`ij dZddlZddlmZddlmZmZddlZddlm Z ddl m Z ddl mZdd lmZdd lmZdd lmZmZmZmZd d lmZejeZeedGddeZeedGddeZGdde jZ Gdde jZ!Gdde jZ"Gdde jZ#d@dej$d e%d!e&d"ej$fd#Z'Gd$d%e jZ(Gd&d'eZ)Gd(d)e jZ*Gd*d+e j+Z,Gd,d-e jZ-Gd.d/e jZ.eGd0d1eZ/eGd2d3e/Z0d4ej$d5e1d"ej$fd6Z2d4ej$d7e1d8e1d"ej$fd9Z3Gd:d;e jZ4ed<Gd=d>e/Z5gd?Z6dS)AzPyTorch SegGpt model.N) dataclass)OptionalUnion)nn) functional)ACT2FN)GradientCheckpointingLayer)PreTrainedModel) ModelOutputauto_docstringlogging torch_int) SegGptConfigz1 Output type of [`SegGptEncoderOutput`]. ) custom_introceZdZUdZejed<dZee ejed<dZ ee ejed<dZ ee ejed<dS)SegGptEncoderOutputay last_hidden_state (`torch.FloatTensor` of shape `(batch_size, patch_height, patch_width, hidden_size)`): Sequence of hidden-states at the output of the last layer of the model. hidden_states (`tuple[torch.FloatTensor]`, `optional`, returned when `config.output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of shape `(batch_size, patch_height, patch_width, hidden_size)`. attentions (`tuple[torch.FloatTensor]`, `optional`, returned when `config.output_attentions=True`): Tuple of *torch.FloatTensor* (one for each layer) of shape `(batch_size, num_heads, seq_len, seq_len)`. intermediate_hidden_states (`tuple[torch.FloatTensor]`, *optional*, returned when `config.intermediate_hidden_state_indices` is set): Tuple of `torch.FloatTensor` of shape `(batch_size, patch_height, patch_width, hidden_size)`. Each element in the Tuple corresponds to the output of the layer specified in `config.intermediate_hidden_state_indices`. Additionally, each feature passes through a LayerNorm. last_hidden_stateN hidden_states attentionsintermediate_hidden_states) __name__ __module__ __qualname____doc__torch FloatTensor__annotations__rrtuplerr~/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/transformers/models/seggpt/modeling_seggpt.pyrr#s  ((((82D/ L&% 4, , ,w  T_Q' ' '5DOA4F+F+FwVwwewwDO\]L^wwaeapqraswww __\22::1aAFF r")rrrrr.rI __classcell__r<s@r#r)r)ZsV j j j j j       r"r)c eZdZdZdeddffd ZdededejfdZ dd ejd ejd e ej d e e dejf d Z xZS)SegGptEmbeddingszX Construct the embeddings from patch, position embeddings for input and prompt. r;returnNc8ttjt jddd|j|_tjt jddd|j|_tjt jddd|j|_ tjt jddd|j|_ tjt jddd|j|_ t||_ |j|jzdzdz}tjt jd||j|_tj|j|_dS)Nrr@)r-r.r Parameterrzerosr2 mask_tokensegment_token_inputsegment_token_prompttype_token_semantictype_token_instancer)patch_embeddingspretrain_image_sizer0randnposition_embeddingsDropouthidden_dropout_probdropout)r:r; num_positionsr<s r#r.zSegGptEmbeddings.__init__s7 ,u{1aF@SUWXX``abdeghjklle_# O#**1aA66 6"**1feR@@ @r"rDprompt_pixel_valuesbool_masked_posembedding_typecX||}||}|j\}}} } |j||| d} |d| d|| d} |d| z z| | zz}||nd}||| } ||jz}||j z}|| z}|| z}|dkr|j }n |dkr|j }ntd|||z}||z}tj||fd}|S)NrarinstancesemanticzBEmbedding type should be either 'semantic' or 'instance', but got rdim)rWrArRexpand unsqueezetype_asrjrmrSrTrUrVrBrcat)r:rDrnrorpinput_embeddingsprompt_embeddingsrE patch_height patch_width_rRw