Uh*XdZddlZddlZddlZddlmZmZmZmZddl Z ddl Z ddl m Z ddl m Z ddlmZmZmZddlmZdd lmZdd lmZmZmZdd lmZdd lmZmZmZdd l m!Z!m"Z"m#Z#ddl$m%Z%e"jLe'Z(dZ)Gdde jTZ+Gdde jTZ,Gdde jTZ-Gdde jTZ.e!GddeZ/e!Gdde/Z0e!dGdd e/eZ1e!d!Gd"d#e/Z2gd$Z3y)%zPyTorch OpenAI ImageGPT model.N)AnyOptionalTupleUnion)nn)autocast)BCEWithLogitsLossCrossEntropyLossMSELoss)ACT2FN)GenerationMixin))BaseModelOutputWithPastAndCrossAttentions!CausalLMOutputWithCrossAttentions SequenceClassifierOutputWithPast)PreTrainedModel)Conv1D find_pruneable_heads_and_indicesprune_conv1d_layer)auto_docstringlogging torch_float)ImageGPTConfigc ddl}ddl}t j j|}tjdj||jj|}g}g}|D]v\} } tjdj| | |jj|| } |j| |j| jxt||D]\} } | dd} | j!d} t#d| Ds| d d vr4tjd jdj%| j|} | d d vr t'| d } | D]W} |j)d| r|j!d| }n| g}|ddk(s|ddk(r t'| d} n|ddk(r t'| d} n|ddk(s|ddk(rt'| |d} t'| d} n|ddvrt'| d} t'| d} nt+| dk(r,| ddk(r$|ddk(rt'| |d} t'| d} nQ|ddk(rt'| d} t'| d} n0|ddk(rt'| d} t'| d} nt'| |d} t+|d k\sEt-|d}| |} Zt+| dkDr| ddk(s| d dk(s| d dk(s| d dk(rn | j.| j.k(sJ tjd!j| | d d"k(rbt5j6| j9|j:|j:j<| j>ddd|j:f<| d d#k(rot5j6| j9|j:|j:j<| j>dd|j:d |j:zf<-| d d$k(ret5j6| j9|j:|j:j<| j>ddd |j:zdf<t+| dk(rP| ddk(rH| d dk(r@t5j6| j9|j:|j:| _| d dk(rt5j6| | _| d dk(r7t5j6| | j>d|j@dz ddf<[| d dk(r$t5j6| | j>d <t5j6| | _|S#t$rtj dwxYw#t0$r1}|xj2| j.| j.fz c_d}~wwxYw)%z0 Load tf checkpoints in a pytorch model rNzLoading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see https://www.tensorflow.org/install/ for installation instructions.z(Converting TensorFlow checkpoint from {}z"Loading TF weight {} with shape {}/c3$K|]}|dv yw))adam_vadam_mAdamWeightDecayOptimizerAdamWeightDecayOptimizer_1 global_stepN).0ns /var/www/catia.catastroantioquia-mas.com/valormas/lib/python3.12/site-packages/transformers/models/imagegpt/modeling_imagegpt.py z.load_tf_weights_in_imagegpt..Qs  n n )_stepz Skipping {})wtet transformerz [A-Za-z]+\d+z(\d+)wgweightbbiaswpewte)q_projk_projv_projc_attnr rattnc_projr,lm_headsoszInitialize PyTorch weight {}r5r6r7)!re tensorflow ImportErrorloggererrorospathabspathinfoformattrainlist_variables load_variableappendsqueezezipsplitanyjoingetattr fullmatchlenintshapeAssertionErrorargstorch from_numpyreshapen_embdTdata vocab_size)modelconfigimagegpt_checkpoint_pathr>tftf_path init_varsnamesarraysnamerUarraypointerm_name scope_namesnumes r'load_tf_weights_in_imagegptrn0s ggoo67G KK:AA'JK''0I E F ' e 8??eLM&&w5 T emmo& ' 5&)L3 eABxzz#    "X " KK ,,SXXd^< =  88 #g}5G 'F||OV4 hhx8 %h 1~$ A#(=!'84Q3&!'62Q5(KNe,C!';q>:!'84Q#AA!'84!'84TaDGv$5+a.H:T!';q>:!'84Q6)!'95!'84Q5(!'51!'84!';q>:;1$+a.)!#,; '> t9q=T!W.