U 9%e?@s.dZddlZddlZddlmZmZmZmZddlZddlm Z ddl m Z m Z m Z ddlmZddlmZdd lmZmZmZmZmZmZdd lmZdd lmZmZmZmZm Z m!Z!m"Z"d d l#m$Z$e!%e&Z'dZ(dZ)Gddde j*Z+Gddde j*Z,Gddde j*Z-Gddde j*Z.Gddde j*Z/Gddde j*Z0Gddde j*Z1Gddde j*Z2Gd d!d!e j*Z3Gd"d#d#eZ4Gd$d%d%e4Z5d&Z6d'Z7d(Z8ed)e6Gd*d+d+e4Z9ed,e6Gd-d.d.e4Z:ed/e6Gd0d1d1e4Z;ed2e6Gd3d4d4e4Zcs&ttt||_||_dS)ze Construct a layernorm module in the UMT5 style. No bias and no subtraction of mean. N)super__init__r Parametertorchonesweightvariance_epsilon)self hidden_sizeeps __class__e/var/www/html/Darija-Ai-API/env/lib/python3.8/site-packages/transformers/models/umt5/modeling_umt5.pyr 8s zUMT5LayerNorm.__init__cCs\|tjdjddd}|t||j}|jjtj tj fkrR||jj}|j|S)NT)Zkeepdim) tor"Zfloat32powmeanZrsqrtr%r$dtypefloat16Zbfloat16)r& hidden_statesZvariancer+r+r,forward@s zUMT5LayerNorm.forward)r__name__ __module__ __qualname__r r5 __classcell__r+r+r)r,r7srcs*eZdZedfdd ZddZZS)UMT5DenseActDenseconfigcsTttj|j|jdd|_tj|j|jdd|_t|j |_ t |j |_ dSNFbias)rr rLineard_modeld_ffwiwoDropout dropout_ratedropoutr dense_act_fnactr&r=r)r+r,r Rs  zUMT5DenseActDense.__init__cCsl||}||}||}t|jjtjr^|j|jjjkr^|jjjtj kr^| |jjj}||}|SN) rDrJrH isinstancerEr$r"Tensorr2int8r/r&r4r+r+r,r5Ys    zUMT5DenseActDense.forwardr7r8r9rr r5r:r+r+r)r,r;Qsr;cs*eZdZedfdd ZddZZS)UMT5DenseGatedActDenser<csjttj|j|jdd|_tj|j|jdd|_tj|j|jdd|_t |j |_ t |j |_dSr>)rr rrArBrCwi_0wi_1rErFrGrHr rIrJrKr)r+r,r is  zUMT5DenseGatedActDense.__init__cCsz|||}||}||}||}t|jjtjrl|j |jjj krl|jjj tj krl| |jjj }||}|SrL) rJrSrTrHrMrEr$r"rNr2rOr/)r&r4Z hidden_geluZ hidden_linearr+r+r,r5qs   zUMT5DenseGatedActDense.forwardrQr+r+r)r,rRhsrRcs*eZdZedfdd ZddZZS) UMT5LayerFFr<csJt|jrt||_n t||_t|j|jd|_ t |j |_ dS)Nr()rr Z is_gated_actrRDenseReluDenser;rrBlayer_norm_epsilon layer_normrrFrGrHrKr)r+r,r s    zUMT5LayerFF.__init__cCs&||}||}|||}|SrL)rYrWrH)r&r4Zforwarded_statesr+r+r,r5s  zUMT5LayerFF.forwardrQr+r+r)r,rUs rUcs~eZdZdZdfdd ZejejdddZdd Zdd d Z deje eje e eje eje ejd ddZ Z S) UMT5Attentionz7 T5's attention using relative_attention_bias. Fcst|j|_||_|j|_|j|_|j|_|j|_|j |_ |j |_ |j |j|_ tj|j|j dd|_tj|j|j dd|_tj|j|j dd|_tj|j |jdd|_|jrt|j|j |_t|_dSr>)rr is_decoderhas_relative_attention_biasrelative_attention_num_bucketsrelative_attention_max_distancerBd_kvkey_value_proj_dim num_headsn_headsrGrHZ inner_dimrrAqkvo Embeddingrelative_attention_biassetZ pruned_heads)r&r=r\r)r+r,r s" zUMT5Attention.__init__) projectionreturncCs6|dd|j|jf}||dddd}|S)Nr.rr-rr )sizerbr`viewpermute)r&rjZnew_projection_shapeZnew_projectionr+r+r,_shapeszUMT5Attention._shapec Csd}|j}|j}|jsB|d}||dktj|7}t|}nt|t| }|d}||k}t | |t ||}|||}||tj}t|t ||d}|t |||7}|S)a Adapted from Mesh Tensorflow: https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593 Translate relative position to a bucket number for relative attention. The relative position is defined as memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for