o wi|@sdZddlmZmZmZddlZddlZddlmZddlm Z m Z ddl m Z ddl mZdd lmZmZmZdd lmZmZdd lmZmZdd lmZmZmZd dlmZee Z!Gdddej"Z#Gdddej"Z$ d6dej"dej%dej%dej%deej%de&de&fddZ'Gdddej"Z(Gdd d ej"Z)Gd!d"d"ej"Z*Gd#d$d$ej"Z+Gd%d&d&ej"Z,Gd'd(d(eZ-Gd)d*d*ej"Z.Gd+d,d,ej"Z/eGd-d.d.eZ0eGd/d0d0e0Z1ed1d2Gd3d4d4e0Z2gd5Z3dS)7zPyTorch ViViT model.)CallableOptionalUnionN)nn)CrossEntropyLossMSELoss)ACT2FN)GradientCheckpointingLayer)BaseModelOutputBaseModelOutputWithPoolingImageClassifierOutput)ALL_ATTENTION_FUNCTIONSPreTrainedModel) find_pruneable_heads_and_indicesprune_linear_layer)auto_docstringlogging torch_int) VivitConfigcs0eZdZdZfddZddefddZZS) VivitTubeletEmbeddingsa Construct Vivit Tubelet embeddings. This module turns a batch of videos of shape (batch_size, num_frames, num_channels, height, width) into a tensor of shape (batch_size, seq_len, hidden_size) to be consumed by a Transformer encoder. The seq_len (the number of patches) equals (number of frames // tubelet_size[0]) * (height // tubelet_size[1]) * (width // tubelet_size[2]). cs|t|j|_|j|_|j|_|j|jd|j|jd|j|jd|_|j|_t j |j |j|j|jd|_ dS)Nrr) kernel_sizestride) super__init__ num_frames image_size tubelet_size patch_size num_patches hidden_size embed_dimrConv3d num_channels projectionselfconfig __class__e/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/vivit/modeling_vivit.pyr/s  zVivitTubeletEmbeddings.__init__Finterpolate_pos_encodingc Cs|j\}}}}}|s+||jks||jkr+td|d|d|jdd|jdd |ddddd }||}|ddd}|S) NzImage image size (*z) doesn't match model (rrz).rr)shaper ValueErrorpermuter&flatten transpose) r( pixel_valuesr. batch_sizerr%heightwidthxr,r,r-forward?s( zVivitTubeletEmbeddings.forwardF)__name__ __module__ __qualname____doc__rboolr; __classcell__r,r,r*r-r$s rcsNeZdZdZfddZdejdededejfdd Zdd e fd d Z Z S)VivitEmbeddingsz Vivit Embeddings. Creates embeddings from a video using VivitTubeletEmbeddings, adds CLS token and positional embeddings. csttttdd|j|_t||_ ttd|j j d|j|_ t |j |_|jdd|_||_dS)Nr)rrr Parametertorchzerosr" cls_tokenrpatch_embeddingsr!position_embeddingsDropouthidden_dropout_probdropoutrr r)r'r*r,r-rWs   zVivitEmbeddings.__init__ embeddingsr8r9returnc Cs|jdd}|jjdd}tjs||kr||kr|jS|jddddf}|jddddf}|jd}||jd} ||jd} t|d} |d| | |}|dddd}t j j || | fdd d }|dddd dd|}tj ||fdd S) a This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution images. This method is also adapted to support torch.jit tracing. Adapted from: - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211 rNrg?rrbicubicF)sizemode align_cornersdim)r1rIrEjit is_tracingr rreshaper3r functional interpolateviewcat) r(rMr8r9r! num_positionsclass_pos_embedpatch_pos_embedrU new_height new_widthsqrt_num_positionsr,r,r-r.es(   z(VivitEmbeddings.interpolate_pos_encodingFr.c Csr|j\}}}}}|j||d}|j|ddg} tj| |fdd}|r-|||||}n||j}||}|S)Nr.rrT) r1rHrGtilerEr\r.rIrL) r(r6r.r7rr%r8r9rM cls_tokensr,r,r-r;s  zVivitEmbeddings.forwardr<) r=r>r?r@rrETensorintr.rAr;rBr,r,r*r-rCPs  (rCmodulequerykeyvalueattention_maskscalingrLc Ks|t||dd|}tjj|dtjd|j}tjj |||j d}|dur,||}t||} | dd } | |fS)NrO)rUdtype)ptrainingrr) rEmatmulr5rrYsoftmaxfloat32torprLrr contiguous) rirjrkrlrmrnrLkwargs attn_weights attn_outputr,r,r-eager_attention_forwards  r{c sveZdZdeddffdd ZdejdejfddZ dd eejd e de e ejejfe ejffd d Z Z S)VivitSelfAttentionr)rNNcst|j|jdkrt|dstd|jd|jd||_|j|_t|j|j|_|j|j|_ |j |_ |jd|_ d|_ tj|j|j |jd|_tj|j|j |jd|_tj|j|j |jd|_dS) Nrembedding_sizezThe hidden size z4 is not a multiple of the number of attention heads .gF)bias)rrr"num_attention_headshasattrr2r)rgattention_head_size all_head_sizeattention_probs_dropout_prob dropout_probrn is_causalrLinearqkv_biasrjrkrlr'r*r,r-rs"   zVivitSelfAttention.__init__r:cCs6|dd|j|jf}||}|ddddS)NrOrrrr)rQrrr[r3)r(r: new_x_shaper,r,r-transpose_for_scoress z'VivitSelfAttention.transpose_for_scoresF head_maskoutput_attentionsc Cs|||}|||}|||}t}|jjdkr4|jjdkr.|r.tdnt |jj}|||||||j |j |j sCdn|j d\}} |dd|jf} || }|rc|| f} | S|f} | S)Neagersdpaz`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.rh)rrnrLro)rrkrlrjr{r)_attn_implementationlogger warning_oncerrrnrrrrQrrX) r( hidden_statesrr key_layer value_layer query_layerattention_interface context_layerattention_probsnew_context_layer_shapeoutputsr,r,r-r;s4     zVivitSelfAttention.forwardNF)r=r>r?rrrErfrrrArtupler;rBr,r,r*r-r|sr|csFeZdZdZdeddffdd Zdejdejdejfd d ZZ S) VivitSelfOutputz The residual connection is defined in VivitLayer instead of here (as is the case with other models), due to the layernorm applied before each block. r)rNNcs.tt|j|j|_t|j|_dSN) rrrrr"denserJrKrLr'r*r,r-r zVivitSelfOutput.__init__r input_tensorcCs||}||}|SrrrLr(rrr,r,r-r; s  zVivitSelfOutput.forward) r=r>r?r@rrrErfr;rBr,r,r*r-rs$rc s~eZdZdeddffdd ZdeeddfddZ  dd ej d e ej d e de e ej ej fe ej ffd dZZS)VivitAttentionr)rNNcs*tt||_t||_t|_dSr)rrr| attentionroutputset pruned_headsr'r*r,r-rs    zVivitAttention.__init__headscCst|dkrdSt||jj|jj|j\}}t|jj||j_t|jj||j_t|jj ||j_ t|j j |dd|j _ |jjt||j_|jj|jj|j_ |j ||_dS)NrrrT)lenrrrrrrrjrkrlrrrunion)r(rindexr,r,r- prune_headss zVivitAttention.prune_headsFrrrcCs4||||}||d|}|f|dd}|S)Nrr)rr)r(rrr self_outputsattention_outputrr,r,r-r;*szVivitAttention.forwardr)r=r>r?rrrrgrrErfrrArrr;rBr,r,r*r-rsrc$eZdZfddZddZZS)VivitIntermediatecsRtt|j|j|_t|j|_ t |j t r#t |j |_dS|j |_dSr)rrrrr"intermediate_sizerrJrKrL isinstance hidden_actstrr intermediate_act_fnr'r*r,r-r9s   zVivitIntermediate.__init__cCs"||}||}||}|Sr)rrrL)r(rr,r,r-r;Bs   zVivitIntermediate.forwardr=r>r?rr;rBr,r,r*r-r8s  rcr) VivitOutputcs.tt|j|j|_t|j|_ dSr) rrrrrr"rrJrKrLr'r*r,r-rKrzVivitOutput.__init__cCs ||}||}||}|Srrrr,r,r-r;Ps  zVivitOutput.forwardrr,r,r*r-rJ rcs*eZdZdZfddZdddZZS) VivitLayerzNThis corresponds to the EncoderBlock class in the scenic/vivit implementation.csbt|j|_d|_t||_t||_t||_ t j |j |j d|_t j |j |j d|_dS)Nreps)rrchunk_size_feed_forward seq_len_dimrrr intermediaterrr LayerNormr"layer_norm_epslayernorm_beforelayernorm_afterr'r*r,r-r]s    zVivitLayer.