Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2025 Google LLC and HuggingFace Inc. team.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15"""TimesFM model configuration"""16 17from ...configuration_utils import PretrainedConfig18from ...utils import logging19 20 21logger = logging.get_logger(__name__)22 23 24class TimesFmConfig(PretrainedConfig):25    r"""26    This is the configuration class to store the configuration of a [`TimesFmModelForPrediction`] or a [`TFTimesFmModel`]. It is used to27    instantiate a TimesFM model according to the specified arguments, defining the model architecture. Instantiating a28    configuration with the defaults will yield a similar configuration to that of the TimesFM29    [google/timesfm-2.0-500m-pytorch](https://huggingface.co/google/timesfm-2.0-500m-pytorch) architecture.30 31    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the32    documentation from [`PretrainedConfig`] for more information.33 34    Arguments:35        patch_length (`int`, *optional*, defaults to 32):36            The length of one patch in the input sequence.37        context_length (`int`, *optional*, defaults to 512):38            The length of the input context.39        horizon_length (`int`, *optional*, defaults to 128):40            The length of the prediction horizon.41        freq_size (`int`, *optional*, defaults to 3):42            The number of frequency embeddings.43        num_hidden_layers (`int`, *optional*, defaults to 50):44            Number of Transformer layers.45        hidden_size (`int`, *optional*, defaults to 1280):46            Size of the hidden layers in the feed-forward networks.47        intermediate_size (`int`, *optional*, defaults to 1280):48            Dimension of the MLP representations.49        head_dim (`int`, *optional*, defaults to 80):50            Size of the key, query, value projections per attention head. The `inner_dim` of the projection layer will51            be defined as `num_attention_heads * head_dim`.52        num_attention_heads (`int`, *optional*, defaults to 16):53            Number of attention heads for each attention layer in the Transformer encoder.54        tolerance (`float`, *optional*, defaults to 1e-06):55            The tolerance for the quantile loss.56        rms_norm_eps (`float`, *optional*, defaults to 1e-06):57            The epsilon used by the RMS normalization layers.58        quantiles (`list[float]`, *optional*, defaults to `[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]`):59            The quantiles to predict.60        pad_val (`float`, *optional*, defaults to 1123581321.0):61            The value used to pad the predictions.62        attention_dropout (`float`, *optional*, defaults to 0.0):63            The dropout probability for the attention scores.64        use_positional_embedding (`bool`, *optional*, defaults to `False`):65            Whether to add positional embeddings.66        initializer_range (`float`, *optional*, defaults to 0.02):67            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.68        min_timescale (`int`, *optional*, defaults to 1):69            The start of the geometric positional index. Determines the periodicity of70            the added signal.71        max_timescale (`int`, *optional*, defaults to 10000):72            The end of the geometric positional index. Determines the frequency of the73            added signal.74    """75 76    model_type = "timesfm"77    keys_to_ignore_at_inference = []78    is_encoder_decoder = False79 80    def __init__(81        self,82        patch_length: int = 32,83        context_length: int = 512,84        horizon_length: int = 128,85        freq_size: int = 3,86        num_hidden_layers: int = 50,87        hidden_size: int = 1280,88        intermediate_size: int = 1280,89        head_dim: int = 80,90        num_attention_heads: int = 16,91        tolerance: float = 1e-6,92        rms_norm_eps: float = 1e-6,93        quantiles: list[float] = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9],94        pad_val: float = 1123581321.0,95        attention_dropout: float = 0.0,96        use_positional_embedding: bool = False,97        initializer_range: float = 0.02,98        min_timescale: int = 1,99        max_timescale: int = 10_000,100        **kwargs,101    ):102        self.patch_length = patch_length103        self.context_length = context_length104        self.horizon_length = horizon_length105        self.quantiles = quantiles106        self.pad_val = pad_val107        self.freq_size = freq_size108        self.hidden_size = hidden_size109        self.intermediate_size = intermediate_size110        self.head_dim = head_dim111        self.num_hidden_layers = num_hidden_layers112        self.num_attention_heads = num_attention_heads113        self.tolerance = tolerance114        self.rms_norm_eps = rms_norm_eps115        self.attention_dropout = attention_dropout116        self.use_positional_embedding = use_positional_embedding117        self.initializer_range = initializer_range118        self.min_timescale = min_timescale119        self.max_timescale = max_timescale120 121        super().__init__(122            is_encoder_decoder=self.is_encoder_decoder,123            **kwargs,124        )125 126 127__all__ = ["TimesFmConfig"]128