$r(f2DRTYHY]abd]ein]n  }} 333  299$?@ 8x /4/?/? fmm]c]j]j@k/l/n/nGLLOfmmO+ , "X !AFAQAQ fmmV]];Ba LLFMMA ,=== >"X !383C3CEMMRXR_R_agananDo3p3r3rGLLA -// 0 Y!^Q6 1d1g6I ++EMM&--,WXGL "X  ++E2GL "X 7<7G7G7NGLL06,,q00!3 4 "X $//6GLL  ++E2GLYL3\ LC   Q   P" 7==%++66 s#V?V< V9< W6,W11W6cheZdZddeedeffd ZdejdejfdZ xZ S)ImageGPTLayerNorm hidden_sizeepsct|||_tjt j ||_yN)super__init__rrr ParameterrXTensorr0)selfrqrr __class__s r'rvzImageGPTLayerNorm.__init__s. ll5<< #<= tensorreturnc|tjtjtj|dd|jzz }||j z}|S)Nr*T)axiskeepdim)rXsqrtmeansquarerrr0)ryr|s r'forwardzImageGPTLayerNorm.forwardsK%**UZZ V0D2W[%\_c_g_g%ghh$++% r{)gh㈵>) __name__ __module__ __qualname__rrTfloatrvrXrxr __classcell__rzs@r'rprps5>E#J>U> ellu||r{rpc"eZdZddeedeeffd ZdZddZddZ dZ dZ dd e jd eed ee jd ee jd ee jdee jdeedeedefdZxZS)ImageGPTAttentionis_cross_attention layer_idxc t||j}|jdt j t j ||ftjjdd||d|jdt jdd|j|_ |j|_ |j|jz|_|j|_|j|jz|jk7r&t!d|jd |jd |j"|_||_|j&|_||_|j*|_|j$rNt-d |jz|j|_t-|j|j|_n(t-d |jz|j|_t-|j|j|_t5j6|j8|_t5j6|j<|_tA|_!y) Nr2dtyperF) persistent masked_biasgz=`embed_dim` must be divisible by num_heads (got `embed_dim`: z and `num_heads`: z).r=r )"rurvmax_position_embeddingsregister_bufferrXtrilonesboolviewr|rq embed_dimnum_attention_heads num_headshead_dim split_size ValueErrorscale_attn_weightsrscale_attn_by_inverse_layer_idxrreorder_and_upcast_attnrr8q_attnr:rDropout attn_pdrop attn_dropout resid_pdrop resid_dropoutset pruned_heads)ryr`rr max_positionsrzs r'rvzImageGPTAttention.__init__s 66    JJuzz=-"@ S T Y Y1m]    ]ELL,>5Q++33$..8 .. ==4>> )T^^ ;OPTP^P^O_`NN#2'  #)";";"406/U/U,"'-'E'E$  " " T^^!3T^^DDK @DK T^^!3T^^DDKT^^T^^< JJv'8'89ZZ(:(:;Er{cFt|dk(ryt||j|j|j\}}t j |||jz|d|jzzg}t|j|d|_ t|j|d|_ |j|jz|jt|z z|_|jt|z |_|jj||_y)Nrr=rdim) rSrrrrrXcatrrr8r:union)ryheadsindex index_attns r' prune_headszImageGPTAttention.prune_headss u:? 7t~~t}}^b^o^op uYYut'>T__I\@]^_ )jaH (eC  ??dnn<RUV[R\A\]#e*4 --33E:r{cDtj||jdd}|jr |t |j ddzz }|j r|t|jdzz }|js|j d|j d}}|jdddd||z |d|f} tj|jj} tj| |j|j} tj | || }|||z}t#j$d|}|j'|j}|j)|}|||z}tj||} | |fS)Nr*?rrdevicer)rXmatmul transposerrsizerrrrr2finforminr|rwhererSoftmaxtyper) ryquerykeyvalueattention_mask head_mask attn_weights query_length key_length causal_mask mask_value attn_outputs r'_attnzImageGPTAttention._attnst||E3==R+@A  " "'+ejjn6K*LLL  / /'%0B*CCL&&',zz"~sxx|*L))Aq*|*Cj*PR]S]R]$]^K\%7%78<sEKKM#2&)^)DD i(~~aAq))r{c|jddddj}|jdd||zfz}|j|S)zS Merges attn_head_size dim and num_attn_heads dim into hidden_size rr=rr Nr)r contiguousrrrs r' _merge_headszImageGPTAttention._merge_headsFsO1a+668KKM#2&)n*D)FF {{9%%r{ hidden_states layer_pastrrencoder_hidden_statesencoder_attention_mask use_cacheoutput_attentionsr}c |Zt|ds td|j|} |j|j |j d\} } |}n0|j|j |j d\} } } |j | |j|j} |j | |j|j} |j | |j|j} |7|\} } tj| | fd} tj| | fd} |dur| | f}nd}|jr|j| | | ||\}}n|j| | | ||\}}|j||j|j}|j|}|j!