small absolute relative_position and larger buckets for larger absolute relative_positions. All relative positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket. This should allow for more graceful generalization to longer sequences than the model has been trained on Args: relative_position: an int32 Tensor bidirectional: a boolean - whether the attention is bidirectional num_buckets: an integer max_distance: an integer Returns: a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets) rr-r)r]r^r[r/r"longabsminZ zeros_likelogfloatmathZ full_likewhere) r&relative_positionZrelative_bucketsZ num_bucketsZ max_distanceZ max_exactZis_smallZ log_ratioZrelative_position_if_larger+r+r,_relative_position_buckets&   z'UMT5Attention._relative_position_bucketNc Cs|dkr|jjj}tj|tj|ddddf}tj|tj|ddddf}||}||}||}|dddgd}|S)z%Compute binned relative position biasN)r2devicer-rr) rhr$ryr"ZarangerprxrnZ unsqueeze) r& query_lengthZ key_lengthryZcontext_positionZmemory_positionrwZrelative_position_bucketvaluesr+r+r, compute_biass   zUMT5Attention.compute_bias)r4encoder_hidden_statespast_key_valueattention_masklayer_head_maskcCs|dk }|jdd\}}|dk r&|n|} |r\|r\|djd| jdkr\|d} |d} nX||| } ||| } |dk r|stj|d| gdd} tj|d| gdd} |||} t| | dd} |j r|}|dk r||djd7}|j || d| j d}n(tj d|j|| df| j | j|jd}|dk rr|dddd| d dddf}|dk r||}|jr| | f}| |7} tjj| dd| }tjj||j|jd }|dk r||}t|| }|dddd ||d}||}|||fS) Nr-rrdimr.ry)ryr2Z requires_grad)ptrainingr )shaperordrer"catrcmatmulZ transposer\r|rlryZzerosrbr2rr[rZ functionalZsoftmaxrtZtype_asrHrn contiguousrmrf)r&r4r}r~rrZis_cross_attention batch_size seq_lengthZcurrent_statesZ key_statesZ value_statesZ query_statesZattention_scoresrzZ position_biasZ attn_weightsZcontext_statesZ attn_outputr+r+r,r5sN    (    zUMT5Attention.forward)F)N)NNNN)r7r8r9__doc__r r"rNrorxr|rrr5r:r+r+r)r,rZs /  rZcs&eZdZfddZdddZZS)UMT5LayerSelfAttentioncs<tt|dd|_t|j|jd|_t |j |_ dS)NTr\rV) rr rZ SelfAttentionrrBrXrYrrFrGrHrKr)r+r,r @s zUMT5LayerSelfAttention.__init__NcCsD||}|j||||d}|||d}|f|dd}|S)Nrrr~rr)rYrrH)r&r4rrr~normed_hidden_statesattention_outputoutputsr+r+r,r5Fs zUMT5LayerSelfAttention.forward)NNNr6r+r+r)r,r?s  rcs&eZdZfddZdddZZS)UMT5LayerCrossAttentioncs<tt|dd|_t|j|jd|_t |j |_ dS)NFrrV) rr rZEncDecAttentionrrBrXrYrrFrGrHrKr)r+r,r Zs z UMT5LayerCrossAttention.__init__Nc CsF||}|j|||||d}|||d}|f|dd} | S)Nr}rrr~rr)rYrrH) r&r4r}rrr~rrZ layer_outputrr+r+r,r5`s zUMT5LayerCrossAttention.forward)NNNNr6r+r+r)r,rYs  rcs&eZdZfddZdddZZS) UMT5BlockcsVt|j|_t|_|jt||jrB|jt||jt |dSrL) rr r[r ModuleListlayerappendrrrUrKr)r+r,r vs  zUMT5Block.