Aluode/PerceptionLabPortable

1# coding=utf-82# 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 at8#9#     http://www.apache.org/licenses/LICENSE-2.010#11# Unless required by applicable law or agreed to in writing, software12# 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 and15# limitations under the License.16"""XLM-RoBERTa configuration"""17 18from collections import OrderedDict19from collections.abc import Mapping20 21from ...configuration_utils import PretrainedConfig22from ...onnx import OnnxConfig23from ...utils import logging24 25 26logger = logging.get_logger(__name__)27 28 29class XLMRobertaConfig(PretrainedConfig):30    r"""31    This is the configuration class to store the configuration of a [`XLMRobertaModel`] or a [`TFXLMRobertaModel`]. It32    is used to instantiate a XLM-RoBERTa model according to the specified arguments, defining the model architecture.33    Instantiating a configuration with the defaults will yield a similar configuration to that of the XLMRoBERTa34    [FacebookAI/xlm-roberta-base](https://huggingface.co/FacebookAI/xlm-roberta-base) architecture.35 36    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the37    documentation from [`PretrainedConfig`] for more information.38 39 40    Args:41        vocab_size (`int`, *optional*, defaults to 30522):42            Vocabulary size of the XLM-RoBERTa model. Defines the number of different tokens that can be represented by43            the `inputs_ids` passed when calling [`XLMRobertaModel`] or [`TFXLMRobertaModel`].44        hidden_size (`int`, *optional*, defaults to 768):45            Dimensionality of the encoder layers and the pooler layer.46        num_hidden_layers (`int`, *optional*, defaults to 12):47            Number of hidden layers in the Transformer encoder.48        num_attention_heads (`int`, *optional*, defaults to 12):49            Number of attention heads for each attention layer in the Transformer encoder.50        intermediate_size (`int`, *optional*, defaults to 3072):51            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.52        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):53            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,54            `"relu"`, `"silu"` and `"gelu_new"` are supported.55        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):56            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.57        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):58            The dropout ratio for the attention probabilities.59        max_position_embeddings (`int`, *optional*, defaults to 512):60            The maximum sequence length that this model might ever be used with. Typically set this to something large61            just in case (e.g., 512 or 1024 or 2048).62        type_vocab_size (`int`, *optional*, defaults to 2):63            The vocabulary size of the `token_type_ids` passed when calling [`XLMRobertaModel`] or64            [`TFXLMRobertaModel`].65        initializer_range (`float`, *optional*, defaults to 0.02):66            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.67        layer_norm_eps (`float`, *optional*, defaults to 1e-12):68            The epsilon used by the layer normalization layers.69        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):70            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For71            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to72            [Self-Attention with Relative Position Representations (Shaw et al.)](https://huggingface.co/papers/1803.02155).73            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models74            with Better Relative Position Embeddings (Huang et al.)](https://huggingface.co/papers/2009.13658).75        is_decoder (`bool`, *optional*, defaults to `False`):76            Whether the model is used as a decoder or not. If `False`, the model is used as an encoder.77        use_cache (`bool`, *optional*, defaults to `True`):78            Whether or not the model should return the last key/values attentions (not used by all models). Only79            relevant if `config.is_decoder=True`.80        classifier_dropout (`float`, *optional*):81            The dropout ratio for the classification head.82 83    Examples:84 85    ```python86    >>> from transformers import XLMRobertaConfig, XLMRobertaModel87 88    >>> # Initializing a XLM-RoBERTa FacebookAI/xlm-roberta-base style configuration89    >>> configuration = XLMRobertaConfig()90 91    >>> # Initializing a model (with random weights) from the FacebookAI/xlm-roberta-base style configuration92    >>> model = XLMRobertaModel(configuration)93 94    >>> # Accessing the model configuration95    >>> configuration = model.config96    ```"""97 98    model_type = "xlm-roberta"99 100    def __init__(101        self,102        vocab_size=30522,103        hidden_size=768,104        num_hidden_layers=12,105        num_attention_heads=12,106        intermediate_size=3072,107        hidden_act="gelu",108        hidden_dropout_prob=0.1,109        attention_probs_dropout_prob=0.1,110        max_position_embeddings=512,111        type_vocab_size=2,112        initializer_range=0.02,113        layer_norm_eps=1e-12,114        pad_token_id=1,115        bos_token_id=0,116        eos_token_id=2,117        position_embedding_type="absolute",118        use_cache=True,119        classifier_dropout=None,120        **kwargs,121    ):122        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)123 124        self.vocab_size = vocab_size125        self.hidden_size = hidden_size126        self.num_hidden_layers = num_hidden_layers127        self.num_attention_heads = num_attention_heads128        self.hidden_act = hidden_act129        self.intermediate_size = intermediate_size130        self.hidden_dropout_prob = hidden_dropout_prob131        self.attention_probs_dropout_prob = attention_probs_dropout_prob132        self.max_position_embeddings = max_position_embeddings133        self.type_vocab_size = type_vocab_size134        self.initializer_range = initializer_range135        self.layer_norm_eps = layer_norm_eps136        self.position_embedding_type = position_embedding_type137        self.use_cache = use_cache138        self.classifier_dropout = classifier_dropout139 140 141# Copied from transformers.models.roberta.configuration_roberta.RobertaOnnxConfig with Roberta->XLMRoberta142class XLMRobertaOnnxConfig(OnnxConfig):143    @property144    def inputs(self) -> Mapping[str, Mapping[int, str]]:145        if self.task == "multiple-choice":146            dynamic_axis = {0: "batch", 1: "choice", 2: "sequence"}147        else:148            dynamic_axis = {0: "batch", 1: "sequence"}149        return OrderedDict(150            [151                ("input_ids", dynamic_axis),152                ("attention_mask", dynamic_axis),153            ]154        )155 156 157__all__ = ["XLMRobertaConfig", "XLMRobertaOnnxConfig"]158