|}||f}|r||fz }|S)NrzIf class is used as cross attention, the weights `q_attn` have to be defined. Please make sure to instantiate class with `ImageGPTAttention(..., is_cross_attention=True)`.r=rrT)hasattrrrr8rNrrrrrXrrrrrr:r)ryrrrrrrrrrrrpast_key past_valuepresentrroutputss r'rzImageGPTAttention.forwardNs ! ,4* t KK .E%:;AA$//WXAYJC3N $ M : @ @VW @ X E3!!%GT^^T]]C!!%G  !#- Hj))XsO4CIIz51r:E  ElGG  ' '(,(G(GsTY[ikt(u %K(, 5#unV_(` %K'' T^^T]]S kk+. ((5 (    &Gr{)FN)NNNNNNNFF)rrrrrrTrvrrrrrrXrxtuplerrrs@r'rrs)"8D>)"V^_bVc)"V ;$)L2)h*&&*15,08<9=$),13||3TN3!. 3 ELL) 3 ( 5 3!) 63D>3$D>3 3r{rcVeZdZfdZdej dej fdZxZS) ImageGPTMLPct||j}t|||_t|||_t |j|_tj|j|_ yrt) rurvrqrc_fcr:r activation_functionactrrrdropout)ryintermediate_sizer`rrzs r'rvzImageGPTMLP.__init__s_ && ,i8 Y(9: &445zz&"4"45 r{rr}c|j|}|j|}|j|}|j|}|Srt)rrr:r)ryrs r'rzImageGPTMLP.forwards@ -0 /  M2  ]3 r{)rrrrvrXrxrrrs@r'rrs#6U\\ellr{rceZdZd fd Z d dej deedeej deej deej deej deed eed efd Z xZ S) ImageGPTBlockct||j}|j |jnd|z}t ||j |_t|||_t ||j |_ |jr/t|d||_ t ||j |_ t|||_y)NrrrT)rr)rurvrqn_innerrplayer_norm_epsilonln_1rr9ln_2add_cross_attentioncrossattention ln_cross_attnrmlp)ryr`rrq inner_dimrzs r'rvzImageGPTBlock.__init__s (( &,nn&@FNNa+o %kv7P7PQ %f B %kv7P7PQ  % %"3Ft_h"iD !2;FD]D]!^D y&1r{rrrrrrrrr}c |} |j|}|j||||||} | d} | dd} | | z}|Wt|dstd|d|} |j |}|j ||||||} | d} | | z}| | ddz} |} |j |}|j|}| |z}|f|r| z} | S| ddz} | S) N)rrrrrrrr z'If `encoder_hidden_states` are passed, z` has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`)rrrrrr=)rr9rrr r r r )ryrrrrrrrrresidual attn_outputsrrcross_attn_outputsfeed_forward_hidden_statess r'rzImageGPTBlock.forwardsN! -0 yy !)/ ! #1o qr"#h. ,4!12 =dVDZZ%H ..}=M!%!4!4-#&;'="3 "5" -Q/K${2M 212 66G  -0 %)XXm%<" #==  "gLAH Lr{rtr) rrrrvrXrxrrrrrrs@r'rrs 2$&*15,08<9=$),18||8TN8!. 8 ELL) 8 ( 5 8!) 68D>8$D>8 8r{rc>eZdZeZeZdZdZdZ dgZ fdZ dZ xZ S)ImageGPTPreTrainedModelr- input_idsTrc$t||i|yrt)rurv)ryinputskwargsrzs r'rvz ImageGPTPreTrainedModel.__init__s &+F+r{c t|tjtfrl|jj j d|jj|j|jj jnt|tjry|jj j d|jj|jg|jj |jjn5t|tr%|jj jd|jD]m\}}d|vs d|vs|j j d|jjt!j"d|jj$zz oy)zInitialize the weights.g)rstdNrr:r0r=) isinstancerLinearrr0r]normal_r`initializer_ranger2zero_ Embedding padding_idxrpfill_named_parametersmathrn_layer)rymodulergps r' _init_weightsz%ImageGPTPreTrainedModel._init_weightssQ fryy&1 2 MM   & &CT[[5R5R & S{{&   &&(  - MM   & &CT[[5R5R & S!!