__init__NFc Cs|dk r|ddnd} |jd|||| d\}} } |jtjkr|t|jj} tt|| d| }tj || |d}d}d}|j o|dk }|r |dk r|ddnd}|jd|||||d\}}}|jtjkrt|jj} tt|| d| }tj || |d}| |7} |jd |}|jtjkrxt|jj} tt|| d| }tj || |d}|| f}| r|| |f7}|S) Nr-rri)rrmaxrrrr.) rr2r"r3Zfinforrvisinfanyclampr[)r&r4rr}encoder_attention_maskrcross_attn_layer_head_maskr~ use_cacheoutput_attentionsZself_attn_past_key_valueZself_attn_weightsZpresent_key_valueZ max_dtypeZ clamp_valueZcross_attn_present_key_valueZcross_attn_weightsZdo_cross_attentionZcross_attn_past_key_valuerr+r+r,r5sN    zUMT5Block.forward)NNNNNNFFr6r+r+r)r,rus rcs:eZdZdZedfdd ZejejdddZZ S)UMT5ClassificationHeadz-Head for sentence-level classification tasks.r<csBtt|j|j|_tj|jd|_t|j|j |_ dS)N)r) rr rrArBdenserFZclassifier_dropoutrH num_labelsout_projrKr)r+r,r s zUMT5ClassificationHead.__init__)r4rkcCs6||}||}t|}||}||}|SrL)rHrr"tanhrrPr+r+r,r5s      zUMT5ClassificationHead.forward) r7r8r9rrr r"rNr5r:r+r+r)r,rsrc@sNeZdZdZeZdZdZdgZdgZ e ddZ dd Z dd d Z d dZdS)UMT5PreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. transformerTrrEcCs$tt}tt}|||d}|S)N)decoder_input_ids input_idsdecoder_attention_mask)r"Ztensorrr)r&rZ input_mask dummy_inputsr+r+r,rs  z UMT5PreTrainedModel.dummy_inputscCs|jj}t|tr(|jj|dnt|ttt t fr|j jjj d|ddt |dr||jjs||jjjj d|ddt |dr|jjjj d||jjdd|jjjnt|tr\|jjjj d||jjddt |jdr|jjdk r|jjj|jjjj d||jjddt |jdr|jjdk r|jjjnPt|tr|jjjj d||jjddt |jdr|jjdk r|jjj|jjjj d||jjddt |jdr|jjdk r|jjjnt|tr|jjjj d||jjddt |jdrV|jjdk rV|jjj|jjjj d||jjddt |jdr|jjdk r|jjj|jjjj d||jjddt |jdr|jjdk r|jjjnt|tr|jj}|jj}|jj }|j!jjj d|||dd|j"jjj d||dd|j#jjj d||dd|j$jjj d|||dd|j%r|j&jjj d||dddS) zInitialize the weightsg?g)r1Zstdlm_head qa_outputsr@N)'r=Zinitializer_factorrMrr$dataZfill_ UMT5ModelUMT5ForConditionalGenerationUMT5EncoderModelUMT5ForQuestionAnsweringsharedZnormal_hasattrtie_word_embeddingsrrrBr@Zzero_rrrr;rDrErCrRrSrTrZr_rarcrdrerfr\rh)r&modulefactorrBr`rbr+r+r, _init_weightssf                 z!UMT5PreTrainedModel._init_weightsFcCst|ttfr||_dSrL)rMrZ UMT5Stackgradient_checkpointing)r&rvaluer+r+r,_set_gradient_checkpointing/sz/UMT5PreTrainedModel._set_gradient_checkpointingcCs|jj}|jj}|dkr tdt|rbt|jddd|}tj||dddfgdd}n4| |j}|dddf |dddf<||d<|dkrtd| |d k||S) Nzself.model.config.decoder_start_token_id has to be defined. In UMT5 it is usually set to the pad_token_id.See UMT5 docs for more information.r.)r.rr).rz1self.model.config.pad_token_id has to be defined.) r=decoder_start_token_id pad_token_id ValueErrorrr"fullrrZ new_zeroscloneZ masked_fill_)r&rrrZshifted_input_idsr+r+r, _shift_right3s    z UMT5PreTrainedModel._shift_rightN)F)r7r8r9rr config_classZbase_model_prefixZsupports_gradient_checkpointingZ_no_split_modulesZ_keep_in_fp32_modulespropertyrrrrr+r+r+r,rs > rc s8eZdZd fdd ZddZddZd dd ZZS) rNcslt||_j|_tfddtjD|_t j j d|_ t j|_d|_|dS)Ncsg|] }tqSr+)r).0ir<r+r, Tsz&UMT5Stack.__init__..rVF)rr embed_tokensr[rrrange num_layersblockrrBrXfinal_layer_normrFrGrHr post_init)r&r=rr)r<r,r Ps  zUMT5Stack.