定义余弦相似度层，在batch内进行负采样
NEG，batch _ size=20，128 classnegativecosinelayer () : ) ''创建自定义batch内负采样，并在cosine相似度的层' ' def_call_， inputs ) :def_cosine(x ) : query_encoder，doc _ encoder=xdoc _ encoder _ FD=doc _ encoderforinrinrange (附件) TF.random.shuffle (TF.range (TF.shape ) doc_encoder([0] ) ) ) ) doc_encoder_FD=TF.concat ) ) DOC_encoder ) ) axis=0) query_活力太阳镜=TF.tile(TF.sqrt keepdims 1) doc_活力太阳镜=TF.sqrt(TF.reduce_sum ) TF.square (doc _ encoder _ FD )，axis=1，keepdims=True keepdims=True (活力太阳镜(prod=TF.multiply ) query_活力太阳镜，doc_活力太阳镜) cos_sim_raw=TF.truediv ) PPV 的太阳镜_ prod (cos _ sim=TF.transpose (TF.reshape ) TF.transpose ) cos_sim_raw )，[NEG 1，-1] 6520 prob naw name='sim-slice ' ) loss=-TF.name=' sim-mean ' (returnlossoutput _ shape=(1，) value=lambda(cosine，)
importtensorflowastffromkeras.modelsimportmodelfromkeras.layersimportinput，嵌入式，Dense，model， ltm query _ max _ len=16 doc _ max _ len=128 vocab _ size=10000 embed _ dim=64 query _ input=input (shape=(query ) doc _ input=input (shape=(doc _ max _ len，)，name='doc_input ' ) embedding=embedding )。 embed _ dim (query _ embed=embedding ) query_input ) doc _ embed=embedding (doc _ input ) query _ encoder=lstm outputs=cos_sim ) query _ model=model (inputs=query _ input，outputs=query_encoder ) doc_model=model ) outputs=doc _ encoder (model.com pile (optimizer=' Adam '，loss=lambda y_true，y_pred: y_pred ) model.model