__init__NFcCs`|j||||d}|d}|dd}||}||}||}|||}|f|}|S)N)rrr)rrrrr)r(rrrself_attention_outputsrr layer_outputr,r,r-r;gs     zVivitLayer.forwardr)r=r>r?r@rr;rBr,r,r*r-rZs  rcs.eZdZfddZ    dddZZS) VivitEncodercs:t|_tfddtjD|_d|_dS)Ncsg|]}tqSr,)r).0_r)r,r- sz)VivitEncoder.__init__..F) rrr)r ModuleListrangenum_hidden_layerslayergradient_checkpointingr'r*rr-rs   zVivitEncoder.__init__NFTc Cs|rdnd}|r dnd}t|jD](\}} |r||f}|dur$||nd} | || |} | d}|r9|| df}q|rA||f}|sOtdd|||fDSt|||dS)Nr,rrcss|] }|dur|VqdSrr,)rvr,r,r- sz'VivitEncoder.forward..)last_hidden_stater attentions) enumeraterrr ) r(rrroutput_hidden_states return_dictall_hidden_statesall_self_attentionsi layer_modulelayer_head_mask layer_outputsr,r,r-r;s(     zVivitEncoder.forward)NFFTrr,r,r*r-rs  rcr) VivitPoolercs*tt|j|j|_t|_dSr)rrrrr"rTanh activationr'r*r,r-rs zVivitPooler.__init__cCs(|dddf}||}||}|S)Nr)rr)r(rfirst_token_tensor pooled_outputr,r,r-r;s  zVivitPooler.forwardrr,r,r*r-rrrc@s8eZdZeZdZdZdZgZdZ dZ dZ dZ ddZ dS)VivitPreTrainedModelvivitr6TcCst|tjtjfr#|jjjd|jjd|j dur!|j j dSdSt|tj rF|jjjd|jjd|j durD|jj|j  dSdSt|tj r[|j j |jjddSt|trn|jj |jj dSdS)zInitialize the weightsrh)meanstdNg?)rrrr$weightdatanormal_r)initializer_rangerzero_ Embedding padding_idxrfill_rCrGrI)r(rir,r,r- _init_weightss"       z"VivitPreTrainedModel._init_weightsN)r=r>r?r config_classbase_model_prefixmain_input_namesupports_gradient_checkpointing_no_split_modules_supports_sdpa_supports_flash_attn_2_supports_flex_attn_supports_attention_backendrr,r,r,r-rs rcseZdZdfdd ZddZddZe     dd eej d eej d ee d ee de dee de e ej e ffddZZS) VivitModelTcsXt|||_t||_t||_tj|j |j d|_ |r#t |nd|_ |dS)zv add_pooling_layer (bool, *optional*, defaults to `True`): Whether to add a pooling layer rN)rrr)rCrMrencoderrrr"r layernormrpooler post_init)r(r)add_pooling_layerr*r,r-rs    zVivitModel.__init__cCs|jjSr)rMrH)r(r,r,r-get_input_embeddingsszVivitModel.get_input_embeddingscCs*|D]\}}|jj|j|qdS)z Prunes heads of the model. Args: heads_to_prune: dict of {layer_num: list of heads to prune in this layer} N)itemsrrrr)r(heads_to_prunerrr,r,r- _prune_headsszVivitModel._prune_headsNFr6rrrr.rrNc Cs|dur|n|jj}|dur|n|jj}|dur|n|jj}|dur&td|||jj}|j||d}|j|||||d}|d} | | } |j durR| | nd} |s`| | f|ddSt | | |j |j dS)a Examples: ```python >>> import av >>> import numpy as np >>> from transformers import VivitImageProcessor, VivitModel >>> from huggingface_hub import hf_hub_download >>> np.random.seed(0) >>> def read_video_pyav(container, indices): ... ''' ... Decode the video with PyAV decoder. ... Args: ... container (`av.container.input.InputContainer`): PyAV container. ... indices (`list[int]`): List of frame indices to decode. ... Returns: ... result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3). ... ''' ... frames = [] ... container.seek(0) ... start_index = indices[0] ... end_index = indices[-1] ... for i, frame in