# coding=utf-8
# Copyright 2018 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, 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_utils import ACT2FN, FlaxPreTrainedModel
from ...utils import logging
from .configuration_roberta import RobertaConfig
logger = logging.get_logger(__name__)
_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")
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.nn.Module
<https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html>`__ subclass. Use it as a regular Flax
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.PreTrainedModel.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.FlaxBertLayerNorm with Bert->Roberta
class FlaxRobertaLayerNorm(nn.Module):
"""
Layer normalization (https://arxiv.org/abs/1607.06450). Operates on the last axis of the input data.
"""
hidden_size: int
epsilon: float = 1e-6
dtype: jnp.dtype = jnp.float32
use_bias: bool = True
scale: bool = True
scale_init: Callable[..., np.ndarray] = jax.nn.initializers.ones
bias_init: Callable[..., np.ndarray] = jax.nn.initializers.zeros
def setup(self):
self.weight = self.param("weight", self.scale_init, (self.hidden_size,))
self.bias = self.param("bias", self.scale_init, (self.hidden_size,))
def __call__(self, x):
"""
Applies layer normalization on the input. It normalizes the activations of the layer for each given example in
a batch independently, rather than across a batch like Batch Normalization. i.e. applies a transformation that
maintains the mean activation within each example close to 0 and the activation standard deviation close to 1
Args:
x: the inputs
Returns:
Normalized inputs (the same shape as inputs).
"""
mean = jnp.mean(x, axis=-1, keepdims=True)
mean2 = jnp.mean(jax.lax.square(x), axis=-1, keepdims=True)
var = mean2 - jax.lax.square(mean)
mul = jax.lax.rsqrt(var + self.epsilon)
if self.scale:
mul = mul * jnp.asarray(self.weight)
y = (x - mean) * mul
if self.use_bias:
y = y + jnp.asarray(self.bias)
return y
# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertEmbedding with Bert->Roberta
class FlaxRobertaEmbedding(nn.Module):
"""
Specify a new class for doing the embedding stuff as Flax's one use 'embedding' for the parameter name and PyTorch
use 'weight'
"""
vocab_size: int
hidden_size: int
initializer_range: float
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
def setup(self):
init_fn: Callable[..., np.ndarray] = jax.nn.initializers.normal(stddev=self.initializer_range)
self.embeddings = self.param("weight", init_fn, (self.vocab_size, self.hidden_size))
def __call__(self, input_ids):
return jnp.take(self.embeddings, input_ids, axis=0)
# 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 = FlaxRobertaEmbedding(
self.config.vocab_size,
self.config.hidden_size,
initializer_range=self.config.initializer_range,
dtype=self.dtype,
)
self.position_embeddings = FlaxRobertaEmbedding(
self.config.max_position_embeddings,
self.config.hidden_size,
initializer_range=self.config.initializer_range,
dtype=self.dtype,
)
self.token_type_embeddings = FlaxRobertaEmbedding(
self.config.type_vocab_size,
self.config.hidden_size,
initializer_range=self.config.initializer_range,
dtype=self.dtype,
)
self.LayerNorm = FlaxRobertaLayerNorm(hidden_size=self.config.hidden_size, 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(jnp.atleast_2d(input_ids.astype("i4")))
position_embeds = self.position_embeddings(jnp.atleast_2d(position_ids.astype("i4")))
token_type_embeddings = self.token_type_embeddings(jnp.atleast_2d(token_type_ids.astype("i4")))
# Sum all embeddings
hidden_states = inputs_embeds + jnp.broadcast_to(position_embeds, inputs_embeds.shape) + token_type_embeddings
# 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):
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.dropout_rate > 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,
)
return attn_output.reshape(attn_output.shape[:2] + (-1,))
# 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 = FlaxRobertaLayerNorm(hidden_size=self.config.hidden_size)
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):
# 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_output = self.self(hidden_states, attention_mask, deterministic=deterministic)
hidden_states = self.output(attn_output, hidden_states, deterministic=deterministic)
return hidden_states
# 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 = FlaxRobertaLayerNorm(hidden_size=self.config.hidden_size, 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):
attention_output = self.attention(hidden_states, attention_mask, deterministic=deterministic)
hidden_states = self.intermediate(attention_output)
hidden_states = self.output(hidden_states, attention_output, deterministic=deterministic)
return hidden_states
# 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):
for layer in self.layers:
hidden_states = layer(hidden_states, attention_mask, deterministic=deterministic)
return hidden_states
# 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):
return self.layer(hidden_states, attention_mask, deterministic=deterministic)
# 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 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"
def init(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
input_ids, attention_mask, token_type_ids, position_ids = self._check_inputs(
jnp.zeros(input_shape, dtype="i4"), None, None, None
)
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"]
def _check_inputs(self, input_ids, attention_mask, token_type_ids, position_ids):
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)
return input_ids, attention_mask, token_type_ids, position_ids
[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):
"""
The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
cross-attention is added between the self-attention layers, following the architecture described in `Attention is
all you need`_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
Kaiser and Illia Polosukhin.
"""
def __init__(
self,
config: RobertaConfig,
input_shape: Tuple = (1, 1),
seed: int = 0,
dtype: jnp.dtype = jnp.float32,
**kwargs
):
module = FlaxRobertaModule(config, dtype=dtype, **kwargs)
super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype)
[docs] @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
def __call__(
self,
input_ids,
token_type_ids=None,
attention_mask=None,
position_ids=None,
params: dict = None,
dropout_rng: PRNGKey = None,
train: bool = False,
):
input_ids, attention_mask, token_type_ids, position_ids = self._check_inputs(
input_ids, attention_mask, token_type_ids, position_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,
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):
hidden_states = self.embeddings(
input_ids, token_type_ids, position_ids, attention_mask, deterministic=deterministic
)
hidden_states = self.encoder(hidden_states, attention_mask, deterministic=deterministic)
if not self.add_pooling_layer:
return hidden_states
pooled = self.pooler(hidden_states)
return hidden_states, pooled