Upload configuration_roberta.py
Browse files- configuration_roberta.py +151 -0
configuration_roberta.py
ADDED
@@ -0,0 +1,151 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
|
3 |
+
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
4 |
+
#
|
5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
6 |
+
# you may not use this file except in compliance with the License.
|
7 |
+
# You may obtain a copy of the License at
|
8 |
+
#
|
9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
10 |
+
#
|
11 |
+
# Unless required by applicable law or agreed to in writing, software
|
12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
14 |
+
# See the License for the specific language governing permissions and
|
15 |
+
# limitations under the License.
|
16 |
+
""" RoBERTa configuration"""
|
17 |
+
from collections import OrderedDict
|
18 |
+
from typing import Mapping
|
19 |
+
|
20 |
+
from transformers import PretrainedConfig
|
21 |
+
from transformers.onnx import OnnxConfig
|
22 |
+
from transformers.utils import logging
|
23 |
+
|
24 |
+
|
25 |
+
logger = logging.get_logger(__name__)
|
26 |
+
|
27 |
+
|
28 |
+
class RobertaConfig(PretrainedConfig):
|
29 |
+
r"""
|
30 |
+
This is the configuration class to store the configuration of a [`RobertaModel`] or a [`TFRobertaModel`]. It is
|
31 |
+
used to instantiate a RoBERTa model according to the specified arguments, defining the model architecture.
|
32 |
+
Instantiating a configuration with the defaults will yield a similar configuration to that of the RoBERTa
|
33 |
+
[FacebookAI/roberta-base](https://huggingface.co/FacebookAI/roberta-base) architecture.
|
34 |
+
|
35 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
36 |
+
documentation from [`PretrainedConfig`] for more information.
|
37 |
+
|
38 |
+
|
39 |
+
Args:
|
40 |
+
vocab_size (`int`, *optional*, defaults to 50265):
|
41 |
+
Vocabulary size of the RoBERTa model. Defines the number of different tokens that can be represented by the
|
42 |
+
`inputs_ids` passed when calling [`RobertaModel`] or [`TFRobertaModel`].
|
43 |
+
hidden_size (`int`, *optional*, defaults to 768):
|
44 |
+
Dimensionality of the encoder layers and the pooler layer.
|
45 |
+
num_hidden_layers (`int`, *optional*, defaults to 12):
|
46 |
+
Number of hidden layers in the Transformer encoder.
|
47 |
+
num_attention_heads (`int`, *optional*, defaults to 12):
|
48 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
49 |
+
intermediate_size (`int`, *optional*, defaults to 3072):
|
50 |
+
Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
|
51 |
+
hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
|
52 |
+
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
|
53 |
+
`"relu"`, `"silu"` and `"gelu_new"` are supported.
|
54 |
+
hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
|
55 |
+
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
|
56 |
+
attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
|
57 |
+
The dropout ratio for the attention probabilities.
|
58 |
+
max_position_embeddings (`int`, *optional*, defaults to 512):
|
59 |
+
The maximum sequence length that this model might ever be used with. Typically set this to something large
|
60 |
+
just in case (e.g., 512 or 1024 or 2048).
|
61 |
+
type_vocab_size (`int`, *optional*, defaults to 2):
|
62 |
+
The vocabulary size of the `token_type_ids` passed when calling [`RobertaModel`] or [`TFRobertaModel`].
|
63 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
64 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
65 |
+
layer_norm_eps (`float`, *optional*, defaults to 1e-12):
|
66 |
+
The epsilon used by the layer normalization layers.
|
67 |
+
position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
|
68 |
+
Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
|
69 |
+
positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
|
70 |
+
[Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
|
71 |
+
For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
|
72 |
+
with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
|
73 |
+
is_decoder (`bool`, *optional*, defaults to `False`):
|
74 |
+
Whether the model is used as a decoder or not. If `False`, the model is used as an encoder.
|
75 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
76 |
+
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
77 |
+
relevant if `config.is_decoder=True`.
|
78 |
+
classifier_dropout (`float`, *optional*):
|
79 |
+
The dropout ratio for the classification head.
|
80 |
+
|
81 |
+
Examples:
|
82 |
+
|
83 |
+
```python
|
84 |
+
>>> from transformers import RobertaConfig, RobertaModel
|
85 |
+
|
86 |
+
>>> # Initializing a RoBERTa configuration
|
87 |
+
>>> configuration = RobertaConfig()
|
88 |
+
|
89 |
+
>>> # Initializing a model (with random weights) from the configuration
|
90 |
+
>>> model = RobertaModel(configuration)
|
91 |
+
|
92 |
+
>>> # Accessing the model configuration
|
93 |
+
>>> configuration = model.config
|
94 |
+
```"""
|
95 |
+
|
96 |
+
model_type = "roberta"
|
97 |
+
|
98 |
+
def __init__(
|
99 |
+
self,
|
100 |
+
vocab_size=50265,
|
101 |
+
hidden_size=768,
|
102 |
+
num_hidden_layers=12,
|
103 |
+
num_attention_heads=12,
|
104 |
+
intermediate_size=3072,
|
105 |
+
hidden_act="gelu",
|
106 |
+
hidden_dropout_prob=0.1,
|
107 |
+
attention_probs_dropout_prob=0.1,
|
108 |
+
max_position_embeddings=512,
|
109 |
+
type_vocab_size=2,
|
110 |
+
initializer_range=0.02,
|
111 |
+
layer_norm_eps=1e-12,
|
112 |
+
pad_token_id=1,
|
113 |
+
bos_token_id=0,
|
114 |
+
eos_token_id=2,
|
115 |
+
position_embedding_type="absolute",
|
116 |
+
use_cache=True,
|
117 |
+
classifier_dropout=None,
|
118 |
+
**kwargs,
|
119 |
+
):
|
120 |
+
super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
|
121 |
+
|
122 |
+
self.vocab_size = vocab_size
|
123 |
+
self.hidden_size = hidden_size
|
124 |
+
self.num_hidden_layers = num_hidden_layers
|
125 |
+
self.num_attention_heads = num_attention_heads
|
126 |
+
self.hidden_act = hidden_act
|
127 |
+
self.intermediate_size = intermediate_size
|
128 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
129 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
130 |
+
self.max_position_embeddings = max_position_embeddings
|
131 |
+
self.type_vocab_size = type_vocab_size
|
132 |
+
self.initializer_range = initializer_range
|
133 |
+
self.layer_norm_eps = layer_norm_eps
|
134 |
+
self.position_embedding_type = position_embedding_type
|
135 |
+
self.use_cache = use_cache
|
136 |
+
self.classifier_dropout = classifier_dropout
|
137 |
+
|
138 |
+
|
139 |
+
class RobertaOnnxConfig(OnnxConfig):
|
140 |
+
@property
|
141 |
+
def inputs(self) -> Mapping[str, Mapping[int, str]]:
|
142 |
+
if self.task == "multiple-choice":
|
143 |
+
dynamic_axis = {0: "batch", 1: "choice", 2: "sequence"}
|
144 |
+
else:
|
145 |
+
dynamic_axis = {0: "batch", 1: "sequence"}
|
146 |
+
return OrderedDict(
|
147 |
+
[
|
148 |
+
("input_ids", dynamic_axis),
|
149 |
+
("attention_mask", dynamic_axis),
|
150 |
+
]
|
151 |
+
)
|