o ei$@s&ddlmZGdddeZdgZdS))PreTrainedConfigc1seZdZdZdZ                  d+deedBdeedBdeedBdededededededededede dede d ed!ed"ed#ed$ed%e dBd&ed'ed(ef0fd)d* Z Z S),EfficientLoFTRConfigad This is the configuration class to store the configuration of a [`EfficientLoFTRFromKeypointMatching`]. It is used to instantiate a EfficientLoFTR 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 EfficientLoFTR [zju-community/efficientloftr](https://huggingface.co/zju-community/efficientloftr) architecture. Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PreTrainedConfig`] for more information. Args: stage_num_blocks (`List`, *optional*, defaults to [1, 2, 4, 14]): The number of blocks in each stages out_features (`List`, *optional*, defaults to [64, 64, 128, 256]): The number of channels in each stage stage_stride (`List`, *optional*, defaults to [2, 1, 2, 2]): The stride used in each stage hidden_size (`int`, *optional*, defaults to 256): The dimension of the descriptors. activation_function (`str`, *optional*, defaults to `"relu"`): The activation function used in the backbone q_aggregation_kernel_size (`int`, *optional*, defaults to 4): The kernel size of the aggregation of query states in the fusion network kv_aggregation_kernel_size (`int`, *optional*, defaults to 4): The kernel size of the aggregation of key and value states in the fusion network q_aggregation_stride (`int`, *optional*, defaults to 4): The stride of the aggregation of query states in the fusion network kv_aggregation_stride (`int`, *optional*, defaults to 4): The stride of the aggregation of key and value states in the fusion network num_attention_layers (`int`, *optional*, defaults to 4): Number of attention layers in the LocalFeatureTransformer num_attention_heads (`int`, *optional*, defaults to 8): The number of heads in the GNN layers. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. attention_bias (`bool`, *optional*, defaults to `False`): Whether to use a bias in the query, key, value and output projection layers during attention. mlp_activation_function (`str`, *optional*, defaults to `"leaky_relu"`): Activation function used in the attention mlp layer. coarse_matching_skip_softmax (`bool`, *optional*, defaults to `False`): Whether to skip softmax or not at the coarse matching step. coarse_matching_threshold (`float`, *optional*, defaults to 0.2): The threshold for the minimum score required for a match. coarse_matching_temperature (`float`, *optional*, defaults to 0.1): The temperature to apply to the coarse similarity matrix coarse_matching_border_removal (`int`, *optional*, defaults to 2): The size of the border to remove during coarse matching fine_kernel_size (`int`, *optional*, defaults to 8): Kernel size used for the fine feature matching batch_norm_eps (`float`, *optional*, defaults to 1e-05): The epsilon used by the batch normalization layers rope_parameters (`RopeParameters`, *optional*): Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE with longer `max_position_embeddings`. fine_matching_slice_dim (`int`, *optional*, defaults to 8): The size of the slice used to divide the fine features for the first and second fine matching stages. fine_matching_regress_temperature (`float`, *optional*, defaults to 10.0): The temperature to apply to the fine similarity matrix initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. Examples: ```python >>> from transformers import EfficientLoFTRConfig, EfficientLoFTRForKeypointMatching >>> # Initializing a EfficientLoFTR configuration >>> configuration = EfficientLoFTRConfig() >>> # Initializing a model from the EfficientLoFTR configuration >>> model = EfficientLoFTRForKeypointMatching(configuration) >>> # Accessing the model configuration >>> configuration = model.config ``` efficientloftrNreluF leaky_relu皙?皙?h㈵>$@{Gz?stage_num_blocks out_features stage_stride hidden_sizeactivation_functionq_aggregation_kernel_sizekv_aggregation_kernel_sizeq_aggregation_stridekv_aggregation_stridenum_attention_layersnum_attention_headsattention_dropoutattention_biasmlp_activation_functioncoarse_matching_skip_softmaxcoarse_matching_thresholdcoarse_matching_temperaturecoarse_matching_border_removalfine_kernel_sizebatch_norm_epsrope_parametersfine_matching_slice_dim!fine_matching_regress_temperatureinitializer_rangec s|dur|ngd_|dur|ngd_|dur|ngd_dgjdd_ddtjjD_fddtjD_fd dtt jD_ t t jdd_ |_jjdkrztd jd jd|_|_|_|_| _| _| _| _| _jd _|_|_|_|_|_|_|_ |_!|_"| _#|_$|_%|&d dt'j(di|dS)N)r r)r r)r r )@r+rr)cSs$g|]\}}|gdg|dqS)r)).0stride num_blocksr.r.}/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/transformers/models/efficientloftr/configuration_efficientloftr.py sz1EfficientLoFTRConfig.__init__..cs g|] \}}j|g|qSr.)r)r/ stage_idxr1selfr.r2r3scs*g|]}j|gj|ddqS)Nr-)stage_in_channelsstage_block_out_channels)r/r4r5r.r2r3szMhidden_size should be equal to the last value in out_features. hidden_size = z, out_features = r partial_rotary_factorg@r.))rrrr7zipstage_block_stride enumerater8rangelenstage_block_in_channelslistreversedfine_fusion_dimsr ValueErrorrrrrrrrrrintermediate_sizerrr r!r"r#r$r&r'num_key_value_headsr(r% setdefaultsuper__init__)r6rrrrrrrrrrrrrrrr r!r"r#r$r%r&r'r(kwargs __class__r5r2rHasV      zEfficientLoFTRConfig.__init__)NNNrrrrrrrrr Fr Fr r r rrNrrr) __name__ __module__ __qualname____doc__ model_typer@intstrfloatbooldictrH __classcell__r.r.rJr2rsL        rN)configuration_utilsrr__all__r.r.r.r2s  #