U &c@s`dZddlZddlZddlmZddlmZmZddl m Z ddl m Z m Z mZmZddlmZeeZd d iZGd d d ejjjZGd ddejjjZGdddejjjZGdddejjjZGdddejjjZeGdddejjjZGddde ZdZ dZ!ede GdddeZ"GdddejjjZ#ed e Gd!d"d"eZ$dS)#z TF 2.0 Transformer XL model. N)TransfoXLConfig)add_start_docstrings add_start_docstrings_to_callable)TFAdaptiveSoftmaxMask)TFPreTrainedModelget_initializerkeras_serializable shape_list) BatchEncodingztransfo-xl-wt103z7https://cdn.huggingface.co/transfo-xl-wt103-tf_model.h5cs&eZdZfddZdddZZS)TFPositionalEmbeddingc s.tjf|ddtd|d||_dS)Nri'rg@)super__init__tfrangeinv_freq)selfZdembkwargs __class__G/tmp/pip-unpacked-wheel-ymerj3tt/transformers/modeling_tf_transfo_xl.pyr'szTFPositionalEmbedding.__init__NcCsrtd||j}tt|t|gd}|dk rXt|dddddfd|dgS|dddddfSdS)Nzi,j->ijr)reinsumrconcatsincosZtile)rpos_seqbszZ sinusoid_inppos_embrrrcall,s $zTFPositionalEmbedding.call)N__name__ __module__ __qualname__rr __classcell__rrrrr &s r cs(eZdZdfdd Zd ddZZS) TFPositionwiseFFFh㈵>{Gz?c stjf|||_||_||_tjjj|t |tj j dd|_ tjj ||_tjjj|t |dd|_tjj ||_tjjj|dd|_||_dS)Nz CoreNet_._0)kernel_initializerZ activationnamez CoreNet_._3)r)r* layer_normepsilonr*)r rd_modeld_innerdropoutrkeraslayersDensernnZrelulayer_1Dropoutdrop_1layer_2drop_2LayerNormalizationr+ pre_lnorm)rr.r/r0r;layer_norm_epsiloninit_stdrrrrr7szTFPositionwiseFF.__init__cCs|jrJ||}||}|j||d}||}|j||d}||}n>||}|j||d}||}|j||d}|||}|S)Ntraining)r;r+r5r7r8r9)rinpr?core_outoutputrrrr Is      zTFPositionwiseFF.call)Fr'r()Fr!rrrrr&6sr&c s<eZdZdfdd Zfdd Zd d Zdd d ZZS)"TFRelPartialLearnableMultiHeadAttnrNFr'r(c stjf|| |_||_||_||_||_tjj j d||t |ddd|_ tjj ||_tjj ||_tjj j |t |ddd|_tjj j| dd|_d|d |_| |_| dk r| dk r| |_| |_n d|_d|_tjj j |j|jt |dd d|_dS) NFqkv_net)r)Zuse_biasr*o_netr+r,r?r_net)r routput_attentionsn_headr.d_headr0rr1r2r3rrEr6dropdropattrFr:r+scaler;r_r_biasr_w_biasrH)rrJr.rKr0rMtgt_lenext_lenmem_lenr;rOrPrIr<r=rrrrrbsD  z+TFRelPartialLearnableMultiHeadAttn.__init__cs\|jdks|jdkrL|j|j|jfdddd|_|j|j|jfdddd|_t|dS)NzerosTrOshapeZ initializerZ trainabler*rP)rOrP add_weightrJrKr buildr input_shaperrrrXs  z(TFRelPartialLearnableMultiHeadAttn.buildcCst|}t|ddgddgddgddgg}t||dd|d|d|dg}t|ddddgddddg}t||}|S)NrrrDr)r rpadreshapeslice)rxZx_sizerrr _rel_shifts $( z-TFRelPartialLearnableMultiHeadAttn._rel_shiftcCs|\}}}}}t|dt|dt|d}} } |dk rt||gd} |jrf||| } n || } ||} tj| ddd\}}}|| d}nB|jr|||} n ||} ||} tj| ddd\}}}t|d}t||| |j |j f}t||| |j |j f}t||| |j |j f}t| | |j |j f} ||j }t d||}||j }t d|| }||}||}||j}|dk r|ddddddf}|d|d|}tjj|dd}|j||d }|dk r||}t d ||}t|}t||d|d|j |j f}||}|j||d }|jrb||g}n|||g}|jr|||S) NrrrDrZaxiszibnd,jbnd->ijbnzibnd,jnd->ijbngꌠ9Y>)Fr>zijbn,jbnd->ibnd)r rrr;rEr+rHsplitr]rJrKrPrrOr`rNr4ZsoftmaxrMrFrLrIappend)rinputsr?wr attn_maskmems head_maskqlenZrlenrcatZw_headsZr_head_kZw_head_qZw_head_kZw_head_vklenZ rw_head_qZACZ rr_head_qZBDZ attn_scoreZ attn_mask_tZ attn_probZattn_vecZattn_vec_sizesZattn_outoutputsrrrr sZ(           "   z'TFRelPartialLearnableMultiHeadAttn.call) rNNNFNNFr'r()F)r"r#r$rrXr`r r%rrrrrCas5  rCc s(eZdZd fdd Zd dd ZZS) !TFRelPartialLearnableDecoderLayerNFr'r(c sRtjf|t|||||||| | | | || |dd|_t|||| ||dd|_dS)Ndec_attn) rQrRrSrMr;rPrOr=rIr<r*pos_ff)r;r=r<r*)r rrCrpr&rq)rrJr.rKr/r0rQrRrSrMr;rPrOrIr<r=rrrrrs6z*TFRelPartialLearnableDecoderLayer.