- ""6#5#56<<>  1 2 MM   $ $S )..0 sGD!4H$4Cdkk.K.KdiiXY\`\g\g\o\oXoNp.pr sr{)rrrr config_classrnload_tf_weightsbase_model_prefixmain_input_namesupports_gradient_checkpointing_no_split_modulesrvr)rrs@r'rrs2!L1O%!O&*#(),sr{rc"eZdZdeffd ZdZdZdZe dde e jde e e e jde e jd e e jd e e jd e e jd e e jd e e jde e jde e de e de e de e dedee effdZxZS) ImageGPTModelr`c vt|||j|_t j |j |j|_t j |j|j|_ t j|j|_ t jt|jDcgc]}t!||c}|_t%|j|j&|_d|_d|_d|_|j1ycc}w)NrrF)rurvrqrrr!r^r4rr3r embd_pdropdrop ModuleListrangenum_hidden_layersrhrprln_fmodel_parallel device_mapgradient_checkpointing post_init)ryr`irzs r'rvzImageGPTModel.__init__ s  ++<< 1 14>>B<< > >OJJv001 ERXRjRjLklq f Blm%dnn&:S:ST $&+#  ms D6c|jSrtr4rys r'get_input_embeddingsz"ImageGPTModel.get_input_embeddingss xxr{c||_yrtr@rynew_embeddingss r'set_input_embeddingsz"ImageGPTModel.set_input_embeddingss !r{c|jD]-\}}|j|jj|/y)zv Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} N)itemsr8r9r)ryheads_to_prunelayerrs r' _prune_headszImageGPTModel._prune_heads"s<+002 2LE5 FF5M   * *5 1 2r{rpast_key_valuesrtoken_type_ids position_idsr inputs_embedsrrrroutput_hidden_states return_dictrr}c $d|vr8tjdt| td|j d}| | n|j j } | | n|j j} | | n|j j} | | n|j j} | | td|G|j|||j}|jd|d}|jd}n0|#|jdd}|jd}n td| |jn |j}||jd|d}|%d}tdgt!|j"z}n|ddjd }|>t%j&||d|zt$j(| }|j+d}|z|dkr td |j|d}|ddddddf}|j-|j. }d |z t%j0|j.j2z}|j j4rE|C|j\}}}||f}| t%j6||} |j9| } nd} |j;||j j<}||j?|}|jA|}||j-|jz$||j?|}$|z$|jC$$|$jdfz}|jDr%|jFr| rtHjKdd} | rdnd}| rdnd}| r|j j4rdnd}| rdnd}tMtO|j"|D]\}\}} |jPrt$jRjU$j| t$fd| D} ||j-$j}tW|t$jXr|j-$j}| r|$fz}|jDr3|jFr'|j[|j\$d||||| | | }!n|$| ||||| | | }!|!d$| dur ||!dfz}| r0||!| rdndfz}|j j4r ||!| rdndfz}|jPsS|j^jaD]J\}"}#||#dk(sdtc|"z|jdk7s+$j-dtc|"dzz$L|jg$$$j|$| r|$fz}| std$||||fDSti$||||S)a input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`): `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary. If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as `input_ids`. Indices can be obtained using [`AutoImageProcessor`]. See [`ImageGPTImageProcessor.__call__`] for details. Examples: ```python >>> from transformers import AutoImageProcessor, ImageGPTModel >>> from PIL import Image >>> import requests >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> image_processor = AutoImageProcessor.from_pretrained("openai/imagegpt-small") >>> model = ImageGPTModel.from_pretrained("openai/imagegpt-small") >>> inputs = image_processor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> last_hidden_states = outputs.last_hidden_state ``` pixel_values`The `pixel_values` argument is deprecated and will be removed in v4.47, use `input_ids` instead.N_You cannot pass both `pixel_values` and `input_ids`. Please make sure to only pass `input_ids`.zDYou cannot specify both input_ids and inputs_embeds at the same timer*rz5You have to specify either input_ids or inputs_embedsrrz$batch_size has to be defined and > 0rr)rzZ`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...Fr$c3TK|]}|jj!ywrt)tor)r% past_staters r'r(z(ImageGPTModel.forward..s &hzz}}]5I5I'J&hs%()rrrrrrrTrr=r zcuda:c3$K|]}|| ywrtr$)r%vs r'r(z(ImageGPTModel.forward..s=r))last_hidden_staterLr attentionscross_attentions)5warningswarn FutureWarningrpopr`rrPruse_return_dict%warn_if_padding_and_no_attention_maskrrrUrrrSr8rXarangelong unsqueezerWrrrr rinvert_attention_mask get_head_maskr&r4r3r4r<trainingrA warning_once enumeraterMr:cuda set_devicerrx_gradient_checkpointing_func__call__r;rHstr last_devicer9r)%ryrrLrrMrNrrOrrrrrPrQr input_shape batch_sizer past_lengthencoder_batch_sizeencoder_sequence_lengthrencoder_hidden_shapeposition_embedstoken_type_embeds output_shapepresentsall_self_attentionsall_cross_attentionsall_hidden_statesr>blockrrrrZrs% @r'rzImageGPTModel.forward)s8^ V # MMr  $ u >2I1B1N-TXT_T_TqTq$8$D $++JjJj "+!6IDKKcd d  "  6 6y. Q#..*K!r;r?;I"+J  &',,.s3K&,,Q/JTU U%.%:!! @T@T  %+00[_EN  "K#TFS[$89O)!,Q/44R8K   << [_{5RZ_ZdZdmstL'11!4L  %Q !GHH+00R@N ,AtT1,<=N,..TZZ.@N!N2ekk$**6M6Q6QQN ;; * */D/P=R=W=W=Y :  7$68O#P %-).4HQW)X&%)%?%?@V%W "%) " &&y$++2E2EF   HHY/M((<0%(:(:=;O;O(PP  % $ 8 ),==M -0 "m&8&8&<%>>  & &4==##p" "2$5b4%64;;;Z;Zr`d"6BD&/DFFO0L&M4 O "A"z"" %%m&:&:;)!&&h]g&h!hJ!-%3%6%6}7K7K%LNi6 ) ]-A-A BI#$58H$H!**t}};;NN!"aL)*%  !)#1'l*?+A'&7 $AJMD #wqzm3 &9W)QYZ=[<]&]#;;22+?7PY1_`CaBc+c("" OO113ODAqAbEzgA&6$:J:J&J(5(8(83q1u:9M(N Oe4 Ol -0 * **L9  1]4D D '3DFY[op  9+$+*1   r{) NNNNNNNNNNNNN)rrrrrvrBrFrKrrrXrxrrrrrrrrs@r'r1r1sv~&"2-1@D1515/3,0048<9=$(,0/3&*d ELL)d "%ell(;"<=d !. d !. d u||, d ELL)d   -d  ( 5d !) 6d D>d $D>d 'tnd d^d d u?? @!d d r{r1z The ImageGPT Model transformer with a language modeling head on top (linear layer with weights tied to the input embeddings). ) custom_introc$JeZdZdgZdeffd ZdZdZe dde e jde e e e jde e jd e e jd e e jd e e jd e e jd e e jde e jde e jde e de e de e de e dedee eff dZede e e jde jde e e jfdZxZS)ImageGPTForCausalImageModelingzlm_head.weightr`ct||t||_t j |j |jdz d|_d|_ d|_ |jy)NrFr2) rurvr1r-rrr[r^r;r:r;r=ryr`rzs r'rvz'ImageGPTForCausalImageModeling.