pos_embedtype_embeddingrHs r#rIzSegGptEmbeddings.forwards 00>> 112EFF3C3I0 L+q_++J kSUVV  % %b ) ) 1 1* = = E Eb,Xcef g g-Q7*q.H+9+E:11, LL ,d.FF-0II,i7- 9 Z ' '!5NN z ) )!5NNrbprrss s+n<->Y 02CD!LLL r")NN)rrrrrr.intrTensorrmr BoolTensorstrrIrJrKs@r#rMrM}s>|>>>>>>> AsA3A5<AAAA,7;(, ++l+#\+"%"23 + ! +  ++++++++r"rMceZdZdZfdZdededejdejfdZdejd ejd ejd ejde eefde eefdejfd Z ddejdejfdZ xZ S)SegGptAttentionz=Multi-head Attention block with relative position embeddings.cRt|j|j}}t |t jjr|n||f}t |t jjr|n||f}|d|jz|d|jzf}|j|j z}|j |_ |dz|_ tj |j|jdz|j |_tj |j|j|_|j|_|jr|t#dtjt'jd|dzdz ||_tjt'jd|dzdz ||_dSdS)NrrgrbiaszBInput size must be provided if using relative positional encoding.r@)r-r.r/r0r3r4r5r6r2num_attention_headsscalerLinearqkv_biasqkvproj use_relative_position_embeddingsrBrPrrQ rel_pos_h rel_pos_w)r:r;r/r0 input_sizehead_dimr<s r#r.zSegGptAttention.__init__s !'!2F4EJ #-j+/:R#S#SqZZZdfpYq #-j+/:R#S#SqZZZdfpYq  mv'88*Q-6K\:\] %)CC#)#= t^ 9V/1Ca1Gfo^^^If0&2DEE 060W-  0 X! !efff \%+a*Q-6G!6KX*V*VWWDN\%+a*Q-6G!6KX*V*VWWDNNN  X Xr"q_sizek_sizerel_posrNcntdt||zdz }tj|d|jddddd|d}|d|dd}tj|dddft||z dz}tj|dddft||z dz}||z |dz t||z dzz}|| S) a Get relative positional embeddings according to the relative positions of query and key sizes. Args: q_size (int): size of the query. k_size (int): size of key k. rel_pos (`torch.Tensor`): relative position embeddings (L, channel). Returns: Extracted positional embeddings according to relative positions. r@rrralinear)rcrdN?) rmaxrhrirjrArCrarangelong) r:rrr max_rel_distrel_pos_resizedq_coordsk_coordsrelative_coordss r# get_rel_poszSegGptAttention.get_rel_poss2 1s66222Q677 - OOAw}Q/ 4 4 < bhwkzbhwc,wkc->bhwkN)rrArjreinsum)r:rrrrrr query_height query_width key_height key_widthrelative_position_heightrelative_position_widthrEr~rureshaped_queryrel_hrel_ws r#add_decomposed_rel_posz&SegGptAttention.add_decomposed_rel_pos s$>%+! k & I#'#3#3L*i#X#X "&"2"2; 9"U"U"[ Asz<cRR -~?WXX -~?VWW||J k:yYYeAAAqqq!!!QQQ,--aaaAAAtQQQ6F0GG||J {(BJQZDZ[[ r"Frc |j\}}}}|||||zd|jdddddd}|d||jz||zdd\}} } ||jz| ddz} |jr(| | ||j |j ||f||f} tj j| tjd|j} |rC| ||j||zd} | ||jz||zd} nd} | | z||j||d} | ddddd|||d} || } | | fS) Nrrar@rr)dtyperu)rArrjrrCunbindr transposerrrrrrrsoftmaxfloat32torviewr)r:routput_attentionsrErFrGr~rrkeyvalue attn_weightsattn_weights_reshaped attn_outputs r#rIzSegGptAttention.forward:s'4':$ FE1 HH] # # WZ%D4Lb Q Q WQ1a # #  KK:8P+PRX[`R`bdeellmnoosE *cmmB.C.CC  0 66eT^T^fe_W]_dVeLx*22FDYZZZr"Fr ensemble_condfeature_ensemblerc||||}|d}|dd}|r|jddz|kr||jddzd\}} |dkra|jddz} | d| d} | dd| } | j|j} n*| dd| } tj|| gd}| ||z}|} | |}| |}| | |z}|f|z}|S) N)rrrr@rtraT)rukeepdim) rrrAsplitrjmean expand_asrryrrr) r:rrrrself_attention_outputsattention_outputoutputspromptinputs num_promptsresiduals r#rIzSegGptLayer.forwards"&  ! !- 0 0/"0" " 2!4(,  B 0 6q 9Q >- O O-334D4J14MQR4RXY3ZZNFF!!.4Q71< ;;;D99CCFKK'6D99CCFKK$y&&)9qAAA '788=H  ,,];; //  4>>-#@#@@  "W,r")FF)rrrrrr.rrrboolrr rIrJrKs@r#rrs[|[U[t[[[[[["'"' ##|## #  # uU\5</0% 2EE F ########r"rcleZdZdeddffd Z d dejded ed ed edee e ff d Z xZ S) SegGptEncoderr;rNNczt|_dtjdjjdDtjfdtjD|_ tj j j |_d|_dS)Nc6g|]}|Sr!)item).0xs r# z*SegGptEncoder.