o ۷i-@s$dZddlmZddlmZmZddlZddlmZddlm Z ddl m Z dd l m Z dd lmZmZmZmZdd lmZdd lmZd dlmZeeZeeddGdddeZeGddde ZdddZGdddejZ GdddejZ!eddGdddeZ"ddgZ#dS) zPyTorch VitPose model.) dataclass)OptionalUnionN)nn)BackboneOutput)PreTrainedModel)Unpack) ModelOutputTransformersKwargsauto_docstringlogging) load_backbone)can_return_tuple) VitPoseConfigz6 Class for outputs of pose estimation models. ) custom_introc@sjeZdZUdZdZeejed<dZ eejed<dZ ee ejdfed<dZ ee ejdfed<dS)VitPoseEstimatorOutputaH loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Loss is not supported at this moment. See https://github.com/ViTAE-Transformer/ViTPose/tree/main/mmpose/models/losses for further detail. heatmaps (`torch.FloatTensor` of shape `(batch_size, num_keypoints, height, width)`): Heatmaps as predicted by the model. hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each stage) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states (also called feature maps) of the model at the output of each stage. Nlossheatmaps. hidden_states attentions) __name__ __module__ __qualname____doc__rrtorch FloatTensor__annotations__rrtuplerr r b/home/ubuntu/vllm_env/lib/python3.10/site-packages/transformers/models/vitpose/modeling_vitpose.pyr%s  rc@s@eZdZUeed<dZdZdZdee j e j e j ffddZ dS) VitPosePreTrainedModelconfigvit pixel_valuesTmodulecCst|tjtjfr0tjj|jjt j d|j j d|jj |j_|jdur.|jjdSdSt|tjrE|jj|jjddSdS)zInitialize the weightsg)meanstdNg?) isinstancerLinearConv2dinit trunc_normal_weightdatatorfloat32r#initializer_rangedtypebiaszero_ LayerNormfill_)selfr&r r r! _init_weightsDs    z$VitPosePreTrainedModel._init_weightsN)rrrrrbase_model_prefixmain_input_namesupports_gradient_checkpointingrrr*r+r6r9r r r r!r"=s "r"gaussian-heatmapc Cs|dvrtd|jdkrtd|j\}}}}d}|dkr7d}|dddddd f |dddddd f<||d |||}|}|D]$\} } |dd| d f|dd| d f<|dd| d f|dd| d f<qH|||||f}|d }|S) aFlip the flipped heatmaps back to the original form. Args: output_flipped (`torch.tensor` of shape `(batch_size, num_keypoints, height, width)`): The output heatmaps obtained from the flipped images. flip_pairs (`torch.Tensor` of shape `(num_keypoints, 2)`): Pairs of keypoints which are mirrored (for example, left ear -- right ear). target_type (`str`, *optional*, defaults to `"gaussian-heatmap"`): Target type to use. Can be gaussian-heatmap or combined-target. gaussian-heatmap: Classification target with gaussian distribution. combined-target: The combination of classification target (response map) and regression target (offset map). Paper ref: Huang et al. The Devil is in the Details: Delving into Unbiased Data Processing for Human Pose Estimation (CVPR 2020). Returns: torch.Tensor: heatmaps that flipped back to the original image )r=combined-targetz9target_type should be gaussian-heatmap or combined-targetzCoutput_flipped should be [batch_size, num_keypoints, height, width]rr>rN.) ValueErrorndimshapereshapeclonetolistflip) output_flipped flip_pairs target_type batch_size num_keypointsheightwidthchannelsoutput_flipped_backleftrightr r r! flip_backSs" . " rScsHeZdZdZdeffdd Zd dejdeejdejfd d Z Z S) VitPoseSimpleDecoderz Simple decoding head consisting of a ReLU activation, 4x upsampling and a 3x3 convolution, turning the feature maps into heatmaps. r#csHtt|_tj|jddd|_tj|j j |j dddd|_ dS)NbilinearF) scale_factormode align_cornersrr kernel_sizestridepadding) super__init__rReLU activationUpsamplerV upsamplingr+backbone_config hidden_size num_labelsconvr8r# __class__r r!r^s   zVitPoseSimpleDecoder.