__init__s[  (0yy0A0AA0EER $ r{c|jSrtr;rAs r'get_output_embeddingsz4ImageGPTForCausalImageModeling.get_output_embeddings%s ||r{c||_yrtrrDs r'set_output_embeddingsz4ImageGPTForCausalImageModeling.set_output_embeddings(s % r{rrLrrMrNrrOrrlabelsrrrPrQrr}c d|vr8tjdt| td|j d}||n|j j }|j||||||||| | | | | }|d}|j|}d}| r|dddddfj}| dd dfj}t}||jd|jd|jd}|s|f|d dz}||f|zS|St|||j|j|j |j" S) a% input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`): `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary. If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as `input_ids`. Indices can be obtained using [`AutoImageProcessor`]. See [`ImageGPTImageProcessor.__call__`] for details. labels (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*): Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` Examples: ```python >>> from transformers import AutoImageProcessor, ImageGPTForCausalImageModeling >>> import torch >>> import matplotlib.pyplot as plt >>> import numpy as np >>> image_processor = AutoImageProcessor.from_pretrained("openai/imagegpt-small") >>> model = ImageGPTForCausalImageModeling.from_pretrained("openai/imagegpt-small") >>> device = torch.device("cuda" if torch.cuda.is_available() else "cpu") >>> model.to(device) # doctest: +IGNORE_RESULT >>> # unconditional generation of 8 images >>> batch_size = 4 >>> context = torch.full((batch_size, 1), model.config.vocab_size - 1) # initialize with SOS token >>> context = context.to(device) >>> output = model.generate( ... input_ids=context, max_length=model.config.n_positions + 1, temperature=1.0, do_sample=True, top_k=40 ... ) >>> clusters = image_processor.clusters >>> height = image_processor.size["height"] >>> width = image_processor.size["width"] >>> samples = output[:, 1:].detach().cpu().numpy() >>> samples_img = [ ... np.reshape(np.rint(127.5 * (clusters[s] + 1.0)), [height, width, 3]).astype(np.uint8) for s in samples ... ] # convert color cluster tokens back to pixels >>> f, axes = plt.subplots(1, batch_size, dpi=300) >>> for img, ax in zip(samples_img, axes): # doctest: +IGNORE_RESULT ... ax.axis("off") ... ax.imshow(img) ```rSrTNrU) rLrrMrNrrOrrrrrPrQr.r*r)losslogitsrLrr\r])r^r_r`rrar`rbr-r;rr rrrrLrr\r])ryrrLrrMrNrrOrrrrrrPrQrtransformer_outputsr lm_logitsr shift_logits shift_labelsloss_fctoutputs r'rz&ImageGPTForCausalImageModeling.forward+sL V # MMr  $ u >2I%0%L!#qr'?557L')HL--b,2C2CB2GH,J[J[\^J_`D\$7$;;F)-)9TGf$ Ev E0/??