__init__..s rrrAqvvxxrrrr"rcpu)rc<g|]}t|Sr!)r)rir;dprs r#r z*SegGptEncoder.__init__..s'$j$j$jQ[Q%@%@$j$j$jr"rF)r-r.r;rlinspacernum_hidden_layersr ModuleListrangelayersrr2r layernormgradient_checkpointing)r:r;r r<s `@r#r.zSegGptEncoder.__init__s  rr63H&Jbkp!q!q!qrrrm$j$j$j$j$j%PVPhJiJi$j$j$jkk f&8f>STTT&+###r"FTrrroutput_hidden_states return_dictc:|rdnd}|rdnd}g}t|jD]\} } |r||fz}|jj| krdnd} | || ||} | d}| |jjkr4|d|jddz||jddzdzdz}| |jjvr(||||r || dfz}|r||fz}|std||||fDSt||||S)Nr!r@rrr`c3K|]}||V dSrr!)rvs r# z(SegGptEncoder.forward..s0= ===r")rrrr) enumeraterr; merge_indexrA!intermediate_hidden_state_indicesappendrr r) r:rrrrrall_hidden_statesall_self_attentionsrr  layer_moduler layer_outputss r#rIzSegGptEncoder.forwards#7@BBD$5?bb4%'"(55 P POA|# I$58H$H!"&!81! channels_last)r data_formatc ztj|fd|i||dvrtd|||_dS)Nr)r%channels_firstzUnsupported data format: )r-r.NotImplementedErrorr&)r:normalized_shaperr&kwargsr<s r#r.zSegGptLayerNorm.__init__sY)==s=f=== A A A%&O+&O&OPP P&r"featuresrNc|jdkrR|dddd}t|}|dddd}n!t|}|S)z Args: features: Tensor of shape (batch_size, channels, height, width) OR (batch_size, height, width, channels) r(rr@rr)r&rCr-rI)r:r,r<s r#rIzSegGptLayerNorm.forward sw  / / /''1a33Hwwx00H''1a33HHwwx00Hr") rrrrr.rrrIrJrKs@r#r$r$s 15/'''''''            r"r$c4eZdZfdZdejfdZxZS)SegGptDecoderHeadcJttj|j|jdd|_t |j|jd|_t|j |_ tj|jddd|_ dS)Nrr)r+paddingr()r*rr&T)r+r) r-r.rr8decoder_hidden_sizeconvr$rrr ract_fctheadrs r#r.zSegGptDecoderHead.__init__s I  &  &     )#7V=R`p   f/0 If8!QUVVV r"rc||}||}||}||}|Sr)r3rr4r5rs r#rIzSegGptDecoderHead.forward(sL -00 }55  ]33  -00 r")rrrr.rrrIrJrKs@r#r/r/s[ W W W W WU%6r"r/cXeZdZfdZdejdejfdZdejfdZxZS) SegGptDecoderc:ttj|jt |jz|jdz|jzd|_ t||_ |j|_|j|_||_ dS)Nr@Tr) r-r.rrr2lenrr0r2 decoder_embedr/ decoder_predr;rs r#r.zSegGptDecoder.__init__2s Y  V%M!N!N N  q 6#= =    .f55 +#)#=  r"rrNc|j\}}}}|||||j|j|j}|dddddd}||d||jz||jzf}|S) Nrrrr@rrarA)rArjr0r2rC)r:rrEr|r}r~s r#_reshape_hidden_statesz$SegGptDecoder._reshape_hidden_states>s3@3F0 L+q%--  k4?DOUYUm  &--aAq!Q?? %--r<$/#A;QUQ`C`a.  r"c||}||}||}|Sr)r;r@r<rs r#rIzSegGptDecoder.forwardJsA**=99 33MBB ))-88 r") rrrr.rrr@rIrJrKs@r#r8r81s~      E4E %J[    U%6r"r8cHeZdZUeed<dZdZdZddgZde j dd fd Z d S) SegGptPreTrainedModelr;modelrDTrMrmodulerNNc|jj}t|tjtjfrtj|jj tj d| |jj |j_ |j |jj dSdSt|tjt"fr?|jj |jj ddSt|t&rtj|jj tj d| |jj |j_ tj|jj tj d| |jj |j_ dSt|t,rItj|jj tj d| |jj |j_ tjj|j|tjj|j|tjj|j|tjj|j|tjj|j|dSdS)zInitialize the weightsr)rstdNr)rG)r;initializer_ranger3rrr8init trunc_normal_weightdatarrrrrzero_rr$fill_rrrrMrZnormal_rRrSrTrUrV)r:rErGs r# _init_weightsz#SegGptPreTrainedModel._init_weightsZsk+ fry")4 5 5# G"$!6!6v}7I7L7LU]7[7[bekn!6!o!o!r!r #""FM {&  &&((((('&  ? @ @ G K  " " $ $ $ M  $ $S ) ) ) ) )  0 0 G$&G$9$9 %((77%:%%b!'((   ! %'G$9$9 %((77%:%%b!'((   ! ! !  0 1 1 G.0g.C.C*/225=AA/D//b+122  & + HM ! !