o Zhg$@sRdZddlmZddlmZddlmZmZee Z GdddeeZ dgZ dS)zHiera model configuration)PretrainedConfig)logging)BackboneConfigMixin*get_aligned_output_features_output_indicescseZdZdZdZddiZdddgddgddgd d gd gd gd d d ddgddggddd dddddddddddffdd ZZS) HieraConfigac 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 [facebook/hiera-base-224](https://huggingface.co/facebook/hiera-base-224) architecture. Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: embed_dim (`int`, *optional*, defaults to 96): Dimensionality of patch embedding. image_size (`list(int)`, *optional*, defaults to `[224, 224]`): The size (resolution) of input in the format (height, width) for images and (frames, height, width) for videos. patch_size (`list(int)`, *optional*, defaults to `[7, 7]`): The size (resolution) of each patch. patch_stride (`list(int)`, *optional*, defaults to `[4, 4]`): The stride of the patch. patch_padding (`list(int)`, *optional*, defaults to `[3, 3]`): The padding of the patch. mlp_ratio (`float`, *optional*, defaults to 4.0): The ratio of mlp hidden dim to embedding dim. 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. embed_dim_multiplier (`float`, *optional*, defaults to 2.0): The multiplier to the dimensionality of patch embedding in each layer of the Transformer encoder. num_query_pool (`int`, *optional*, defaults to 3): The number of query pool stages. query_stride (`list(int)`, *optional*, defaults to `[2, 2]`): The stride of the query pool. masked_unit_size (`list(int)`, *optional*, defaults to `[8, 8]`): The size of the masked unit. masked_unit_attention (`list(bool)`, *optional*, defaults to `[True, True, False, False]`): Whether to use masked unit attention in each layer of the Transformer encoder. drop_path_rate (`float`, *optional*, defaults to 0.0): The drop path rate. num_channels (`int`, *optional*, defaults to 3): The number of input channels. hidden_act (`str`, *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. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices and the zero_initializer for initializing all bias vectors. layer_norm_init (`float`, *optional*, defaults to 1.0): The initial weight value for layer normalization layers. layer_norm_eps (`float`, *optional*, defaults to 1e-06): The epsilon used by the layer normalization layers. decoder_hidden_size (`int`, *optional*): Dimensionality of decoder embeddings for MAE pretraining. decoder_depth (`int`, *optional*): Depth of the decoder for MAE pretraining. decoder_num_heads (`int`, *optional*): Number of attention heads in each layer of the decoder for MAE pretraining. normalize_pixel_loss (`bool`, *optional*, defaults to `True`): Whether to normalize the pixel loss by the number of pixels. mask_ratio (`float`, *optional*, defaults to 0.6): The ratio of masked tokens in the input. out_features (`List[str]`, *optional*): If used as backbone, list of features to output. Can be any of `"stem"`, `"stage1"`, `"stage2"`, etc. (depending on how many stages the model has). If unset and `out_indices` is set, will default to the corresponding stages. If unset and `out_indices` is unset, will default to the last stage. Must be in the same order as defined in the `stage_names` attribute. out_indices (`List[int]`, *optional*): If used as backbone, list of indices of features to output. Can be any of 0, 1, 2, etc. (depending on how many stages the model has). If unset and `out_features` is set, will default to the corresponding stages. If unset and `out_features` is unset, will default to the last stage. Must be in the same order as defined in the `stage_names` attribute. Example: ```python >>> from transformers import HieraConfig, HieraModel >>> # Initializing a Hiera hiera-base-patch16-224 style configuration >>> configuration = HieraConfig() >>> # Initializing a model (with random weights) from the hiera-base-patch16-224 style configuration >>> model = HieraModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```ZhieraZnum_hidden_layers num_layers`rg@)rr)r r g@r r)TTFFgZgelug{Gz?g?gư>NTg333333?c sxtjd i|| d| dt|ddkr/td| dd| ddt|dd| t|krBtd| dt|d||_||_||_||_||_||_ ||_ ||_ t||_ | |_ | |_| |_| |_| |_||_||_||_||_||_||_||_||_||_||_||_t|| t|d|_d gd d tdt|dD|_ t!|||j d \|_"|_#dS)Nrzmasked_unit_size[0] (z() must be divisible by query_stride[0] (z/) raised to the power of the number of layers ()znum_query_pool (z*) must be less than the number of layers (stemcSsg|]}d|qS)Zstage).0idxrr\/var/www/auris/lib/python3.10/site-packages/transformers/models/hiera/configuration_hiera.py sz(HieraConfig.__init__..) out_features out_indices stage_namesr)$super__init__len ValueError embed_dim image_size patch_size patch_stride patch_padding mlp_ratiodepths num_headsrembed_dim_multipliernum_query_pool query_stridemasked_unit_sizemasked_unit_attentiondrop_path_rate num_channels hidden_actinitializer_rangelayer_norm_initlayer_norm_epsdecoder_hidden_size decoder_depthdecoder_num_headsnormalize_pixel_loss mask_ratiointZ hidden_sizerangerrZ _out_featuresZ _out_indices)selfrr r!r"r#r$r%r&r'r(r)r*r+r,r-r.r/r0r1r2r3r4r5r6rrkwargs __class__rrrvsR    $zHieraConfig.__init__)__name__ __module__ __qualname____doc__Z model_typeZ attribute_mapr __classcell__rrr;rrs>XrN) r@Zconfiguration_utilsrutilsrZutils.backbone_utilsrrZ get_loggerr=loggerr__all__rrrrs    *