Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2020 The Trax 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"""Reformer model configuration"""17 18from ...configuration_utils import PretrainedConfig19from ...utils import logging20 21 22logger = logging.get_logger(__name__)23 24 25class ReformerConfig(PretrainedConfig):26    r"""27    This is the configuration class to store the configuration of a [`ReformerModel`]. It is used to instantiate a28    Reformer model according to the specified arguments, defining the model architecture. Instantiating a configuration29    with the defaults will yield a similar configuration to that of the ReFormer30    [google/reformer-crime-and-punishment](https://huggingface.co/google/reformer-crime-and-punishment) architecture.31 32    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the33    documentation from [`PretrainedConfig`] for more information.34 35    Args:36        attention_head_size (`int`, *optional*, defaults to 64):37            Dimensionality of the projected key, query and value vectors38        attn_layers (`list[str]`, *optional*, defaults to `["local", "lsh", "local", "lsh", "local", "lsh"]`):39            List of attention layer types in ascending order. It can be chosen between a LSHSelfAttention layer40            (`"lsh"`) and a LocalSelfAttention layer (`"local"`).41 42            For more information on LSHSelfAttention layer, see [LSH Self Attention](reformer#lsh-self-attention). For43            more information on LocalSelfAttention layer, see [Local Self Attention](reformer#local-self-attention).44        axial_pos_embds (`bool`, *optional*, defaults to `True`):45            Whether or not to use axial position embeddings. For more information on how axial position embeddings46            work, see [Axial Position Encodings](reformer#axial-positional-encodings).47        axial_norm_std (`float`, *optional*, defaults to 1.0):48            The standard deviation of the normal_initializer for initializing the weight matrices of the axial49            positional encodings.50        axial_pos_shape (`list[int]`, *optional*, defaults to `[64, 64]`):51            The position dims of the axial position encodings. During training, the product of the position dims has to52            be equal to the sequence length.53 54            For more information on how axial position embeddings work, see [Axial Position55            Encodings](reformer#axial-positional-encodings).56        axial_pos_embds_dim (`list[int]`, *optional*, defaults to `[64, 192]`):57            The embedding dims of the axial position encodings. The sum of the embedding dims has to be equal to the58            hidden size.59 60            For more information on how axial position embeddings work, see [Axial Position61            Encodings](reformer#axial-positional-encodings).62        chunk_size_lm_head (`int`, *optional*, defaults to 0):63            The chunk size of the final language model feed forward head layer. A chunk size of 0 means that the feed64            forward layer is not chunked. A chunk size of n means that the feed forward layer processes n <65            sequence_length embeddings at a time.66 67            For more information on feed forward chunking, see [How does Feed Forward Chunking68            work?](../glossary#feed-forward-chunking).69        eos_token_id (`int`, *optional*, defaults to 2):70            The token id for the end-of-sentence token.71        feed_forward_size (`int`, *optional*, defaults to 512):72            Dimensionality of the feed_forward layer in the residual attention block.73        hash_seed (`int`, *optional*):74            Seed that can be used to make local sensitive hashing in `LSHSelfAttention` deterministic. This should only75            be set for testing purposed. For evaluation and training purposes `hash_seed` should be left as `None` to76            ensure fully random rotations in local sensitive hashing scheme.77        hidden_act (`str` or `Callable`, *optional*, defaults to `"relu"`):78            The non-linear activation function (function or string) in the feed forward layer in the residual attention79            block. If string, `"gelu"`, `"relu"`, `"silu"` and `"gelu_new"` are supported.80        hidden_dropout_prob (`float`, *optional*, defaults to 0.05):81            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.82        hidden_size (`int`, *optional*, defaults to 256):83            Dimensionality of the output hidden states of the residual attention blocks.84        initializer_range (`float`, *optional*, defaults to 0.02):85            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.86        is_decoder (`bool`, *optional*, defaults to `False`):87            Whether or not to use a causal mask in addition to the `attention_mask` passed to [`ReformerModel`]. When88            using the Reformer for causal language modeling, this argument should be set to `True`.89        layer_norm_eps (`float`, *optional*, defaults to 1e-12):90            The epsilon used by the layer normalization layers.91        local_chunk_length (`int`, *optional*, defaults to 64):92            Length of chunk which attends to itself in `LocalSelfAttention`. Chunking reduces memory complexity from93            sequence length x sequence length (self attention) to chunk length x chunk length x sequence length / chunk94            length (chunked self attention).95        local_num_chunks_before (`int`, *optional*, defaults to 1):96            Number of previous neighbouring chunks to attend to in `LocalSelfAttention` layer to itself.97        local_num_chunks_after (`int`, *optional*, defaults to 0):98            Number of following neighbouring chunks to attend to in `LocalSelfAttention` layer in addition to itself.99        local_attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):100            The dropout ratio for the attention probabilities in `LocalSelfAttention`.101        lsh_attn_chunk_length (`int`, *optional*, defaults to 64):102            Length of chunk which attends to itself in `LSHSelfAttention`. Chunking reduces memory complexity from103            sequence length x sequence length (self attention) to chunk length x chunk length x sequence length / chunk104            length (chunked self attention).105        lsh_num_chunks_before (`int`, *optional*, defaults to 1):106            