&"3 ! = = = HM ! !&"<# ! F F F HM ! !&"=3 ! G G G HM ! !&"<# ! F F F HM ! !&"<# ! F F F F F G Gr") rrrrrbase_model_prefixmain_input_namesupports_gradient_checkpointing_no_split_modulesrModulerPr!r"r#rCrCRsj $O&*#+];&GBI&G$&G&G&G&G&G&Gr"rCcFeZdZdeffd ZdefdZdeee efddfdZ e dde j d e j d e j d ee jd eed eedee jdeedeedeedeeeffdZxZS) SegGptModelr;ct|||_t||_t ||_|dSr)r-r.r;rMrHrencoder post_initrs r#r.zSegGptModel.__init__sX     *622$V,,  r"rNc|jjSr)rHrWrs r#get_input_embeddingsz SegGptModel.get_input_embeddingss //r"heads_to_pruneNc|D]/\}}|jj|j|0dS)z 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)itemsrYlayerr prune_heads)r:r]r`headss r# _prune_headszSegGptModel._prune_headssU +0022 C CLE5 L u % / ; ;E B B B B C Cr"rDrn prompt_masksrorrplabelsrrrc X||n |jj}| | n |jj} | | n |jj} ||nd}|jjjjj} | | }| | }tj ||fd}|tj ||fdntj ||fd}||t d||jjj} tj| dztj|j} tj| | dzz tj|j}tj | |g}|d}|||||}||||| | }|S) a prompt_pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Prompt pixel values. Prompt pixel values can be obtained using [`AutoImageProcessor`]. See [`SegGptImageProcessor.__call__`] for details. prompt_masks (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Prompt mask. Prompt mask can be obtained using [`AutoImageProcessor`]. See [`SegGptImageProcessor.__call__`] for details. bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, num_patches)`, *optional*): Boolean masked positions. Indicates which patches are masked (1) and which aren't (0). feature_ensemble (`bool`, *optional*): Boolean indicating whether to use feature ensemble or not. If `True`, the model will use feature ensemble if we have at least two prompts. If `False`, the model will not use feature ensemble. This argument should be considered when doing few-shot inference on an input image i.e. more than one prompt for the same image. embedding_type (`str`, *optional*): Embedding type. Indicates whether the prompt is a semantic or instance embedding. Can be either instance or semantic. labels (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`, `optional`): Ground truth mask for input images. Examples: ```python >>> from transformers import SegGptImageProcessor, SegGptModel >>> from PIL import Image >>> import requests >>> image_input_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_2.jpg" >>> image_prompt_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_1.jpg" >>> mask_prompt_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_1_target.png" >>> image_input = Image.open(requests.get(image_input_url, stream=True).raw) >>> image_prompt = Image.open(requests.get(image_prompt_url, stream=True).raw) >>> mask_prompt = Image.open(requests.get(mask_prompt_url, stream=True).raw).convert("L") >>> checkpoint = "BAAI/seggpt-vit-large" >>> model = SegGptModel.from_pretrained(checkpoint) >>> image_processor = SegGptImageProcessor.from_pretrained(checkpoint) >>> inputs = image_processor(images=image_input, prompt_images=image_prompt, prompt_masks=mask_prompt, return_tensors="pt") >>> outputs = model(**inputs) >>> list(outputs.last_hidden_state.shape) [1, 56, 28, 1024] ``` NFr@rtzLabels were provided, but bool_masked_pos were not. It will be set to default value. If you're training the model, make sure to provide a bool_masked_pos.rr)rpro)rrrr)r;rruse_return_dictrHrWr9rKrrrrylogger warning_oncer7rQrronesrwrY)r:rDrnrdrorrprerrrexpected_dtyper7bool_masked_pos_zerosbool_masked_pos_onesembedding_outputencoder_outputss r#rIzSegGptModel.forwardsv2C1N--TXT_Tq$8$D $+Jj &1%@T(UVVO-77::O?? -n^m+  ,, -/!5# '  r"NNNNNNN)rrrrr.r)r\dictrlistrcr rrrrrrrrr rrIrJrKs@r#rWrWs|0&;0000C4T#Y+?CDCCCC 7;+/(,.2,0/3&*kklk#\kl k "%"23 k #4. k! k*+k$D>k'tnkd^k u)) *kkk^kkkkkr"rWtensorr0c|j\}}}}||z}||z}|||||||f}|dddddd}||||z|dzdzf}|S)Nr?rr@rrr>r)rArjrC)rsr0rEr1rFrGr|r}s r#patchifyru s.4l+J feZ'L:%K ^^:|\:Wbdn"o^ p pF ^^Aq!Q1 - -F ^^:|k/I:WX=[\K\"]^ ^ ^F Mr"r|r}c ||jd}t|jddz dz}||z|jdkr$td|jdd|d|d ||||||df }|dd ddd d }||d||z||zf }|S)Nrrarr`rzNumber of patches z does not match patch height (z ) and width (r?r?r>r@r)rArrBrjrC)rsr|r}rEr0s r# unpatchifyrwsaJfl2&*s233Jk!V\!_44 za z zP\ z zkv z z z   ^^:|[*V`bc"d^ e eF ^^Aq!Q1 - -F ^^:q,2K[[eMe"f^ g gF Mr"c^eZdZfdZdejdejdejdejfdZxZS) SegGptLosscxt|j|_|j|_dSr)r-r.betar0rs r#r.zSegGptLoss.__init__%s0 K  +r"rdr'reroctj||fd}|dddddfdd|jdzdz}t ||jd|jz|jd|jz}t j||d|j}||z | z }|S)aNComputes the L1 loss between the predicted masks and the ground truth masks. Args: prompt_masks (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Pixel values from mask prompt. pred_masks (`torch.FloatTensor` of shape `(batch_size, num_channels, 2*height, width)`): Predicted masks. labels (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Ground truth mask for input images. bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, num_patches)`): Boolean masked positions. Indicates which patches are masked (1) and which aren't (0). Returns: `torch.FloatTensor`: The mean L1 loss between the predicted masks and the ground truth masks. r@rtNrrnone) reductionr{) rryrepeatr0rwrArhsmooth_l1_lossr{sum)r:rdr'rero ground_truthmaskr&s r#rIzSegGptLoss.forward*s2y,!7Q??? qqq!!!Tz*11!Q8JQ8NOO$ 21 5 H,J\]^J_cgcrJrss LFQUQZ[[[t   ""TXXZZ/ r") rrrr.rrrrIrJrKs@r#ryry$s~,,,,, !'!%!! ! ) !!!!!!!!r"ryzM SegGpt model with a decoder on top for one-shot image segmentation. ceZdZdeffd Ze ddejdejdejdeej dee d ee d eej d ee d ee d ee de eeffdZxZS)SegGptForImageSegmentationr;ct|||_t||_t ||_|dSr)r-r.r;rWrDr8decoderrZrs r#r.z#SegGptForImageSegmentation.__init__TsX      (( $V,,  r"NrDrnrdrorrprerrrrNc 0||n |jj}| | n |jj} | | n |jj} ||jjjj} tj | dztj |j } tj | | dzz tj |j } tj | | g}|d}|||||||||| |  }| r|jn|d}tj |d}||}d}|"t#|j}|||||}| s/|f}| r ||dfz}|r| rdnd}|||fz}||f|z}|St%|||j|j S) aY prompt_pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Prompt pixel values. Prompt pixel values can be obtained using [`AutoImageProcessor`]. See [`SegGptImageProcessor.