@idZddlZddlmZmZddlZ ddlZddlm Z dZ ddl mZejeZGdd ZGd d e re j&ZyeZy#e $rdZ ejdY\wxYw) z5Pattern encoder with learnable preference projection.N)ListUnionTFz9PyTorch not available, PreferenceEncoder will be disabled)SentenceTransformercReZdZdZddefdZdeeeefdejfdZ y) PatternEncoderz:Base pattern encoder using sentence transformers (frozen). model_namect||_|jj|_tj d|jy)zzInitialize pattern encoder. Args: model_name: Name of the sentence transformer model z*PatternEncoder initialized with dimension N)rmodel get_sentence_embedding_dimension embedding_dimloggerinfo)selfrs H/home/ubuntu/codebase/yexijia/保研/colocation_mvp/embedding/encoder.py__init__zPatternEncoder.__init__s@ )4 !ZZHHJ @ASAS@TUVtextreturnct|tr|jj|S|jj|S)zEncode pattern text to vector. Args: text: Pattern text (e.g., "Park Zoo") or list of patterns Returns: Vector representation(s) ) isinstancestrr encode)rrs rrzPatternEncoder.encode!s8 dC ::$$T* *::$$T* *rN)z&sentence-transformers/all-MiniLM-L6-v2) __name__ __module__ __qualname____doc__rrrrnpndarrayrrrrrs7DW3W +5d3i0 +RZZ +rrceZdZdZd dededeffd ZdZdejdejfd Z de e jejfde jfd Z xZS) PreferenceEncoderzLearnable preference encoder for projecting embeddings to preference space. Architecture: Linear(dim -> hidden) -> ReLU -> Linear(hidden -> dim) -> Normalize input_dim hidden_dim output_dimcxts tdtt|||}||_||_||_tj|||_ tj|_ tj|||_ |jtj!d|d|d|y)zInitialize preference encoder. Args: input_dim: Input embedding dimension hidden_dim: Hidden layer dimension output_dim: Output dimension (default: same as input_dim) z)PyTorch is required for PreferenceEncoderNzPreferenceEncoder initialized: z -> )TORCH_AVAILABLE ImportErrorsuperr!rr"r#r$nnLinearfc1ReLUrelufc2_initialize_weightsr r)rr"r#r$ __class__s rrzPreferenceEncoder.__init__7sIJ J /1  "J"$$99Y 3GGI 99Z4   " 5i[ZLPTU_T`abrctjj|jjtjj |jj tjj|jjtjj |jj y)zInitialize network weights.N)r)initxavier_uniform_r+weightzeros_biasr.)rs rr/z%PreferenceEncoder._initialize_weightsUsf 0 txx}}% 0 txx}}%rxrc|j|}|j|}|j|}tjj j |dd}|S)zForward pass. Args: x: Input tensor of shape (batch_size, input_dim) Returns: Projected and normalized tensor of shape (batch_size, output_dim) )pdim)r+r-r.torchr) functional normalize)rr7s rforwardzPreferenceEncoder.forward\sS HHQK IIaL HHQK HH   ) )!qa ) 8rct|tjrtj|}t |j dk(r|jd}|jtj5|j|}dddj ddk(r|jdjS|jS#1swYJxYw)zEncode input to preference space (for inference). Args: x: Input array or tensor Returns: Projected vector as numpy array r:rN) rrrr= FloatTensorlenshape unsqueezeevalno_gradr@squeezenumpy)rr7outputs rrzPreferenceEncoder.encodens a $!!!$A qww<1  AA ]]_\\!_F <<?a >>!$**, ,<<> ! _s =CC!)iN)rrrrintrr/r=Tensorr@rrrr __classcell__)r0s@rr!r!0sp c#ccPSc<&%,,$"bjj%,,67"BJJ"rr!)rrIrtypingrrloggingr=torch.nnr)r&r'warningsentence_transformersr getLoggerrr rModuleobjectr!rrrrWs;QO 6   8 $++8T"_ T"&T"KQOGOOOPQs A!!A=<A=