__init__c CsN|\}}}}}|j|||||g|d}|j|d|d} | g|dd} | S)Nr>rr)rprq) rrdr?Zdec_inprf dec_attn_maskrhriZ attn_outputsZ ff_outputrmrrrr ,s z&TFRelPartialLearnableDecoderLayer.call) NNNroFNNFr'r()Fr!rrrrrns0rncs2eZdZd fdd ZfddZdd ZZS) TFAdaptiveEmbeddingrr(Fc stjf|||_||_||_||g|_||_||_|d|_dg|j|_ g|_ g|_ |dkrjt nft t|jD]V} |j | |j | d} } ||| } |j tjjj| | | t|d| dqxdS)NrGrrzemb_layers_._{})Zembeddings_initializerr*)r rn_tokend_embedr=cutoffsdiv_vald_proj emb_scale cutoff_ends emb_layers emb_projsNotImplementedErrorrlenrcrr1r2Z Embeddingrformat) rrtrurxrvrwr=sample_softmaxril_idxr_idxd_emb_irrrr7s0   zTFAdaptiveEmbedding.__init__c s`tt|jD]@}|j|j|}|j|j||jft |j dd |dqt |dS)NTzemb_projs_._{}rU)rr~rvrurwr|rcrWrxrr=rr rX)rrZrrrrrrXWszTFAdaptiveEmbedding.buildc Cs|jdkrtnt|d}tt|d|jg}tt|j D]}|j ||j |d}}||k||k@}t |||}|j ||} t d| |j|} tjt|tjd} |t| | tjt|tjd7}qBt||jg} t|| } | |j9} | S)Nr)rrz id,de->ie)Zdtype)rwr}rr]rTr rxrr~rvrzZ boolean_maskr{rr|castwhereZint64Z scatter_ndry) rr@Zinp_flatZemb_flatrrrZmask_iZinp_iZemb_iZmask_idxZ embed_shapeembedrrrr ds   $  zTFAdaptiveEmbedding.call)rr(Fr"r#r$rrXr r%rrrrrs6s rscsneZdZeZfddZfddZddZddZd d Z d d Z d dZ ddZ ddZ dddZZS)TFTransfoXLMainLayerc stjf||j|_|j|_|j|_|j|_|j|_|j|_|j |_ |j |_ t |j|j|j|j |j |jdd|_tjj|j|_|j|_|j|_|j|_|j|_|j|j|j|_|j|_g|_|jdkrPt|jD]p}|jt|j|j|j |j|j|j|j|j|j |j!|j rdn|j"|j r,dn|j#|j|j$|jd%|dqnt&|j'|_'|j(|_(|jdkrt)|jdd|_*nt&dS)Nword_emb)rwr=r*rz layers_._{}) rQrRrSrMr;rPrOrIr<r=r*rr*)+r rrIoutput_hidden_states vocab_sizertrur.rJrKuntie_rrsrvrwr=rrr1r2r6r0rLn_layerrQrSrRZmax_klen attn_typerrcrnr/rMr;rPrOr<rr} same_length clamp_lenr r)rconfigrrrrrrsl   zTFTransfoXLMainLayer.__init__csN|js>|j|j|jfdddd|_|j|j|jfdddd|_t|dS)NrTTrPrUrO)rrWrJrKrPrOr rXrYrrrrXs  zTFTransfoXLMainLayer.buildcCs|jSNrrrrrget_input_embeddingssz)TFTransfoXLMainLayer.get_input_embeddingscCs|jSrr)rZnew_num_tokensrrr_resize_token_embeddingssz-TFTransfoXLMainLayer._resize_token_embeddingscCs d|_dS)Nr)rrrrrbackward_compatiblesz(TFTransfoXLMainLayer.backward_compatiblecCs||_||_||_dSr)rQrSrRrrQrRrSrrr reset_lengthsz!TFTransfoXLMainLayer.reset_lengthcCstdSr)r})rZheadsrrr _prune_headssz!TFTransfoXLMainLayer._prune_headscCsH|jdkr@g}t|jD]"}t|j||jg}||q|SdSdS)Nr)rSrrrrTr.rc)rrrhremptyrrr init_memss  zTFTransfoXLMainLayer.init_memsc Cs|dkr dSt|t|ks$tdg}|td|d|j}td||j}tt|D]:}tj||||gdd} t| | | ||q\|S)Nzlen(hids) != len(mems)rra) r~AssertionErrormaxrRrSrrrZ stop_gradientrc) rhidsrhmlenrjnew_memsZend_idxZbeg_idxrrkrrr _update_memss z!TFTransfoXLMainLayer._update_memsNFcCst|ttfrt|d}t|dkr*|dn|}t|dkrB|dn|}t|dkrZ|dn|}t|dkstdnVt|ttfr|d}|d|}|d |}|d |}t|dkstdn|}|dk r|dk rtd n\|dk r t j |d d }t |\}}n6|dk r8t j |dd }t |dd\}}ntd|dkrT| |}|dk rdt n dg|j}|dk r|} n ||} |dk rt |ddnd} | |} t ||g} t j| dd} t j| dd}t || g}t || |gd}|jrPt j| dd}t |ddd|f|||dd|dfgd}g}g}|jdkr&t | ddd}|jdkrt ||j}||}|j| |d}|j||d}t|jD]b\}}|||dkrdn||}|||||||g|d}|d}|jr||dqnt |j||d}| ||| |}t j |dd |g}|j!r||tdd|D}|||jrtdd|D}|||S)Nrrr[rDToo many inputs. input_idsrhri inputs_embedszDYou cannot specify both input_ids and inputs_embeds at the same time)rrpermrrr[z5You have to specify either input_ids or inputs_embedsrgr>css|]}tj|ddVqdS)rrNr transpose.0trrr jsz,TFTransfoXLMainLayer.call..css|]}tj|ddVqdS))r[rDrrrNrrrrrrns)" isinstancetuplelistr~rdictr get ValueErrorrrr rr}rrZonesZlinalgZ band_partrTrrrrrZminimumrrL enumerater2rcrIrr)rrdrhrirr?rrjrrrrlrgZmask_uZmask_diaZ attn_mask_padrrZmask_lrZ attentionsrrrArZlayerZmems_iZ layer_outputsrrmrrrr s             8       zTFTransfoXLMainLayer.call)NNNF)r"r#r$r config_classrrXrrrrrrrr r%rrrrr~s D  rc@seZdZdZeZeZdZdS)TFTransfoXLPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. transformerN) r"r#r$__doc__rr*TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAPZpretrained_model_archive_mapZbase_model_prefixrrrrrssra .. note:: TF 2.0 models accepts two formats as inputs: - having all inputs as keyword arguments (like PyTorch models), or - having all inputs as a list, tuple or dict in the first positional arguments. This second option is useful when using :obj:`tf.keras.Model.fit()` method which currently requires having all the tensors in the first argument of the model call function: :obj:`model(inputs)`. If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument : - a single Tensor with input_ids only and nothing else: :obj:`model(inputs_ids)` - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring: :obj:`model([input_ids, attention_mask])` or :obj:`model([input_ids, attention_mask, token_type_ids])` - a dictionary with one or several input Tensors associated to the input names given in the docstring: :obj:`model({'input_ids': input_ids, 'token_type_ids': token_type_ids})` Parameters: config (:class:`~transformers.TransfoXLConfig`): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights. a Args: input_ids (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using :class:`transformers.TransfoXLTokenizer`. See :func:`transformers.PreTrainedTokenizer.encode` and :func:`transformers.PreTrainedTokenizer.encode_plus` for details. `What are input IDs? <../glossary.html#input-ids>`__ mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`): Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model (see `mems` output below). Can be used to speed up sequential decoding. The token ids which have their mems given to this model should not be passed as input ids as they have already been computed. head_mask (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`): Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``: :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**. input_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`): Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. z]The bare Bert Model transformer outputing raw hidden-states without any specific head on top.cs,eZdZfddZeeddZZS)TFTransfoXLModelcs&tj|f||t|dd|_dS)Nrr)r rrr)rrrdrrrrrszTFTransfoXLModel.__init__cKs|j|f|}|S)a Return: :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.TransfoXLConfig`) and inputs: last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the last layer of the model. mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`): Contains pre-computed hidden-states (key and values in the attention blocks). Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model should not be passed as input ids as they have already been computed. hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_hidden_states=True``): Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape :obj:`(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the initial embedding outputs. attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_attentions=True``): Tuple of :obj:`tf.Tensor` (one for each layer) of shape :obj:`(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Examples:: import