Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2021 NVIDIA and The HuggingFace Inc. team. All rights reserved.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"""SegFormer model configuration"""16 17import warnings18from collections import OrderedDict19from collections.abc import Mapping20 21from packaging import version22 23from ...configuration_utils import PretrainedConfig24from ...onnx import OnnxConfig25from ...utils import logging26 27 28logger = logging.get_logger(__name__)29 30 31class SegformerConfig(PretrainedConfig):32    r"""33    This is the configuration class to store the configuration of a [`SegformerModel`]. It is used to instantiate an34    SegFormer model according to the specified arguments, defining the model architecture. Instantiating a35    configuration with the defaults will yield a similar configuration to that of the SegFormer36    [nvidia/segformer-b0-finetuned-ade-512-512](https://huggingface.co/nvidia/segformer-b0-finetuned-ade-512-512)37    architecture.38 39    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the40    documentation from [`PretrainedConfig`] for more information.41 42    Args:43        num_channels (`int`, *optional*, defaults to 3):44            The number of input channels.45        num_encoder_blocks (`int`, *optional*, defaults to 4):46            The number of encoder blocks (i.e. stages in the Mix Transformer encoder).47        depths (`list[int]`, *optional*, defaults to `[2, 2, 2, 2]`):48            The number of layers in each encoder block.49        sr_ratios (`list[int]`, *optional*, defaults to `[8, 4, 2, 1]`):50            Sequence reduction ratios in each encoder block.51        hidden_sizes (`list[int]`, *optional*, defaults to `[32, 64, 160, 256]`):52            Dimension of each of the encoder blocks.53        patch_sizes (`list[int]`, *optional*, defaults to `[7, 3, 3, 3]`):54            Patch size before each encoder block.55        strides (`list[int]`, *optional*, defaults to `[4, 2, 2, 2]`):56            Stride before each encoder block.57        num_attention_heads (`list[int]`, *optional*, defaults to `[1, 2, 5, 8]`):58            Number of attention heads for each attention layer in each block of the Transformer encoder.59        mlp_ratios (`list[int]`, *optional*, defaults to `[4, 4, 4, 4]`):60            Ratio of the size of the hidden layer compared to the size of the input layer of the Mix FFNs in the61            encoder blocks.62        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):63            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,64            `"relu"`, `"selu"` and `"gelu_new"` are supported.65        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):66            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.67        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):68            The dropout ratio for the attention probabilities.69        classifier_dropout_prob (`float`, *optional*, defaults to 0.1):70            The dropout probability before the classification head.71        initializer_range (`float`, *optional*, defaults to 0.02):72            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.73        drop_path_rate (`float`, *optional*, defaults to 0.1):74            The dropout probability for stochastic depth, used in the blocks of the Transformer encoder.75        layer_norm_eps (`float`, *optional*, defaults to 1e-06):76            The epsilon used by the layer normalization layers.77        decoder_hidden_size (`int`, *optional*, defaults to 256):78            The dimension of the all-MLP decode head.79        semantic_loss_ignore_index (`int`, *optional*, defaults to 255):80            The index that is ignored by the loss function of the semantic segmentation model.81 82    Example:83 84    ```python85    >>> from transformers import SegformerModel, SegformerConfig86 87    >>> # Initializing a SegFormer nvidia/segformer-b0-finetuned-ade-512-512 style configuration88    >>> configuration = SegformerConfig()89 90    >>> # Initializing a model from the nvidia/segformer-b0-finetuned-ade-512-512 style configuration91    >>> model = SegformerModel(configuration)92 93    >>> # Accessing the model configuration94    >>> configuration = model.config95    ```"""96 97    model_type = "segformer"98 99    def __init__(100        self,101        num_channels=3,102        num_encoder_blocks=4,103        depths=[2, 2, 2, 2],104        sr_ratios=[8, 4, 2, 1],105        hidden_sizes=[32, 64, 160, 256],106        patch_sizes=[7, 3, 3, 3],107        strides=[4, 2, 2, 2],108        num_attention_heads=[1, 2, 5, 8],109        mlp_ratios=[4, 4, 4, 4],110        hidden_act="gelu",111        hidden_dropout_prob=0.0,112        attention_probs_dropout_prob=0.0,113        classifier_dropout_prob=0.1,114        initializer_range=0.02,115        drop_path_rate=0.1,116        layer_norm_eps=1e-6,117        decoder_hidden_size=256,118        semantic_loss_ignore_index=255,119        **kwargs,120    ):121        super().__init__(**kwargs)122 123        if "reshape_last_stage" in kwargs and kwargs["reshape_last_stage"] is False:124            warnings.warn(125                "Reshape_last_stage is set to False in this config. This argument is deprecated and will soon be"126                " removed, as the behaviour will default to that of reshape_last_stage = True.",127                FutureWarning,128            )129 130        self.num_channels = num_channels131        self.num_encoder_blocks = num_encoder_blocks132        self.depths = depths133        self.sr_ratios = sr_ratios134        self.hidden_sizes = hidden_sizes135        self.patch_sizes = patch_sizes136        self.strides = strides137        self.mlp_ratios = mlp_ratios138        self.num_attention_heads = num_attention_heads139        self.hidden_act = hidden_act140        self.hidden_dropout_prob = hidden_dropout_prob141        self.attention_probs_dropout_prob = attention_probs_dropout_prob142        self.classifier_dropout_prob = classifier_dropout_prob143        self.initializer_range = initializer_range144        self.drop_path_rate = drop_path_rate145        self.layer_norm_eps = layer_norm_eps146        self.decoder_hidden_size = decoder_hidden_size147        self.reshape_last_stage = kwargs.get("reshape_last_stage", True)148        self.semantic_loss_ignore_index = semantic_loss_ignore_index149 150 151class SegformerOnnxConfig(OnnxConfig):152    torch_onnx_minimum_version = version.parse("1.11")153 154    @property155    def inputs(self) -> Mapping[str, Mapping[int, str]]:156        return OrderedDict(157            [158                ("pixel_values", {0: "batch", 1: "num_channels", 2: "height", 3: "width"}),159            ]160        )161 162    @property163    def atol_for_validation(self) -> float:164        return 1e-4165 166    @property167    def default_onnx_opset(self) -> int:168        return 12169 170 171__all__ = ["SegformerConfig", "SegformerOnnxConfig"]172