__init__cCs|jSrLrr&r+r+r,get_input_embeddings\szUMT5Stack.get_input_embeddingscCs ||_dSrLrr&Znew_embeddingsr+r+r,set_input_embeddings_szUMT5Stack.set_input_embeddingsc % sdk r n|jjdk r n|jj| dk r4| n|jj} | dk rH| n|jj} |dk r|dk r|jrjdnd} td| d| dn`|dk r|}|d|d}n>|dk r|dd}n$|jrdnd} td| d| d |dkr|j dkrtd | |}|\}}|dk r6|d d j d |n|}d kr\|js\td|d|dkrxt j |||j d}|jr|dkr|dk r|j d}t j |||j t jd}|dkrdgt|j}|||}|jr(|dk r(|\}}}||f}|dkrt j ||j d}||}nd}|jrP|jrPrPtdd|||jj}|||jj}rzdnd}| rdnd}rdnd}r|jrdnd}||}tt|j|D]\}\}} ||}!||}"| r||f}|jr0|jr0fdd}#t|#||||||!|"d}$n"||||||!|"| d }$|$d }rf||$df7}r||$d f7}|jr||$df7}q||}||}| r||f}| stdd|||||fDSt|||||dS)NZdecoder_zYou cannot specify both zinput_ids and zinputs_embeds at the same timer.zYou have to specify either z input_ids or inputs_embedsz.create_custom_forward..custom_forwardr+)rrrr)rr,create_custom_forwardsz0UMT5Stack.forward..create_custom_forward)rr}rrrr~rrr css|]}|dk r|VqdSrLr+)rrer+r+r, sz$UMT5Stack.forward..)last_hidden_statepast_key_valuesr4 attentionscross_attentions) r=rroutput_hidden_statesuse_return_dictr[rrlrmrrr"r#ryrplenrZget_extended_attention_maskZinvert_attention_maskrrloggerZ warning_onceZ get_head_maskrrH enumeratezipr rrr)%r&rrr}rr head_maskcross_attn_head_maskrrrr return_dictZerr_msg_prefixZ input_shaperrZmask_seq_lengthZencoder_seq_lengthZextended_attention_maskZencoder_batch_sizeZencoder_sequence_length_Zencoder_hidden_shapeZencoder_extended_attention_maskZpresent_key_value_statesZall_hidden_statesZall_attentionsZall_cross_attentionsr4rZ layer_moduler~rrrZ layer_outputsr+rr,r5bs   $               zUMT5Stack.forward)N) NNNNNNNNNNNN)r7r8r9r rrr5r:r+r+r)r,rOs  ra The UMT5 model was proposed in [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu. It's an encoder decoder transformer pre-trained in a text-to-text denoising generative setting. This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`UMT5Config`]): 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 [`~PreTrainedModel.from_pretrained`] method to load the model weights. ah Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. UMT5 is a model with relative position embeddings so you should be able to pad the inputs on both the right and the left. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for detail. [What are input IDs?](../glossary#input-ids) To know more on how to prepare `input_ids` for pretraining take a look a [UMT5 Training](./umt5#training). attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are decoder input IDs?](../glossary#decoder-input-ids) UMT5 uses the `pad_token_id` as the starting token for `decoder_input_ids` generation. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). To know more on how to prepare `decoder_input_ids` for pretraining take a look at [UMT5 Training](./umt5#training). decoder_attention_mask (`torch.BoolTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default. head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules in the encoder. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. decoder_head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules in the decoder. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. cross_attn_head_mask (`torch.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*): Tuple consists of (`last_hidden_state`, `optional`: *hidden_states*, `optional`: *attentions*) `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)` is a sequence of hidden states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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. decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be input (see `past_key_values`). This is useful if you want more control over how to convert `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix. If `decoder_input_ids` and `decoder_inputs_embeds` are both unset, `decoder_inputs_embeds` takes the value of `inputs_embeds`. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. a Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. UMT5 is a model with relative position embeddings so you should be able to pad the inputs on both the right and the left. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for detail. To know more on how to prepare `input_ids` for pretraining take a look a [UMT5 Training](./umt5#training). attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. z^The bare UMT5 Model transformer outputting raw hidden-states without any specific head on top.cs eZdZdZdZeZddgZfddZddZ d d Z d d Z d dZ ddZ eeeeeddeejeejeejeejeejeejeejeeeejeeeejeejeejeeeeeeeeeeejefdddZZS)rao Examples: ```python >>> from transformers import UMT5Model, AutoTokenizer >>> model = UMT5Model.from_pretrained("google/umt5-small") >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> noisy_text = "UN Offizier sagt, dass weiter werden muss in Syrien." >>> label = " verhandelt" >>> inputs = tokenizer(inputs, return_tensors="pt") >>> labels = tokenizer(label=label, return_tensors="pt") >>> outputs = model(input_ids=inputs["input_ids"], decoder_input_ids=labels["input_ids"]) >>> hidden_states = outputs.last_hidden_state ```Zuumt5encoder.embed_tokens.weightdecoder.embed_tokens.weightcst|t|j|j|_t|}d|_ d|_ d|_ t ||j|_ t|}d|_ d|_ |j|_t ||j|_|dSNFT)rr rrg vocab_sizerBrcopydeepcopyr[ris_encoder_decoderrencodernum_decoder_layersrdecoderrr&r=encoder_configZdecoder_configr)r+r,r s   zUMT5Model.__init__cCs|jSrLrrr+r+r,rszUMT5Model.get_input_embeddingscCs"||_|j||j|dSrLrrrrrr+r+r,rs zUMT5Model.set_input_embeddingscCs|jSrLrrr+r+r, get_encoderszUMT5Model.get_encodercCs|jSrLrrr+r+r, get_decoderszUMT5Model.get_decodercCs*|D]\}}|jj|j|qdS) Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel N)itemsrrZ attention prune_headsr&Zheads_to_prunerZheadsr+r+r, _prune_headsszUMT5Model._prune_heads output_typerN)rrrrrdecoder_head_maskrencoder_outputsrrdecoder_inputs_embedsrrrrrkcCs| dk r | n|jj} |dk r |n|jj}|dkrJ|j||| || ||d}nH|rt|tst|dt|dkrt|dndt|dkr|dndd}|d}|j||| | ||||| | ||d }|s||St|j |j |j |j |j |j |j |j dS) a Returns: Example: ```python >>> from transformers import AutoTokenizer, UMT5Model >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> model = UMT5Model.from_pretrained("google/umt5-small") >>> input_ids = tokenizer( ... "Studies have been shown that owning a dog is good for you", return_tensors="pt" ... ).input_ids # Batch size 1 >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids # Batch size 1 >>> # preprocess: Prepend decoder_input_ids with start token which is pad token for UMT5Model. >>> # This is not needed for torch's UMT5ForConditionalGeneration as it does this internally using labels arg. >>> decoder_input_ids = model._shift_right(decoder_input_ids) >>> # forward pass >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids) >>> last_hidden_states = outputs.last_hidden_state ```Nrrrrrrrrrr-rr4r rrrrr}rrrrrrr)rrdecoder_hidden_statesdecoder_attentionsrencoder_last_hidden_stater}encoder_attentions)r=rrrrMr rrrrrr4rr)r&rrrrrrrr rrr rrrrr4decoder_outputsr+r+r,r5sZ, zUMT5Model.forward)NNNNNNNNNNNNNNN)r7r8r9r model_typerr_tied_weights_keysr rrrrrrUMT5_INPUTS_DOCSTRINGrr_CONFIG_FOR_DOCrr" LongTensor FloatTensor BoolTensorrNrboolrr5r:r+r+r)r,rsZ  rz2UMT5 Model with a `language modeling` head on top.cs>eZdZdZdZdddgZfddZdd Zd d Zd d Z ddZ ddZ ddZ e eeeedd eejeejeejeejeejeejeejeeeejeeeejeejeejeejeeeeeeeeeeejefdddZd!ddZejdddZeddZZ S)"ra Examples: ```python >>> from transformers import UMT5ForConditionalGeneration, AutoTokenizer >>> model = UMT5ForConditionalGeneration.from_pretrained("google/umt5-small") >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> article = "UN Offizier sagt, dass weiter verhandelt werden muss in Syrien." >>> summary = "Weiter Verhandlung in Syrien." >>> inputs = tokenizer(article, text_target=summary, return_tensors="pt") >>> outputs = model(**inputs) >>> loss = outputs.loss ```umt5rrzlm_head.weightcst||j|_t|j|j|_t |}d|_ d|_ d|_ t ||j|_t |}d|_ d|_ |j|_t ||j|_tj|j|jdd|_|dS)NFTr?)rr rB model_dimrrgrrrrr[rrrrrrrrArrrr)r+r,r bs   z%UMT5ForConditionalGeneration.__init__cCs|jSrLrrr+r+r,rzsz1UMT5ForConditionalGeneration.get_input_embeddingscCs"||_|j||j|dSrLrrr+r+r,r~s z1UMT5ForConditionalGeneration.set_input_embeddingscCs ||_dSrLrrr+r+r,set_output_embeddingssz2UMT5ForConditionalGeneration.set_output_embeddingscCs|jSrLrrr+r+r,get_output_embeddingssz2UMT5ForConditionalGeneration.get_output_embeddingscCs|jSrLrrr+r+r,rsz(UMT5ForConditionalGeneration.get_encodercCs|jSrLrrr+r+r,rsz(UMT5ForConditionalGeneration.get_decoderrN)rrrrrrrr rrr labelsrrrrrkcCs| dk r | n|jj} |dk r |n|jj}|dkrJ|j||| ||||d}nH|rt|tst|dt|dkrt|dndt|dkr|dndd}|d}| dk r|dkr| dkr|| }|j||| | ||||| |||d }|d}|jj r||j d}| |}d}| dk rHt d d }| |j} ||d |d | d }|s||f|dd|}|dk rx|f|S|St|||j|j|j|j|j|j|jd S) a labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ..., config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` Returns: Examples: ```python >>> from transformers import AutoTokenizer, UMT5ForConditionalGeneration >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> model = UMT5ForConditionalGeneration.from_pretrained("google/umt5-small") >>> # training >>> input_ids = tokenizer("The walks in park", return_tensors="pt").input_ids >>> labels = tokenizer(" cute dog the ", return_tensors="pt").input_ids >>> outputs = model(input_ids=input_ids, labels=labels) >>> loss = outputs.loss >>> logits = outputs.logits >>> # inference >>> input_ids = tokenizer("Studies have shown that good for you", return_tensors="pt").input_ids >>> outputs = model.generate(input_ids) >>> tokenizer.decode(outputs[0], skip_special_tokens=True) ```Nr rrr-r r rrZ ignore_indexr. losslogitsrrrrrr}r)r=rrrrMr rrrrrrrr/ryrmrlrrr4rrr)r&rrrrrrrr rrr r rrrrr4rsequence_outputZ lm_logitsr#loss_fctoutputr+r+r,r5st1      z$UMT5ForConditionalGeneration.forwardc Ks4|dk r|ddddf}||| ||||||d S)Nr.) rrr rrrrrrr+) r&rrrrrrrrr kwargsr+r+r,prepare_inputs_for_generationsz:UMT5ForConditionalGeneration.prepare_inputs_for_generation)r cCs ||SrL)r)r&r r+r+r,%prepare_decoder_input_ids_from_labels/szBUMT5ForConditionalGeneration.prepare_decoder_input_ids_from_labelscs.d}|D] }|tfdd|Df7}q|S)Nr+c3s"|]}|d|jVqdS)rN)Z index_selectr/ry)rZ past_statebeam_idxr+r,r7sz>UMT5ForConditionalGeneration._reorder_cache..r)rr,Zreordered_pastZ layer_pastr+r+r,_reorder_cache2s z+UMT5ForConditionalGeneration._reorder_cache)NNNNNNNNNNNNNNNN)NNNNNNNN)!r7r8r9rrrr rrrrrrrrrrrrr"rrrrNrrrr5r)r* staticmethodr-r:r+r+r)r,rMsv    rzhThe bare UMT5 Model transformer outputting encoder's raw hidden-states without any specific head on top.c seZdZdZdZdgZfddZddZdd Zd d Z d d Z e e e eeddeejeejeejeejeeeeeeeeejefdddZZS)ra Examples: ```python >>> from transformers import UMT5EncoderModel, AutoTokenizer >>> model = UMT5EncoderModel.from_pretrained("google/umt5-small") >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> article = "UN Offizier sagt, dass weiter verhandelt werden muss in Syrien." >>> input_ids = tokenizer(article, return_tensors="pt").input_ids >>> outputs = model(input_ids) >>> hidden_state = outputs.last_hidden_state ```rrcsNt|t|j|j|_t|}d|_ d|_ t ||j|_ | dSNF)rr rrgrrBrrrrrrrr)r&r=rr)r+r,r Ss  zUMT5EncoderModel.__init__cCs|jSrLrrr+r+r,r`sz%UMT5EncoderModel.get_input_embeddingscCs||_|j|dSrL)rrrrr+r+r,rdsz%UMT5EncoderModel.set_input_embeddingscCs|jSrLrrr+r+r,riszUMT5EncoderModel.get_encodercCs0|D]"\}}|jj|jdj|qdS)rrN)rrrrrrrr+r+r,rmszUMT5EncoderModel._prune_headsrN)rrrrrrrrkc Cs0|dk r |n|jj}|j|||||||d}|S)a5 Returns: Example: ```python >>> from transformers import AutoTokenizer, UMT5EncoderModel >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> model = UMT5EncoderModel.from_pretrained("google/umt5-small") >>> input_ids = tokenizer( ... "Studies have been shown that owning a dog is good