__init__N hidden_staterIreturncCs4||}||}||}|durt||}|SN)r`rbrfrSr8rjrIrr r r!forwards    zVitPoseSimpleDecoder.forwardrl rrrrrr^rTensorrrn __classcell__r r rhr!rT{s* rTcsBeZdZdZdeffdd Zd dejdeejfdd Z Z S) VitPoseClassicDecoderz Classic decoding head consisting of a 2 deconvolutional blocks, followed by a 1x1 convolution layer, turning the feature maps into heatmaps. r#csttj|jjdddddd|_td|_t |_ tjddddddd|_ td|_ t |_ tjd|jdddd|_dS) Nr?rF)rZr[r\r4rrY)r]r^rConvTranspose2drcrddeconv1 BatchNorm2d batchnorm1r_relu1deconv2 batchnorm2relu2r+rerfrgrhr r!r^s     zVitPoseClassicDecoder.__init__NrjrIcCs\||}||}||}||}||}||}||}|dur,t||}|Srl)rvrxryrzr{r|rfrSrmr r r!rns        zVitPoseClassicDecoder.forwardrlror r rhr!rrs$rrz? The VitPose model with a pose estimation head on top. csleZdZdeffdd Zee   d dejde ejde ejde ejd e e d e f d d Z ZS)VitPoseForPoseEstimationr#cs|t|t||_t|jjdstdt|jjds!tdt|jjds,td|jr3t|nt ||_ | dS)Nrdz0The backbone should have a hidden_size attribute image_sizez0The backbone should have an image_size attribute patch_sizez/The backbone should have a patch_size attribute) r]r^rbackbonehasattrr#rAuse_simple_decoderrTrrhead post_initrgrhr r!r^s   z!VitPoseForPoseEstimation.__init__Nr% dataset_indexrIlabelskwargsrkc Ksd}|dur td|jj|fd|i|}|jd}|jd} |jjjd|jjjd} |jjjd|jjjd} | ddd}| | d| |  }|j ||d} t || |j|jd S) ac dataset_index (`torch.Tensor` of shape `(batch_size,)`): Index to use in the Mixture-of-Experts (MoE) blocks of the backbone. This corresponds to the dataset index used during training, e.g. For the single dataset index 0 refers to the corresponding dataset. For the multiple datasets index 0 refers to dataset A (e.g. MPII) and index 1 refers to dataset B (e.g. CrowdPose). flip_pairs (`torch.tensor`, *optional*): Whether to mirror pairs of keypoints (for example, left ear -- right ear). Examples: ```python >>> from transformers import AutoImageProcessor, VitPoseForPoseEstimation >>> import torch >>> from PIL import Image >>> import requests >>> processor = AutoImageProcessor.from_pretrained("usyd-community/vitpose-base-simple") >>> model = VitPoseForPoseEstimation.from_pretrained("usyd-community/vitpose-base-simple") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> boxes = [[[412.8, 157.61, 53.05, 138.01], [384.43, 172.21, 15.12, 35.74]]] >>> inputs = processor(image, boxes=boxes, return_tensors="pt") >>> with torch.no_grad(): ... outputs = model(**inputs) >>> heatmaps = outputs.heatmaps ```NzTraining is not yet supportedrr@rrrt)rI)rrrr)NotImplementedErrorrforward_with_filtered_kwargs feature_mapsrCr#rcr~rpermuterD contiguousrrrr) r8r%rrIrrroutputssequence_outputrK patch_height patch_widthrr r r!rns.'  z VitPoseForPoseEstimation.forward)NNN)rrrrr^rr rrprr r rrnrqr r rhr!r}s(r})r=)$r dataclassesrtypingrrrrmodeling_outputsrmodeling_utilsrprocessing_utilsr utilsr r r r utils.backbone_utilsr utils.genericrconfiguration_vitposer get_loggerrloggerrr"rSModulerTrrr}__all__r r r r!s8          (& V