-;;*550AA   r{beam_idxc,tfd|DS)a This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct beam_idx at every generation step. c3FK|]}tfd|Dyw)c3tK|]/}|jdj|j1yw)rN) index_selectrWr)r%rXrs r'r(zJImageGPTForCausalImageModeling._reorder_cache...s.jQ[*))!X[[9J9J-KLjs58Nr)r%rrs r'r(z@ImageGPTForCausalImageModeling._reorder_cache..s%  j_ij j s!r)rLrs `r'_reorder_cachez-ImageGPTForCausalImageModeling._reorder_caches -   r{)NNNNNNNNNNNNNN)rrr_tied_weights_keysrrvrrrrrXrxrrrrrr staticmethodrrrs@r'rrs++ ~ &-1@D1515/3,0048<9=)-$(,0/3&*{ ELL){ "%ell(;"<={ !. { !. { u||, { ELL){   -{  ( 5{ !) 6{ &{ D>{ $D>{ 'tn{ d^{ !{ " u77 8#{ { z  uU\\23  ?D||  uU\\" #    r{rz The ImageGPT Model transformer with an image classification head on top (linear layer). [`ImageGPTForImageClassification`] average-pools the hidden states in order to do the classification. c eZdZdeffd Ze ddeejdee e ejdeejdeejdeejdeejd eejd eejd ee d ee d ee dee de de e e ffdZxZS)ImageGPTForImageClassificationr`ct|||j|_t||_t j |j|jd|_|jy)NFr) rurv num_labelsr1r-rrr[scorer=rs r'rvz'ImageGPTForImageClassification.__init__sR   ++(0YYv}}dooEJ  r{rrLrrMrNrrOrrrrPrQrr}c d| vr8tjdt| td| j d}| | n|j j } |j|||||||| | | |  }|d}|jd}|j|}d}||j j|jdk(rd |j _ nl|jdkDrL|jtjk(s|jtjk(rd |j _ nd |j _ |j jd k(rIt!}|jdk(r&||j#|j#}n|||}n|j jd k(r=t%}||j'd |j|j'd }n,|j jd k(rt)}|||}| s|f|ddz}||f|zS|St+|||j,|j.|j0 S)ax input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`): `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary. If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as `input_ids`. Indices can be obtained using [`AutoImageProcessor`]. See [`ImageGPTImageProcessor.__call__`] for details. 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 regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). Examples: ```python >>> from transformers import AutoImageProcessor, ImageGPTForImageClassification >>> from PIL import Image >>> import requests >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> image_processor = AutoImageProcessor.from_pretrained("openai/imagegpt-small") >>> model = ImageGPTForImageClassification.from_pretrained("openai/imagegpt-small") >>> inputs = image_processor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> logits = outputs.logits ```rSrTNrU) rLrrMrNrrOrrrPrQrrr regressionsingle_label_classificationmulti_label_classificationr*)rrrLrr\)r^r_r`rrar`rbr-rr problem_typerrrXrerTr rLr rr rrLrr\)ryrrLrrMrNrrOrrrrPrQrrrpooled_hidden_statesrrrrs r'rz&ImageGPTForImageClassification.forwardsGd V # MMr  $ u >2I%0%s $D>s 'tns d^s s  u66 7s s r{r)rrr1rrn)4__doc__r%rCr^typingrrrrrXtorch.utils.checkpointrtorch.cuda.amprtorch.nnr r r activationsr generationrmodeling_outputsrrrmodeling_utilsr pytorch_utilsrrrutilsrrrconfiguration_imagegptr get_loggerrrArnModulerprrrrr1rr__all__r$r{r'rsR% .. #AA!) .YY 3   H %iX  X Xv"))"HBIIHV#so#s#sLF +F F R ^ %