__call__`] for details. prompt_masks (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Prompt mask. Prompt mask can be obtained using [`AutoImageProcessor`]. See [`SegGptImageProcessor.__call__`] for details. bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, num_patches)`, *optional*): Boolean masked positions. Indicates which patches are masked (1) and which aren't (0). feature_ensemble (`bool`, *optional*): Boolean indicating whether to use feature ensemble or not. If `True`, the model will use feature ensemble if we have at least two prompts. If `False`, the model will not use feature ensemble. This argument should be considered when doing few-shot inference on an input image i.e. more than one prompt for the same image. embedding_type (`str`, *optional*): Embedding type. Indicates whether the prompt is a semantic or instance embedding. Can be either instance or semantic. labels (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`, `optional`): Ground truth mask for input images. Examples: ```python >>> from transformers import SegGptImageProcessor, SegGptForImageSegmentation >>> from PIL import Image >>> import requests >>> image_input_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_2.jpg" >>> image_prompt_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_1.jpg" >>> mask_prompt_url = "https://raw.githubusercontent.com/baaivision/Painter/main/SegGPT/SegGPT_inference/examples/hmbb_1_target.png" >>> image_input = Image.open(requests.get(image_input_url, stream=True).raw) >>> image_prompt = Image.open(requests.get(image_prompt_url, stream=True).raw) >>> mask_prompt = Image.open(requests.get(mask_prompt_url, stream=True).raw).convert("L") >>> checkpoint = "BAAI/seggpt-vit-large" >>> model = SegGptForImageSegmentation.from_pretrained(checkpoint) >>> image_processor = SegGptImageProcessor.from_pretrained(checkpoint) >>> inputs = image_processor(images=image_input, prompt_images=image_prompt, prompt_masks=mask_prompt, return_tensors="pt") >>> outputs = model(**inputs) >>> result = image_processor.post_process_semantic_segmentation(outputs, target_sizes=[(image_input.height, image_input.width)])[0] >>> print(list(result.shape)) [170, 297] ``` Nr@rr) rDrnrdrorrprerrrrartr)r&r'rr)r;rrrgrDrHrWr7rrQrrrjryrwrrryr%rr)r:rDrnrdrorrprerrrr7rlrmrrr'r&loss_fnridxs r#rIz"SegGptForImageSegmentation.forward^sv2C1N--TXT_Tq$8$D $+Jj &1%@T(UVVO-77::O**% 3%+-)/!5#   LW%gW%G%G\cdf\g"%*Y/Ir%R%R%R"\\"<==    --G7<V_MMD  ]F# 071:-/  2/6aaQ73q 'tnq d^q  u33 4q q q ^q q q q q r"r)rWrCr)rF)7rcollections.abcr4 dataclassesrtypingrrrrtorch.nnrrh activationsr modeling_layersr modeling_utilsr utilsr r rrconfiguration_seggptr get_loggerrrhrr%rUr)rMrrrrrrrrrrr$r/r8rCrWrrurwryr__all__r!r"r#rs*!!!!!!"""""""" $$$$$$!!!!!!999999------DDDDDDDDDDDD......  H % %  JJJJJ+JJ   J,  :::::K::   :*     BI   FRRRRRryRRRjK4K4K4K4K4biK4K4K4^        U\eTV[Vb* % % % % %RY % % %,,,,,,,,,^9 9 9 9 9 BI9 9 9 zbl4 0BIB-G-G-G-G-GO-G-G-G`BBBBB'BBBJ U\ s u|     u| 3 S U\    ''''''''T } } } } } !6} }  } @ Q P Pr"