""" Hiera model configuration"""

import math

from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging

logger = logging.get_logger(__name__)

# HIERA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
#     "hoge/hoge": ("/config.json"),
# }


class HieraConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`HieraModel`]. It is used to instantiate a Hiera
    model according to the specified arguments, defining the model architecture. Instantiating a
    configuration with the defaults will yield a similar configuration to that of the Hiera
    [/]()
    architecture.

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    Args:
        image_size (`int`, *optional*, defaults to 224):
            The size (resolution) of each image.
        patch_size (`list(int)`, *optional*, defaults to [7, 7]):
            The size (resolution) of each patch.
        stride_size (`list(int)`, *optional*, defaults to [4, 4]):
            The size (resolution) of each stride.
        padding_size (`list(int)`, *optional*, defaults to [3, 3]):
            The size (resolution) of each padding.
        num_channels (`int`, *optional*, defaults to 3):
            The number of input channels.
        embed_dim (`int`, *optional*, defaults to 96):
            Dimensionality of patch embedding.
        depths (`list(int)`, *optional*, defaults to `[2, 3, 16, 3]`):
            Depth of each layer in the Transformer encoder.
        num_heads (`list(int)`, *optional*, defaults to `[1, 2, 4, 8]`):
            Number of attention heads in each layer of the Transformer encoder.
        q_pool (`int`, *optional*, defaults to 3):
            Number of q_pool stages.
        q_stride (`list(int)`, *optional*, defaults to [2, 2]):
            Size of stride of q_pool,
        mask_unit_size (`list(int)`, *optional*, defaults to [8, 8]):
            Size of mask unit in attention.
        mask_unit_attention (`list(bool)`, *optional*, defaults to [True, True, False, False]):
            Whether or not to enable mask unit attention in each stage.
        separate_positional_embeds (`bool`, *optional*, defaults to False):
            Whether or not to use separeted positional embeddings.
        mlp_ratio (`float`, *optional*, defaults to 4.0):
            Ratio of MLP hidden dimensionality to embedding dimensionality.
        drop_path_rate (`float`, *optional*, defaults to 0.1):
            Stochastic depth rate.
        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
            The non-linear activation function (function or string) in the encoder. If string, `"gelu"`, `"relu"`,
            `"selu"` and `"gelu_new"` are supported.
        layer_norm_eps (`float`, *optional*, defaults to 1e-05):
            The epsilon used by the layer normalization layers.
        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
            The dropout probability for all fully connected layers in the embeddings and encoder.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        initializer_bias (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all bias matrices.

    Example:

    ```python
    >>> from transformers import HieraConfig, HieraModel

    >>> # Initializing a Hiera / style configuration
    >>> configuration = HieraConfig()

    >>> # Initializing a model (with random weights) from the / style configuration
    >>> model = HieraModel(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "hiera"

    attribute_map = {}

    def __init__(
        self,
        image_size=224,
        patch_size=[7, 7],
        stride_size=[4, 4],
        padding_size=[3, 3],
        num_channels=3,
        embed_dim=96,
        depths=[2, 3, 16, 3],
        num_heads=[1, 2, 4, 8],
        q_pool=3,  # number of q_pool stages
        q_stride=[2, 2],
        mask_unit_size=[8, 8],
        mask_unit_attention=[True, True, False, False],
        separate_positional_embeds=False,
        mlp_ratio=4.0,
        drop_path_rate=0.0,
        hidden_act="gelu",
        layer_norm_eps=1e-6,
        hidden_dropout_prob=0.0,
        initializer_range=0.02,
        initializer_bias=0.02,
        **kwargs,
    ):
        super().__init__(**kwargs)

        self.image_size = image_size
        self.patch_size = patch_size
        self.stride_size = stride_size
        self.padding_size = padding_size
        self.num_channels = num_channels
        self.embed_dim = embed_dim
        self.depths = depths
        self.num_layers = len(depths)
        self.num_heads = num_heads
        self.mlp_ratio = mlp_ratio
        self.hidden_dropout_prob = hidden_dropout_prob
        self.drop_path_rate = drop_path_rate
        self.hidden_act = hidden_act
        self.layer_norm_eps = layer_norm_eps

        assert q_pool < len(depths), "q_pool must be less than depths"

        self.mask_unit_size = mask_unit_size
        self.flat_mask_unit_size = int(math.prod(mask_unit_size))
        self.mask_unit_attention = mask_unit_attention
        self.q_pool = q_pool
        self.q_stride = q_stride
        self.flat_q_stride = int(math.prod(q_stride))
        self.separate_positional_embeds = separate_positional_embeds

        self.initializer_range = initializer_range
        self.initializer_bias = initializer_bias