outputs[1:]
return FlaxSequenceClassifierOutput(
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
[docs]@add_start_docstrings(
"""
Roberta Model transformer with a sequence classification/regression head on top (a linear layer on top of the
pooled output) e.g. for GLUE tasks.
""",
ROBERTA_START_DOCSTRING,
)
class FlaxRobertaForSequenceClassification(FlaxRobertaPreTrainedModel):
module_class = FlaxRobertaForSequenceClassificationModule
append_call_sample_docstring(
FlaxRobertaForSequenceClassification,
_TOKENIZER_FOR_DOC,
_CHECKPOINT_FOR_DOC,
FlaxSequenceClassifierOutput,
_CONFIG_FOR_DOC,
)
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertForMultipleChoiceModule with Bert->Roberta, with self.bert->self.roberta
class FlaxRobertaForMultipleChoiceModule(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32
def setup(self):
self.roberta = FlaxRobertaModule(config=self.config, dtype=self.dtype)
self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
self.classifier = nn.Dense(1, dtype=self.dtype)
def __call__(
self,
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
num_choices = input_ids.shape[1]
input_ids = input_ids.reshape(-1, input_ids.shape[-1]) if input_ids is not None else None
attention_mask = attention_mask.reshape(-1, attention_mask.shape[-1]) if attention_mask is not None else None
token_type_ids = token_type_ids.reshape(-1, token_type_ids.shape[-1]) if token_type_ids is not None else None
position_ids = position_ids.reshape(-1, position_ids.shape[-1]) if position_ids is not None else None
# Model
outputs = self.roberta(
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic=deterministic,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
pooled_output = outputs[1]
pooled_output = self.dropout(pooled_output, deterministic=deterministic)
logits = self.classifier(pooled_output)
reshaped_logits = logits.reshape(-1, num_choices)
if not return_dict:
return (reshaped_logits,) + outputs[2:]
return FlaxMultipleChoiceModelOutput(
logits=reshaped_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
overwrite_call_docstring(
FlaxRobertaForMultipleChoice, ROBERTA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
)
append_call_sample_docstring(
FlaxRobertaForMultipleChoice,
_TOKENIZER_FOR_DOC,
_CHECKPOINT_FOR_DOC,
FlaxMultipleChoiceModelOutput,
_CONFIG_FOR_DOC,
)
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertForTokenClassificationModule with Bert->Roberta, with self.bert->self.roberta
class FlaxRobertaForTokenClassificationModule(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32
def setup(self):
self.roberta = FlaxRobertaModule(config=self.config, dtype=self.dtype, add_pooling_layer=False)
self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
self.classifier = nn.Dense(self.config.num_labels, dtype=self.dtype)
def __call__(
self,
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
# Model
outputs = self.roberta(
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic=deterministic,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
hidden_states = outputs[0]
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
logits = self.classifier(hidden_states)
if not return_dict:
return (logits,) + outputs[1:]
return FlaxTokenClassifierOutput(
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
[docs]@add_start_docstrings(
"""
Roberta Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
Named-Entity-Recognition (NER) tasks.
""",
ROBERTA_START_DOCSTRING,
)
class FlaxRobertaForTokenClassification(FlaxRobertaPreTrainedModel):
module_class = FlaxRobertaForTokenClassificationModule
append_call_sample_docstring(
FlaxRobertaForTokenClassification,
_TOKENIZER_FOR_DOC,
_CHECKPOINT_FOR_DOC,
FlaxTokenClassifierOutput,
_CONFIG_FOR_DOC,
)
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertForQuestionAnsweringModule with Bert->Roberta, with self.bert->self.roberta
class FlaxRobertaForQuestionAnsweringModule(nn.Module):
config: RobertaConfig
dtype: jnp.dtype = jnp.float32
def setup(self):
self.roberta = FlaxRobertaModule(config=self.config, dtype=self.dtype, add_pooling_layer=False)
self.qa_outputs = nn.Dense(self.config.num_labels, dtype=self.dtype)
def __call__(
self,
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
return_dict: bool = True,
):
# Model
outputs = self.roberta(
input_ids,
attention_mask,
token_type_ids,
position_ids,
deterministic=deterministic,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
hidden_states = outputs[0]
logits = self.qa_outputs(hidden_states)
start_logits, end_logits = logits.split(self.config.num_labels, axis=-1)
start_logits = start_logits.squeeze(-1)
end_logits = end_logits.squeeze(-1)
if not return_dict:
return (start_logits, end_logits) + outputs[1:]
return FlaxQuestionAnsweringModelOutput(
start_logits=start_logits,
end_logits=end_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
[docs]@add_start_docstrings(
"""
Roberta Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
""",
ROBERTA_START_DOCSTRING,
)
class FlaxRobertaForQuestionAnswering(FlaxRobertaPreTrainedModel):
module_class = FlaxRobertaForQuestionAnsweringModule
append_call_sample_docstring(
FlaxRobertaForQuestionAnswering,
_TOKENIZER_FOR_DOC,
_CHECKPOINT_FOR_DOC,
FlaxQuestionAnsweringModelOutput,
_CONFIG_FOR_DOC,
)