Number of previous neighbouring chunks to attend to in `LSHSelfAttention` layer to itself.107        lsh_num_chunks_after (`int`, *optional*, defaults to 0):108            Number of following neighbouring chunks to attend to in `LSHSelfAttention` layer to itself.109        lsh_attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):110            The dropout ratio for the attention probabilities in `LSHSelfAttention`.111        max_position_embeddings (`int`, *optional*, defaults to 4096):112            The maximum sequence length that this model might ever be used with. Typically set this to something large113            just in case (e.g., 512 or 1024 or 2048).114        num_attention_heads (`int`, *optional*, defaults to 12):115            Number of attention heads for each attention layer in the Transformer encoder.116        num_buckets (`int` or `list[int]`, *optional*):117            Number of buckets, the key query vectors can be "hashed into" using the locality sensitive hashing scheme.118            Each query key vector is hashed into a hash in `1, ..., num_buckets`. The number of buckets can also be119            factorized into a list for improved memory complexity. In this case, each query key vector is hashed into a120            hash in `1-1, 1-2, ..., num_buckets[0]-1, ..., num_buckets[0]-num_buckets[1]` if `num_buckets` is121            factorized into two factors. The number of buckets (or the product the factors) should approximately equal122            sequence length / lsh_chunk_length. If `num_buckets` not set, a good value is calculated on the fly.123        num_hashes (`int`, *optional*, defaults to 1):124            Number of hashing rounds (e.g., number of random rotations) in Local Sensitive Hashing scheme. The higher125            `num_hashes`, the more accurate the `LSHSelfAttention` becomes, but also the more memory and time intensive126            the hashing becomes.127        pad_token_id (`int`, *optional*, defaults to 0):128            The token id for the padding token.129        vocab_size (`int`, *optional*, defaults to 320):\130            Vocabulary size of the Reformer model. Defines the number of different tokens that can be represented by131            the `inputs_ids` passed when calling [`ReformerModel`].132        tie_word_embeddings (`bool`, *optional*, defaults to `False`):133            Whether to tie input and output embeddings.134        use_cache (`bool`, *optional*, defaults to `True`):135            Whether or not the model should return the last key/values attentions (not used by all models).136        classifier_dropout (`float`, *optional*):137            The dropout ratio for the classification head.138 139    Examples:140 141    ```python142    >>> from transformers import ReformerConfig, ReformerModel143 144    >>> # Initializing a Reformer configuration145    >>> configuration = ReformerConfig()146 147    >>> # Initializing a Reformer model (with random weights)148    >>> model = ReformerModel(configuration)149 150    >>> # Accessing the model configuration151    >>> configuration = model.config152    ```153"""154 155    model_type = "reformer"156    keys_to_ignore_at_inference = ["past_buckets_states"]157    attribute_map = {}158 159    def __init__(160        self,161        attention_head_size=64,162        attn_layers=["local", "lsh", "local", "lsh", "local", "lsh"],163        axial_norm_std=1.0,164        axial_pos_embds=True,165        axial_pos_shape=[64, 64],166        axial_pos_embds_dim=[64, 192],167        chunk_size_lm_head=0,168        eos_token_id=2,169        feed_forward_size=512,170        hash_seed=None,171        hidden_act="relu",172        hidden_dropout_prob=0.05,173        hidden_size=256,174        initializer_range=0.02,175        is_decoder=False,176        layer_norm_eps=1e-12,177        local_num_chunks_before=1,178        local_num_chunks_after=0,179        local_attention_probs_dropout_prob=0.05,180        local_attn_chunk_length=64,181        lsh_attn_chunk_length=64,182        lsh_attention_probs_dropout_prob=0.0,183        lsh_num_chunks_before=1,184        lsh_num_chunks_after=0,185        max_position_embeddings=4096,186        num_attention_heads=12,187        num_buckets=None,188        num_hashes=1,189        pad_token_id=0,190        vocab_size=320,191        tie_word_embeddings=False,192        use_cache=True,193        classifier_dropout=None,194        **kwargs,195    ):196        self.hash_seed = hash_seed197        self.vocab_size = vocab_size198        self.attention_head_size = attention_head_size199        self.hidden_size = hidden_size200        self.num_attention_heads = num_attention_heads201        self.num_hashes = num_hashes202        self.num_hidden_layers = len(attn_layers)203        self.num_buckets = tuple(num_buckets) if isinstance(num_buckets, list) else num_buckets204        self.lsh_attn_chunk_length = lsh_attn_chunk_length205        self.local_attn_chunk_length = local_attn_chunk_length206        self.lsh_num_chunks_after = lsh_num_chunks_after207        self.lsh_num_chunks_before = lsh_num_chunks_before208        self.local_num_chunks_after = local_num_chunks_after209        self.local_num_chunks_before = local_num_chunks_before210        self.hidden_act = hidden_act211        self.feed_forward_size = feed_forward_size212        self.hidden_dropout_prob = hidden_dropout_prob213        self.lsh_attention_probs_dropout_prob = lsh_attention_probs_dropout_prob214        self.local_attention_probs_dropout_prob = local_attention_probs_dropout_prob215        self.max_position_embeddings = max_position_embeddings216        self.initializer_range = initializer_range217        self.layer_norm_eps = layer_norm_eps218        self.axial_pos_embds = axial_pos_embds219        self.axial_pos_shape = tuple(axial_pos_shape)220        self.axial_pos_embds_dim = tuple(axial_pos_embds_dim)221        self.axial_norm_std = axial_norm_std222        self.chunk_size_lm_head = chunk_size_lm_head223        self.attn_layers = attn_layers224        self.use_cache = use_cache225        self.classifier_dropout = classifier_dropout226        super().__init__(227            pad_token_id=pad_token_id,228            eos_token_id=eos_token_id,229            is_decoder=is_decoder,230            tie_word_embeddings=tie_word_embeddings,231            **kwargs,232        )233 234 235__all__ = ["ReformerConfig"]236