o i1@shddlmZddlmZmZmZmZddlmZm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZm Z m!Z!ddl"m#Z$ddl%m&Z&m'Z'ddl(m)Z)m*Z*ddl+m,Z,dd l-m.Z.dd l/m0Z0dd l1m2Z2d d l3m3Z3d dl4m5Z5ddl6m7Z7dZ8 d>dede9dee:deee0e*ffddZ;ddZ< d>de9de:de9dee:deee0e*ff ddZ=  d?de9de:de9d e:dee:deee0e*ff d!d"Z> d>de9de:de9d#eee)e)e)fe)fdee:deee0e*ff d$d%Z? d>deee0ee*fd&eee0e*fdee:deee0e*ffd'd(Z@defd)d*ZA d>d+e:d,eeBdee:deee0e*ffd-d.ZC d>deee0ee*fde9dee:deee0e*ffd/d0ZDdd1deee0ee*fd2ee*e*fd3eee*e*fdeee0e*ffd4d5ZEdefd6d7ZF d>dede9d8e9d9e9d:e9d;e9dee:deee0e*ffd>rT fill_defaults resize_layer)attrr2r< multi_labelrC resize_output)r define_operatorsr NEG_VALUErr rrr"r,r'r opsr@rintr>attrsrD) r1rErFrIr2rXr<rCrAr9r9r:rJos0       rJ linear_modelc CsL|jd }tttdV|d}t|}t||d}t|d}|t ?|?t ?t ||?t d??}|r=|dnd} |rHt || d} n t|| dt?} ||B| ?} | d|Wdn1sgwY| dd ur|dur| dtt|| d |d | d || d || d || | jd<t| _| S)Nr\rV|r2r2nIrhr+r0Fr<attention_layer maxout_layer norm_layer)rcr r_rr maybe_get_dimrr r rrrrrr r r>has_dimr@rrbget_refinit_ensemble_textcatr!) r0rdr2r1widthrkrlrm cnn_model nO_doubler<rAr9r9r:build_text_classifier_v2s@         rucCs|d}||t|}|dd||dd||dd||dd||dd|t||||S)Nr0rkr2rlrhrmrp initializer/r@ init_chainrAXYr0 tok2vec_widthr9r9r:rqs   rqrrdropoutcCstttd2t|t?t|?t?tt ||d?t ||?}|r-|t |?}|t ?}Wd|S1s=wY|S)N)rVz**r+) r r_rrr-rrrrrr rr )rrr}r2rAr9r9r:build_text_classifier_lowdatas(     r~cCs|d}t|t||dt||dd}tdti|r$t|d}nt|dt ?}||?}Wdn1s:wY| ddurS|durS| dt t ||d|| |jd<|S) Nr2)rhr2)r0nonlinear_layer key_transformrVrGFr<r\)rn3_build_parametric_attention_with_residual_nonlinearr rr r_rrr r ror@rrbr>rc)r0r1r2rrparametric_attentionr<rAr9r9r:%build_textcat_parametric_attention_v1s"       r)rrrcCstttdO|d}t||d}t|d}|t?|?t?t ||?t d??}t |_ | d|| d|| d|| d || d ||WdS1sZwYdS) Nrer2)r2rrirjr0rkrrrm)r r_rrrnrr rrrr2_init_parametric_attention_with_residual_nonlinearr!r>)r0rrrrrkrmrr9r9r:rs,        $rcCs|d}||t|}|dd||dd||dd||dd||dd||dd||dd|t||||S)Nr0rkr2rrhrrmrvryr9r9r:rs   rr4r5r6r7cCszddd}g}|r|t|r|t|r|t|r'|tt|s0ttjt dt ic|t ?t |?} |d} | durP| t|nd} |rmt|| d} t|d<t| tt|dd } | | ?}nt|| d} t| tt|dd } | | ?t?}|d | tt| d |jd <Wdn1swY|d ||dur|dtt|| |jd<|S)aBuild a model that classifies pooled `Doc` representations. Pooling is performed using reductions. Reductions are concatenated when multiple reductions are used. tok2vec (Model): the tok2vec layer to pool over. exclusive_classes (bool): Whether or not classes are mutually exclusive. use_reduce_first (bool): Pool by using the hidden representation of the first token of a `Doc`. use_reduce_last (bool): Pool by using the hidden representation of the last token of a `Doc`. use_reduce_max (bool): Pool by taking the maximum values of the hidden representations of a `Doc`. use_reduce_mean (bool): Pool by taking the mean of all hidden representations of a `Doc`. nO (Optional[int]): Number of classes. rrSrVr2NrgrTrWrYr<r]r^r0r\)appendrrrrlenrNr(E1057r r_rrrrnrr`rrr"r r r>rDrcr@rrb)r0r1r4r5r6r7r2rX reductionscnn nO_tok2vecrhr<rCrAr9r9r:r8#sZ             r8)N)rLN)H functoolsrtypingrrrr thinc.apirrr r r r r rrrrrrrrrrrrrrrrrrr thinc.layers.chainr!rxthinc.layers.resizabler"r# thinc.typesr$r%rcr'errorsr(tokensr)utilr*r, staticvectorsr-r0r/r`boolrbr;rDrKrRrJrurqfloatr~rrrr8r9r9r9r:s p             " )