tensorflow as tf from transformers import TransfoXLTokenizer, TFTransfoXLModel tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103') model = TFTransfoXLModel.from_pretrained('transfo-xl-wt103') input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1 outputs = model(input_ids) last_hidden_states, mems = outputs[:2] )r)rrdrrmrrrr s#zTFTransfoXLModel.call)r"r#r$rrTRANSFO_XL_INPUTS_DOCSTRINGr r%rrrrrs rcs0eZdZfddZfddZddZZS)TFTransfoXLLMHeadc s tjf||j|_||_dSr)r rrinput_embeddings)rrrrrrrrszTFTransfoXLLMHead.__init__cs(|j|jfdddd|_t|dS)NrTTbiasrU)rWrrr rXrYrrrrXszTFTransfoXLLMHead.buildcCs|j|dd}||j}|S)NZlinear)mode)rr)rZ hidden_statesrrrr s zTFTransfoXLLMHead.callrrrrrrs  rzThe Transformer-XL Model with a language modeling head on top (adaptive softmax with weights tied to the adaptive input embeddings)csNeZdZfddZddZddZddZeedd d Z d dZ Z S)TFTransfoXLLMHeadModelcsXt|t|dd|_|j|_|jdks4tdt|j|j|j |j |j dd|_ dS)NrrrzSampling from the softmax is not implemented yet. Please look at issue: #3310: https://github.com/huggingface/transformers/issues/3310crit)rwr*) r rrrrrrrrur.rvrwr)rrrrrrs zTFTransfoXLLMHeadModel.__init__cCs t|jjdkr|jjdSdS)z9 Double-check if you are using adaptive softmax. rrN)r~rZ out_layersrrrrget_output_embeddingss z,TFTransfoXLLMHeadModel.get_output_embeddingscCs|j|||dSr)rrrrrrr sz#TFTransfoXLLMHeadModel.reset_lengthcCs |j|Sr)rr)rrrrrrsz TFTransfoXLLMHeadModel.init_memsNFcCs~t|ttfr|d}t|dkr*|dn|}t|dkrB|dn|}t|dkrZ|dn|}t|dkrr|dn|}t|dkstdn^t|tr|d}|d |}|d |}|d |}|d |}t|dkstdn|}|d k r t|d d\}} nt|d d\}} |j||||g|d} | d} | d d | d f} | dd } |j | |g|d}|g| } | S)a Return: :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.TransfoXLConfig`) and inputs: prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`): Contains pre-computed hidden-states (key and values in the attention blocks). Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model should not be passed as input ids as they have already been computed. hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_hidden_states=True``): Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape :obj:`(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the initial embedding outputs. attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_attentions=True``): Tuple of :obj:`tf.Tensor` (one for each layer) of shape :obj:`(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Examples:: import tensorflow as tf from transformers import TransfoXLTokenizer, TFTransfoXLLMHeadModel tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103') model = TFTransfoXLLMHeadModel.from_pretrained('transfo-xl-wt103') input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1 outputs = model(input_ids) prediction_scores, mems = outputs[:2] rrr[rDrrrrhrirlabelsNr>) rrrr~rrrr rr)rrdrhrirrr?rrrQZtransformer_outputsZ last_hiddenZpred_hidrmZsoftmax_outputrrrr s2#         zTFTransfoXLLMHeadModel.callcKsd|i}|r||d<|S)Nrdrhr)rrdZpastZ model_kwargsrrrprepare_inputs_for_generationWsz4TFTransfoXLLMHeadModel.prepare_inputs_for_generation)NNNNF) r"r#r$rrrrrrr rr%rrrrrs  Cr)%rloggingZ tensorflowrZconfiguration_transfo_xlrZ file_utilsrrZ modeling_tf_transfo_xl_utilitiesrZmodeling_tf_utilsrrr r Ztokenization_utilsr getLoggerr"loggerrr1r2ZLayerr r&rCrnrsrrZTRANSFO_XL_START_DOCSTRINGrrrrrrrrsD    +;Hu ,