for you", return_tensors="pt" ... ).input_ids # Batch size 1 >>> outputs = model(input_ids=input_ids) >>> last_hidden_states = outputs.last_hidden_state ```Nr )r=rr) r&rrrrrrrr r+r+r,r5us zUMT5EncoderModel.forward)NNNNNNN)r7r8r9rrrr rrrrrUMT5_ENCODER_INPUTS_DOCSTRINGrr rrr"rrrrrr5r:r+r+r)r,r<s6  rz UMT5 model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. cseZdZdgZddgZedfdd Zeee e e dd e j ee jee j ee j ee jee jee jeee jee jee jee j eeeeeeeeeee fd d d ZZS) UMT5ForSequenceClassificationzFdecoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weightrrr<cs2t|t||_t||_|d|_dSr/)rr rrrclassification_headrZmodel_parallelrKr)r+r,r s    z&UMT5ForSequenceClassification.__init__rN)rrrrrrrr rr r rrrrrkcCsx|dk r |n|jj}| dk r d} |dkrB| dk rBtd|jj|dkrl| dkrl|dkrbtd||}|j||||||||| | | | ||d}|d}||jj  |j }t t |ddkrtd|j\}}}||ddf|d |ddd ddf}||}d}| dk r | |j } |jjdkr|jjdkrRd |j_n<|jjdkr| jt jks|| jt jkrd |j_nd |j_|jjd krt}|jjdkr||| }n ||| }nP|jjd krt}||d |jj| d }n|jjd kr t}||| }|sP|f|dd}|dk rL|f|S|St|||j|j|j|j |j!|j"|j#d S)aD labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). Returns: NFz8Passing input embeddings is currently not supported for If no `decoder_input_ids` or `decoder_inputs_embeds` are passed, `input_ids` cannot be `None`. Please pass either `input_ids` or `decoder_input_ids` or `decoder_inputs_embeds`.) rrrrrrr rr rrrrrrz7All examples must have the same number of tokens.r.Z regressionZsingle_label_classificationZmulti_label_classificationr")$r=rNotImplementedErrorr*r7rrreqZ eos_token_idr/ryrr"Zunique_consecutivesumrrmr2Z problem_typerr2rpintr squeezerrrrrrrrr}r)r&rrrrrrrr rr r rrrrrr%Zeos_maskrrr'Zsentence_representationr$r#r&r'r+r+r,r5s   ,    *   z%UMT5ForSequenceClassification.forward)NNNNNNNNNNNNNNN)r7r8r9Z"_keys_to_ignore_on_load_unexpectedrrr rrrrrr"rrrNrrrrrr5r:r+r+r)r,r1sL    r1z UMT5 Model with a span classification head on top for extractive question-answering tasks like SQuAD (linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`). cseZdZddgZfddZddZddZd d Zd d Ze e e e e d deejeejeejeejeejeejeejeeeejeejeejeejeejeeeeeeeeeeeje fdddZZS)rrrcst||j|_t|j|j|_t |}d|_ d|_ d|_ t ||j|_t |}d|_ d|_ |j|_t ||j|_|j|_t|j|j|_|dSr)rr rBrrrgrrrrr[rrrrrrrrrArrrr)r+r,r 1s    z!UMT5ForQuestionAnswering.__init__cCs|jSrLrrr+r+r,rJsz-UMT5ForQuestionAnswering.get_input_embeddingscCs"||_|j||j|dSrLrrr+r+r,rNs z-UMT5ForQuestionAnswering.set_input_embeddingscCs|jSrLrrr+r+r,rTsz$UMT5ForQuestionAnswering.get_encodercCs|jSrLrrr+r+r,rXsz$UMT5ForQuestionAnswering.get_decoderrN)rrrrrrrr start_positions end_positionsrr rrrrrkcCsn|dk r |n|jj}| dk r | n|jj} | dk r<| dk rrrrutypingrrrrr"rZtorch.nnrrr Ztorch.utils.checkpointr Z activationsr Zmodeling_outputsr rrrrrZmodeling_utilsrutilsrrrrrrrZconfiguration_umt5rZ get_loggerr7rrZ_CHECKPOINT_FOR_DOCModulerr;rRrUrZrrrrrrZUMT5_START_DOCSTRINGrr0rrrr1rr+r+r+r,sn     $  (Su>]&&oa