enumerate(container.decode(video=0)): ... if i > end_index: ... break ... if i >= start_index and i in indices: ... frames.append(frame) ... return np.stack([x.to_ndarray(format="rgb24") for x in frames]) >>> def sample_frame_indices(clip_len, frame_sample_rate, seg_len): ... ''' ... Sample a given number of frame indices from the video. ... Args: ... clip_len (`int`): Total number of frames to sample. ... frame_sample_rate (`int`): Sample every n-th frame. ... seg_len (`int`): Maximum allowed index of sample's last frame. ... Returns: ... indices (`list[int]`): List of sampled frame indices ... ''' ... converted_len = int(clip_len * frame_sample_rate) ... end_idx = np.random.randint(converted_len, seg_len) ... start_idx = end_idx - converted_len ... indices = np.linspace(start_idx, end_idx, num=clip_len) ... indices = np.clip(indices, start_idx, end_idx - 1).astype(np.int64) ... return indices >>> # video clip consists of 300 frames (10 seconds at 30 FPS) >>> file_path = hf_hub_download( ... repo_id="nielsr/video-demo", filename="eating_spaghetti.mp4", repo_type="dataset" ... ) >>> container = av.open(file_path) >>> # sample 32 frames >>> indices = sample_frame_indices(clip_len=32, frame_sample_rate=1, seg_len=container.streams.video[0].frames) >>> video = read_video_pyav(container=container, indices=indices) >>> image_processor = VivitImageProcessor.from_pretrained("google/vivit-b-16x2-kinetics400") >>> model = VivitModel.from_pretrained("google/vivit-b-16x2-kinetics400") >>> # prepare video for the model >>> inputs = image_processor(list(video), return_tensors="pt") >>> # forward pass >>> outputs = model(**inputs) >>> last_hidden_states = outputs.last_hidden_state >>> list(last_hidden_states.shape) [1, 3137, 768] ```Nz You have to specify pixel_valuesrc)rrrrrr)r pooler_outputrr)r)rruse_return_dictr2 get_head_maskrrMrrrr rr) r(r6rrrr.rembedding_outputencoder_outputssequence_outputrr,r,r-r;s4T zVivitModel.forward)T)NNNNFN)r=r>r?rrrrrrE FloatTensorrArrr r;rBr,r,r*r-rs4 ra ViViT Transformer model with a video classification head on top (a linear layer on top of the final hidden state of the [CLS] token) e.g. for Kinetics-400. Note that it's possible to fine-tune ViT on higher resolution images than the ones it has been trained on, by setting `interpolate_pos_encoding` to `True` in the forward of the model. This will interpolate the pre-trained position embeddings to the higher resolution. ) custom_introcseZdZfddZe       ddeejdeejdeejdee d ee d e d ee d e e eje ffd dZ ZS)VivitForVideoClassificationcsRt||j|_t|dd|_|jdkrt|j|jnt|_ | dS)NF)rr) rr num_labelsrrrrr"Identity classifierrr'r*r,r-rs $ z$VivitForVideoClassification.__init__NFr6rlabelsrrr.rrNcCs|dur|n|jj}|j||||||d}|d} || dddddf} d} |durQ|jdkrAt} | | d|d} nt} | | d|j|d} |sg| f|dd} | dure| f| S| St| | |j |j dS)a labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the image 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 >>> import av >>> import numpy as np >>> import torch >>> from transformers import VivitImageProcessor, VivitForVideoClassification >>> from huggingface_hub import hf_hub_download >>> np.random.seed(0) >>> def read_video_pyav(container, indices): ... ''' ... Decode the video with PyAV decoder. ... Args: ... container (`av.container.input.InputContainer`): PyAV container. ... indices (`list[int]`): List of frame indices to decode. ... Returns: ... result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3). ... ''' ... frames = [] ... container.seek(0) ... start_index = indices[0] ... end_index = indices[-1] ... for i, frame in enumerate(container.decode(video=0)): ... if i > end_index: ... break ... if i >= start_index and i in indices: ... frames.append(frame) ... return np.stack([x.to_ndarray(format="rgb24") for x in frames]) >>> def sample_frame_indices(clip_len, frame_sample_rate, seg_len): ... ''' ... Sample a given number of frame indices from the video. ... Args: ... clip_len (`int`): Total number of frames to sample. ... frame_sample_rate (`int`): Sample every n-th frame. ... seg_len (`int`): Maximum allowed index of sample's last frame. ... Returns: ... indices (`list[int]`): List of sampled frame indices ... ''' ... converted_len = int(clip_len * frame_sample_rate) ... end_idx = np.random.randint(converted_len, seg_len) ... start_idx = end_idx - converted_len ... indices = np.linspace(start_idx, end_idx, num=clip_len) ... indices = np.clip(indices, start_idx, end_idx - 1).astype(np.int64) ... return indices >>> # video clip consists of 300 frames (10 seconds at 30 FPS) >>> file_path = hf_hub_download( ... repo_id="nielsr/video-demo", filename="eating_spaghetti.mp4", repo_type="dataset" ... ) >>> container = av.open(file_path) >>> # sample 32 frames >>> indices = sample_frame_indices(clip_len=32, frame_sample_rate=4, seg_len=container.streams.video[0].frames) >>> video = read_video_pyav(container=container, indices=indices) >>> image_processor = VivitImageProcessor.from_pretrained("google/vivit-b-16x2-kinetics400") >>> model = VivitForVideoClassification.from_pretrained("google/vivit-b-16x2-kinetics400") >>> inputs = image_processor(list(video), return_tensors="pt") >>> with torch.no_grad(): ... outputs = model(**inputs) ... logits = outputs.logits >>> # model predicts one of the 400 Kinetics-400 classes >>> predicted_label = logits.argmax(-1).item() >>> print(model.config.id2label[predicted_label]) LABEL_116 ```N)rrrr.rrrrOr)losslogitsrr) r)rrr r rr[rr rr)r(r6rrrrr.rrrrrloss_fctrr,r,r-r;s6]  z#VivitForVideoClassification.forward)NNNNNFN)r=r>r?rrrrEr LongTensorrArrr r;rBr,r,r*r-r ss6   r )rrr )rh)4r@typingrrrrEtorch.utils.checkpointrtorch.nnrr activationsr modeling_layersr modeling_outputsr r r modeling_utilsrr pytorch_utilsrrutilsrrrconfiguration_vivitr get_loggerr=rModulerrCrffloatr{r|rrrrrrrrrr __all__r,r,r,r-sf     ,W ?''+